1 //===-- DependenceAnalysis.cpp - DA Implementation --------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // DependenceAnalysis is an LLVM pass that analyses dependences between memory
10 // accesses. Currently, it is an (incomplete) implementation of the approach
11 // described in
12 //
13 // Practical Dependence Testing
14 // Goff, Kennedy, Tseng
15 // PLDI 1991
16 //
17 // There's a single entry point that analyzes the dependence between a pair
18 // of memory references in a function, returning either NULL, for no dependence,
19 // or a more-or-less detailed description of the dependence between them.
20 //
21 // Currently, the implementation cannot propagate constraints between
22 // coupled RDIV subscripts and lacks a multi-subscript MIV test.
23 // Both of these are conservative weaknesses;
24 // that is, not a source of correctness problems.
25 //
26 // Since Clang linearizes some array subscripts, the dependence
27 // analysis is using SCEV->delinearize to recover the representation of multiple
28 // subscripts, and thus avoid the more expensive and less precise MIV tests. The
29 // delinearization is controlled by the flag -da-delinearize.
30 //
31 // We should pay some careful attention to the possibility of integer overflow
32 // in the implementation of the various tests. This could happen with Add,
33 // Subtract, or Multiply, with both APInt's and SCEV's.
34 //
35 // Some non-linear subscript pairs can be handled by the GCD test
36 // (and perhaps other tests).
37 // Should explore how often these things occur.
38 //
39 // Finally, it seems like certain test cases expose weaknesses in the SCEV
40 // simplification, especially in the handling of sign and zero extensions.
41 // It could be useful to spend time exploring these.
42 //
43 // Please note that this is work in progress and the interface is subject to
44 // change.
45 //
46 //===----------------------------------------------------------------------===//
47 // //
48 // In memory of Ken Kennedy, 1945 - 2007 //
49 // //
50 //===----------------------------------------------------------------------===//
51
52 #include "llvm/Analysis/DependenceAnalysis.h"
53 #include "llvm/ADT/STLExtras.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/LoopInfo.h"
57 #include "llvm/Analysis/ScalarEvolution.h"
58 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
59 #include "llvm/Analysis/ValueTracking.h"
60 #include "llvm/Config/llvm-config.h"
61 #include "llvm/IR/InstIterator.h"
62 #include "llvm/IR/Module.h"
63 #include "llvm/IR/Operator.h"
64 #include "llvm/InitializePasses.h"
65 #include "llvm/Support/CommandLine.h"
66 #include "llvm/Support/Debug.h"
67 #include "llvm/Support/ErrorHandling.h"
68 #include "llvm/Support/raw_ostream.h"
69
70 using namespace llvm;
71
72 #define DEBUG_TYPE "da"
73
74 //===----------------------------------------------------------------------===//
75 // statistics
76
77 STATISTIC(TotalArrayPairs, "Array pairs tested");
78 STATISTIC(SeparableSubscriptPairs, "Separable subscript pairs");
79 STATISTIC(CoupledSubscriptPairs, "Coupled subscript pairs");
80 STATISTIC(NonlinearSubscriptPairs, "Nonlinear subscript pairs");
81 STATISTIC(ZIVapplications, "ZIV applications");
82 STATISTIC(ZIVindependence, "ZIV independence");
83 STATISTIC(StrongSIVapplications, "Strong SIV applications");
84 STATISTIC(StrongSIVsuccesses, "Strong SIV successes");
85 STATISTIC(StrongSIVindependence, "Strong SIV independence");
86 STATISTIC(WeakCrossingSIVapplications, "Weak-Crossing SIV applications");
87 STATISTIC(WeakCrossingSIVsuccesses, "Weak-Crossing SIV successes");
88 STATISTIC(WeakCrossingSIVindependence, "Weak-Crossing SIV independence");
89 STATISTIC(ExactSIVapplications, "Exact SIV applications");
90 STATISTIC(ExactSIVsuccesses, "Exact SIV successes");
91 STATISTIC(ExactSIVindependence, "Exact SIV independence");
92 STATISTIC(WeakZeroSIVapplications, "Weak-Zero SIV applications");
93 STATISTIC(WeakZeroSIVsuccesses, "Weak-Zero SIV successes");
94 STATISTIC(WeakZeroSIVindependence, "Weak-Zero SIV independence");
95 STATISTIC(ExactRDIVapplications, "Exact RDIV applications");
96 STATISTIC(ExactRDIVindependence, "Exact RDIV independence");
97 STATISTIC(SymbolicRDIVapplications, "Symbolic RDIV applications");
98 STATISTIC(SymbolicRDIVindependence, "Symbolic RDIV independence");
99 STATISTIC(DeltaApplications, "Delta applications");
100 STATISTIC(DeltaSuccesses, "Delta successes");
101 STATISTIC(DeltaIndependence, "Delta independence");
102 STATISTIC(DeltaPropagations, "Delta propagations");
103 STATISTIC(GCDapplications, "GCD applications");
104 STATISTIC(GCDsuccesses, "GCD successes");
105 STATISTIC(GCDindependence, "GCD independence");
106 STATISTIC(BanerjeeApplications, "Banerjee applications");
107 STATISTIC(BanerjeeIndependence, "Banerjee independence");
108 STATISTIC(BanerjeeSuccesses, "Banerjee successes");
109
110 static cl::opt<bool>
111 Delinearize("da-delinearize", cl::init(true), cl::Hidden, cl::ZeroOrMore,
112 cl::desc("Try to delinearize array references."));
113 static cl::opt<bool> DisableDelinearizationChecks(
114 "da-disable-delinearization-checks", cl::init(false), cl::Hidden,
115 cl::ZeroOrMore,
116 cl::desc(
117 "Disable checks that try to statically verify validity of "
118 "delinearized subscripts. Enabling this option may result in incorrect "
119 "dependence vectors for languages that allow the subscript of one "
120 "dimension to underflow or overflow into another dimension."));
121
122 //===----------------------------------------------------------------------===//
123 // basics
124
125 DependenceAnalysis::Result
run(Function & F,FunctionAnalysisManager & FAM)126 DependenceAnalysis::run(Function &F, FunctionAnalysisManager &FAM) {
127 auto &AA = FAM.getResult<AAManager>(F);
128 auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F);
129 auto &LI = FAM.getResult<LoopAnalysis>(F);
130 return DependenceInfo(&F, &AA, &SE, &LI);
131 }
132
133 AnalysisKey DependenceAnalysis::Key;
134
135 INITIALIZE_PASS_BEGIN(DependenceAnalysisWrapperPass, "da",
136 "Dependence Analysis", true, true)
137 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
138 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
139 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
140 INITIALIZE_PASS_END(DependenceAnalysisWrapperPass, "da", "Dependence Analysis",
141 true, true)
142
143 char DependenceAnalysisWrapperPass::ID = 0;
144
DependenceAnalysisWrapperPass()145 DependenceAnalysisWrapperPass::DependenceAnalysisWrapperPass()
146 : FunctionPass(ID) {
147 initializeDependenceAnalysisWrapperPassPass(*PassRegistry::getPassRegistry());
148 }
149
createDependenceAnalysisWrapperPass()150 FunctionPass *llvm::createDependenceAnalysisWrapperPass() {
151 return new DependenceAnalysisWrapperPass();
152 }
153
runOnFunction(Function & F)154 bool DependenceAnalysisWrapperPass::runOnFunction(Function &F) {
155 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
156 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
157 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
158 info.reset(new DependenceInfo(&F, &AA, &SE, &LI));
159 return false;
160 }
161
getDI() const162 DependenceInfo &DependenceAnalysisWrapperPass::getDI() const { return *info; }
163
releaseMemory()164 void DependenceAnalysisWrapperPass::releaseMemory() { info.reset(); }
165
getAnalysisUsage(AnalysisUsage & AU) const166 void DependenceAnalysisWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
167 AU.setPreservesAll();
168 AU.addRequiredTransitive<AAResultsWrapperPass>();
169 AU.addRequiredTransitive<ScalarEvolutionWrapperPass>();
170 AU.addRequiredTransitive<LoopInfoWrapperPass>();
171 }
172
173 // Used to test the dependence analyzer.
174 // Looks through the function, noting instructions that may access memory.
175 // Calls depends() on every possible pair and prints out the result.
176 // Ignores all other instructions.
dumpExampleDependence(raw_ostream & OS,DependenceInfo * DA)177 static void dumpExampleDependence(raw_ostream &OS, DependenceInfo *DA) {
178 auto *F = DA->getFunction();
179 for (inst_iterator SrcI = inst_begin(F), SrcE = inst_end(F); SrcI != SrcE;
180 ++SrcI) {
181 if (SrcI->mayReadOrWriteMemory()) {
182 for (inst_iterator DstI = SrcI, DstE = inst_end(F);
183 DstI != DstE; ++DstI) {
184 if (DstI->mayReadOrWriteMemory()) {
185 OS << "Src:" << *SrcI << " --> Dst:" << *DstI << "\n";
186 OS << " da analyze - ";
187 if (auto D = DA->depends(&*SrcI, &*DstI, true)) {
188 D->dump(OS);
189 for (unsigned Level = 1; Level <= D->getLevels(); Level++) {
190 if (D->isSplitable(Level)) {
191 OS << " da analyze - split level = " << Level;
192 OS << ", iteration = " << *DA->getSplitIteration(*D, Level);
193 OS << "!\n";
194 }
195 }
196 }
197 else
198 OS << "none!\n";
199 }
200 }
201 }
202 }
203 }
204
print(raw_ostream & OS,const Module *) const205 void DependenceAnalysisWrapperPass::print(raw_ostream &OS,
206 const Module *) const {
207 dumpExampleDependence(OS, info.get());
208 }
209
210 PreservedAnalyses
run(Function & F,FunctionAnalysisManager & FAM)211 DependenceAnalysisPrinterPass::run(Function &F, FunctionAnalysisManager &FAM) {
212 OS << "'Dependence Analysis' for function '" << F.getName() << "':\n";
213 dumpExampleDependence(OS, &FAM.getResult<DependenceAnalysis>(F));
214 return PreservedAnalyses::all();
215 }
216
217 //===----------------------------------------------------------------------===//
218 // Dependence methods
219
220 // Returns true if this is an input dependence.
isInput() const221 bool Dependence::isInput() const {
222 return Src->mayReadFromMemory() && Dst->mayReadFromMemory();
223 }
224
225
226 // Returns true if this is an output dependence.
isOutput() const227 bool Dependence::isOutput() const {
228 return Src->mayWriteToMemory() && Dst->mayWriteToMemory();
229 }
230
231
232 // Returns true if this is an flow (aka true) dependence.
isFlow() const233 bool Dependence::isFlow() const {
234 return Src->mayWriteToMemory() && Dst->mayReadFromMemory();
235 }
236
237
238 // Returns true if this is an anti dependence.
isAnti() const239 bool Dependence::isAnti() const {
240 return Src->mayReadFromMemory() && Dst->mayWriteToMemory();
241 }
242
243
244 // Returns true if a particular level is scalar; that is,
245 // if no subscript in the source or destination mention the induction
246 // variable associated with the loop at this level.
247 // Leave this out of line, so it will serve as a virtual method anchor
isScalar(unsigned level) const248 bool Dependence::isScalar(unsigned level) const {
249 return false;
250 }
251
252
253 //===----------------------------------------------------------------------===//
254 // FullDependence methods
255
FullDependence(Instruction * Source,Instruction * Destination,bool PossiblyLoopIndependent,unsigned CommonLevels)256 FullDependence::FullDependence(Instruction *Source, Instruction *Destination,
257 bool PossiblyLoopIndependent,
258 unsigned CommonLevels)
259 : Dependence(Source, Destination), Levels(CommonLevels),
260 LoopIndependent(PossiblyLoopIndependent) {
261 Consistent = true;
262 if (CommonLevels)
263 DV = std::make_unique<DVEntry[]>(CommonLevels);
264 }
265
266 // The rest are simple getters that hide the implementation.
267
268 // getDirection - Returns the direction associated with a particular level.
getDirection(unsigned Level) const269 unsigned FullDependence::getDirection(unsigned Level) const {
270 assert(0 < Level && Level <= Levels && "Level out of range");
271 return DV[Level - 1].Direction;
272 }
273
274
275 // Returns the distance (or NULL) associated with a particular level.
getDistance(unsigned Level) const276 const SCEV *FullDependence::getDistance(unsigned Level) const {
277 assert(0 < Level && Level <= Levels && "Level out of range");
278 return DV[Level - 1].Distance;
279 }
280
281
282 // Returns true if a particular level is scalar; that is,
283 // if no subscript in the source or destination mention the induction
284 // variable associated with the loop at this level.
isScalar(unsigned Level) const285 bool FullDependence::isScalar(unsigned Level) const {
286 assert(0 < Level && Level <= Levels && "Level out of range");
287 return DV[Level - 1].Scalar;
288 }
289
290
291 // Returns true if peeling the first iteration from this loop
292 // will break this dependence.
isPeelFirst(unsigned Level) const293 bool FullDependence::isPeelFirst(unsigned Level) const {
294 assert(0 < Level && Level <= Levels && "Level out of range");
295 return DV[Level - 1].PeelFirst;
296 }
297
298
299 // Returns true if peeling the last iteration from this loop
300 // will break this dependence.
isPeelLast(unsigned Level) const301 bool FullDependence::isPeelLast(unsigned Level) const {
302 assert(0 < Level && Level <= Levels && "Level out of range");
303 return DV[Level - 1].PeelLast;
304 }
305
306
307 // Returns true if splitting this loop will break the dependence.
isSplitable(unsigned Level) const308 bool FullDependence::isSplitable(unsigned Level) const {
309 assert(0 < Level && Level <= Levels && "Level out of range");
310 return DV[Level - 1].Splitable;
311 }
312
313
314 //===----------------------------------------------------------------------===//
315 // DependenceInfo::Constraint methods
316
317 // If constraint is a point <X, Y>, returns X.
318 // Otherwise assert.
getX() const319 const SCEV *DependenceInfo::Constraint::getX() const {
320 assert(Kind == Point && "Kind should be Point");
321 return A;
322 }
323
324
325 // If constraint is a point <X, Y>, returns Y.
326 // Otherwise assert.
getY() const327 const SCEV *DependenceInfo::Constraint::getY() const {
328 assert(Kind == Point && "Kind should be Point");
329 return B;
330 }
331
332
333 // If constraint is a line AX + BY = C, returns A.
334 // Otherwise assert.
getA() const335 const SCEV *DependenceInfo::Constraint::getA() const {
336 assert((Kind == Line || Kind == Distance) &&
337 "Kind should be Line (or Distance)");
338 return A;
339 }
340
341
342 // If constraint is a line AX + BY = C, returns B.
343 // Otherwise assert.
getB() const344 const SCEV *DependenceInfo::Constraint::getB() const {
345 assert((Kind == Line || Kind == Distance) &&
346 "Kind should be Line (or Distance)");
347 return B;
348 }
349
350
351 // If constraint is a line AX + BY = C, returns C.
352 // Otherwise assert.
getC() const353 const SCEV *DependenceInfo::Constraint::getC() const {
354 assert((Kind == Line || Kind == Distance) &&
355 "Kind should be Line (or Distance)");
356 return C;
357 }
358
359
360 // If constraint is a distance, returns D.
361 // Otherwise assert.
getD() const362 const SCEV *DependenceInfo::Constraint::getD() const {
363 assert(Kind == Distance && "Kind should be Distance");
364 return SE->getNegativeSCEV(C);
365 }
366
367
368 // Returns the loop associated with this constraint.
getAssociatedLoop() const369 const Loop *DependenceInfo::Constraint::getAssociatedLoop() const {
370 assert((Kind == Distance || Kind == Line || Kind == Point) &&
371 "Kind should be Distance, Line, or Point");
372 return AssociatedLoop;
373 }
374
setPoint(const SCEV * X,const SCEV * Y,const Loop * CurLoop)375 void DependenceInfo::Constraint::setPoint(const SCEV *X, const SCEV *Y,
376 const Loop *CurLoop) {
377 Kind = Point;
378 A = X;
379 B = Y;
380 AssociatedLoop = CurLoop;
381 }
382
setLine(const SCEV * AA,const SCEV * BB,const SCEV * CC,const Loop * CurLoop)383 void DependenceInfo::Constraint::setLine(const SCEV *AA, const SCEV *BB,
384 const SCEV *CC, const Loop *CurLoop) {
385 Kind = Line;
386 A = AA;
387 B = BB;
388 C = CC;
389 AssociatedLoop = CurLoop;
390 }
391
setDistance(const SCEV * D,const Loop * CurLoop)392 void DependenceInfo::Constraint::setDistance(const SCEV *D,
393 const Loop *CurLoop) {
394 Kind = Distance;
395 A = SE->getOne(D->getType());
396 B = SE->getNegativeSCEV(A);
397 C = SE->getNegativeSCEV(D);
398 AssociatedLoop = CurLoop;
399 }
400
setEmpty()401 void DependenceInfo::Constraint::setEmpty() { Kind = Empty; }
402
setAny(ScalarEvolution * NewSE)403 void DependenceInfo::Constraint::setAny(ScalarEvolution *NewSE) {
404 SE = NewSE;
405 Kind = Any;
406 }
407
408 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
409 // For debugging purposes. Dumps the constraint out to OS.
dump(raw_ostream & OS) const410 LLVM_DUMP_METHOD void DependenceInfo::Constraint::dump(raw_ostream &OS) const {
411 if (isEmpty())
412 OS << " Empty\n";
413 else if (isAny())
414 OS << " Any\n";
415 else if (isPoint())
416 OS << " Point is <" << *getX() << ", " << *getY() << ">\n";
417 else if (isDistance())
418 OS << " Distance is " << *getD() <<
419 " (" << *getA() << "*X + " << *getB() << "*Y = " << *getC() << ")\n";
420 else if (isLine())
421 OS << " Line is " << *getA() << "*X + " <<
422 *getB() << "*Y = " << *getC() << "\n";
423 else
424 llvm_unreachable("unknown constraint type in Constraint::dump");
425 }
426 #endif
427
428
429 // Updates X with the intersection
430 // of the Constraints X and Y. Returns true if X has changed.
431 // Corresponds to Figure 4 from the paper
432 //
433 // Practical Dependence Testing
434 // Goff, Kennedy, Tseng
435 // PLDI 1991
intersectConstraints(Constraint * X,const Constraint * Y)436 bool DependenceInfo::intersectConstraints(Constraint *X, const Constraint *Y) {
437 ++DeltaApplications;
438 LLVM_DEBUG(dbgs() << "\tintersect constraints\n");
439 LLVM_DEBUG(dbgs() << "\t X ="; X->dump(dbgs()));
440 LLVM_DEBUG(dbgs() << "\t Y ="; Y->dump(dbgs()));
441 assert(!Y->isPoint() && "Y must not be a Point");
442 if (X->isAny()) {
443 if (Y->isAny())
444 return false;
445 *X = *Y;
446 return true;
447 }
448 if (X->isEmpty())
449 return false;
450 if (Y->isEmpty()) {
451 X->setEmpty();
452 return true;
453 }
454
455 if (X->isDistance() && Y->isDistance()) {
456 LLVM_DEBUG(dbgs() << "\t intersect 2 distances\n");
457 if (isKnownPredicate(CmpInst::ICMP_EQ, X->getD(), Y->getD()))
458 return false;
459 if (isKnownPredicate(CmpInst::ICMP_NE, X->getD(), Y->getD())) {
460 X->setEmpty();
461 ++DeltaSuccesses;
462 return true;
463 }
464 // Hmmm, interesting situation.
465 // I guess if either is constant, keep it and ignore the other.
466 if (isa<SCEVConstant>(Y->getD())) {
467 *X = *Y;
468 return true;
469 }
470 return false;
471 }
472
473 // At this point, the pseudo-code in Figure 4 of the paper
474 // checks if (X->isPoint() && Y->isPoint()).
475 // This case can't occur in our implementation,
476 // since a Point can only arise as the result of intersecting
477 // two Line constraints, and the right-hand value, Y, is never
478 // the result of an intersection.
479 assert(!(X->isPoint() && Y->isPoint()) &&
480 "We shouldn't ever see X->isPoint() && Y->isPoint()");
481
482 if (X->isLine() && Y->isLine()) {
483 LLVM_DEBUG(dbgs() << "\t intersect 2 lines\n");
484 const SCEV *Prod1 = SE->getMulExpr(X->getA(), Y->getB());
485 const SCEV *Prod2 = SE->getMulExpr(X->getB(), Y->getA());
486 if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2)) {
487 // slopes are equal, so lines are parallel
488 LLVM_DEBUG(dbgs() << "\t\tsame slope\n");
489 Prod1 = SE->getMulExpr(X->getC(), Y->getB());
490 Prod2 = SE->getMulExpr(X->getB(), Y->getC());
491 if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2))
492 return false;
493 if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) {
494 X->setEmpty();
495 ++DeltaSuccesses;
496 return true;
497 }
498 return false;
499 }
500 if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) {
501 // slopes differ, so lines intersect
502 LLVM_DEBUG(dbgs() << "\t\tdifferent slopes\n");
503 const SCEV *C1B2 = SE->getMulExpr(X->getC(), Y->getB());
504 const SCEV *C1A2 = SE->getMulExpr(X->getC(), Y->getA());
505 const SCEV *C2B1 = SE->getMulExpr(Y->getC(), X->getB());
506 const SCEV *C2A1 = SE->getMulExpr(Y->getC(), X->getA());
507 const SCEV *A1B2 = SE->getMulExpr(X->getA(), Y->getB());
508 const SCEV *A2B1 = SE->getMulExpr(Y->getA(), X->getB());
509 const SCEVConstant *C1A2_C2A1 =
510 dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1A2, C2A1));
511 const SCEVConstant *C1B2_C2B1 =
512 dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1B2, C2B1));
513 const SCEVConstant *A1B2_A2B1 =
514 dyn_cast<SCEVConstant>(SE->getMinusSCEV(A1B2, A2B1));
515 const SCEVConstant *A2B1_A1B2 =
516 dyn_cast<SCEVConstant>(SE->getMinusSCEV(A2B1, A1B2));
517 if (!C1B2_C2B1 || !C1A2_C2A1 ||
518 !A1B2_A2B1 || !A2B1_A1B2)
519 return false;
520 APInt Xtop = C1B2_C2B1->getAPInt();
521 APInt Xbot = A1B2_A2B1->getAPInt();
522 APInt Ytop = C1A2_C2A1->getAPInt();
523 APInt Ybot = A2B1_A1B2->getAPInt();
524 LLVM_DEBUG(dbgs() << "\t\tXtop = " << Xtop << "\n");
525 LLVM_DEBUG(dbgs() << "\t\tXbot = " << Xbot << "\n");
526 LLVM_DEBUG(dbgs() << "\t\tYtop = " << Ytop << "\n");
527 LLVM_DEBUG(dbgs() << "\t\tYbot = " << Ybot << "\n");
528 APInt Xq = Xtop; // these need to be initialized, even
529 APInt Xr = Xtop; // though they're just going to be overwritten
530 APInt::sdivrem(Xtop, Xbot, Xq, Xr);
531 APInt Yq = Ytop;
532 APInt Yr = Ytop;
533 APInt::sdivrem(Ytop, Ybot, Yq, Yr);
534 if (Xr != 0 || Yr != 0) {
535 X->setEmpty();
536 ++DeltaSuccesses;
537 return true;
538 }
539 LLVM_DEBUG(dbgs() << "\t\tX = " << Xq << ", Y = " << Yq << "\n");
540 if (Xq.slt(0) || Yq.slt(0)) {
541 X->setEmpty();
542 ++DeltaSuccesses;
543 return true;
544 }
545 if (const SCEVConstant *CUB =
546 collectConstantUpperBound(X->getAssociatedLoop(), Prod1->getType())) {
547 const APInt &UpperBound = CUB->getAPInt();
548 LLVM_DEBUG(dbgs() << "\t\tupper bound = " << UpperBound << "\n");
549 if (Xq.sgt(UpperBound) || Yq.sgt(UpperBound)) {
550 X->setEmpty();
551 ++DeltaSuccesses;
552 return true;
553 }
554 }
555 X->setPoint(SE->getConstant(Xq),
556 SE->getConstant(Yq),
557 X->getAssociatedLoop());
558 ++DeltaSuccesses;
559 return true;
560 }
561 return false;
562 }
563
564 // if (X->isLine() && Y->isPoint()) This case can't occur.
