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/Statistic.h"
54 #include "llvm/Analysis/AliasAnalysis.h"
55 #include "llvm/Analysis/Delinearization.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/IR/InstIterator.h"
61 #include "llvm/IR/Module.h"
62 #include "llvm/InitializePasses.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/ErrorHandling.h"
66 #include "llvm/Support/raw_ostream.h"
67 
68 using namespace llvm;
69 
70 #define DEBUG_TYPE "da"
71 
72 //===----------------------------------------------------------------------===//
73 // statistics
74 
75 STATISTIC(TotalArrayPairs, "Array pairs tested");
76 STATISTIC(SeparableSubscriptPairs, "Separable subscript pairs");
77 STATISTIC(CoupledSubscriptPairs, "Coupled subscript pairs");
78 STATISTIC(NonlinearSubscriptPairs, "Nonlinear subscript pairs");
79 STATISTIC(ZIVapplications, "ZIV applications");
80 STATISTIC(ZIVindependence, "ZIV independence");
81 STATISTIC(StrongSIVapplications, "Strong SIV applications");
82 STATISTIC(StrongSIVsuccesses, "Strong SIV successes");
83 STATISTIC(StrongSIVindependence, "Strong SIV independence");
84 STATISTIC(WeakCrossingSIVapplications, "Weak-Crossing SIV applications");
85 STATISTIC(WeakCrossingSIVsuccesses, "Weak-Crossing SIV successes");
86 STATISTIC(WeakCrossingSIVindependence, "Weak-Crossing SIV independence");
87 STATISTIC(ExactSIVapplications, "Exact SIV applications");
88 STATISTIC(ExactSIVsuccesses, "Exact SIV successes");
89 STATISTIC(ExactSIVindependence, "Exact SIV independence");
90 STATISTIC(WeakZeroSIVapplications, "Weak-Zero SIV applications");
91 STATISTIC(WeakZeroSIVsuccesses, "Weak-Zero SIV successes");
92 STATISTIC(WeakZeroSIVindependence, "Weak-Zero SIV independence");
93 STATISTIC(ExactRDIVapplications, "Exact RDIV applications");
94 STATISTIC(ExactRDIVindependence, "Exact RDIV independence");
95 STATISTIC(SymbolicRDIVapplications, "Symbolic RDIV applications");
96 STATISTIC(SymbolicRDIVindependence, "Symbolic RDIV independence");
97 STATISTIC(DeltaApplications, "Delta applications");
98 STATISTIC(DeltaSuccesses, "Delta successes");
99 STATISTIC(DeltaIndependence, "Delta independence");
100 STATISTIC(DeltaPropagations, "Delta propagations");
101 STATISTIC(GCDapplications, "GCD applications");
102 STATISTIC(GCDsuccesses, "GCD successes");
103 STATISTIC(GCDindependence, "GCD independence");
104 STATISTIC(BanerjeeApplications, "Banerjee applications");
105 STATISTIC(BanerjeeIndependence, "Banerjee independence");
106 STATISTIC(BanerjeeSuccesses, "Banerjee successes");
107 
108 static cl::opt<bool>
109     Delinearize("da-delinearize", cl::init(true), cl::Hidden,
110                 cl::desc("Try to delinearize array references."));
111 static cl::opt<bool> DisableDelinearizationChecks(
112     "da-disable-delinearization-checks", cl::Hidden,
113     cl::desc(
114         "Disable checks that try to statically verify validity of "
115         "delinearized subscripts. Enabling this option may result in incorrect "
116         "dependence vectors for languages that allow the subscript of one "
117         "dimension to underflow or overflow into another dimension."));
118 
119 static cl::opt<unsigned> MIVMaxLevelThreshold(
120     "da-miv-max-level-threshold", cl::init(7), cl::Hidden,
121     cl::desc("Maximum depth allowed for the recursive algorithm used to "
122              "explore MIV direction vectors."));
123 
124 //===----------------------------------------------------------------------===//
125 // basics
126 
127 DependenceAnalysis::Result
128 DependenceAnalysis::run(Function &F, FunctionAnalysisManager &FAM) {
129   auto &AA = FAM.getResult<AAManager>(F);
130   auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F);
131   auto &LI = FAM.getResult<LoopAnalysis>(F);
132   return DependenceInfo(&F, &AA, &SE, &LI);
133 }
134 
135 AnalysisKey DependenceAnalysis::Key;
136 
137 INITIALIZE_PASS_BEGIN(DependenceAnalysisWrapperPass, "da",
138                       "Dependence Analysis", true, true)
139 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
140 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
141 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
142 INITIALIZE_PASS_END(DependenceAnalysisWrapperPass, "da", "Dependence Analysis",
143                     true, true)
144 
145 char DependenceAnalysisWrapperPass::ID = 0;
146 
147 DependenceAnalysisWrapperPass::DependenceAnalysisWrapperPass()
148     : FunctionPass(ID) {
149   initializeDependenceAnalysisWrapperPassPass(*PassRegistry::getPassRegistry());
150 }
151 
152 FunctionPass *llvm::createDependenceAnalysisWrapperPass() {
153   return new DependenceAnalysisWrapperPass();
154 }
155 
156 bool DependenceAnalysisWrapperPass::runOnFunction(Function &F) {
157   auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
158   auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
159   auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
160   info.reset(new DependenceInfo(&F, &AA, &SE, &LI));
161   return false;
162 }
163 
164 DependenceInfo &DependenceAnalysisWrapperPass::getDI() const { return *info; }
165 
166 void DependenceAnalysisWrapperPass::releaseMemory() { info.reset(); }
167 
168 void DependenceAnalysisWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
169   AU.setPreservesAll();
170   AU.addRequiredTransitive<AAResultsWrapperPass>();
171   AU.addRequiredTransitive<ScalarEvolutionWrapperPass>();
172   AU.addRequiredTransitive<LoopInfoWrapperPass>();
173 }
174 
175 // Used to test the dependence analyzer.