565 assert(!(X->isLine() && Y->isPoint()) && "This case should never occur");
566
567 if (X->isPoint() && Y->isLine()) {
568 LLVM_DEBUG(dbgs() << "\t intersect Point and Line\n");
569 const SCEV *A1X1 = SE->getMulExpr(Y->getA(), X->getX());
570 const SCEV *B1Y1 = SE->getMulExpr(Y->getB(), X->getY());
571 const SCEV *Sum = SE->getAddExpr(A1X1, B1Y1);
572 if (isKnownPredicate(CmpInst::ICMP_EQ, Sum, Y->getC()))
573 return false;
574 if (isKnownPredicate(CmpInst::ICMP_NE, Sum, Y->getC())) {
575 X->setEmpty();
576 ++DeltaSuccesses;
577 return true;
578 }
579 return false;
580 }
581
582 llvm_unreachable("shouldn't reach the end of Constraint intersection");
583 return false;
584 }
585
586
587 //===----------------------------------------------------------------------===//
588 // DependenceInfo methods
589
590 // For debugging purposes. Dumps a dependence to OS.
dump(raw_ostream & OS) const591 void Dependence::dump(raw_ostream &OS) const {
592 bool Splitable = false;
593 if (isConfused())
594 OS << "confused";
595 else {
596 if (isConsistent())
597 OS << "consistent ";
598 if (isFlow())
599 OS << "flow";
600 else if (isOutput())
601 OS << "output";
602 else if (isAnti())
603 OS << "anti";
604 else if (isInput())
605 OS << "input";
606 unsigned Levels = getLevels();
607 OS << " [";
608 for (unsigned II = 1; II <= Levels; ++II) {
609 if (isSplitable(II))
610 Splitable = true;
611 if (isPeelFirst(II))
612 OS << 'p';
613 const SCEV *Distance = getDistance(II);
614 if (Distance)
615 OS << *Distance;
616 else if (isScalar(II))
617 OS << "S";
618 else {
619 unsigned Direction = getDirection(II);
620 if (Direction == DVEntry::ALL)
621 OS << "*";
622 else {
623 if (Direction & DVEntry::LT)
624 OS << "<";
625 if (Direction & DVEntry::EQ)
626 OS << "=";
627 if (Direction & DVEntry::GT)
628 OS << ">";
629 }
630 }
631 if (isPeelLast(II))
632 OS << 'p';
633 if (II < Levels)
634 OS << " ";
635 }
636 if (isLoopIndependent())
637 OS << "|<";
638 OS << "]";
639 if (Splitable)
640 OS << " splitable";
641 }
642 OS << "!\n";
643 }
644
645 // Returns NoAlias/MayAliass/MustAlias for two memory locations based upon their
646 // underlaying objects. If LocA and LocB are known to not alias (for any reason:
647 // tbaa, non-overlapping regions etc), then it is known there is no dependecy.
648 // Otherwise the underlying objects are checked to see if they point to
649 // different identifiable objects.
underlyingObjectsAlias(AliasAnalysis * AA,const DataLayout & DL,const MemoryLocation & LocA,const MemoryLocation & LocB)650 static AliasResult underlyingObjectsAlias(AliasAnalysis *AA,
651 const DataLayout &DL,
652 const MemoryLocation &LocA,
653 const MemoryLocation &LocB) {
654 // Check the original locations (minus size) for noalias, which can happen for
655 // tbaa, incompatible underlying object locations, etc.
656 MemoryLocation LocAS(LocA.Ptr, LocationSize::unknown(), LocA.AATags);
657 MemoryLocation LocBS(LocB.Ptr, LocationSize::unknown(), LocB.AATags);
658 if (AA->alias(LocAS, LocBS) == NoAlias)
659 return NoAlias;
660
661 // Check the underlying objects are the same
662 const Value *AObj = GetUnderlyingObject(LocA.Ptr, DL);
663 const Value *BObj = GetUnderlyingObject(LocB.Ptr, DL);
664
665 // If the underlying objects are the same, they must alias
666 if (AObj == BObj)
667 return MustAlias;
668
669 // We may have hit the recursion limit for underlying objects, or have
670 // underlying objects where we don't know they will alias.
671 if (!isIdentifiedObject(AObj) || !isIdentifiedObject(BObj))
672 return MayAlias;
673
674 // Otherwise we know the objects are different and both identified objects so
675 // must not alias.
676 return NoAlias;
677 }
678
679
680 // Returns true if the load or store can be analyzed. Atomic and volatile
681 // operations have properties which this analysis does not understand.
682 static
isLoadOrStore(const Instruction * I)683 bool isLoadOrStore(const Instruction *I) {
684 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
685 return LI->isUnordered();
686 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
687 return SI->isUnordered();
688 return false;
689 }
690
691
692 // Examines the loop nesting of the Src and Dst
693 // instructions and establishes their shared loops. Sets the variables
694 // CommonLevels, SrcLevels, and MaxLevels.
695 // The source and destination instructions needn't be contained in the same
696 // loop. The routine establishNestingLevels finds the level of most deeply
697 // nested loop that contains them both, CommonLevels. An instruction that's
698 // not contained in a loop is at level = 0. MaxLevels is equal to the level
699 // of the source plus the level of the destination, minus CommonLevels.
700 // This lets us allocate vectors MaxLevels in length, with room for every
701 // distinct loop referenced in both the source and destination subscripts.
702 // The variable SrcLevels is the nesting depth of the source instruction.
703 // It's used to help calculate distinct loops referenced by the destination.
704 // Here's the map from loops to levels:
705 // 0 - unused
706 // 1 - outermost common loop
707 // ... - other common loops
708 // CommonLevels - innermost common loop
709 // ... - loops containing Src but not Dst
710 // SrcLevels - innermost loop containing Src but not Dst
711 // ... - loops containing Dst but not Src
712 // MaxLevels - innermost loops containing Dst but not Src
713 // Consider the follow code fragment:
714 // for (a = ...) {
715 // for (b = ...) {
716 // for (c = ...) {
717 // for (d = ...) {
718 // A[] = ...;
719 // }
720 // }
721 // for (e = ...) {
722 // for (f = ...) {
723 // for (g = ...) {
724 // ... = A[];
725 // }
726 // }
727 // }
728 // }
729 // }
730 // If we're looking at the possibility of a dependence between the store
731 // to A (the Src) and the load from A (the Dst), we'll note that they
732 // have 2 loops in common, so CommonLevels will equal 2 and the direction
733 // vector for Result will have 2 entries. SrcLevels = 4 and MaxLevels = 7.
734 // A map from loop names to loop numbers would look like
735 // a - 1
736 // b - 2 = CommonLevels
737 // c - 3
738 // d - 4 = SrcLevels
739 // e - 5
740 // f - 6
741 // g - 7 = MaxLevels
establishNestingLevels(const Instruction * Src,const Instruction * Dst)742 void DependenceInfo::establishNestingLevels(const Instruction *Src,
743 const Instruction *Dst) {
744 const BasicBlock *SrcBlock = Src->getParent();
745 const BasicBlock *DstBlock = Dst->getParent();
746 unsigned SrcLevel = LI->getLoopDepth(SrcBlock);
747 unsigned DstLevel = LI->getLoopDepth(DstBlock);
748 const Loop *SrcLoop = LI->getLoopFor(SrcBlock);
749 const Loop *DstLoop = LI->getLoopFor(DstBlock);
750 SrcLevels = SrcLevel;
751 MaxLevels = SrcLevel + DstLevel;
752 while (SrcLevel > DstLevel) {
753 SrcLoop = SrcLoop->getParentLoop();
754 SrcLevel--;
755 }
756 while (DstLevel > SrcLevel) {
757 DstLoop = DstLoop->getParentLoop();
758 DstLevel--;
759 }
760 while (SrcLoop != DstLoop) {
761 SrcLoop = SrcLoop->getParentLoop();
762 DstLoop = DstLoop->getParentLoop();
763 SrcLevel--;
764 }
765 CommonLevels = SrcLevel;
766 MaxLevels -= CommonLevels;
767 }
768
769
770 // Given one of the loops containing the source, return
771 // its level index in our numbering scheme.
mapSrcLoop(const Loop * SrcLoop) const772 unsigned DependenceInfo::mapSrcLoop(const Loop *SrcLoop) const {
773 return SrcLoop->getLoopDepth();
774 }
775
776
777 // Given one of the loops containing the destination,
778 // return its level index in our numbering scheme.
mapDstLoop(const Loop * DstLoop) const779 unsigned DependenceInfo::mapDstLoop(const Loop *DstLoop) const {
780 unsigned D = DstLoop->getLoopDepth();
781 if (D > CommonLevels)
782 return D - CommonLevels + SrcLevels;
783 else
784 return D;
785 }
786
787
788 // Returns true if Expression is loop invariant in LoopNest.
isLoopInvariant(const SCEV * Expression,const Loop * LoopNest) const789 bool DependenceInfo::isLoopInvariant(const SCEV *Expression,
790 const Loop *LoopNest) const {
791 if (!LoopNest)
792 return true;
793 return SE->isLoopInvariant(Expression, LoopNest) &&
794 isLoopInvariant(Expression, LoopNest->getParentLoop());
795 }
796
797
798
799 // Finds the set of loops from the LoopNest that
800 // have a level <= CommonLevels and are referred to by the SCEV Expression.
collectCommonLoops(const SCEV * Expression,const Loop * LoopNest,SmallBitVector & Loops) const801 void DependenceInfo::collectCommonLoops(const SCEV *Expression,
802 const Loop *LoopNest,
803 SmallBitVector &Loops) const {
804 while (LoopNest) {
805 unsigned Level = LoopNest->getLoopDepth();
806 if (Level <= CommonLevels && !SE->isLoopInvariant(Expression, LoopNest))
807 Loops.set(Level);
808 LoopNest = LoopNest->getParentLoop();
809 }
810 }
811
unifySubscriptType(ArrayRef<Subscript * > Pairs)812 void DependenceInfo::unifySubscriptType(ArrayRef<Subscript *> Pairs) {
813
814 unsigned widestWidthSeen = 0;
815 Type *widestType;
816
817 // Go through each pair and find the widest bit to which we need
818 // to extend all of them.
819 for (Subscript *Pair : Pairs) {
820 const SCEV *Src = Pair->Src;
821 const SCEV *Dst = Pair->Dst;
822 IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
823 IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
824 if (SrcTy == nullptr || DstTy == nullptr) {
825 assert(SrcTy == DstTy && "This function only unify integer types and "
826 "expect Src and Dst share the same type "
827 "otherwise.");
828 continue;
829 }
830 if (SrcTy->getBitWidth() > widestWidthSeen) {
831 widestWidthSeen = SrcTy->getBitWidth();
832 widestType = SrcTy;
833 }
834 if (DstTy->getBitWidth() > widestWidthSeen) {
835 widestWidthSeen = DstTy->getBitWidth();
836 widestType = DstTy;
837 }
838 }
839
840
841 assert(widestWidthSeen > 0);
842
843 // Now extend each pair to the widest seen.
844 for (Subscript *Pair : Pairs) {
845 const SCEV *Src = Pair->Src;
846 const SCEV *Dst = Pair->Dst;
847 IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
848 IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
849 if (SrcTy == nullptr || DstTy == nullptr) {
850 assert(SrcTy == DstTy && "This function only unify integer types and "
851 "expect Src and Dst share the same type "
852 "otherwise.");
853 continue;
854 }
855 if (SrcTy->getBitWidth() < widestWidthSeen)
856 // Sign-extend Src to widestType
857 Pair->Src = SE->getSignExtendExpr(Src, widestType);
858 if (DstTy->getBitWidth() < widestWidthSeen) {
859 // Sign-extend Dst to widestType
860 Pair->Dst = SE->getSignExtendExpr(Dst, widestType);
861 }
862 }
863 }
864
865 // removeMatchingExtensions - Examines a subscript pair.
866 // If the source and destination are identically sign (or zero)
867 // extended, it strips off the extension in an effect to simplify
868 // the actual analysis.
removeMatchingExtensions(Subscript * Pair)869 void DependenceInfo::removeMatchingExtensions(Subscript *Pair) {
870 const SCEV *Src = Pair->Src;
871 const SCEV *Dst = Pair->Dst;
872 if ((isa<SCEVZeroExtendExpr>(Src) && isa<SCEVZeroExtendExpr>(Dst)) ||
873 (isa<SCEVSignExtendExpr>(Src) && isa<SCEVSignExtendExpr>(Dst))) {
874 const SCEVCastExpr *SrcCast = cast<SCEVCastExpr>(Src);
875 const SCEVCastExpr *DstCast = cast<SCEVCastExpr>(Dst);
876 const SCEV *SrcCastOp = SrcCast->getOperand();
877 const SCEV *DstCastOp = DstCast->getOperand();
878 if (SrcCastOp->getType() == DstCastOp->getType()) {
879 Pair->Src = SrcCastOp;
880 Pair->Dst = DstCastOp;
881 }
882 }
883 }
884
885 // Examine the scev and return true iff it's linear.
886 // Collect any loops mentioned in the set of "Loops".
checkSubscript(const SCEV * Expr,const Loop * LoopNest,SmallBitVector & Loops,bool IsSrc)887 bool DependenceInfo::checkSubscript(const SCEV *Expr, const Loop *LoopNest,
888 SmallBitVector &Loops, bool IsSrc) {
889 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
890 if (!AddRec)
891 return isLoopInvariant(Expr, LoopNest);
892 const SCEV *Start = AddRec->getStart();
893 const SCEV *Step = AddRec->getStepRecurrence(*SE);
894 const SCEV *UB = SE->getBackedgeTakenCount(AddRec->getLoop());
895 if (!isa<SCEVCouldNotCompute>(UB)) {
896 if (SE->getTypeSizeInBits(Start->getType()) <
897 SE->getTypeSizeInBits(UB->getType())) {
898 if (!AddRec->getNoWrapFlags())
899 return false;
900 }
901 }
902 if (!isLoopInvariant(Step, LoopNest))
903 return false;
904 if (IsSrc)
905 Loops.set(mapSrcLoop(AddRec->getLoop()));
906 else
907 Loops.set(mapDstLoop(AddRec->getLoop()));
908 return checkSubscript(Start, LoopNest, Loops, IsSrc);
909 }
910
911 // Examine the scev and return true iff it's linear.
912 // Collect any loops mentioned in the set of "Loops".
checkSrcSubscript(const SCEV * Src,const Loop * LoopNest,SmallBitVector & Loops)913 bool DependenceInfo::checkSrcSubscript(const SCEV *Src, const Loop *LoopNest,
914 SmallBitVector &Loops) {
915 return checkSubscript(Src, LoopNest, Loops, true);
916 }
917
918 // Examine the scev and return true iff it's linear.
919 // Collect any loops mentioned in the set of "Loops".
checkDstSubscript(const SCEV * Dst,const Loop * LoopNest,SmallBitVector & Loops)920 bool DependenceInfo::checkDstSubscript(const SCEV *Dst, const Loop *LoopNest,
921 SmallBitVector &Loops) {
922 return checkSubscript(Dst, LoopNest, Loops, false);
923 }
924
925
926 // Examines the subscript pair (the Src and Dst SCEVs)
927 // and classifies it as either ZIV, SIV, RDIV, MIV, or Nonlinear.
928 // Collects the associated loops in a set.
929 DependenceInfo::Subscript::ClassificationKind
classifyPair(const SCEV * Src,const Loop * SrcLoopNest,const SCEV * Dst,const Loop * DstLoopNest,SmallBitVector & Loops)930 DependenceInfo::classifyPair(const SCEV *Src, const Loop *SrcLoopNest,
931 const SCEV *Dst, const Loop *DstLoopNest,
932 SmallBitVector &Loops) {
933 SmallBitVector SrcLoops(MaxLevels + 1);
934 SmallBitVector DstLoops(MaxLevels + 1);
935 if (!checkSrcSubscript(Src, SrcLoopNest, SrcLoops))
936 return Subscript::NonLinear;
937 if (!checkDstSubscript(Dst, DstLoopNest, DstLoops))
938 return Subscript::NonLinear;
939 Loops = SrcLoops;
940 Loops |= DstLoops;
941 unsigned N = Loops.count();
942 if (N == 0)
943 return Subscript::ZIV;
944 if (N == 1)
945 return Subscript::SIV;
946 if (N == 2 && (SrcLoops.count() == 0 ||
947 DstLoops.count() == 0 ||
948 (SrcLoops.count() == 1 && DstLoops.count() == 1)))
949 return Subscript::RDIV;
950 return Subscript::MIV;
951 }
952
953
954 // A wrapper around SCEV::isKnownPredicate.
955 // Looks for cases where we're interested in comparing for equality.
956 // If both X and Y have been identically sign or zero extended,
957 // it strips off the (confusing) extensions before invoking
958 // SCEV::isKnownPredicate. Perhaps, someday, the ScalarEvolution package
959 // will be similarly updated.
960 //
961 // If SCEV::isKnownPredicate can't prove the predicate,
962 // we try simple subtraction, which seems to help in some cases
963 // involving symbolics.
isKnownPredicate(ICmpInst::Predicate Pred,const SCEV * X,const SCEV * Y) const964 bool DependenceInfo::isKnownPredicate(ICmpInst::Predicate Pred, const SCEV *X,
965 const SCEV *Y) const {
966 if (Pred == CmpInst::ICMP_EQ ||
967 Pred == CmpInst::ICMP_NE) {
968 if ((isa<SCEVSignExtendExpr>(X) &&
969 isa<SCEVSignExtendExpr>(Y)) ||
970 (isa<SCEVZeroExtendExpr>(X) &&
971 isa<SCEVZeroExtendExpr>(Y))) {
972 const SCEVCastExpr *CX = cast<SCEVCastExpr>(X);
973 const SCEVCastExpr *CY = cast<SCEVCastExpr>(Y);
974 const SCEV *Xop = CX->getOperand();
975 const SCEV *Yop = CY->getOperand();
976 if (Xop->getType() == Yop->getType()) {
977 X = Xop;
978 Y = Yop;
979 }
980 }
981 }
982 if (SE->isKnownPredicate(Pred, X, Y))
983 return true;
984 // If SE->isKnownPredicate can't prove the condition,
985 // we try the brute-force approach of subtracting
986 // and testing the difference.
987 // By testing with SE->isKnownPredicate first, we avoid
988 // the possibility of overflow when the arguments are constants.
989 const SCEV *Delta = SE->getMinusSCEV(X, Y);
990 switch (Pred) {
991 case CmpInst::ICMP_EQ:
992 return Delta->isZero();
993 case CmpInst::ICMP_NE:
994 return SE->isKnownNonZero(Delta);
995 case CmpInst::ICMP_SGE:
996 return SE->isKnownNonNegative(Delta);
997 case CmpInst::ICMP_SLE:
998 return SE->isKnownNonPositive(Delta);
999 case CmpInst::ICMP_SGT:
1000 return SE->isKnownPositive(Delta);
1001 case CmpInst::ICMP_SLT:
1002 return SE->isKnownNegative(Delta);
1003 default:
1004 llvm_unreachable("unexpected predicate in isKnownPredicate");
1005 }
1006 }
1007
1008 /// Compare to see if S is less than Size, using isKnownNegative(S - max(Size, 1))
1009 /// with some extra checking if S is an AddRec and we can prove less-than using
1010 /// the loop bounds.
isKnownLessThan(const SCEV * S,const SCEV * Size) const1011 bool DependenceInfo::isKnownLessThan(const SCEV *S, const SCEV *Size) const {
1012 // First unify to the same type
1013 auto *SType = dyn_cast<IntegerType>(S->getType());
1014 auto *SizeType = dyn_cast<IntegerType>(Size->getType());
1015 if (!SType || !SizeType)
1016 return false;
1017 Type *MaxType =
1018 (SType->getBitWidth() >= SizeType->getBitWidth()) ? SType : SizeType;
1019 S = SE->getTruncateOrZeroExtend(S, MaxType);
1020 Size = SE->getTruncateOrZeroExtend(Size, MaxType);
1021
1022 // Special check for addrecs using BE taken count
1023 const SCEV *Bound = SE->getMinusSCEV(S, Size);
1024 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Bound)) {
1025 if (AddRec->isAffine()) {
1026 const SCEV *BECount = SE->getBackedgeTakenCount(AddRec->getLoop());
1027 if (!isa<SCEVCouldNotCompute>(BECount)) {
1028 const SCEV *Limit = AddRec->evaluateAtIteration(BECount, *SE);
1029 if (SE->isKnownNegative(Limit))
1030 return true;
1031 }
1032 }
1033 }
1034
1035 // Check using normal isKnownNegative
1036 const SCEV *LimitedBound =
1037 SE->getMinusSCEV(S, SE->getSMaxExpr(Size, SE->getOne(Size->getType())));
1038 return SE->isKnownNegative(LimitedBound);
1039 }
1040
isKnownNonNegative(const SCEV * S,const Value * Ptr) const1041 bool DependenceInfo::isKnownNonNegative(const SCEV *S, const Value *Ptr) const {
1042 bool Inbounds = false;
1043 if (auto *SrcGEP = dyn_cast<GetElementPtrInst>(Ptr))
1044 Inbounds = SrcGEP->isInBounds();
1045 if (Inbounds) {
1046 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
1047 if (AddRec->isAffine()) {
1048 // We know S is for Ptr, the operand on a load/store, so doesn't wrap.
1049 // If both parts are NonNegative, the end result will be NonNegative
1050 if (SE->isKnownNonNegative(AddRec->getStart()) &&
1051 SE->isKnownNonNegative(AddRec->getOperand(1)))
1052 return true;
1053 }
1054 }
1055 }
1056
1057 return SE->isKnownNonNegative(S);
1058 }
1059
1060 // All subscripts are all the same type.
1061 // Loop bound may be smaller (e.g., a char).
1062 // Should zero extend loop bound, since it's always >= 0.
1063 // This routine collects upper bound and extends or truncates if needed.
1064 // Truncating is safe when subscripts are known not to wrap. Cases without
1065 // nowrap flags should have been rejected earlier.
1066 // Return null if no bound available.
collectUpperBound(const Loop * L,Type * T) const1067 const SCEV *DependenceInfo::collectUpperBound(const Loop *L, Type *T) const {
1068 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
1069 const SCEV *UB = SE->getBackedgeTakenCount(L);
1070 return SE->getTruncateOrZeroExtend(UB, T);
1071 }
1072 return nullptr;
1073 }
1074
1075
1076 // Calls collectUpperBound(), then attempts to cast it to SCEVConstant.
1077 // If the cast fails, returns NULL.
collectConstantUpperBound(const Loop * L,Type * T) const1078 const SCEVConstant *DependenceInfo::collectConstantUpperBound(const Loop *L,
1079 Type *T) const {
1080 if (const SCEV *UB = collectUpperBound(L, T))
1081 return dyn_cast<SCEVConstant>(UB);
1082 return nullptr;
1083 }
1084
1085
1086 // testZIV -
1087 // When we have a pair of subscripts of the form [c1] and [c2],
1088 // where c1 and c2 are both loop invariant, we attack it using
1089 // the ZIV test. Basically, we test by comparing the two values,
1090 // but there are actually three possible results:
1091 // 1) the values are equal, so there's a dependence
1092 // 2) the values are different, so there's no dependence
1093 // 3) the values might be equal, so we have to assume a dependence.
1094 //
1095 // Return true if dependence disproved.
testZIV(const SCEV * Src,const SCEV * Dst,FullDependence & Result) const1096 bool DependenceInfo::testZIV(const SCEV *Src, const SCEV *Dst,
1097 FullDependence &Result) const {
1098 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
1099 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
1100 ++ZIVapplications;
1101 if (isKnownPredicate(CmpInst::ICMP_EQ, Src, Dst)) {
1102 LLVM_DEBUG(dbgs() << " provably dependent\n");
1103 return false; // provably dependent
1104 }
1105 if (isKnownPredicate(CmpInst::ICMP_NE, Src, Dst)) {
1106 LLVM_DEBUG(dbgs() << " provably independent\n");
1107 ++ZIVindependence;
1108 return true; // provably independent
1109 }
1110 LLVM_DEBUG(dbgs() << " possibly dependent\n");
1111 Result.Consistent = false;
1112 return false; // possibly dependent
1113 }
1114
1115
1116 // strongSIVtest -
1117 // From the paper, Practical Dependence Testing, Section 4.2.1
1118 //
1119 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 + a*i],
1120 // where i is an induction variable, c1 and c2 are loop invariant,
1121 // and a is a constant, we can solve it exactly using the Strong SIV test.
1122 //
1123 // Can prove independence. Failing that, can compute distance (and direction).
1124 // In the presence of symbolic terms, we can sometimes make progress.
1125 //
1126 // If there's a dependence,
1127 //
1128 // c1 + a*i = c2 + a*i'
1129 //
1130 // The dependence distance is
1131 //
1132 // d = i' - i = (c1 - c2)/a
1133 //
1134 // A dependence only exists if d is an integer and abs(d) <= U, where U is the
1135 // loop's upper bound. If a dependence exists, the dependence direction is
1136 // defined as
1137 //
1138 // { < if d > 0
1139 // direction = { = if d = 0
1140 // { > if d < 0
1141 //
1142 // Return true if dependence disproved.
strongSIVtest(const SCEV * Coeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1143 bool DependenceInfo::strongSIVtest(const SCEV *Coeff, const SCEV *SrcConst,
1144 const SCEV *DstConst, const Loop *CurLoop,
1145 unsigned Level, FullDependence &Result,
1146 Constraint &NewConstraint) const {
1147 LLVM_DEBUG(dbgs() << "\tStrong SIV test\n");
1148 LLVM_DEBUG(dbgs() << "\t Coeff = " << *Coeff);
1149 LLVM_DEBUG(dbgs() << ", " << *Coeff->getType() << "\n");
1150 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst);
1151 LLVM_DEBUG(dbgs() << ", " << *SrcConst->getType() << "\n");
1152 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst);
1153 LLVM_DEBUG(dbgs() << ", " << *DstConst->getType() << "\n");
1154 ++StrongSIVapplications;
1155 assert(0 < Level && Level <= CommonLevels && "level out of range");
1156 Level--;
1157
1158 const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst);
1159 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta);
1160 LLVM_DEBUG(dbgs() << ", " << *Delta->getType() << "\n");
1161
1162 // check that |Delta| < iteration count
1163 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1164 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound);
1165 LLVM_DEBUG(dbgs() << ", " << *UpperBound->getType() << "\n");
1166 const SCEV *AbsDelta =
1167 SE->isKnownNonNegative(Delta) ? Delta : SE->getNegativeSCEV(Delta);
1168 const SCEV *AbsCoeff =
1169 SE->isKnownNonNegative(Coeff) ? Coeff : SE->getNegativeSCEV(Coeff);
1170 const SCEV *Product = SE->getMulExpr(UpperBound, AbsCoeff);
1171 if (isKnownPredicate(CmpInst::ICMP_SGT, AbsDelta, Product)) {
1172 // Distance greater than trip count - no dependence
1173 ++StrongSIVindependence;
1174 ++StrongSIVsuccesses;
1175 return true;
1176 }
1177 }
1178
1179 // Can we compute distance?
1180 if (isa<SCEVConstant>(Delta) && isa<SCEVConstant>(Coeff)) {
1181 APInt ConstDelta = cast<SCEVConstant>(Delta)->getAPInt();
1182 APInt ConstCoeff = cast<SCEVConstant>(Coeff)->getAPInt();
1183 APInt Distance = ConstDelta; // these need to be initialized
1184 APInt Remainder = ConstDelta;
1185 APInt::sdivrem(ConstDelta, ConstCoeff, Distance, Remainder);
1186 LLVM_DEBUG(dbgs() << "\t Distance = " << Distance << "\n");
1187 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n");
1188 // Make sure Coeff divides Delta exactly
1189 if (Remainder != 0) {
1190 // Coeff doesn't divide Distance, no dependence
1191 ++StrongSIVindependence;
1192 ++StrongSIVsuccesses;
1193 return true;
1194 }
1195 Result.DV[Level].Distance = SE->getConstant(Distance);
1196 NewConstraint.setDistance(SE->getConstant(Distance), CurLoop);
1197 if (Distance.sgt(0))
1198 Result.DV[Level].Direction &= Dependence::DVEntry::LT;
1199 else if (Distance.slt(0))
1200 Result.DV[Level].Direction &= Dependence::DVEntry::GT;
1201 else
1202 Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
1203 ++StrongSIVsuccesses;
1204 }
1205 else if (Delta->isZero()) {
1206 // since 0/X == 0
1207 Result.DV[Level].Distance = Delta;
1208 NewConstraint.setDistance(Delta, CurLoop);
1209 Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
1210 ++StrongSIVsuccesses;
1211 }
1212 else {
1213 if (Coeff->isOne()) {
1214 LLVM_DEBUG(dbgs() << "\t Distance = " << *Delta << "\n");
1215 Result.DV[Level].Distance = Delta; // since X/1 == X
1216 NewConstraint.setDistance(Delta, CurLoop);
1217 }
1218 else {
1219 Result.Consistent = false;
1220 NewConstraint.setLine(Coeff,
1221 SE->getNegativeSCEV(Coeff),
1222 SE->getNegativeSCEV(Delta), CurLoop);
1223 }
1224
1225 // maybe we can get a useful direction
1226 bool DeltaMaybeZero = !SE->isKnownNonZero(Delta);
1227 bool DeltaMaybePositive = !SE->isKnownNonPositive(Delta);
1228 bool DeltaMaybeNegative = !SE->isKnownNonNegative(Delta);
1229 bool CoeffMaybePositive = !SE->isKnownNonPositive(Coeff);
1230 bool CoeffMaybeNegative = !SE->isKnownNonNegative(Coeff);
1231 // The double negatives above are confusing.
1232 // It helps to read !SE->isKnownNonZero(Delta)
1233 // as "Delta might be Zero"
1234 unsigned NewDirection = Dependence::DVEntry::NONE;
1235 if ((DeltaMaybePositive && CoeffMaybePositive) ||
1236 (DeltaMaybeNegative && CoeffMaybeNegative))
1237 NewDirection = Dependence::DVEntry::LT;
1238 if (DeltaMaybeZero)
1239 NewDirection |= Dependence::DVEntry::EQ;
1240 if ((DeltaMaybeNegative && CoeffMaybePositive) ||
1241 (DeltaMaybePositive && CoeffMaybeNegative))
1242 NewDirection |= Dependence::DVEntry::GT;
1243 if (NewDirection < Result.DV[Level].Direction)
1244 ++StrongSIVsuccesses;
1245 Result.DV[Level].Direction &= NewDirection;
1246 }
1247 return false;
1248 }
1249
1250
1251 // weakCrossingSIVtest -
1252 // From the paper, Practical Dependence Testing, Section 4.2.2
1253 //
1254 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 - a*i],
1255 // where i is an induction variable, c1 and c2 are loop invariant,
1256 // and a is a constant, we can solve it exactly using the
1257 // Weak-Crossing SIV test.
1258 //
1259 // Given c1 + a*i = c2 - a*i', we can look for the intersection of
1260 // the two lines, where i = i', yielding
1261 //
1262 // c1 + a*i = c2 - a*i
1263 // 2a*i = c2 - c1
1264 // i = (c2 - c1)/2a
1265 //
1266 // If i < 0, there is no dependence.
1267 // If i > upperbound, there is no dependence.
1268 // If i = 0 (i.e., if c1 = c2), there's a dependence with distance = 0.
1269 // If i = upperbound, there's a dependence with distance = 0.
1270 // If i is integral, there's a dependence (all directions).
1271 // If the non-integer part = 1/2, there's a dependence (<> directions).
1272 // Otherwise, there's no dependence.
1273 //
1274 // Can prove independence. Failing that,
1275 // can sometimes refine the directions.
1276 // Can determine iteration for splitting.
1277 //
1278 // Return true if dependence disproved.
weakCrossingSIVtest(const SCEV * Coeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint,const SCEV * & SplitIter) const1279 bool DependenceInfo::weakCrossingSIVtest(
1280 const SCEV *Coeff, const SCEV *SrcConst, const SCEV *DstConst,
1281 const Loop *CurLoop, unsigned Level, FullDependence &Result,
1282 Constraint &NewConstraint, const SCEV *&SplitIter) const {
1283 LLVM_DEBUG(dbgs() << "\tWeak-Crossing SIV test\n");
1284 LLVM_DEBUG(dbgs() << "\t Coeff = " << *Coeff << "\n");
1285 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1286 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1287 ++WeakCrossingSIVapplications;
1288 assert(0 < Level && Level <= CommonLevels && "Level out of range");
1289 Level--;
1290 Result.Consistent = false;
1291 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1292 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1293 NewConstraint.setLine(Coeff, Coeff, Delta, CurLoop);
1294 if (Delta->isZero()) {
1295 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT);
1296 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT);
1297 ++WeakCrossingSIVsuccesses;
1298 if (!Result.DV[Level].Direction) {
1299 ++WeakCrossingSIVindependence;
1300 return true;
1301 }
1302 Result.DV[Level].Distance = Delta; // = 0
1303 return false;
1304 }
1305 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(Coeff);
1306 if (!ConstCoeff)
1307 return false;
1308
1309 Result.DV[Level].Splitable = true;
1310 if (SE->isKnownNegative(ConstCoeff)) {
1311 ConstCoeff = dyn_cast<SCEVConstant>(SE->getNegativeSCEV(ConstCoeff));
1312 assert(ConstCoeff &&
1313 "dynamic cast of negative of ConstCoeff should yield constant");
1314 Delta = SE->getNegativeSCEV(Delta);
1315 }
1316 assert(SE->isKnownPositive(ConstCoeff) && "ConstCoeff should be positive");
1317
1318 // compute SplitIter for use by DependenceInfo::getSplitIteration()
1319 SplitIter = SE->getUDivExpr(
1320 SE->getSMaxExpr(SE->getZero(Delta->getType()), Delta),
1321 SE->getMulExpr(SE->getConstant(Delta->getType(), 2), ConstCoeff));
1322 LLVM_DEBUG(dbgs() << "\t Split iter = " << *SplitIter << "\n");
1323
1324 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1325 if (!ConstDelta)
1326 return false;
1327
1328 // We're certain that ConstCoeff > 0; therefore,
1329 // if Delta < 0, then no dependence.
1330 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1331 LLVM_DEBUG(dbgs() << "\t ConstCoeff = " << *ConstCoeff << "\n");
1332 if (SE->isKnownNegative(Delta)) {
1333 // No dependence, Delta < 0
1334 ++WeakCrossingSIVindependence;
1335 ++WeakCrossingSIVsuccesses;
1336 return true;
1337 }
1338
1339 // We're certain that Delta > 0 and ConstCoeff > 0.
1340 // Check Delta/(2*ConstCoeff) against upper loop bound
1341 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1342 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n");
1343 const SCEV *ConstantTwo = SE->getConstant(UpperBound->getType(), 2);
1344 const SCEV *ML = SE->getMulExpr(SE->getMulExpr(ConstCoeff, UpperBound),
1345 ConstantTwo);
1346 LLVM_DEBUG(dbgs() << "\t ML = " << *ML << "\n");
1347 if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, ML)) {
1348 // Delta too big, no dependence
1349 ++WeakCrossingSIVindependence;
1350 ++WeakCrossingSIVsuccesses;
1351 return true;
1352 }
1353 if (isKnownPredicate(CmpInst::ICMP_EQ, Delta, ML)) {
1354 // i = i' = UB
1355 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT);
1356 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT);
1357 ++WeakCrossingSIVsuccesses;
1358 if (!Result.DV[Level].Direction) {
1359 ++WeakCrossingSIVindependence;
1360 return true;
1361 }
1362 Result.DV[Level].Splitable = false;
1363 Result.DV[Level].Distance = SE->getZero(Delta->getType());
1364 return false;
1365 }
1366 }
1367
1368 // check that Coeff divides Delta
1369 APInt APDelta = ConstDelta->getAPInt();
1370 APInt APCoeff = ConstCoeff->getAPInt();
1371 APInt Distance = APDelta; // these need to be initialzed
1372 APInt Remainder = APDelta;
1373 APInt::sdivrem(APDelta, APCoeff, Distance, Remainder);
1374 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n");
1375 if (Remainder != 0) {
1376 // Coeff doesn't divide Delta, no dependence
1377 ++WeakCrossingSIVindependence;
1378 ++WeakCrossingSIVsuccesses;
1379 return true;
1380 }
1381 LLVM_DEBUG(dbgs() << "\t Distance = " << Distance << "\n");
1382
1383 // if 2*Coeff doesn't divide Delta, then the equal direction isn't possible
1384 APInt Two = APInt(Distance.getBitWidth(), 2, true);
1385 Remainder = Distance.srem(Two);
1386 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n");
1387 if (Remainder != 0) {
1388 // Equal direction isn't possible
1389 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::EQ);
1390 ++WeakCrossingSIVsuccesses;
1391 }
1392 return false;
1393 }
1394
1395
1396 // Kirch's algorithm, from
1397 //
1398 // Optimizing Supercompilers for Supercomputers
1399 // Michael Wolfe
1400 // MIT Press, 1989
1401 //
1402 // Program 2.1, page 29.
1403 // Computes the GCD of AM and BM.
1404 // Also finds a solution to the equation ax - by = gcd(a, b).
1405 // Returns true if dependence disproved; i.e., gcd does not divide Delta.
findGCD(unsigned Bits,const APInt & AM,const APInt & BM,const APInt & Delta,APInt & G,APInt & X,APInt & Y)1406 static bool findGCD(unsigned Bits, const APInt &AM, const APInt &BM,
1407 const APInt &Delta, APInt &G, APInt &X, APInt &Y) {
1408 APInt A0(Bits, 1, true), A1(Bits, 0, true);
1409 APInt B0(Bits, 0, true), B1(Bits, 1, true);
1410 APInt G0 = AM.abs();
1411 APInt G1 = BM.abs();
1412 APInt Q = G0; // these need to be initialized
1413 APInt R = G0;
1414 APInt::sdivrem(G0, G1, Q, R);
1415 while (R != 0) {
1416 APInt A2 = A0 - Q*A1; A0 = A1; A1 = A2;
1417 APInt B2 = B0 - Q*B1; B0 = B1; B1 = B2;
1418 G0 = G1; G1 = R;
1419 APInt::sdivrem(G0, G1, Q, R);
1420 }
1421 G = G1;
1422 LLVM_DEBUG(dbgs() << "\t GCD = " << G << "\n");
1423 X = AM.slt(0) ? -A1 : A1;
1424 Y = BM.slt(0) ? B1 : -B1;
1425
1426 // make sure gcd divides Delta
1427 R = Delta.srem(G);
1428 if (R != 0)
1429 return true; // gcd doesn't divide Delta, no dependence
1430 Q = Delta.sdiv(G);
1431 X *= Q;
1432 Y *= Q;
1433 return false;
1434 }
1435
floorOfQuotient(const APInt & A,const APInt & B)1436 static APInt floorOfQuotient(const APInt &A, const APInt &B) {
1437 APInt Q = A; // these need to be initialized
1438 APInt R = A;
1439 APInt::sdivrem(A, B, Q, R);
1440 if (R == 0)
1441 return Q;
1442 if ((A.sgt(0) && B.sgt(0)) ||
1443 (A.slt(0) && B.slt(0)))
1444 return Q;
1445 else
1446 return Q - 1;
1447 }
1448
ceilingOfQuotient(const APInt & A,const APInt & B)1449 static APInt ceilingOfQuotient(const APInt &A, const APInt &B) {
1450 APInt Q = A; // these need to be initialized
1451 APInt R = A;
1452 APInt::sdivrem(A, B, Q, R);
1453 if (R == 0)
1454 return Q;
1455 if ((A.sgt(0) && B.sgt(0)) ||
1456 (A.slt(0) && B.slt(0)))
1457 return Q + 1;
1458 else
1459 return Q;
1460 }
1461
1462
1463 static
maxAPInt(APInt A,APInt B)1464 APInt maxAPInt(APInt A, APInt B) {
1465 return A.sgt(B) ? A : B;
1466 }
1467
1468
1469 static
minAPInt(APInt A,APInt B)1470 APInt minAPInt(APInt A, APInt B) {
1471 return A.slt(B) ? A : B;
1472 }
1473
1474
1475 // exactSIVtest -
1476 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*i],
1477 // where i is an induction variable, c1 and c2 are loop invariant, and a1
1478 // and a2 are constant, we can solve it exactly using an algorithm developed
1479 // by Banerjee and Wolfe. See Section 2.5.3 in
1480 //
1481 // Optimizing Supercompilers for Supercomputers
1482 // Michael Wolfe
1483 // MIT Press, 1989
1484 //
1485 // It's slower than the specialized tests (strong SIV, weak-zero SIV, etc),
1486 // so use them if possible. They're also a bit better with symbolics and,
1487 // in the case of the strong SIV test, can compute Distances.
1488 //
1489 // Return true if dependence disproved.
exactSIVtest(const SCEV * SrcCoeff,const SCEV * DstCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1490 bool DependenceInfo::exactSIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff,
1491 const SCEV *SrcConst, const SCEV *DstConst,
1492 const Loop *CurLoop, unsigned Level,
1493 FullDependence &Result,
1494 Constraint &NewConstraint) const {
1495 LLVM_DEBUG(dbgs() << "\tExact SIV test\n");
1496 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << " = AM\n");
1497 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << " = BM\n");
1498 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1499 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1500 ++ExactSIVapplications;
1501 assert(0 < Level && Level <= CommonLevels && "Level out of range");
1502 Level--;
1503 Result.Consistent = false;
1504 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1505 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1506 NewConstraint.setLine(SrcCoeff, SE->getNegativeSCEV(DstCoeff),
1507 Delta, CurLoop);
1508 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1509 const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1510 const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1511 if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff)
1512 return false;
1513
1514 // find gcd
1515 APInt G, X, Y;
1516 APInt AM = ConstSrcCoeff->getAPInt();
1517 APInt BM = ConstDstCoeff->getAPInt();
1518 unsigned Bits = AM.getBitWidth();
1519 if (findGCD(Bits, AM, BM, ConstDelta->getAPInt(), G, X, Y)) {
1520 // gcd doesn't divide Delta, no dependence
1521 ++ExactSIVindependence;
1522 ++ExactSIVsuccesses;
1523 return true;
1524 }
1525
1526 LLVM_DEBUG(dbgs() << "\t X = " << X << ", Y = " << Y << "\n");
1527
1528 // since SCEV construction normalizes, LM = 0
1529 APInt UM(Bits, 1, true);
1530 bool UMvalid = false;
1531 // UM is perhaps unavailable, let's check
1532 if (const SCEVConstant *CUB =
1533 collectConstantUpperBound(CurLoop, Delta->getType())) {
1534 UM = CUB->getAPInt();
1535 LLVM_DEBUG(dbgs() << "\t UM = " << UM << "\n");
1536 UMvalid = true;
1537 }
1538
1539 APInt TU(APInt::getSignedMaxValue(Bits));
1540 APInt TL(APInt::getSignedMinValue(Bits));
1541
1542 // test(BM/G, LM-X) and test(-BM/G, X-UM)
1543 APInt TMUL = BM.sdiv(G);
1544 if (TMUL.sgt(0)) {
1545 TL = maxAPInt(TL, ceilingOfQuotient(-X, TMUL));
1546 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1547 if (UMvalid) {
1548 TU = minAPInt(TU, floorOfQuotient(UM - X, TMUL));
1549 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1550 }
1551 }
1552 else {
1553 TU = minAPInt(TU, floorOfQuotient(-X, TMUL));
1554 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1555 if (UMvalid) {
1556 TL = maxAPInt(TL, ceilingOfQuotient(UM - X, TMUL));
1557 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1558 }
1559 }
1560
1561 // test(AM/G, LM-Y) and test(-AM/G, Y-UM)
1562 TMUL = AM.sdiv(G);
1563 if (TMUL.sgt(0)) {
1564 TL = maxAPInt(TL, ceilingOfQuotient(-Y, TMUL));
1565 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1566 if (UMvalid) {
1567 TU = minAPInt(TU, floorOfQuotient(UM - Y, TMUL));
1568 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1569 }
1570 }
1571 else {
1572 TU = minAPInt(TU, floorOfQuotient(-Y, TMUL));
1573 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1574 if (UMvalid) {
1575 TL = maxAPInt(TL, ceilingOfQuotient(UM - Y, TMUL));
1576 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1577 }
1578 }
1579 if (TL.sgt(TU)) {
1580 ++ExactSIVindependence;
1581 ++ExactSIVsuccesses;
1582 return true;
1583 }
1584
1585 // explore directions
1586 unsigned NewDirection = Dependence::DVEntry::NONE;
1587
1588 // less than
1589 APInt SaveTU(TU); // save these
1590 APInt SaveTL(TL);
1591 LLVM_DEBUG(dbgs() << "\t exploring LT direction\n");
1592 TMUL = AM - BM;
1593 if (TMUL.sgt(0)) {
1594 TL = maxAPInt(TL, ceilingOfQuotient(X - Y + 1, TMUL));
1595 LLVM_DEBUG(dbgs() << "\t\t TL = " << TL << "\n");
1596 }
1597 else {
1598 TU = minAPInt(TU, floorOfQuotient(X - Y + 1, TMUL));
1599 LLVM_DEBUG(dbgs() << "\t\t TU = " << TU << "\n");
1600 }
1601 if (TL.sle(TU)) {
1602 NewDirection |= Dependence::DVEntry::LT;
1603 ++ExactSIVsuccesses;
1604 }
1605
1606 // equal
1607 TU = SaveTU; // restore
1608 TL = SaveTL;
1609 LLVM_DEBUG(dbgs() << "\t exploring EQ direction\n");
1610 if (TMUL.sgt(0)) {
1611 TL = maxAPInt(TL, ceilingOfQuotient(X - Y, TMUL));
1612 LLVM_DEBUG(dbgs() << "\t\t TL = " << TL << "\n");
1613 }
1614 else {
1615 TU = minAPInt(TU, floorOfQuotient(X - Y, TMUL));
1616 LLVM_DEBUG(dbgs() << "\t\t TU = " << TU << "\n");
1617 }
1618 TMUL = BM - AM;
1619 if (TMUL.sgt(0)) {
1620 TL = maxAPInt(TL, ceilingOfQuotient(Y - X, TMUL));
1621 LLVM_DEBUG(dbgs() << "\t\t TL = " << TL << "\n");
1622 }
1623 else {
1624 TU = minAPInt(TU, floorOfQuotient(Y - X, TMUL));
1625 LLVM_DEBUG(dbgs() << "\t\t TU = " << TU << "\n");
1626 }
1627 if (TL.sle(TU)) {
1628 NewDirection |= Dependence::DVEntry::EQ;
1629 ++ExactSIVsuccesses;
1630 }
1631
1632 // greater than
1633 TU = SaveTU; // restore
1634 TL = SaveTL;
1635 LLVM_DEBUG(dbgs() << "\t exploring GT direction\n");
1636 if (TMUL.sgt(0)) {
1637 TL = maxAPInt(TL, ceilingOfQuotient(Y - X + 1, TMUL));
1638 LLVM_DEBUG(dbgs() << "\t\t TL = " << TL << "\n");
1639 }
1640 else {
1641 TU = minAPInt(TU, floorOfQuotient(Y - X + 1, TMUL));
1642 LLVM_DEBUG(dbgs() << "\t\t TU = " << TU << "\n");
1643 }
1644 if (TL.sle(TU)) {
1645 NewDirection |= Dependence::DVEntry::GT;
1646 ++ExactSIVsuccesses;
1647 }
1648
1649 // finished
1650 Result.DV[Level].Direction &= NewDirection;
1651 if (Result.DV[Level].Direction == Dependence::DVEntry::NONE)
1652 ++ExactSIVindependence;
1653 return Result.DV[Level].Direction == Dependence::DVEntry::NONE;
1654 }
1655
1656
1657
1658 // Return true if the divisor evenly divides the dividend.
1659 static
isRemainderZero(const SCEVConstant * Dividend,const SCEVConstant * Divisor)1660 bool isRemainderZero(const SCEVConstant *Dividend,
1661 const SCEVConstant *Divisor) {
1662 const APInt &ConstDividend = Dividend->getAPInt();
1663 const APInt &ConstDivisor = Divisor->getAPInt();
1664 return ConstDividend.srem(ConstDivisor) == 0;
1665 }
1666
1667
1668 // weakZeroSrcSIVtest -
1669 // From the paper, Practical Dependence Testing, Section 4.2.2
1670 //
1671 // When we have a pair of subscripts of the form [c1] and [c2 + a*i],
1672 // where i is an induction variable, c1 and c2 are loop invariant,
1673 // and a is a constant, we can solve it exactly using the
1674 // Weak-Zero SIV test.
1675 //
1676 // Given
1677 //
1678 // c1 = c2 + a*i
1679 //
1680 // we get
1681 //
1682 // (c1 - c2)/a = i
1683 //
1684 // If i is not an integer, there's no dependence.
1685 // If i < 0 or > UB, there's no dependence.
1686 // If i = 0, the direction is >= and peeling the
1687 // 1st iteration will break the dependence.
1688 // If i = UB, the direction is <= and peeling the
1689 // last iteration will break the dependence.
1690 // Otherwise, the direction is *.
1691 //
1692 // Can prove independence. Failing that, we can sometimes refine
1693 // the directions. Can sometimes show that first or last
1694 // iteration carries all the dependences (so worth peeling).
1695 //
1696 // (see also weakZeroDstSIVtest)
1697 //
1698 // Return true if dependence disproved.
weakZeroSrcSIVtest(const SCEV * DstCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1699 bool DependenceInfo::weakZeroSrcSIVtest(const SCEV *DstCoeff,
1700 const SCEV *SrcConst,
1701 const SCEV *DstConst,
1702 const Loop *CurLoop, unsigned Level,
1703 FullDependence &Result,
1704 Constraint &NewConstraint) const {
1705 // For the WeakSIV test, it's possible the loop isn't common to
1706 // the Src and Dst loops. If it isn't, then there's no need to
1707 // record a direction.
1708 LLVM_DEBUG(dbgs() << "\tWeak-Zero (src) SIV test\n");
1709 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << "\n");
1710 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1711 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1712 ++WeakZeroSIVapplications;
1713 assert(0 < Level && Level <= MaxLevels && "Level out of range");
1714 Level--;
1715 Result.Consistent = false;
1716 const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst);
1717 NewConstraint.setLine(SE->getZero(Delta->getType()), DstCoeff, Delta,
1718 CurLoop);
1719 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1720 if (isKnownPredicate(CmpInst::ICMP_EQ, SrcConst, DstConst)) {
1721 if (Level < CommonLevels) {
1722 Result.DV[Level].Direction &= Dependence::DVEntry::GE;
1723 Result.DV[Level].PeelFirst = true;
1724 ++WeakZeroSIVsuccesses;
1725 }
1726 return false; // dependences caused by first iteration
1727 }
1728 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1729 if (!ConstCoeff)
1730 return false;
1731 const SCEV *AbsCoeff =
1732 SE->isKnownNegative(ConstCoeff) ?
1733 SE->getNegativeSCEV(ConstCoeff) : ConstCoeff;
1734 const SCEV *NewDelta =
1735 SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta;
1736
1737 // check that Delta/SrcCoeff < iteration count
1738 // really check NewDelta < count*AbsCoeff
1739 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1740 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n");
1741 const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound);
1742 if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) {
1743 ++WeakZeroSIVindependence;
1744 ++WeakZeroSIVsuccesses;
1745 return true;
1746 }
1747 if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) {
1748 // dependences caused by last iteration
1749 if (Level < CommonLevels) {
1750 Result.DV[Level].Direction &= Dependence::DVEntry::LE;
1751 Result.DV[Level].PeelLast = true;
1752 ++WeakZeroSIVsuccesses;
1753 }
1754 return false;
1755 }
1756 }
1757
1758 // check that Delta/SrcCoeff >= 0
1759 // really check that NewDelta >= 0
1760 if (SE->isKnownNegative(NewDelta)) {
1761 // No dependence, newDelta < 0
1762 ++WeakZeroSIVindependence;
1763 ++WeakZeroSIVsuccesses;
1764 return true;
1765 }
1766
1767 // if SrcCoeff doesn't divide Delta, then no dependence
1768 if (isa<SCEVConstant>(Delta) &&
1769 !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) {
1770 ++WeakZeroSIVindependence;
1771 ++WeakZeroSIVsuccesses;
1772 return true;
1773 }
1774 return false;
1775 }
1776
1777
1778 // weakZeroDstSIVtest -
1779 // From the paper, Practical Dependence Testing, Section 4.2.2
1780 //
1781 // When we have a pair of subscripts of the form [c1 + a*i] and [c2],
1782 // where i is an induction variable, c1 and c2 are loop invariant,
1783 // and a is a constant, we can solve it exactly using the
1784 // Weak-Zero SIV test.
1785 //
1786 // Given
1787 //
1788 // c1 + a*i = c2
1789 //
1790 // we get
1791 //
1792 // i = (c2 - c1)/a
1793 //
1794 // If i is not an integer, there's no dependence.
1795 // If i < 0 or > UB, there's no dependence.
1796 // If i = 0, the direction is <= and peeling the
1797 // 1st iteration will break the dependence.
1798 // If i = UB, the direction is >= and peeling the
1799 // last iteration will break the dependence.
1800 // Otherwise, the direction is *.
1801 //
1802 // Can prove independence. Failing that, we can sometimes refine
1803 // the directions. Can sometimes show that first or last
1804 // iteration carries all the dependences (so worth peeling).
1805 //
1806 // (see also weakZeroSrcSIVtest)
1807 //
1808 // Return true if dependence disproved.
weakZeroDstSIVtest(const SCEV * SrcCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1809 bool DependenceInfo::weakZeroDstSIVtest(const SCEV *SrcCoeff,
1810 const SCEV *SrcConst,
1811 const SCEV *DstConst,
1812 const Loop *CurLoop, unsigned Level,
1813 FullDependence &Result,
1814 Constraint &NewConstraint) const {
1815 // For the WeakSIV test, it's possible the loop isn't common to the
1816 // Src and Dst loops. If it isn't, then there's no need to record a direction.
1817 LLVM_DEBUG(dbgs() << "\tWeak-Zero (dst) SIV test\n");
1818 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << "\n");
1819 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1820 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1821 ++WeakZeroSIVapplications;
1822 assert(0 < Level && Level <= SrcLevels && "Level out of range");
1823 Level--;
1824 Result.Consistent = false;
1825 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1826 NewConstraint.setLine(SrcCoeff, SE->getZero(Delta->getType()), Delta,
1827 CurLoop);
1828 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1829 if (isKnownPredicate(CmpInst::ICMP_EQ, DstConst, SrcConst)) {
1830 if (Level < CommonLevels) {
1831 Result.DV[Level].Direction &= Dependence::DVEntry::LE;
1832 Result.DV[Level].PeelFirst = true;
1833 ++WeakZeroSIVsuccesses;
1834 }
1835 return false; // dependences caused by first iteration
1836 }
1837 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1838 if (!ConstCoeff)
1839 return false;
1840 const SCEV *AbsCoeff =
1841 SE->isKnownNegative(ConstCoeff) ?
1842 SE->getNegativeSCEV(ConstCoeff) : ConstCoeff;
1843 const SCEV *NewDelta =
1844 SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta;
1845
1846 // check that Delta/SrcCoeff < iteration count
1847 // really check NewDelta < count*AbsCoeff
1848 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1849 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n");
1850 const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound);
1851 if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) {
1852 ++WeakZeroSIVindependence;
1853 ++WeakZeroSIVsuccesses;
1854 return true;
1855 }
1856 if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) {
1857 // dependences caused by last iteration
1858 if (Level < CommonLevels) {
1859 Result.DV[Level].Direction &= Dependence::DVEntry::GE;
1860 Result.DV[Level].PeelLast = true;
1861 ++WeakZeroSIVsuccesses;
1862 }
1863 return false;
1864 }
1865 }
1866
1867 // check that Delta/SrcCoeff >= 0
1868 // really check that NewDelta >= 0
1869 if (SE->isKnownNegative(NewDelta)) {
1870 // No dependence, newDelta < 0
1871 ++WeakZeroSIVindependence;
1872 ++WeakZeroSIVsuccesses;
1873 return true;
1874 }
1875
1876 // if SrcCoeff doesn't divide Delta, then no dependence
1877 if (isa<SCEVConstant>(Delta) &&
1878 !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) {
1879 ++WeakZeroSIVindependence;
1880 ++WeakZeroSIVsuccesses;
1881 return true;
1882 }
1883 return false;
1884 }
1885
1886
1887 // exactRDIVtest - Tests the RDIV subscript pair for dependence.
1888 // Things of the form [c1 + a*i] and [c2 + b*j],
1889 // where i and j are induction variable, c1 and c2 are loop invariant,
1890 // and a and b are constants.
1891 // Returns true if any possible dependence is disproved.
1892 // Marks the result as inconsistent.
1893 // Works in some cases that symbolicRDIVtest doesn't, and vice versa.
exactRDIVtest(const SCEV * SrcCoeff,const SCEV * DstCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * SrcLoop,const Loop * DstLoop,FullDependence & Result) const1894 bool DependenceInfo::exactRDIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff,
1895 const SCEV *SrcConst, const SCEV *DstConst,
1896 const Loop *SrcLoop, const Loop *DstLoop,
1897 FullDependence &Result) const {
1898 LLVM_DEBUG(dbgs() << "\tExact RDIV test\n");
1899 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << " = AM\n");
1900 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << " = BM\n");
1901 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1902 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1903 ++ExactRDIVapplications;
1904 Result.Consistent = false;
1905 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1906 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1907 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1908 const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1909 const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1910 if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff)
1911 return false;
1912
1913 // find gcd
1914 APInt G, X, Y;
1915 APInt AM = ConstSrcCoeff->getAPInt();
1916 APInt BM = ConstDstCoeff->getAPInt();
1917 unsigned Bits = AM.getBitWidth();
1918 if (findGCD(Bits, AM, BM, ConstDelta->getAPInt(), G, X, Y)) {
1919 // gcd doesn't divide Delta, no dependence
1920 ++ExactRDIVindependence;
1921 return true;
1922 }
1923
1924 LLVM_DEBUG(dbgs() << "\t X = " << X << ", Y = " << Y << "\n");
1925
1926 // since SCEV construction seems to normalize, LM = 0
1927 APInt SrcUM(Bits, 1, true);
1928 bool SrcUMvalid = false;
1929 // SrcUM is perhaps unavailable, let's check
1930 if (const SCEVConstant *UpperBound =
1931 collectConstantUpperBound(SrcLoop, Delta->getType())) {
1932 SrcUM = UpperBound->getAPInt();
1933 LLVM_DEBUG(dbgs() << "\t SrcUM = " << SrcUM << "\n");
1934 SrcUMvalid = true;
1935 }
1936
1937 APInt DstUM(Bits, 1, true);
1938 bool DstUMvalid = false;
1939 // UM is perhaps unavailable, let's check
1940 if (const SCEVConstant *UpperBound =
1941 collectConstantUpperBound(DstLoop, Delta->getType())) {
1942 DstUM = UpperBound->getAPInt();
1943 LLVM_DEBUG(dbgs() << "\t DstUM = " << DstUM << "\n");
1944 DstUMvalid = true;
1945 }
1946
1947 APInt TU(APInt::getSignedMaxValue(Bits));
1948 APInt TL(APInt::getSignedMinValue(Bits));
1949
1950 // test(BM/G, LM-X) and test(-BM/G, X-UM)
1951 APInt TMUL = BM.sdiv(G);
1952 if (TMUL.sgt(0)) {
1953 TL = maxAPInt(TL, ceilingOfQuotient(-X, TMUL));
1954 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1955 if (SrcUMvalid) {
1956 TU = minAPInt(TU, floorOfQuotient(SrcUM - X, TMUL));
1957 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1958 }
1959 }
1960 else {
1961 TU = minAPInt(TU, floorOfQuotient(-X, TMUL));
1962 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1963 if (SrcUMvalid) {
1964 TL = maxAPInt(TL, ceilingOfQuotient(SrcUM - X, TMUL));
1965 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1966 }
1967 }
1968
1969 // test(AM/G, LM-Y) and test(-AM/G, Y-UM)
1970 TMUL = AM.sdiv(G);
1971 if (TMUL.sgt(0)) {
1972 TL = maxAPInt(TL, ceilingOfQuotient(-Y, TMUL));
1973 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1974 if (DstUMvalid) {
1975 TU = minAPInt(TU, floorOfQuotient(DstUM - Y, TMUL));
1976 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1977 }
1978 }
1979 else {
1980 TU = minAPInt(TU, floorOfQuotient(-Y, TMUL));
1981 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1982 if (DstUMvalid) {
1983 TL = maxAPInt(TL, ceilingOfQuotient(DstUM - Y, TMUL));
1984 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1985 }
1986 }
1987 if (TL.sgt(TU))
1988 ++ExactRDIVindependence;
1989 return TL.sgt(TU);
1990 }
1991
1992
1993 // symbolicRDIVtest -
1994 // In Section 4.5 of the Practical Dependence Testing paper,the authors
1995 // introduce a special case of Banerjee's Inequalities (also called the
1996 // Extreme-Value Test) that can handle some of the SIV and RDIV cases,
1997 // particularly cases with symbolics. Since it's only able to disprove
1998 // dependence (not compute distances or directions), we'll use it as a
1999 // fall back for the other tests.
2000 //
2001 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j]
2002 // where i and j are induction variables and c1 and c2 are loop invariants,
2003 // we can use the symbolic tests to disprove some dependences, serving as a
2004 // backup for the RDIV test. Note that i and j can be the same variable,
2005 // letting this test serve as a backup for the various SIV tests.
2006 //
2007 // For a dependence to exist, c1 + a1*i must equal c2 + a2*j for some
2008 // 0 <= i <= N1 and some 0 <= j <= N2, where N1 and N2 are the (normalized)
2009 // loop bounds for the i and j loops, respectively. So, ...
2010 //
2011 // c1 + a1*i = c2 + a2*j
2012 // a1*i - a2*j = c2 - c1
2013 //
2014 // To test for a dependence, we compute c2 - c1 and make sure it's in the
2015 // range of the maximum and minimum possible values of a1*i - a2*j.
2016 // Considering the signs of a1 and a2, we have 4 possible cases:
2017 //
2018 // 1) If a1 >= 0 and a2 >= 0, then
2019 // a1*0 - a2*N2 <= c2 - c1 <= a1*N1 - a2*0
2020 // -a2*N2 <= c2 - c1 <= a1*N1
2021 //
2022 // 2) If a1 >= 0 and a2 <= 0, then
2023 // a1*0 - a2*0 <= c2 - c1 <= a1*N1 - a2*N2
2024 // 0 <= c2 - c1 <= a1*N1 - a2*N2
2025 //
2026 // 3) If a1 <= 0 and a2 >= 0, then
2027 // a1*N1 - a2*N2 <= c2 - c1 <= a1*0 - a2*0
2028 // a1*N1 - a2*N2 <= c2 - c1 <= 0
2029 //
2030 // 4) If a1 <= 0 and a2 <= 0, then
2031 // a1*N1 - a2*0 <= c2 - c1 <= a1*0 - a2*N2
2032 // a1*N1 <= c2 - c1 <= -a2*N2
2033 //
2034 // return true if dependence disproved
symbolicRDIVtest(const SCEV * A1,const SCEV * A2,const SCEV * C1,const SCEV * C2,const Loop * Loop1,const Loop * Loop2) const2035 bool DependenceInfo::symbolicRDIVtest(const SCEV *A1, const SCEV *A2,
2036 const SCEV *C1, const SCEV *C2,
2037 const Loop *Loop1,
2038 const Loop *Loop2) const {
2039 ++SymbolicRDIVapplications;
2040 LLVM_DEBUG(dbgs() << "\ttry symbolic RDIV test\n");
2041 LLVM_DEBUG(dbgs() << "\t A1 = " << *A1);
2042 LLVM_DEBUG(dbgs() << ", type = " << *A1->getType() << "\n");
2043 LLVM_DEBUG(dbgs() << "\t A2 = " << *A2 << "\n");
2044 LLVM_DEBUG(dbgs() << "\t C1 = " << *C1 << "\n");
2045 LLVM_DEBUG(dbgs() << "\t C2 = " << *C2 << "\n");
2046 const SCEV *N1 = collectUpperBound(Loop1, A1->getType());
2047 const SCEV *N2 = collectUpperBound(Loop2, A1->getType());
2048 LLVM_DEBUG(if (N1) dbgs() << "\t N1 = " << *N1 << "\n");
2049 LLVM_DEBUG(if (N2) dbgs() << "\t N2 = " << *N2 << "\n");
2050 const SCEV *C2_C1 = SE->getMinusSCEV(C2, C1);
2051 const SCEV *C1_C2 = SE->getMinusSCEV(C1, C2);
2052 LLVM_DEBUG(dbgs() << "\t C2 - C1 = " << *C2_C1 << "\n");
2053 LLVM_DEBUG(dbgs() << "\t C1 - C2 = " << *C1_C2 << "\n");
2054 if (SE->isKnownNonNegative(A1)) {
2055 if (SE->isKnownNonNegative(A2)) {
2056 // A1 >= 0 && A2 >= 0
2057 if (N1) {
2058 // make sure that c2 - c1 <= a1*N1
2059 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2060 LLVM_DEBUG(dbgs() << "\t A1*N1 = " << *A1N1 << "\n");
2061 if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1)) {
2062 ++SymbolicRDIVindependence;
2063 return true;
2064 }
2065 }
2066 if (N2) {
2067 // make sure that -a2*N2 <= c2 - c1, or a2*N2 >= c1 - c2
2068 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2069 LLVM_DEBUG(dbgs() << "\t A2*N2 = " << *A2N2 << "\n");
2070 if (isKnownPredicate(CmpInst::ICMP_SLT, A2N2, C1_C2)) {
2071 ++SymbolicRDIVindependence;
2072 return true;
2073 }
2074 }
2075 }
2076 else if (SE->isKnownNonPositive(A2)) {
2077 // a1 >= 0 && a2 <= 0
2078 if (N1 && N2) {
2079 // make sure that c2 - c1 <= a1*N1 - a2*N2
2080 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2081 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2082 const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2);
2083 LLVM_DEBUG(dbgs() << "\t A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n");
2084 if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1_A2N2)) {
2085 ++SymbolicRDIVindependence;
2086 return true;
2087 }
2088 }
2089 // make sure that 0 <= c2 - c1
2090 if (SE->isKnownNegative(C2_C1)) {
2091 ++SymbolicRDIVindependence;
2092 return true;
2093 }
2094 }
2095 }
2096 else if (SE->isKnownNonPositive(A1)) {
2097 if (SE->isKnownNonNegative(A2)) {
2098 // a1 <= 0 && a2 >= 0
2099 if (N1 && N2) {
2100 // make sure that a1*N1 - a2*N2 <= c2 - c1
2101 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2102 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2103 const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2);
2104 LLVM_DEBUG(dbgs() << "\t A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n");
2105 if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1_A2N2, C2_C1)) {
2106 ++SymbolicRDIVindependence;
2107 return true;
2108 }
2109 }
2110 // make sure that c2 - c1 <= 0
2111 if (SE->isKnownPositive(C2_C1)) {
2112 ++SymbolicRDIVindependence;
2113 return true;
2114 }
2115 }
2116 else if (SE->isKnownNonPositive(A2)) {
2117 // a1 <= 0 && a2 <= 0
2118 if (N1) {
2119 // make sure that a1*N1 <= c2 - c1
2120 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2121 LLVM_DEBUG(dbgs() << "\t A1*N1 = " << *A1N1 << "\n");
2122 if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1, C2_C1)) {
2123 ++SymbolicRDIVindependence;
2124 return true;
2125 }
2126 }
2127 if (N2) {
2128 // make sure that c2 - c1 <= -a2*N2, or c1 - c2 >= a2*N2
2129 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2130 LLVM_DEBUG(dbgs() << "\t A2*N2 = " << *A2N2 << "\n");
2131 if (isKnownPredicate(CmpInst::ICMP_SLT, C1_C2, A2N2)) {
2132 ++SymbolicRDIVindependence;
2133 return true;
2134 }
2135 }
2136 }
2137 }
2138 return false;
2139 }
2140
2141
2142 // testSIV -
2143 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 - a2*i]
2144 // where i is an induction variable, c1 and c2 are loop invariant, and a1 and
2145 // a2 are constant, we attack it with an SIV test. While they can all be
2146 // solved with the Exact SIV test, it's worthwhile to use simpler tests when
2147 // they apply; they're cheaper and sometimes more precise.
2148 //
2149 // Return true if dependence disproved.
testSIV(const SCEV * Src,const SCEV * Dst,unsigned & Level,FullDependence & Result,Constraint & NewConstraint,const SCEV * & SplitIter) const2150 bool DependenceInfo::testSIV(const SCEV *Src, const SCEV *Dst, unsigned &Level,
2151 FullDependence &Result, Constraint &NewConstraint,
2152 const SCEV *&SplitIter) const {
2153 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
2154 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
2155 const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src);
2156 const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst);
2157 if (SrcAddRec && DstAddRec) {
2158 const SCEV *SrcConst = SrcAddRec->getStart();
2159 const SCEV *DstConst = DstAddRec->getStart();
2160 const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2161 const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE);
2162 const Loop *CurLoop = SrcAddRec->getLoop();
2163 assert(CurLoop == DstAddRec->getLoop() &&
2164 "both loops in SIV should be same");
2165 Level = mapSrcLoop(CurLoop);
2166 bool disproven;
2167 if (SrcCoeff == DstCoeff)
2168 disproven = strongSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2169 Level, Result, NewConstraint);
2170 else if (SrcCoeff == SE->getNegativeSCEV(DstCoeff))
2171 disproven = weakCrossingSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2172 Level, Result, NewConstraint, SplitIter);
2173 else
2174 disproven = exactSIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop,
2175 Level, Result, NewConstraint);
2176 return disproven ||
2177 gcdMIVtest(Src, Dst, Result) ||
2178 symbolicRDIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop, CurLoop);
2179 }
2180 if (SrcAddRec) {
2181 const SCEV *SrcConst = SrcAddRec->getStart();
2182 const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2183 const SCEV *DstConst = Dst;
2184 const Loop *CurLoop = SrcAddRec->getLoop();
2185 Level = mapSrcLoop(CurLoop);
2186 return weakZeroDstSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2187 Level, Result, NewConstraint) ||
2188 gcdMIVtest(Src, Dst, Result);
2189 }
2190 if (DstAddRec) {
2191 const SCEV *DstConst = DstAddRec->getStart();
2192 const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE);
2193 const SCEV *SrcConst = Src;
2194 const Loop *CurLoop = DstAddRec->getLoop();
2195 Level = mapDstLoop(CurLoop);
2196 return weakZeroSrcSIVtest(DstCoeff, SrcConst, DstConst,
2197 CurLoop, Level, Result, NewConstraint) ||
2198 gcdMIVtest(Src, Dst, Result);
2199 }
2200 llvm_unreachable("SIV test expected at least one AddRec");
2201 return false;
2202 }
2203
2204
2205 // testRDIV -
2206 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j]
2207 // where i and j are induction variables, c1 and c2 are loop invariant,
2208 // and a1 and a2 are constant, we can solve it exactly with an easy adaptation
2209 // of the Exact SIV test, the Restricted Double Index Variable (RDIV) test.
2210 // It doesn't make sense to talk about distance or direction in this case,
2211 // so there's no point in making special versions of the Strong SIV test or
2212 // the Weak-crossing SIV test.
2213 //
2214 // With minor algebra, this test can also be used for things like
2215 // [c1 + a1*i + a2*j][c2].
2216 //
2217 // Return true if dependence disproved.
testRDIV(const SCEV * Src,const SCEV * Dst,FullDependence & Result) const2218 bool DependenceInfo::testRDIV(const SCEV *Src, const SCEV *Dst,
2219 FullDependence &Result) const {
2220 // we have 3 possible situations here:
2221 // 1) [a*i + b] and [c*j + d]
2222 // 2) [a*i + c*j + b] and [d]
2223 // 3) [b] and [a*i + c*j + d]
2224 // We need to find what we've got and get organized
2225
2226 const SCEV *SrcConst, *DstConst;
2227 const SCEV *SrcCoeff, *DstCoeff;
2228 const Loop *SrcLoop, *DstLoop;
2229
2230 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
2231 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
2232 const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src);
2233 const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst);
2234 if (SrcAddRec && DstAddRec) {
2235 SrcConst = SrcAddRec->getStart();
2236 SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2237 SrcLoop = SrcAddRec->getLoop();
2238 DstConst = DstAddRec->getStart();
2239 DstCoeff = DstAddRec->getStepRecurrence(*SE);
2240 DstLoop = DstAddRec->getLoop();
2241 }
2242 else if (SrcAddRec) {
2243 if (const SCEVAddRecExpr *tmpAddRec =
2244 dyn_cast<SCEVAddRecExpr>(SrcAddRec->getStart())) {
2245 SrcConst = tmpAddRec->getStart();
2246 SrcCoeff = tmpAddRec->getStepRecurrence(*SE);
2247 SrcLoop = tmpAddRec->getLoop();
2248 DstConst = Dst;
2249 DstCoeff = SE->getNegativeSCEV(SrcAddRec->getStepRecurrence(*SE));
2250 DstLoop = SrcAddRec->getLoop();
2251 }
2252 else
2253 llvm_unreachable("RDIV reached by surprising SCEVs");
2254 }
2255 else if (DstAddRec) {
2256 if (const SCEVAddRecExpr *tmpAddRec =
2257 dyn_cast<SCEVAddRecExpr>(DstAddRec->getStart())) {
2258 DstConst = tmpAddRec->getStart();
2259 DstCoeff = tmpAddRec->getStepRecurrence(*SE);
2260 DstLoop = tmpAddRec->getLoop();
2261 SrcConst = Src;
2262 SrcCoeff = SE->getNegativeSCEV(DstAddRec->getStepRecurrence(*SE));
2263 SrcLoop = DstAddRec->getLoop();
2264 }
2265 else
2266 llvm_unreachable("RDIV reached by surprising SCEVs");
2267 }
2268 else
2269 llvm_unreachable("RDIV expected at least one AddRec");
2270 return exactRDIVtest(SrcCoeff, DstCoeff,
2271 SrcConst, DstConst,
2272 SrcLoop, DstLoop,
2273 Result) ||
2274 gcdMIVtest(Src, Dst, Result) ||
2275 symbolicRDIVtest(SrcCoeff, DstCoeff,
2276 SrcConst, DstConst,
2277 SrcLoop, DstLoop);
2278 }
2279
2280
2281 // Tests the single-subscript MIV pair (Src and Dst) for dependence.
2282 // Return true if dependence disproved.
2283 // Can sometimes refine direction vectors.
testMIV(const SCEV * Src,const SCEV * Dst,const SmallBitVector & Loops,FullDependence & Result) const2284 bool DependenceInfo::testMIV(const SCEV *Src, const SCEV *Dst,
2285 const SmallBitVector &Loops,
2286 FullDependence &Result) const {
2287 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
2288 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
2289 Result.Consistent = false;
2290 return gcdMIVtest(Src, Dst, Result) ||
2291 banerjeeMIVtest(Src, Dst, Loops, Result);
2292 }
2293
2294
2295 // Given a product, e.g., 10*X*Y, returns the first constant operand,
2296 // in this case 10. If there is no constant part, returns NULL.
2297 static
getConstantPart(const SCEV * Expr)2298 const SCEVConstant *getConstantPart(const SCEV *Expr) {
2299 if (const auto *Constant = dyn_cast<SCEVConstant>(Expr))
2300 return Constant;
2301 else if (const auto *Product = dyn_cast<SCEVMulExpr>(Expr))
2302 if (const auto *Constant = dyn_cast<SCEVConstant>(Product->getOperand(0)))
2303 return Constant;
2304 return nullptr;
2305 }
2306
2307
2308 //===----------------------------------------------------------------------===//
2309 // gcdMIVtest -
2310 // Tests an MIV subscript pair for dependence.
2311 // Returns true if any possible dependence is disproved.
2312 // Marks the result as inconsistent.
2313 // Can sometimes disprove the equal direction for 1 or more loops,
2314 // as discussed in Michael Wolfe's book,
2315 // High Performance Compilers for Parallel Computing, page 235.
2316 //
2317 // We spend some effort (code!) to handle cases like
2318 // [10*i + 5*N*j + 15*M + 6], where i and j are induction variables,
2319 // but M and N are just loop-invariant variables.
2320 // This should help us handle linearized subscripts;
2321 // also makes this test a useful backup to the various SIV tests.
2322 //
2323 // It occurs to me that the presence of loop-invariant variables
2324 // changes the nature of the test from "greatest common divisor"
2325 // to "a common divisor".
gcdMIVtest(const SCEV * Src,const SCEV * Dst,FullDependence & Result) const2326 bool DependenceInfo::gcdMIVtest(const SCEV *Src, const SCEV *Dst,
2327 FullDependence &Result) const {
2328 LLVM_DEBUG(dbgs() << "starting gcd\n");
2329 ++GCDapplications;
2330 unsigned BitWidth = SE->getTypeSizeInBits(Src->getType());
2331 APInt RunningGCD = APInt::getNullValue(BitWidth);
2332
2333 // Examine Src coefficients.
2334 // Compute running GCD and record source constant.
2335 // Because we're looking for the constant at the end of the chain,
2336 // we can't quit the loop just because the GCD == 1.
2337 const SCEV *Coefficients = Src;
2338 while (const SCEVAddRecExpr *AddRec =
2339 dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2340 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2341 // If the coefficient is the product of a constant and other stuff,
2342 // we can use the constant in the GCD computation.
2343 const auto *Constant = getConstantPart(Coeff);
2344 if (!Constant)
2345 return false;
2346 APInt ConstCoeff = Constant->getAPInt();
2347 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2348 Coefficients = AddRec->getStart();
2349 }
2350 const SCEV *SrcConst = Coefficients;
2351
2352 // Examine Dst coefficients.
2353 // Compute running GCD and record destination constant.
2354 // Because we're looking for the constant at the end of the chain,
2355 // we can't quit the loop just because the GCD == 1.
2356 Coefficients = Dst;
2357 while (const SCEVAddRecExpr *AddRec =
2358 dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2359 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2360 // If the coefficient is the product of a constant and other stuff,
2361 // we can use the constant in the GCD computation.
2362 const auto *Constant = getConstantPart(Coeff);
2363 if (!Constant)
2364 return false;
2365 APInt ConstCoeff = Constant->getAPInt();
2366 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2367 Coefficients = AddRec->getStart();
2368 }
2369 const SCEV *DstConst = Coefficients;
2370
2371 APInt ExtraGCD = APInt::getNullValue(BitWidth);
2372 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
2373 LLVM_DEBUG(dbgs() << " Delta = " << *Delta << "\n");
2374 const SCEVConstant *Constant = dyn_cast<SCEVConstant>(Delta);
2375 if (const SCEVAddExpr *Sum = dyn_cast<SCEVAddExpr>(Delta)) {
2376 // If Delta is a sum of products, we may be able to make further progress.
2377 for (unsigned Op = 0, Ops = Sum->getNumOperands(); Op < Ops; Op++) {
2378 const SCEV *Operand = Sum->getOperand(Op);
2379 if (isa<SCEVConstant>(Operand)) {
2380 assert(!Constant && "Surprised to find multiple constants");
2381 Constant = cast<SCEVConstant>(Operand);
2382 }
2383 else if (const SCEVMulExpr *Product = dyn_cast<SCEVMulExpr>(Operand)) {
2384 // Search for constant operand to participate in GCD;
2385 // If none found; return false.
2386 const SCEVConstant *ConstOp = getConstantPart(Product);
2387 if (!ConstOp)
2388 return false;
2389 APInt ConstOpValue = ConstOp->getAPInt();
2390 ExtraGCD = APIntOps::GreatestCommonDivisor(ExtraGCD,
2391 ConstOpValue.abs());
2392 }
2393 else
2394 return false;
2395 }
2396 }
2397 if (!Constant)
2398 return false;
2399 APInt ConstDelta = cast<SCEVConstant>(Constant)->getAPInt();
2400 LLVM_DEBUG(dbgs() << " ConstDelta = " << ConstDelta << "\n");
2401 if (ConstDelta == 0)
2402 return false;
2403 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ExtraGCD);
2404 LLVM_DEBUG(dbgs() << " RunningGCD = " << RunningGCD << "\n");
2405 APInt Remainder = ConstDelta.srem(RunningGCD);
2406 if (Remainder != 0) {
2407 ++GCDindependence;
2408 return true;
2409 }
2410
2411 // Try to disprove equal directions.
2412 // For example, given a subscript pair [3*i + 2*j] and [i' + 2*j' - 1],
2413 // the code above can't disprove the dependence because the GCD = 1.
2414 // So we consider what happen if i = i' and what happens if j = j'.
2415 // If i = i', we can simplify the subscript to [2*i + 2*j] and [2*j' - 1],
2416 // which is infeasible, so we can disallow the = direction for the i level.
2417 // Setting j = j' doesn't help matters, so we end up with a direction vector
2418 // of [<>, *]
2419 //
2420 // Given A[5*i + 10*j*M + 9*M*N] and A[15*i + 20*j*M - 21*N*M + 5],
2421 // we need to remember that the constant part is 5 and the RunningGCD should
2422 // be initialized to ExtraGCD = 30.
2423 LLVM_DEBUG(dbgs() << " ExtraGCD = " << ExtraGCD << '\n');
2424
2425 bool Improved = false;
2426 Coefficients = Src;
2427 while (const SCEVAddRecExpr *AddRec =
2428 dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2429 Coefficients = AddRec->getStart();
2430 const Loop *CurLoop = AddRec->getLoop();
2431 RunningGCD = ExtraGCD;
2432 const SCEV *SrcCoeff = AddRec->getStepRecurrence(*SE);
2433 const SCEV *DstCoeff = SE->getMinusSCEV(SrcCoeff, SrcCoeff);
2434 const SCEV *Inner = Src;
2435 while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) {
2436 AddRec = cast<SCEVAddRecExpr>(Inner);
2437 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2438 if (CurLoop == AddRec->getLoop())
2439 ; // SrcCoeff == Coeff
2440 else {
2441 // If the coefficient is the product of a constant and other stuff,
2442 // we can use the constant in the GCD computation.
2443 Constant = getConstantPart(Coeff);
2444 if (!Constant)
2445 return false;
2446 APInt ConstCoeff = Constant->getAPInt();
2447 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2448 }
2449 Inner = AddRec->getStart();
2450 }
2451 Inner = Dst;
2452 while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) {
2453 AddRec = cast<SCEVAddRecExpr>(Inner);
2454 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2455 if (CurLoop == AddRec->getLoop())
2456 DstCoeff = Coeff;
2457 else {
2458 // If the coefficient is the product of a constant and other stuff,
2459 // we can use the constant in the GCD computation.
2460 Constant = getConstantPart(Coeff);
2461 if (!Constant)
2462 return false;
2463 APInt ConstCoeff = Constant->getAPInt();
2464 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2465 }
2466 Inner = AddRec->getStart();
2467 }
2468 Delta = SE->getMinusSCEV(SrcCoeff, DstCoeff);
2469 // If the coefficient is the product of a constant and other stuff,
2470 // we can use the constant in the GCD computation.
2471 Constant = getConstantPart(Delta);
2472 if (!Constant)
2473 // The difference of the two coefficients might not be a product
2474 // or constant, in which case we give up on this direction.
2475 continue;
2476 APInt ConstCoeff = Constant->getAPInt();
2477 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2478 LLVM_DEBUG(dbgs() << "\tRunningGCD = " << RunningGCD << "\n");
2479 if (RunningGCD != 0) {
2480 Remainder = ConstDelta.srem(RunningGCD);
2481 LLVM_DEBUG(dbgs() << "\tRemainder = " << Remainder << "\n");
2482 if (Remainder != 0) {
2483 unsigned Level = mapSrcLoop(CurLoop);
2484 Result.DV[Level - 1].Direction &= unsigned(~Dependence::DVEntry::EQ);
2485 Improved = true;
2486 }
2487 }
2488 }
2489 if (Improved)
2490 ++GCDsuccesses;
2491 LLVM_DEBUG(dbgs() << "all done\n");
2492 return false;
2493 }
2494
2495
2496 //===----------------------------------------------------------------------===//
2497 // banerjeeMIVtest -
2498 // Use Banerjee's Inequalities to test an MIV subscript pair.
2499 // (Wolfe, in the race-car book, calls this the Extreme Value Test.)
2500 // Generally follows the discussion in Section 2.5.2 of
2501 //
2502 // Optimizing Supercompilers for Supercomputers
2503 // Michael Wolfe
2504 //
2505 // The inequalities given on page 25 are simplified in that loops are
2506 // normalized so that the lower bound is always 0 and the stride is always 1.
2507 // For example, Wolfe gives
2508 //
2509 // LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2510 //
2511 // where A_k is the coefficient of the kth index in the source subscript,
2512 // B_k is the coefficient of the kth index in the destination subscript,
2513 // U_k is the upper bound of the kth index, L_k is the lower bound of the Kth
2514 // index, and N_k is the stride of the kth index. Since all loops are normalized
2515 // by the SCEV package, N_k = 1 and L_k = 0, allowing us to simplify the
2516 // equation to
2517 //
2518 // LB^<_k = (A^-_k - B_k)^- (U_k - 0 - 1) + (A_k - B_k)0 - B_k 1
2519 // = (A^-_k - B_k)^- (U_k - 1) - B_k
2520 //
2521 // Similar simplifications are possible for the other equations.
2522 //
2523 // When we can't determine the number of iterations for a loop,
2524 // we use NULL as an indicator for the worst case, infinity.
2525 // When computing the upper bound, NULL denotes +inf;
2526 // for the lower bound, NULL denotes -inf.
2527 //
2528 // Return true if dependence disproved.
banerjeeMIVtest(const SCEV * Src,const SCEV * Dst,const SmallBitVector & Loops,FullDependence & Result) const2529 bool DependenceInfo::banerjeeMIVtest(const SCEV *Src, const SCEV *Dst,
2530 const SmallBitVector &Loops,
2531 FullDependence &Result) const {
2532 LLVM_DEBUG(dbgs() << "starting Banerjee\n");
2533 ++BanerjeeApplications;
2534 LLVM_DEBUG(dbgs() << " Src = " << *Src << '\n');
2535 const SCEV *A0;
2536 CoefficientInfo *A = collectCoeffInfo(Src, true, A0);
2537 LLVM_DEBUG(dbgs() << " Dst = " << *Dst << '\n');
2538 const SCEV *B0;
2539 CoefficientInfo *B = collectCoeffInfo(Dst, false, B0);
2540 BoundInfo *Bound = new BoundInfo[MaxLevels + 1];
2541 const SCEV *Delta = SE->getMinusSCEV(B0, A0);
2542 LLVM_DEBUG(dbgs() << "\tDelta = " << *Delta << '\n');
2543
2544 // Compute bounds for all the * directions.
2545 LLVM_DEBUG(dbgs() << "\tBounds[*]\n");
2546 for (unsigned K = 1; K <= MaxLevels; ++K) {
2547 Bound[K].Iterations = A[K].Iterations ? A[K].Iterations : B[K].Iterations;
2548 Bound[K].Direction = Dependence::DVEntry::ALL;
2549 Bound[K].DirSet = Dependence::DVEntry::NONE;
2550 findBoundsALL(A, B, Bound, K);
2551 #ifndef NDEBUG
2552 LLVM_DEBUG(dbgs() << "\t " << K << '\t');
2553 if (Bound[K].Lower[Dependence::DVEntry::ALL])
2554 LLVM_DEBUG(dbgs() << *Bound[K].Lower[Dependence::DVEntry::ALL] << '\t');
2555 else
2556 LLVM_DEBUG(dbgs() << "-inf\t");
2557 if (Bound[K].Upper[Dependence::DVEntry::ALL])
2558 LLVM_DEBUG(dbgs() << *Bound[K].Upper[Dependence::DVEntry::ALL] << '\n');
2559 else
2560 LLVM_DEBUG(dbgs() << "+inf\n");
2561 #endif
2562 }
2563
2564 // Test the *, *, *, ... case.
2565 bool Disproved = false;
2566 if (testBounds(Dependence::DVEntry::ALL, 0, Bound, Delta)) {
2567 // Explore the direction vector hierarchy.
2568 unsigned DepthExpanded = 0;
2569 unsigned NewDeps = exploreDirections(1, A, B, Bound,
2570 Loops, DepthExpanded, Delta);
2571 if (NewDeps > 0) {
2572 bool Improved = false;
2573 for (unsigned K = 1; K <= CommonLevels; ++K) {
2574 if (Loops[K]) {
2575 unsigned Old = Result.DV[K - 1].Direction;
2576 Result.DV[K - 1].Direction = Old & Bound[K].DirSet;
2577 Improved |= Old != Result.DV[K - 1].Direction;
2578 if (!Result.DV[K - 1].Direction) {
2579 Improved = false;
2580 Disproved = true;
2581 break;
2582 }
2583 }
2584 }
2585 if (Improved)
2586 ++BanerjeeSuccesses;
2587 }
2588 else {
2589 ++BanerjeeIndependence;
2590 Disproved = true;
2591 }
2592 }
2593 else {
2594 ++BanerjeeIndependence;
2595 Disproved = true;
2596 }
2597 delete [] Bound;
2598 delete [] A;
2599 delete [] B;
2600 return Disproved;
2601 }
2602
2603
2604 // Hierarchically expands the direction vector
2605 // search space, combining the directions of discovered dependences
2606 // in the DirSet field of Bound. Returns the number of distinct
2607 // dependences discovered. If the dependence is disproved,
2608 // it will return 0.
exploreDirections(unsigned Level,CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,const SmallBitVector & Loops,unsigned & DepthExpanded,const SCEV * Delta) const2609 unsigned DependenceInfo::exploreDirections(unsigned Level, CoefficientInfo *A,
2610 CoefficientInfo *B, BoundInfo *Bound,
2611 const SmallBitVector &Loops,
2612 unsigned &DepthExpanded,
2613 const SCEV *Delta) const {
2614 if (Level > CommonLevels) {
2615 // record result
2616 LLVM_DEBUG(dbgs() << "\t[");
2617 for (unsigned K = 1; K <= CommonLevels; ++K) {
2618 if (Loops[K]) {
2619 Bound[K].DirSet |= Bound[K].Direction;
2620 #ifndef NDEBUG
2621 switch (Bound[K].Direction) {
2622 case Dependence::DVEntry::LT:
2623 LLVM_DEBUG(dbgs() << " <");
2624 break;
2625 case Dependence::DVEntry::EQ:
2626 LLVM_DEBUG(dbgs() << " =");
2627 break;
2628 case Dependence::DVEntry::GT:
2629 LLVM_DEBUG(dbgs() << " >");
2630 break;
2631 case Dependence::DVEntry::ALL:
2632 LLVM_DEBUG(dbgs() << " *");
2633 break;
2634 default:
2635 llvm_unreachable("unexpected Bound[K].Direction");
2636 }
2637 #endif
2638 }
2639 }
2640 LLVM_DEBUG(dbgs() << " ]\n");
2641 return 1;
2642 }
2643 if (Loops[Level]) {
2644 if (Level > DepthExpanded) {
2645 DepthExpanded = Level;
2646 // compute bounds for <, =, > at current level
2647 findBoundsLT(A, B, Bound, Level);
2648 findBoundsGT(A, B, Bound, Level);
2649 findBoundsEQ(A, B, Bound, Level);
2650 #ifndef NDEBUG
2651 LLVM_DEBUG(dbgs() << "\tBound for level = " << Level << '\n');
2652 LLVM_DEBUG(dbgs() << "\t <\t");
2653 if (Bound[Level].Lower[Dependence::DVEntry::LT])
2654 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::LT]
2655 << '\t');
2656 else
2657 LLVM_DEBUG(dbgs() << "-inf\t");
2658 if (Bound[Level].Upper[Dependence::DVEntry::LT])
2659 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::LT]
2660 << '\n');
2661 else
2662 LLVM_DEBUG(dbgs() << "+inf\n");
2663 LLVM_DEBUG(dbgs() << "\t =\t");
2664 if (Bound[Level].Lower[Dependence::DVEntry::EQ])
2665 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::EQ]
2666 << '\t');
2667 else
2668 LLVM_DEBUG(dbgs() << "-inf\t");
2669 if (Bound[Level].Upper[Dependence::DVEntry::EQ])
2670 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::EQ]
2671 << '\n');
2672 else
2673 LLVM_DEBUG(dbgs() << "+inf\n");
2674 LLVM_DEBUG(dbgs() << "\t >\t");
2675 if (Bound[Level].Lower[Dependence::DVEntry::GT])
2676 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::GT]
2677 << '\t');
2678 else
2679 LLVM_DEBUG(dbgs() << "-inf\t");
2680 if (Bound[Level].Upper[Dependence::DVEntry::GT])
2681 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::GT]
2682 << '\n');
2683 else
2684 LLVM_DEBUG(dbgs() << "+inf\n");
2685 #endif
2686 }
2687
2688 unsigned NewDeps = 0;
2689
2690 // test bounds for <, *, *, ...
2691 if (testBounds(Dependence::DVEntry::LT, Level, Bound, Delta))
2692 NewDeps += exploreDirections(Level + 1, A, B, Bound,
2693 Loops, DepthExpanded, Delta);
2694
2695 // Test bounds for =, *, *, ...
2696 if (testBounds(Dependence::DVEntry::EQ, Level, Bound, Delta))
2697 NewDeps += exploreDirections(Level + 1, A, B, Bound,
2698 Loops, DepthExpanded, Delta);
2699
2700 // test bounds for >, *, *, ...
2701 if (testBounds(Dependence::DVEntry::GT, Level, Bound, Delta))
2702 NewDeps += exploreDirections(Level + 1, A, B, Bound,
2703 Loops, DepthExpanded, Delta);
2704
2705 Bound[Level].Direction = Dependence::DVEntry::ALL;
2706 return NewDeps;
2707 }
2708 else
2709 return exploreDirections(Level + 1, A, B, Bound, Loops, DepthExpanded, Delta);
2710 }
2711
2712
2713 // Returns true iff the current bounds are plausible.
testBounds(unsigned char DirKind,unsigned Level,BoundInfo * Bound,const SCEV * Delta) const2714 bool DependenceInfo::testBounds(unsigned char DirKind, unsigned Level,
2715 BoundInfo *Bound, const SCEV *Delta) const {
2716 Bound[Level].Direction = DirKind;
2717 if (const SCEV *LowerBound = getLowerBound(Bound))
2718 if (isKnownPredicate(CmpInst::ICMP_SGT, LowerBound, Delta))
2719 return false;
2720 if (const SCEV *UpperBound = getUpperBound(Bound))
2721 if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, UpperBound))
2722 return false;
2723 return true;
2724 }
2725
2726
2727 // Computes the upper and lower bounds for level K
2728 // using the * direction. Records them in Bound.
2729 // Wolfe gives the equations
2730 //
2731 // LB^*_k = (A^-_k - B^+_k)(U_k - L_k) + (A_k - B_k)L_k
2732 // UB^*_k = (A^+_k - B^-_k)(U_k - L_k) + (A_k - B_k)L_k
2733 //
2734 // Since we normalize loops, we can simplify these equations to
2735 //
2736 // LB^*_k = (A^-_k - B^+_k)U_k
2737 // UB^*_k = (A^+_k - B^-_k)U_k
2738 //
2739 // We must be careful to handle the case where the upper bound is unknown.
2740 // Note that the lower bound is always <= 0
2741 // and the upper bound is always >= 0.
findBoundsALL(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2742 void DependenceInfo::findBoundsALL(CoefficientInfo *A, CoefficientInfo *B,
2743 BoundInfo *Bound, unsigned K) const {
2744 Bound[K].Lower[Dependence::DVEntry::ALL] = nullptr; // Default value = -infinity.
2745 Bound[K].Upper[Dependence::DVEntry::ALL] = nullptr; // Default value = +infinity.
2746 if (Bound[K].Iterations) {
2747 Bound[K].Lower[Dependence::DVEntry::ALL] =
2748 SE->getMulExpr(SE->getMinusSCEV(A[K].NegPart, B[K].PosPart),
2749 Bound[K].Iterations);
2750 Bound[K].Upper[Dependence::DVEntry::ALL] =
2751 SE->getMulExpr(SE->getMinusSCEV(A[K].PosPart, B[K].NegPart),
2752 Bound[K].Iterations);
2753 }
2754 else {
2755 // If the difference is 0, we won't need to know the number of iterations.
2756 if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].NegPart, B[K].PosPart))
2757 Bound[K].Lower[Dependence::DVEntry::ALL] =
2758 SE->getZero(A[K].Coeff->getType());
2759 if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].PosPart, B[K].NegPart))
2760 Bound[K].Upper[Dependence::DVEntry::ALL] =
2761 SE->getZero(A[K].Coeff->getType());
2762 }
2763 }
2764
2765
2766 // Computes the upper and lower bounds for level K
2767 // using the = direction. Records them in Bound.
2768 // Wolfe gives the equations
2769 //
2770 // LB^=_k = (A_k - B_k)^- (U_k - L_k) + (A_k - B_k)L_k
2771 // UB^=_k = (A_k - B_k)^+ (U_k - L_k) + (A_k - B_k)L_k
2772 //
2773 // Since we normalize loops, we can simplify these equations to
2774 //
2775 // LB^=_k = (A_k - B_k)^- U_k
2776 // UB^=_k = (A_k - B_k)^+ U_k
2777 //
2778 // We must be careful to handle the case where the upper bound is unknown.
2779 // Note that the lower bound is always <= 0
2780 // and the upper bound is always >= 0.
findBoundsEQ(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2781 void DependenceInfo::findBoundsEQ(CoefficientInfo *A, CoefficientInfo *B,
2782 BoundInfo *Bound, unsigned K) const {
2783 Bound[K].Lower[Dependence::DVEntry::EQ] = nullptr; // Default value = -infinity.
2784 Bound[K].Upper[Dependence::DVEntry::EQ] = nullptr; // Default value = +infinity.
2785 if (Bound[K].Iterations) {
2786 const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff);
2787 const SCEV *NegativePart = getNegativePart(Delta);
2788 Bound[K].Lower[Dependence::DVEntry::EQ] =
2789 SE->getMulExpr(NegativePart, Bound[K].Iterations);
2790 const SCEV *PositivePart = getPositivePart(Delta);
2791 Bound[K].Upper[Dependence::DVEntry::EQ] =
2792 SE->getMulExpr(PositivePart, Bound[K].Iterations);
2793 }
2794 else {
2795 // If the positive/negative part of the difference is 0,
2796 // we won't need to know the number of iterations.
2797 const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff);
2798 const SCEV *NegativePart = getNegativePart(Delta);
2799 if (NegativePart->isZero())
2800 Bound[K].Lower[Dependence::DVEntry::EQ] = NegativePart; // Zero
2801 const SCEV *PositivePart = getPositivePart(Delta);
2802 if (PositivePart->isZero())
2803 Bound[K].Upper[Dependence::DVEntry::EQ] = PositivePart; // Zero
2804 }
2805 }
2806
2807
2808 // Computes the upper and lower bounds for level K
2809 // using the < direction. Records them in Bound.
2810 // Wolfe gives the equations
2811 //
2812 // LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2813 // UB^<_k = (A^+_k - B_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2814 //
2815 // Since we normalize loops, we can simplify these equations to
2816 //
2817 // LB^<_k = (A^-_k - B_k)^- (U_k - 1) - B_k
2818 // UB^<_k = (A^+_k - B_k)^+ (U_k - 1) - B_k
2819 //
2820 // We must be careful to handle the case where the upper bound is unknown.
findBoundsLT(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2821 void DependenceInfo::findBoundsLT(CoefficientInfo *A, CoefficientInfo *B,
2822 BoundInfo *Bound, unsigned K) const {
2823 Bound[K].Lower[Dependence::DVEntry::LT] = nullptr; // Default value = -infinity.
2824 Bound[K].Upper[Dependence::DVEntry::LT] = nullptr; // Default value = +infinity.
2825 if (Bound[K].Iterations) {
2826 const SCEV *Iter_1 = SE->getMinusSCEV(
2827 Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType()));
2828 const SCEV *NegPart =
2829 getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff));
2830 Bound[K].Lower[Dependence::DVEntry::LT] =
2831 SE->getMinusSCEV(SE->getMulExpr(NegPart, Iter_1), B[K].Coeff);
2832 const SCEV *PosPart =
2833 getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff));
2834 Bound[K].Upper[Dependence::DVEntry::LT] =
2835 SE->getMinusSCEV(SE->getMulExpr(PosPart, Iter_1), B[K].Coeff);
2836 }
2837 else {
2838 // If the positive/negative part of the difference is 0,
2839 // we won't need to know the number of iterations.
2840 const SCEV *NegPart =
2841 getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff));
2842 if (NegPart->isZero())
2843 Bound[K].Lower[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff);
2844 const SCEV *PosPart =
2845 getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff));
2846 if (PosPart->isZero())
2847 Bound[K].Upper[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff);
2848 }
2849 }
2850
2851
2852 // Computes the upper and lower bounds for level K
2853 // using the > direction. Records them in Bound.
2854 // Wolfe gives the equations
2855 //
2856 // LB^>_k = (A_k - B^+_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k
2857 // UB^>_k = (A_k - B^-_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k
2858 //
2859 // Since we normalize loops, we can simplify these equations to
2860 //
2861 // LB^>_k = (A_k - B^+_k)^- (U_k - 1) + A_k
2862 // UB^>_k = (A_k - B^-_k)^+ (U_k - 1) + A_k
2863 //
2864 // We must be careful to handle the case where the upper bound is unknown.
findBoundsGT(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2865 void DependenceInfo::findBoundsGT(CoefficientInfo *A, CoefficientInfo *B,
2866 BoundInfo *Bound, unsigned K) const {
2867 Bound[K].Lower[Dependence::DVEntry::GT] = nullptr; // Default value = -infinity.
2868 Bound[K].Upper[Dependence::DVEntry::GT] = nullptr; // Default value = +infinity.
2869 if (Bound[K].Iterations) {
2870 const SCEV *Iter_1 = SE->getMinusSCEV(
2871 Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType()));
2872 const SCEV *NegPart =
2873 getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart));
2874 Bound[K].Lower[Dependence::DVEntry::GT] =
2875 SE->getAddExpr(SE->getMulExpr(NegPart, Iter_1), A[K].Coeff);
2876 const SCEV *PosPart =
2877 getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart));
2878 Bound[K].Upper[Dependence::DVEntry::GT] =
2879 SE->getAddExpr(SE->getMulExpr(PosPart, Iter_1), A[K].Coeff);
2880 }
2881 else {
2882 // If the positive/negative part of the difference is 0,
2883 // we won't need to know the number of iterations.
2884 const SCEV *NegPart = getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart));
2885 if (NegPart->isZero())
2886 Bound[K].Lower[Dependence::DVEntry::GT] = A[K].Coeff;
2887 const SCEV *PosPart = getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart));
2888 if (PosPart->isZero())
2889 Bound[K].Upper[Dependence::DVEntry::GT] = A[K].Coeff;
2890 }
2891 }
2892
2893
2894 // X^+ = max(X, 0)
getPositivePart(const SCEV * X) const2895 const SCEV *DependenceInfo::getPositivePart(const SCEV *X) const {
2896 return SE->getSMaxExpr(X, SE->getZero(X->getType()));
2897 }
2898
2899
2900 // X^- = min(X, 0)
getNegativePart(const SCEV * X) const2901 const SCEV *DependenceInfo::getNegativePart(const SCEV *X) const {
2902 return SE->getSMinExpr(X, SE->getZero(X->getType()));
2903 }
2904
2905
2906 // Walks through the subscript,
2907 // collecting each coefficient, the associated loop bounds,
2908 // and recording its positive and negative parts for later use.
2909 DependenceInfo::CoefficientInfo *
collectCoeffInfo(const SCEV * Subscript,bool SrcFlag,const SCEV * & Constant) const2910 DependenceInfo::collectCoeffInfo(const SCEV *Subscript, bool SrcFlag,
2911 const SCEV *&Constant) const {
2912 const SCEV *Zero = SE->getZero(Subscript->getType());
2913 CoefficientInfo *CI = new CoefficientInfo[MaxLevels + 1];
2914 for (unsigned K = 1; K <= MaxLevels; ++K) {
2915 CI[K].Coeff = Zero;
2916 CI[K].PosPart = Zero;
2917 CI[K].NegPart = Zero;
2918 CI[K].Iterations = nullptr;
2919 }
2920 while (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Subscript)) {
2921 const Loop *L = AddRec->getLoop();
2922 unsigned K = SrcFlag ? mapSrcLoop(L) : mapDstLoop(L);
2923 CI[K].Coeff = AddRec->getStepRecurrence(*SE);
2924 CI[K].PosPart = getPositivePart(CI[K].Coeff);
2925 CI[K].NegPart = getNegativePart(CI[K].Coeff);
2926 CI[K].Iterations = collectUpperBound(L, Subscript->getType());
2927 Subscript = AddRec->getStart();
2928 }
2929 Constant = Subscript;
2930 #ifndef NDEBUG
2931 LLVM_DEBUG(dbgs() << "\tCoefficient Info\n");
2932 for (unsigned K = 1; K <= MaxLevels; ++K) {
2933 LLVM_DEBUG(dbgs() << "\t " << K << "\t" << *CI[K].Coeff);
2934 LLVM_DEBUG(dbgs() << "\tPos Part = ");
2935 LLVM_DEBUG(dbgs() << *CI[K].PosPart);
2936 LLVM_DEBUG(dbgs() << "\tNeg Part = ");
2937 LLVM_DEBUG(dbgs() << *CI[K].NegPart);
2938 LLVM_DEBUG(dbgs() << "\tUpper Bound = ");
2939 if (CI[K].Iterations)
2940 LLVM_DEBUG(dbgs() << *CI[K].Iterations);
2941 else
2942 LLVM_DEBUG(dbgs() << "+inf");
2943 LLVM_DEBUG(dbgs() << '\n');
2944 }
2945 LLVM_DEBUG(dbgs() << "\t Constant = " << *Subscript << '\n');
2946 #endif
2947 return CI;
2948 }
2949
2950
2951 // Looks through all the bounds info and
2952 // computes the lower bound given the current direction settings
2953 // at each level. If the lower bound for any level is -inf,
2954 // the result is -inf.
getLowerBound(BoundInfo * Bound) const2955 const SCEV *DependenceInfo::getLowerBound(BoundInfo *Bound) const {
2956 const SCEV *Sum = Bound[1].Lower[Bound[1].Direction];
2957 for (unsigned K = 2; Sum && K <= MaxLevels; ++K) {
2958 if (Bound[K].Lower[Bound[K].Direction])
2959 Sum = SE->getAddExpr(Sum, Bound[K].Lower[Bound[K].Direction]);
2960 else
2961 Sum = nullptr;
2962 }
2963 return Sum;
2964 }
2965
2966
2967 // Looks through all the bounds info and
2968 // computes the upper bound given the current direction settings
2969 // at each level. If the upper bound at any level is +inf,
2970 // the result is +inf.
getUpperBound(BoundInfo * Bound) const2971 const SCEV *DependenceInfo::getUpperBound(BoundInfo *Bound) const {
2972 const SCEV *Sum = Bound[1].Upper[Bound[1].Direction];
2973 for (unsigned K = 2; Sum && K <= MaxLevels; ++K) {
2974 if (Bound[K].Upper[Bound[K].Direction])
2975 Sum = SE->getAddExpr(Sum, Bound[K].Upper[Bound[K].Direction]);
2976 else
2977 Sum = nullptr;
2978 }
2979 return Sum;
2980 }
2981
2982
2983 //===----------------------------------------------------------------------===//
2984 // Constraint manipulation for Delta test.
2985
2986 // Given a linear SCEV,
2987 // return the coefficient (the step)
2988 // corresponding to the specified loop.
2989 // If there isn't one, return 0.
2990 // For example, given a*i + b*j + c*k, finding the coefficient
2991 // corresponding to the j loop would yield b.
findCoefficient(const SCEV * Expr,const Loop * TargetLoop) const2992 const SCEV *DependenceInfo::findCoefficient(const SCEV *Expr,
2993 const Loop *TargetLoop) const {
2994 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
2995 if (!AddRec)
2996 return SE->getZero(Expr->getType());
2997 if (AddRec->getLoop() == TargetLoop)
2998 return AddRec->getStepRecurrence(*SE);
2999 return findCoefficient(AddRec->getStart(), TargetLoop);
3000 }
3001
3002
3003 // Given a linear SCEV,
3004 // return the SCEV given by zeroing out the coefficient
3005 // corresponding to the specified loop.
3006 // For example, given a*i + b*j + c*k, zeroing the coefficient
3007 // corresponding to the j loop would yield a*i + c*k.
zeroCoefficient(const SCEV * Expr,const Loop * TargetLoop) const3008 const SCEV *DependenceInfo::zeroCoefficient(const SCEV *Expr,
3009 const Loop *TargetLoop) const {
3010 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
3011 if (!AddRec)
3012 return Expr; // ignore
3013 if (AddRec->getLoop() == TargetLoop)
3014 return AddRec->getStart();
3015 return SE->getAddRecExpr(zeroCoefficient(AddRec->getStart(), TargetLoop),
3016 AddRec->getStepRecurrence(*SE),
3017 AddRec->getLoop(),
3018 AddRec->getNoWrapFlags());
3019 }
3020
3021
3022 // Given a linear SCEV Expr,
3023 // return the SCEV given by adding some Value to the
3024 // coefficient corresponding to the specified TargetLoop.
3025 // For example, given a*i + b*j + c*k, adding 1 to the coefficient
3026 // corresponding to the j loop would yield a*i + (b+1)*j + c*k.
addToCoefficient(const SCEV * Expr,const Loop * TargetLoop,const SCEV * Value) const3027 const SCEV *DependenceInfo::addToCoefficient(const SCEV *Expr,
3028 const Loop *TargetLoop,
3029 const SCEV *Value) const {
3030 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
3031 if (!AddRec) // create a new addRec
3032 return SE->getAddRecExpr(Expr,
3033 Value,
3034 TargetLoop,
3035 SCEV::FlagAnyWrap); // Worst case, with no info.
3036 if (AddRec->getLoop() == TargetLoop) {
3037 const SCEV *Sum = SE->getAddExpr(AddRec->getStepRecurrence(*SE), Value);
3038 if (Sum->isZero())
3039 return AddRec->getStart();
3040 return SE->getAddRecExpr(AddRec->getStart(),
3041 Sum,
3042 AddRec->getLoop(),
3043 AddRec->getNoWrapFlags());
3044 }
3045 if (SE->isLoopInvariant(AddRec, TargetLoop))
3046 return SE->getAddRecExpr(AddRec, Value, TargetLoop, SCEV::FlagAnyWrap);
3047 return SE->getAddRecExpr(
3048 addToCoefficient(AddRec->getStart(), TargetLoop, Value),
3049 AddRec->getStepRecurrence(*SE), AddRec->getLoop(),
3050 AddRec->getNoWrapFlags());
3051 }
3052
3053
3054 // Review the constraints, looking for opportunities
3055 // to simplify a subscript pair (Src and Dst).
3056 // Return true if some simplification occurs.
3057 // If the simplification isn't exact (that is, if it is conservative
3058 // in terms of dependence), set consistent to false.
3059 // Corresponds to Figure 5 from the paper
3060 //
3061 // Practical Dependence Testing
3062 // Goff, Kennedy, Tseng
3063 // PLDI 1991
propagate(const SCEV * & Src,const SCEV * & Dst,SmallBitVector & Loops,SmallVectorImpl<Constraint> & Constraints,bool & Consistent)3064 bool DependenceInfo::propagate(const SCEV *&Src, const SCEV *&Dst,
3065 SmallBitVector &Loops,
3066 SmallVectorImpl<Constraint> &Constraints,
3067 bool &Consistent) {
3068 bool Result = false;
3069 for (unsigned LI : Loops.set_bits()) {
3070 LLVM_DEBUG(dbgs() << "\t Constraint[" << LI << "] is");
3071 LLVM_DEBUG(Constraints[LI].dump(dbgs()));
3072 if (Constraints[LI].isDistance())
3073 Result |= propagateDistance(Src, Dst, Constraints[LI], Consistent);
3074 else if (Constraints[LI].isLine())
3075 Result |= propagateLine(Src, Dst, Constraints[LI], Consistent);
3076 else if (Constraints[LI].isPoint())
3077 Result |= propagatePoint(Src, Dst, Constraints[LI]);
3078 }
3079 return Result;
3080 }
3081
3082
3083 // Attempt to propagate a distance
3084 // constraint into a subscript pair (Src and Dst).
3085 // Return true if some simplification occurs.
3086 // If the simplification isn't exact (that is, if it is conservative
3087 // in terms of dependence), set consistent to false.
propagateDistance(const SCEV * & Src,const SCEV * & Dst,Constraint & CurConstraint,bool & Consistent)3088 bool DependenceInfo::propagateDistance(const SCEV *&Src, const SCEV *&Dst,
3089 Constraint &CurConstraint,
3090 bool &Consistent) {
3091 const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3092 LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n");
3093 const SCEV *A_K = findCoefficient(Src, CurLoop);
3094 if (A_K->isZero())
3095 return false;
3096 const SCEV *DA_K = SE->getMulExpr(A_K, CurConstraint.getD());
3097 Src = SE->getMinusSCEV(Src, DA_K);
3098 Src = zeroCoefficient(Src, CurLoop);
3099 LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n");
3100 LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n");
3101 Dst = addToCoefficient(Dst, CurLoop, SE->getNegativeSCEV(A_K));
3102 LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n");
3103 if (!findCoefficient(Dst, CurLoop)->isZero())
3104 Consistent = false;
3105 return true;
3106 }
3107
3108
3109 // Attempt to propagate a line
3110 // constraint into a subscript pair (Src and Dst).
3111 // Return true if some simplification occurs.
3112 // If the simplification isn't exact (that is, if it is conservative
3113 // in terms of dependence), set consistent to false.
propagateLine(const SCEV * & Src,const SCEV * & Dst,Constraint & CurConstraint,bool & Consistent)3114 bool DependenceInfo::propagateLine(const SCEV *&Src, const SCEV *&Dst,
3115 Constraint &CurConstraint,
3116 bool &Consistent) {
3117 const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3118 const SCEV *A = CurConstraint.getA();
3119 const SCEV *B = CurConstraint.getB();
3120 const SCEV *C = CurConstraint.getC();
3121 LLVM_DEBUG(dbgs() << "\t\tA = " << *A << ", B = " << *B << ", C = " << *C
3122 << "\n");
3123 LLVM_DEBUG(dbgs() << "\t\tSrc = " << *Src << "\n");
3124 LLVM_DEBUG(dbgs() << "\t\tDst = " << *Dst << "\n");
3125 if (A->isZero()) {
3126 const SCEVConstant *Bconst = dyn_cast<SCEVConstant>(B);
3127 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3128 if (!Bconst || !Cconst) return false;
3129 APInt Beta = Bconst->getAPInt();
3130 APInt Charlie = Cconst->getAPInt();
3131 APInt CdivB = Charlie.sdiv(Beta);
3132 assert(Charlie.srem(Beta) == 0 && "C should be evenly divisible by B");
3133 const SCEV *AP_K = findCoefficient(Dst, CurLoop);
3134 // Src = SE->getAddExpr(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB)));
3135 Src = SE->getMinusSCEV(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB)));
3136 Dst = zeroCoefficient(Dst, CurLoop);
3137 if (!findCoefficient(Src, CurLoop)->isZero())
3138 Consistent = false;
3139 }
3140 else if (B->isZero()) {
3141 const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A);
3142 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3143 if (!Aconst || !Cconst) return false;
3144 APInt Alpha = Aconst->getAPInt();
3145 APInt Charlie = Cconst->getAPInt();
3146 APInt CdivA = Charlie.sdiv(Alpha);
3147 assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A");
3148 const SCEV *A_K = findCoefficient(Src, CurLoop);
3149 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA)));
3150 Src = zeroCoefficient(Src, CurLoop);
3151 if (!findCoefficient(Dst, CurLoop)->isZero())
3152 Consistent = false;
3153 }
3154 else if (isKnownPredicate(CmpInst::ICMP_EQ, A, B)) {
3155 const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A);
3156 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3157 if (!Aconst || !Cconst) return false;
3158 APInt Alpha = Aconst->getAPInt();
3159 APInt Charlie = Cconst->getAPInt();
3160 APInt CdivA = Charlie.sdiv(Alpha);
3161 assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A");
3162 const SCEV *A_K = findCoefficient(Src, CurLoop);
3163 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA)));
3164 Src = zeroCoefficient(Src, CurLoop);
3165 Dst = addToCoefficient(Dst, CurLoop, A_K);
3166 if (!findCoefficient(Dst, CurLoop)->isZero())
3167 Consistent = false;
3168 }
3169 else {
3170 // paper is incorrect here, or perhaps just misleading
3171 const SCEV *A_K = findCoefficient(Src, CurLoop);
3172 Src = SE->getMulExpr(Src, A);
3173 Dst = SE->getMulExpr(Dst, A);
3174 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, C));
3175 Src = zeroCoefficient(Src, CurLoop);
3176 Dst = addToCoefficient(Dst, CurLoop, SE->getMulExpr(A_K, B));
3177 if (!findCoefficient(Dst, CurLoop)->isZero())
3178 Consistent = false;
3179 }
3180 LLVM_DEBUG(dbgs() << "\t\tnew Src = " << *Src << "\n");
3181 LLVM_DEBUG(dbgs() << "\t\tnew Dst = " << *Dst << "\n");
3182 return true;
3183 }
3184
3185
3186 // Attempt to propagate a point
3187 // constraint into a subscript pair (Src and Dst).
3188 // Return true if some simplification occurs.
propagatePoint(const SCEV * & Src,const SCEV * & Dst,Constraint & CurConstraint)3189 bool DependenceInfo::propagatePoint(const SCEV *&Src, const SCEV *&Dst,
3190 Constraint &CurConstraint) {
3191 const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3192 const SCEV *A_K = findCoefficient(Src, CurLoop);
3193 const SCEV *AP_K = findCoefficient(Dst, CurLoop);
3194 const SCEV *XA_K = SE->getMulExpr(A_K, CurConstraint.getX());
3195 const SCEV *YAP_K = SE->getMulExpr(AP_K, CurConstraint.getY());
3196 LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n");
3197 Src = SE->getAddExpr(Src, SE->getMinusSCEV(XA_K, YAP_K));
3198 Src = zeroCoefficient(Src, CurLoop);
3199 LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n");
3200 LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n");
3201 Dst = zeroCoefficient(Dst, CurLoop);
3202 LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n");
3203 return true;
3204 }
3205
3206
3207 // Update direction vector entry based on the current constraint.
updateDirection(Dependence::DVEntry & Level,const Constraint & CurConstraint) const3208 void DependenceInfo::updateDirection(Dependence::DVEntry &Level,
3209 const Constraint &CurConstraint) const {
3210 LLVM_DEBUG(dbgs() << "\tUpdate direction, constraint =");
3211 LLVM_DEBUG(CurConstraint.dump(dbgs()));
3212 if (CurConstraint.isAny())
3213 ; // use defaults
3214 else if (CurConstraint.isDistance()) {
3215 // this one is consistent, the others aren't
3216 Level.Scalar = false;
3217 Level.Distance = CurConstraint.getD();
3218 unsigned NewDirection = Dependence::DVEntry::NONE;
3219 if (!SE->isKnownNonZero(Level.Distance)) // if may be zero
3220 NewDirection = Dependence::DVEntry::EQ;
3221 if (!SE->isKnownNonPositive(Level.Distance)) // if may be positive
3222 NewDirection |= Dependence::DVEntry::LT;
3223 if (!SE->isKnownNonNegative(Level.Distance)) // if may be negative
3224 NewDirection |= Dependence::DVEntry::GT;
3225 Level.Direction &= NewDirection;
3226 }
3227 else if (CurConstraint.isLine()) {
3228 Level.Scalar = false;
3229 Level.Distance = nullptr;
3230 // direction should be accurate
3231 }
3232 else if (CurConstraint.isPoint()) {
3233 Level.Scalar = false;
3234 Level.Distance = nullptr;
3235 unsigned NewDirection = Dependence::DVEntry::NONE;
3236 if (!isKnownPredicate(CmpInst::ICMP_NE,
3237 CurConstraint.getY(),
3238 CurConstraint.getX()))
3239 // if X may be = Y
3240 NewDirection |= Dependence::DVEntry::EQ;
3241 if (!isKnownPredicate(CmpInst::ICMP_SLE,
3242 CurConstraint.getY(),
3243 CurConstraint.getX()))
3244 // if Y may be > X
3245 NewDirection |= Dependence::DVEntry::LT;
3246 if (!isKnownPredicate(CmpInst::ICMP_SGE,
3247 CurConstraint.getY(),
3248 CurConstraint.getX()))
3249 // if Y may be < X
3250 NewDirection |= Dependence::DVEntry::GT;
3251 Level.Direction &= NewDirection;
3252 }
3253 else
3254 llvm_unreachable("constraint has unexpected kind");
3255 }
3256
3257 /// Check if we can delinearize the subscripts. If the SCEVs representing the
3258 /// source and destination array references are recurrences on a nested loop,
3259 /// this function flattens the nested recurrences into separate recurrences
3260 /// for each loop level.
tryDelinearize(Instruction * Src,Instruction * Dst,SmallVectorImpl<Subscript> & Pair)3261 bool DependenceInfo::tryDelinearize(Instruction *Src, Instruction *Dst,
3262 SmallVectorImpl<Subscript> &Pair) {
3263 assert(isLoadOrStore(Src) && "instruction is not load or store");
3264 assert(isLoadOrStore(Dst) && "instruction is not load or store");
3265 Value *SrcPtr = getLoadStorePointerOperand(Src);
3266 Value *DstPtr = getLoadStorePointerOperand(Dst);
3267
3268 Loop *SrcLoop = LI->getLoopFor(Src->getParent());
3269 Loop *DstLoop = LI->getLoopFor(Dst->getParent());
3270
3271 // Below code mimics the code in Delinearization.cpp
3272 const SCEV *SrcAccessFn =
3273 SE->getSCEVAtScope(SrcPtr, SrcLoop);
3274 const SCEV *DstAccessFn =
3275 SE->getSCEVAtScope(DstPtr, DstLoop);
3276
3277 const SCEVUnknown *SrcBase =
3278 dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
3279 const SCEVUnknown *DstBase =
3280 dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
3281
3282 if (!SrcBase || !DstBase || SrcBase != DstBase)
3283 return false;
3284
3285 const SCEV *ElementSize = SE->getElementSize(Src);
3286 if (ElementSize != SE->getElementSize(Dst))
3287 return false;
3288
3289 const SCEV *SrcSCEV = SE->getMinusSCEV(SrcAccessFn, SrcBase);
3290 const SCEV *DstSCEV = SE->getMinusSCEV(DstAccessFn, DstBase);
3291
3292 const SCEVAddRecExpr *SrcAR = dyn_cast<SCEVAddRecExpr>(SrcSCEV);
3293 const SCEVAddRecExpr *DstAR = dyn_cast<SCEVAddRecExpr>(DstSCEV);
3294 if (!SrcAR || !DstAR || !SrcAR->isAffine() || !DstAR->isAffine())
3295 return false;
3296
3297 // First step: collect parametric terms in both array references.
3298 SmallVector<const SCEV *, 4> Terms;
3299 SE->collectParametricTerms(SrcAR, Terms);
3300 SE->collectParametricTerms(DstAR, Terms);
3301
3302 // Second step: find subscript sizes.
3303 SmallVector<const SCEV *, 4> Sizes;
3304 SE->findArrayDimensions(Terms, Sizes, ElementSize);
3305
3306 // Third step: compute the access functions for each subscript.
3307 SmallVector<const SCEV *, 4> SrcSubscripts, DstSubscripts;
3308 SE->computeAccessFunctions(SrcAR, SrcSubscripts, Sizes);
3309 SE->computeAccessFunctions(DstAR, DstSubscripts, Sizes);
3310
3311 // Fail when there is only a subscript: that's a linearized access function.
3312 if (SrcSubscripts.size() < 2 || DstSubscripts.size() < 2 ||
3313 SrcSubscripts.size() != DstSubscripts.size())
3314 return false;
3315
3316 int size = SrcSubscripts.size();
3317
3318 // Statically check that the array bounds are in-range. The first subscript we
3319 // don't have a size for and it cannot overflow into another subscript, so is
3320 // always safe. The others need to be 0 <= subscript[i] < bound, for both src
3321 // and dst.
3322 // FIXME: It may be better to record these sizes and add them as constraints
3323 // to the dependency checks.
3324 if (!DisableDelinearizationChecks)
3325 for (int i = 1; i < size; ++i) {
3326 if (!isKnownNonNegative(SrcSubscripts[i], SrcPtr))
3327 return false;
3328
3329 if (!isKnownLessThan(SrcSubscripts[i], Sizes[i - 1]))
3330 return false;
3331
3332 if (!isKnownNonNegative(DstSubscripts[i], DstPtr))
3333 return false;
3334
3335 if (!isKnownLessThan(DstSubscripts[i], Sizes[i - 1]))
3336 return false;
3337 }
3338
3339 LLVM_DEBUG({
3340 dbgs() << "\nSrcSubscripts: ";
3341 for (int i = 0; i < size; i++)
3342 dbgs() << *SrcSubscripts[i];
3343 dbgs() << "\nDstSubscripts: ";
3344 for (int i = 0; i < size; i++)
3345 dbgs() << *DstSubscripts[i];
3346 });
3347
3348 // The delinearization transforms a single-subscript MIV dependence test into
3349 // a multi-subscript SIV dependence test that is easier to compute. So we
3350 // resize Pair to contain as many pairs of subscripts as the delinearization
3351 // has found, and then initialize the pairs following the delinearization.
3352 Pair.resize(size);
3353 for (int i = 0; i < size; ++i) {
3354 Pair[i].Src = SrcSubscripts[i];
3355 Pair[i].Dst = DstSubscripts[i];
3356 unifySubscriptType(&Pair[i]);
3357 }
3358
3359 return true;
3360 }
3361
3362 //===----------------------------------------------------------------------===//
3363
3364 #ifndef NDEBUG
3365 // For debugging purposes, dump a small bit vector to dbgs().
dumpSmallBitVector(SmallBitVector & BV)3366 static void dumpSmallBitVector(SmallBitVector &BV) {
3367 dbgs() << "{";
3368 for (unsigned VI : BV.set_bits()) {
3369 dbgs() << VI;
3370 if (BV.find_next(VI) >= 0)
3371 dbgs() << ' ';
3372 }
3373 dbgs() << "}\n";
3374 }
3375 #endif
3376
invalidate(Function & F,const PreservedAnalyses & PA,FunctionAnalysisManager::Invalidator & Inv)3377 bool DependenceInfo::invalidate(Function &F, const PreservedAnalyses &PA,
3378 FunctionAnalysisManager::Invalidator &Inv) {
3379 // Check if the analysis itself has been invalidated.
3380 auto PAC = PA.getChecker<DependenceAnalysis>();
3381 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
3382 return true;
3383
3384 // Check transitive dependencies.
3385 return Inv.invalidate<AAManager>(F, PA) ||
3386 Inv.invalidate<ScalarEvolutionAnalysis>(F, PA) ||
3387 Inv.invalidate<LoopAnalysis>(F, PA);
3388 }
3389
3390 // depends -
3391 // Returns NULL if there is no dependence.
3392 // Otherwise, return a Dependence with as many details as possible.
3393 // Corresponds to Section 3.1 in the paper
3394 //
3395 // Practical Dependence Testing
3396 // Goff, Kennedy, Tseng
3397 // PLDI 1991
3398 //
3399 // Care is required to keep the routine below, getSplitIteration(),
3400 // up to date with respect to this routine.
3401 std::unique_ptr<Dependence>
depends(Instruction * Src,Instruction * Dst,bool PossiblyLoopIndependent)3402 DependenceInfo::depends(Instruction *Src, Instruction *Dst,
3403 bool PossiblyLoopIndependent) {
3404 if (Src == Dst)
3405 PossiblyLoopIndependent = false;
3406
3407 if (!(Src->mayReadOrWriteMemory() && Dst->mayReadOrWriteMemory()))
3408 // if both instructions don't reference memory, there's no dependence
3409 return nullptr;
3410
3411 if (!isLoadOrStore(Src) || !isLoadOrStore(Dst)) {
3412 // can only analyze simple loads and stores, i.e., no calls, invokes, etc.
3413 LLVM_DEBUG(dbgs() << "can only handle simple loads and stores\n");
3414 return std::make_unique<Dependence>(Src, Dst);
3415 }
3416
3417 assert(isLoadOrStore(Src) && "instruction is not load or store");
3418 assert(isLoadOrStore(Dst) && "instruction is not load or store");
3419 Value *SrcPtr = getLoadStorePointerOperand(Src);
3420 Value *DstPtr = getLoadStorePointerOperand(Dst);
3421
3422 switch (underlyingObjectsAlias(AA, F->getParent()->getDataLayout(),
3423 MemoryLocation::get(Dst),
3424 MemoryLocation::get(Src))) {
3425 case MayAlias:
3426 case PartialAlias:
3427 // cannot analyse objects if we don't understand their aliasing.
3428 LLVM_DEBUG(dbgs() << "can't analyze may or partial alias\n");
3429 return std::make_unique<Dependence>(Src, Dst);
3430 case NoAlias:
3431 // If the objects noalias, they are distinct, accesses are independent.
3432 LLVM_DEBUG(dbgs() << "no alias\n");
3433 return nullptr;
3434 case MustAlias:
3435 break; // The underlying objects alias; test accesses for dependence.
3436 }
3437
3438 // establish loop nesting levels
3439 establishNestingLevels(Src, Dst);
3440 LLVM_DEBUG(dbgs() << " common nesting levels = " << CommonLevels << "\n");
3441 LLVM_DEBUG(dbgs() << " maximum nesting levels = " << MaxLevels << "\n");
3442
3443 FullDependence Result(Src, Dst, PossiblyLoopIndependent, CommonLevels);
3444 ++TotalArrayPairs;
3445
3446 unsigned Pairs = 1;
3447 SmallVector<Subscript, 2> Pair(Pairs);
3448 const SCEV *SrcSCEV = SE->getSCEV(SrcPtr);
3449 const SCEV *DstSCEV = SE->getSCEV(DstPtr);
3450 LLVM_DEBUG(dbgs() << " SrcSCEV = " << *SrcSCEV << "\n");
3451 LLVM_DEBUG(dbgs() << " DstSCEV = " << *DstSCEV << "\n");
3452 Pair[0].Src = SrcSCEV;
3453 Pair[0].Dst = DstSCEV;
3454
3455 if (Delinearize) {
3456 if (tryDelinearize(Src, Dst, Pair)) {
3457 LLVM_DEBUG(dbgs() << " delinearized\n");
3458 Pairs = Pair.size();
3459 }
3460 }
3461
3462 for (unsigned P = 0; P < Pairs; ++P) {
3463 Pair[P].Loops.resize(MaxLevels + 1);
3464 Pair[P].GroupLoops.resize(MaxLevels + 1);
3465 Pair[P].Group.resize(Pairs);
3466 removeMatchingExtensions(&Pair[P]);
3467 Pair[P].Classification =
3468 classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()),
3469 Pair[P].Dst, LI->getLoopFor(Dst->getParent()),
3470 Pair[P].Loops);
3471 Pair[P].GroupLoops = Pair[P].Loops;
3472 Pair[P].Group.set(P);
3473 LLVM_DEBUG(dbgs() << " subscript " << P << "\n");
3474 LLVM_DEBUG(dbgs() << "\tsrc = " << *Pair[P].Src << "\n");
3475 LLVM_DEBUG(dbgs() << "\tdst = " << *Pair[P].Dst << "\n");
3476 LLVM_DEBUG(dbgs() << "\tclass = " << Pair[P].Classification << "\n");
3477 LLVM_DEBUG(dbgs() << "\tloops = ");
3478 LLVM_DEBUG(dumpSmallBitVector(Pair[P].Loops));
3479 }
3480
3481 SmallBitVector Separable(Pairs);
3482 SmallBitVector Coupled(Pairs);
3483
3484 // Partition subscripts into separable and minimally-coupled groups
3485 // Algorithm in paper is algorithmically better;
3486 // this may be faster in practice. Check someday.
3487 //
3488 // Here's an example of how it works. Consider this code:
3489 //
3490 // for (i = ...) {
3491 // for (j = ...) {
3492 // for (k = ...) {
3493 // for (l = ...) {
3494 // for (m = ...) {
3495 // A[i][j][k][m] = ...;
3496 // ... = A[0][j][l][i + j];
3497 // }
3498 // }
3499 // }
3500 // }
3501 // }
3502 //
3503 // There are 4 subscripts here:
3504 // 0 [i] and [0]
3505 // 1 [j] and [j]
3506 // 2 [k] and [l]
3507 // 3 [m] and [i + j]
3508 //
3509 // We've already classified each subscript pair as ZIV, SIV, etc.,
3510 // and collected all the loops mentioned by pair P in Pair[P].Loops.
3511 // In addition, we've initialized Pair[P].GroupLoops to Pair[P].Loops
3512 // and set Pair[P].Group = {P}.
3513 //
3514 // Src Dst Classification Loops GroupLoops Group
3515 // 0 [i] [0] SIV {1} {1} {0}
3516 // 1 [j] [j] SIV {2} {2} {1}
3517 // 2 [k] [l] RDIV {3,4} {3,4} {2}
3518 // 3 [m] [i + j] MIV {1,2,5} {1,2,5} {3}
3519 //
3520 // For each subscript SI 0 .. 3, we consider each remaining subscript, SJ.
3521 // So, 0 is compared against 1, 2, and 3; 1 is compared against 2 and 3, etc.
3522 //
3523 // We begin by comparing 0 and 1. The intersection of the GroupLoops is empty.
3524 // Next, 0 and 2. Again, the intersection of their GroupLoops is empty.
3525 // Next 0 and 3. The intersection of their GroupLoop = {1}, not empty,
3526 // so Pair[3].Group = {0,3} and Done = false (that is, 0 will not be added
3527 // to either Separable or Coupled).
3528 //
3529 // Next, we consider 1 and 2. The intersection of the GroupLoops is empty.
3530 // Next, 1 and 3. The intersection of their GroupLoops = {2}, not empty,
3531 // so Pair[3].Group = {0, 1, 3} and Done = false.
3532 //
3533 // Next, we compare 2 against 3. The intersection of the GroupLoops is empty.
3534 // Since Done remains true, we add 2 to the set of Separable pairs.
3535 //
3536 // Finally, we consider 3. There's nothing to compare it with,
3537 // so Done remains true and we add it to the Coupled set.
3538 // Pair[3].Group = {0, 1, 3} and GroupLoops = {1, 2, 5}.
3539 //
3540 // In the end, we've got 1 separable subscript and 1 coupled group.
3541 for (unsigned SI = 0; SI < Pairs; ++SI) {
3542 if (Pair[SI].Classification == Subscript::NonLinear) {
3543 // ignore these, but collect loops for later
3544 ++NonlinearSubscriptPairs;
3545 collectCommonLoops(Pair[SI].Src,
3546 LI->getLoopFor(Src->getParent()),
3547 Pair[SI].Loops);
3548 collectCommonLoops(Pair[SI].Dst,
3549 LI->getLoopFor(Dst->getParent()),
3550 Pair[SI].Loops);
3551 Result.Consistent = false;
3552 } else if (Pair[SI].Classification == Subscript::ZIV) {
3553 // always separable
3554 Separable.set(SI);
3555 }
3556 else {
3557 // SIV, RDIV, or MIV, so check for coupled group
3558 bool Done = true;
3559 for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) {
3560 SmallBitVector Intersection = Pair[SI].GroupLoops;
3561 Intersection &= Pair[SJ].GroupLoops;
3562 if (Intersection.any()) {
3563 // accumulate set of all the loops in group
3564 Pair[SJ].GroupLoops |= Pair[SI].GroupLoops;
3565 // accumulate set of all subscripts in group
3566 Pair[SJ].Group |= Pair[SI].Group;
3567 Done = false;
3568 }
3569 }
3570 if (Done) {
3571 if (Pair[SI].Group.count() == 1) {
3572 Separable.set(SI);
3573 ++SeparableSubscriptPairs;
3574 }
3575 else {
3576 Coupled.set(SI);
3577 ++CoupledSubscriptPairs;
3578 }
3579 }
3580 }
3581 }
3582
3583 LLVM_DEBUG(dbgs() << " Separable = ");
3584 LLVM_DEBUG(dumpSmallBitVector(Separable));
3585 LLVM_DEBUG(dbgs() << " Coupled = ");
3586 LLVM_DEBUG(dumpSmallBitVector(Coupled));
3587
3588 Constraint NewConstraint;
3589 NewConstraint.setAny(SE);
3590
3591 // test separable subscripts
3592 for (unsigned SI : Separable.set_bits()) {
3593 LLVM_DEBUG(dbgs() << "testing subscript " << SI);
3594 switch (Pair[SI].Classification) {
3595 case Subscript::ZIV:
3596 LLVM_DEBUG(dbgs() << ", ZIV\n");
3597 if (testZIV(Pair[SI].Src, Pair[SI].Dst, Result))
3598 return nullptr;
3599 break;
3600 case Subscript::SIV: {
3601 LLVM_DEBUG(dbgs() << ", SIV\n");
3602 unsigned Level;
3603 const SCEV *SplitIter = nullptr;
3604 if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level, Result, NewConstraint,
3605 SplitIter))
3606 return nullptr;
3607 break;
3608 }
3609 case Subscript::RDIV:
3610 LLVM_DEBUG(dbgs() << ", RDIV\n");
3611 if (testRDIV(Pair[SI].Src, Pair[SI].Dst, Result))
3612 return nullptr;
3613 break;
3614 case Subscript::MIV:
3615 LLVM_DEBUG(dbgs() << ", MIV\n");
3616 if (testMIV(Pair[SI].Src, Pair[SI].Dst, Pair[SI].Loops, Result))
3617 return nullptr;
3618 break;
3619 default:
3620 llvm_unreachable("subscript has unexpected classification");
3621 }
3622 }
3623
3624 if (Coupled.count()) {
3625 // test coupled subscript groups
3626 LLVM_DEBUG(dbgs() << "starting on coupled subscripts\n");
3627 LLVM_DEBUG(dbgs() << "MaxLevels + 1 = " << MaxLevels + 1 << "\n");
3628 SmallVector<Constraint, 4> Constraints(MaxLevels + 1);
3629 for (unsigned II = 0; II <= MaxLevels; ++II)
3630 Constraints[II].setAny(SE);
3631 for (unsigned SI : Coupled.set_bits()) {
3632 LLVM_DEBUG(dbgs() << "testing subscript group " << SI << " { ");
3633 SmallBitVector Group(Pair[SI].Group);
3634 SmallBitVector Sivs(Pairs);
3635 SmallBitVector Mivs(Pairs);
3636 SmallBitVector ConstrainedLevels(MaxLevels + 1);
3637 SmallVector<Subscript *, 4> PairsInGroup;
3638 for (unsigned SJ : Group.set_bits()) {
3639 LLVM_DEBUG(dbgs() << SJ << " ");
3640 if (Pair[SJ].Classification == Subscript::SIV)
3641 Sivs.set(SJ);
3642 else
3643 Mivs.set(SJ);
3644 PairsInGroup.push_back(&Pair[SJ]);
3645 }
3646 unifySubscriptType(PairsInGroup);
3647 LLVM_DEBUG(dbgs() << "}\n");
3648 while (Sivs.any()) {
3649 bool Changed = false;
3650 for (unsigned SJ : Sivs.set_bits()) {
3651 LLVM_DEBUG(dbgs() << "testing subscript " << SJ << ", SIV\n");
3652 // SJ is an SIV subscript that's part of the current coupled group
3653 unsigned Level;
3654 const SCEV *SplitIter = nullptr;
3655 LLVM_DEBUG(dbgs() << "SIV\n");
3656 if (testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level, Result, NewConstraint,
3657 SplitIter))
3658 return nullptr;
3659 ConstrainedLevels.set(Level);
3660 if (intersectConstraints(&Constraints[Level], &NewConstraint)) {
3661 if (Constraints[Level].isEmpty()) {
3662 ++DeltaIndependence;
3663 return nullptr;
3664 }
3665 Changed = true;
3666 }
3667 Sivs.reset(SJ);
3668 }
3669 if (Changed) {
3670 // propagate, possibly creating new SIVs and ZIVs
3671 LLVM_DEBUG(dbgs() << " propagating\n");
3672 LLVM_DEBUG(dbgs() << "\tMivs = ");
3673 LLVM_DEBUG(dumpSmallBitVector(Mivs));
3674 for (unsigned SJ : Mivs.set_bits()) {
3675 // SJ is an MIV subscript that's part of the current coupled group
3676 LLVM_DEBUG(dbgs() << "\tSJ = " << SJ << "\n");
3677 if (propagate(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops,
3678 Constraints, Result.Consistent)) {
3679 LLVM_DEBUG(dbgs() << "\t Changed\n");
3680 ++DeltaPropagations;
3681 Pair[SJ].Classification =
3682 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()),
3683 Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()),
3684 Pair[SJ].Loops);
3685 switch (Pair[SJ].Classification) {
3686 case Subscript::ZIV:
3687 LLVM_DEBUG(dbgs() << "ZIV\n");
3688 if (testZIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
3689 return nullptr;
3690 Mivs.reset(SJ);
3691 break;
3692 case Subscript::SIV:
3693 Sivs.set(SJ);
3694 Mivs.reset(SJ);
3695 break;
3696 case Subscript::RDIV:
3697 case Subscript::MIV:
3698 break;
3699 default:
3700 llvm_unreachable("bad subscript classification");
3701 }
3702 }
3703 }
3704 }
3705 }
3706
3707 // test & propagate remaining RDIVs
3708 for (unsigned SJ : Mivs.set_bits()) {
3709 if (Pair[SJ].Classification == Subscript::RDIV) {
3710 LLVM_DEBUG(dbgs() << "RDIV test\n");
3711 if (testRDIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
3712 return nullptr;
3713 // I don't yet understand how to propagate RDIV results
3714 Mivs.reset(SJ);
3715 }
3716 }
3717
3718 // test remaining MIVs
3719 // This code is temporary.
3720 // Better to somehow test all remaining subscripts simultaneously.
3721 for (unsigned SJ : Mivs.set_bits()) {
3722 if (Pair[SJ].Classification == Subscript::MIV) {
3723 LLVM_DEBUG(dbgs() << "MIV test\n");
3724 if (testMIV(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops, Result))
3725 return nullptr;
3726 }
3727 else
3728 llvm_unreachable("expected only MIV subscripts at this point");
3729 }
3730
3731 // update Result.DV from constraint vector
3732 LLVM_DEBUG(dbgs() << " updating\n");
3733 for (unsigned SJ : ConstrainedLevels.set_bits()) {
3734 if (SJ > CommonLevels)
3735 break;
3736 updateDirection(Result.DV[SJ - 1], Constraints[SJ]);
3737 if (Result.DV[SJ - 1].Direction == Dependence::DVEntry::NONE)
3738 return nullptr;
3739 }
3740 }
3741 }
3742
3743 // Make sure the Scalar flags are set correctly.
3744 SmallBitVector CompleteLoops(MaxLevels + 1);
3745 for (unsigned SI = 0; SI < Pairs; ++SI)
3746 CompleteLoops |= Pair[SI].Loops;
3747 for (unsigned II = 1; II <= CommonLevels; ++II)
3748 if (CompleteLoops[II])
3749 Result.DV[II - 1].Scalar = false;
3750
3751 if (PossiblyLoopIndependent) {
3752 // Make sure the LoopIndependent flag is set correctly.
3753 // All directions must include equal, otherwise no
3754 // loop-independent dependence is possible.
3755 for (unsigned II = 1; II <= CommonLevels; ++II) {
3756 if (!(Result.getDirection(II) & Dependence::DVEntry::EQ)) {
3757 Result.LoopIndependent = false;
3758 break;
3759 }
3760 }
3761 }
3762 else {
3763 // On the other hand, if all directions are equal and there's no
3764 // loop-independent dependence possible, then no dependence exists.
3765 bool AllEqual = true;
3766 for (unsigned II = 1; II <= CommonLevels; ++II) {
3767 if (Result.getDirection(II) != Dependence::DVEntry::EQ) {
3768 AllEqual = false;
3769 break;
3770 }
3771 }
3772 if (AllEqual)
3773 return nullptr;
3774 }
3775
3776 return std::make_unique<FullDependence>(std::move(Result));
3777 }
3778
3779 //===----------------------------------------------------------------------===//
3780 // getSplitIteration -
3781 // Rather than spend rarely-used space recording the splitting iteration
3782 // during the Weak-Crossing SIV test, we re-compute it on demand.
3783 // The re-computation is basically a repeat of the entire dependence test,
3784 // though simplified since we know that the dependence exists.
3785 // It's tedious, since we must go through all propagations, etc.
3786 //
3787 // Care is required to keep this code up to date with respect to the routine
3788 // above, depends().
3789 //
3790 // Generally, the dependence analyzer will be used to build
3791 // a dependence graph for a function (basically a map from instructions
3792 // to dependences). Looking for cycles in the graph shows us loops
3793 // that cannot be trivially vectorized/parallelized.
3794 //
3795 // We can try to improve the situation by examining all the dependences
3796 // that make up the cycle, looking for ones we can break.
3797 // Sometimes, peeling the first or last iteration of a loop will break
3798 // dependences, and we've got flags for those possibilities.
3799 // Sometimes, splitting a loop at some other iteration will do the trick,
3800 // and we've got a flag for that case. Rather than waste the space to
3801 // record the exact iteration (since we rarely know), we provide
3802 // a method that calculates the iteration. It's a drag that it must work
3803 // from scratch, but wonderful in that it's possible.
3804 //
3805 // Here's an example:
3806 //
3807 // for (i = 0; i < 10; i++)
3808 // A[i] = ...
3809 // ... = A[11 - i]
3810 //
3811 // There's a loop-carried flow dependence from the store to the load,
3812 // found by the weak-crossing SIV test. The dependence will have a flag,
3813 // indicating that the dependence can be broken by splitting the loop.
3814 // Calling getSplitIteration will return 5.
3815 // Splitting the loop breaks the dependence, like so:
3816 //
3817 // for (i = 0; i <= 5; i++)
3818 // A[i] = ...
3819 // ... = A[11 - i]
3820 // for (i = 6; i < 10; i++)
3821 // A[i] = ...
3822 // ... = A[11 - i]
3823 //
3824 // breaks the dependence and allows us to vectorize/parallelize
3825 // both loops.
getSplitIteration(const Dependence & Dep,unsigned SplitLevel)3826 const SCEV *DependenceInfo::getSplitIteration(const Dependence &Dep,
3827 unsigned SplitLevel) {
3828 assert(Dep.isSplitable(SplitLevel) &&
3829 "Dep should be splitable at SplitLevel");
3830 Instruction *Src = Dep.getSrc();
3831 Instruction *Dst = Dep.getDst();
3832 assert(Src->mayReadFromMemory() || Src->mayWriteToMemory());
3833 assert(Dst->mayReadFromMemory() || Dst->mayWriteToMemory());
3834 assert(isLoadOrStore(Src));
3835 assert(isLoadOrStore(Dst));
3836 Value *SrcPtr = getLoadStorePointerOperand(Src);
3837 Value *DstPtr = getLoadStorePointerOperand(Dst);
3838 assert(underlyingObjectsAlias(AA, F->getParent()->getDataLayout(),
3839 MemoryLocation::get(Dst),
3840 MemoryLocation::get(Src)) == MustAlias);
3841
3842 // establish loop nesting levels
3843 establishNestingLevels(Src, Dst);
3844
3845 FullDependence Result(Src, Dst, false, CommonLevels);
3846
3847 unsigned Pairs = 1;
3848 SmallVector<Subscript, 2> Pair(Pairs);
3849 const SCEV *SrcSCEV = SE->getSCEV(SrcPtr);
3850 const SCEV *DstSCEV = SE->getSCEV(DstPtr);
3851 Pair[0].Src = SrcSCEV;
3852 Pair[0].Dst = DstSCEV;
3853
3854 if (Delinearize) {
3855 if (tryDelinearize(Src, Dst, Pair)) {
3856 LLVM_DEBUG(dbgs() << " delinearized\n");
3857 Pairs = Pair.size();
3858 }
3859 }
3860
3861 for (unsigned P = 0; P < Pairs; ++P) {
3862 Pair[P].Loops.resize(MaxLevels + 1);
3863 Pair[P].GroupLoops.resize(MaxLevels + 1);
3864 Pair[P].Group.resize(Pairs);
3865 removeMatchingExtensions(&Pair[P]);
3866 Pair[P].Classification =
3867 classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()),
3868 Pair[P].Dst, LI->getLoopFor(Dst->getParent()),
3869 Pair[P].Loops);
3870 Pair[P].GroupLoops = Pair[P].Loops;
3871 Pair[P].Group.set(P);
3872 }
3873
3874 SmallBitVector Separable(Pairs);
3875 SmallBitVector Coupled(Pairs);
3876
3877 // partition subscripts into separable and minimally-coupled groups
3878 for (unsigned SI = 0; SI < Pairs; ++SI) {
3879 if (Pair[SI].Classification == Subscript::NonLinear) {
3880 // ignore these, but collect loops for later
3881 collectCommonLoops(Pair[SI].Src,
3882 LI->getLoopFor(Src->getParent()),
3883 Pair[SI].Loops);
3884 collectCommonLoops(Pair[SI].Dst,
3885 LI->getLoopFor(Dst->getParent()),
3886 Pair[SI].Loops);
3887 Result.Consistent = false;
3888 }
3889 else if (Pair[SI].Classification == Subscript::ZIV)
3890 Separable.set(SI);
3891 else {
3892 // SIV, RDIV, or MIV, so check for coupled group
3893 bool Done = true;
3894 for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) {
3895 SmallBitVector Intersection = Pair[SI].GroupLoops;
3896 Intersection &= Pair[SJ].GroupLoops;
3897 if (Intersection.any()) {
3898 // accumulate set of all the loops in group
3899 Pair[SJ].GroupLoops |= Pair[SI].GroupLoops;
3900 // accumulate set of all subscripts in group
3901 Pair[SJ].Group |= Pair[SI].Group;
3902 Done = false;
3903 }
3904 }
3905 if (Done) {
3906 if (Pair[SI].Group.count() == 1)
3907 Separable.set(SI);
3908 else
3909 Coupled.set(SI);
3910 }
3911 }
3912 }
3913
3914 Constraint NewConstraint;
3915 NewConstraint.setAny(SE);
3916
3917 // test separable subscripts
3918 for (unsigned SI : Separable.set_bits()) {
3919 switch (Pair[SI].Classification) {
3920 case Subscript::SIV: {
3921 unsigned Level;
3922 const SCEV *SplitIter = nullptr;
3923 (void) testSIV(Pair[SI].Src, Pair[SI].Dst, Level,
3924 Result, NewConstraint, SplitIter);
3925 if (Level == SplitLevel) {
3926 assert(SplitIter != nullptr);
3927 return SplitIter;
3928 }
3929 break;
3930 }
3931 case Subscript::ZIV:
3932 case Subscript::RDIV:
3933 case Subscript::MIV:
3934 break;
3935 default:
3936 llvm_unreachable("subscript has unexpected classification");
3937 }
3938 }
3939
3940 if (Coupled.count()) {
3941 // test coupled subscript groups
3942 SmallVector<Constraint, 4> Constraints(MaxLevels + 1);
3943 for (unsigned II = 0; II <= MaxLevels; ++II)
3944 Constraints[II].setAny(SE);
3945 for (unsigned SI : Coupled.set_bits()) {
3946 SmallBitVector Group(Pair[SI].Group);
3947 SmallBitVector Sivs(Pairs);
3948 SmallBitVector Mivs(Pairs);
3949 SmallBitVector ConstrainedLevels(MaxLevels + 1);
3950 for (unsigned SJ : Group.set_bits()) {
3951 if (Pair[SJ].Classification == Subscript::SIV)
3952 Sivs.set(SJ);
3953 else
3954 Mivs.set(SJ);
3955 }
3956 while (Sivs.any()) {
3957 bool Changed = false;
3958 for (unsigned SJ : Sivs.set_bits()) {
3959 // SJ is an SIV subscript that's part of the current coupled group
3960 unsigned Level;
3961 const SCEV *SplitIter = nullptr;
3962 (void) testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level,
3963 Result, NewConstraint, SplitIter);
3964 if (Level == SplitLevel && SplitIter)
3965 return SplitIter;
3966 ConstrainedLevels.set(Level);
3967 if (intersectConstraints(&Constraints[Level], &NewConstraint))
3968 Changed = true;
3969 Sivs.reset(SJ);
3970 }
3971 if (Changed) {
3972 // propagate, possibly creating new SIVs and ZIVs
3973 for (unsigned SJ : Mivs.set_bits()) {
3974 // SJ is an MIV subscript that's part of the current coupled group
3975 if (propagate(Pair[SJ].Src, Pair[SJ].Dst,
3976 Pair[SJ].Loops, Constraints, Result.Consistent)) {
3977 Pair[SJ].Classification =
3978 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()),
3979 Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()),
3980 Pair[SJ].Loops);
3981 switch (Pair[SJ].Classification) {
3982 case Subscript::ZIV:
3983 Mivs.reset(SJ);
3984 break;
3985 case Subscript::SIV:
3986 Sivs.set(SJ);
3987 Mivs.reset(SJ);
3988 break;
3989 case Subscript::RDIV:
3990 case Subscript::MIV:
3991 break;
3992 default:
3993 llvm_unreachable("bad subscript classification");
3994 }
3995 }
3996 }
3997 }
3998 }
3999 }
4000 }
4001 llvm_unreachable("somehow reached end of routine");
4002 return nullptr;
4003 }
4004