176 // Looks through the function, noting instructions that may access memory.
177 // Calls depends() on every possible pair and prints out the result.
178 // Ignores all other instructions.
179 static void dumpExampleDependence(raw_ostream &OS, DependenceInfo *DA) {
180   auto *F = DA->getFunction();
181   for (inst_iterator SrcI = inst_begin(F), SrcE = inst_end(F); SrcI != SrcE;
182        ++SrcI) {
183     if (SrcI->mayReadOrWriteMemory()) {
184       for (inst_iterator DstI = SrcI, DstE = inst_end(F);
185            DstI != DstE; ++DstI) {
186         if (DstI->mayReadOrWriteMemory()) {
187           OS << "Src:" << *SrcI << " --> Dst:" << *DstI << "\n";
188           OS << "  da analyze - ";
189           if (auto D = DA->depends(&*SrcI, &*DstI, true)) {
190             D->dump(OS);
191             for (unsigned Level = 1; Level <= D->getLevels(); Level++) {
192               if (D->isSplitable(Level)) {
193                 OS << "  da analyze - split level = " << Level;
194                 OS << ", iteration = " << *DA->getSplitIteration(*D, Level);
195                 OS << "!\n";
196               }
197             }
198           }
199           else
200             OS << "none!\n";
201         }
202       }
203     }
204   }
205 }
206 
207 void DependenceAnalysisWrapperPass::print(raw_ostream &OS,
208                                           const Module *) const {
209   dumpExampleDependence(OS, info.get());
210 }
211 
212 PreservedAnalyses
213 DependenceAnalysisPrinterPass::run(Function &F, FunctionAnalysisManager &FAM) {
214   OS << "'Dependence Analysis' for function '" << F.getName() << "':\n";
215   dumpExampleDependence(OS, &FAM.getResult<DependenceAnalysis>(F));
216   return PreservedAnalyses::all();
217 }
218 
219 //===----------------------------------------------------------------------===//
220 // Dependence methods
221 
222 // Returns true if this is an input dependence.
223 bool Dependence::isInput() const {
224   return Src->mayReadFromMemory() && Dst->mayReadFromMemory();
225 }
226 
227 
228 // Returns true if this is an output dependence.
229 bool Dependence::isOutput() const {
230   return Src->mayWriteToMemory() && Dst->mayWriteToMemory();
231 }
232 
233 
234 // Returns true if this is an flow (aka true)  dependence.
235 bool Dependence::isFlow() const {
236   return Src->mayWriteToMemory() && Dst->mayReadFromMemory();
237 }
238 
239 
240 // Returns true if this is an anti dependence.
241 bool Dependence::isAnti() const {
242   return Src->mayReadFromMemory() && Dst->mayWriteToMemory();
243 }
244 
245 
246 // Returns true if a particular level is scalar; that is,
247 // if no subscript in the source or destination mention the induction
248 // variable associated with the loop at this level.
249 // Leave this out of line, so it will serve as a virtual method anchor
250 bool Dependence::isScalar(unsigned level) const {
251   return false;
252 }
253 
254 
255 //===----------------------------------------------------------------------===//
256 // FullDependence methods
257 
258 FullDependence::FullDependence(Instruction *Source, Instruction *Destination,
259                                bool PossiblyLoopIndependent,
260                                unsigned CommonLevels)
261     : Dependence(Source, Destination), Levels(CommonLevels),
262       LoopIndependent(PossiblyLoopIndependent) {
263   Consistent = true;
264   if (CommonLevels)
265     DV = std::make_unique<DVEntry[]>(CommonLevels);
266 }
267 
268 // The rest are simple getters that hide the implementation.
269 
270 // getDirection - Returns the direction associated with a particular level.
271 unsigned FullDependence::getDirection(unsigned Level) const {
272   assert(0 < Level && Level <= Levels && "Level out of range");
273   return DV[Level - 1].Direction;
274 }
275 
276 
277 // Returns the distance (or NULL) associated with a particular level.
278 const SCEV *FullDependence::getDistance(unsigned Level) const {
279   assert(0 < Level && Level <= Levels && "Level out of range");
280   return DV[Level - 1].Distance;
281 }
282 
283 
284 // Returns true if a particular level is scalar; that is,
285 // if no subscript in the source or destination mention the induction
286 // variable associated with the loop at this level.
287 bool FullDependence::isScalar(unsigned Level) const {
288   assert(0 < Level && Level <= Levels && "Level out of range");
289   return DV[Level - 1].Scalar;
290 }
291 
292 
293 // Returns true if peeling the first iteration from this loop
294 // will break this dependence.
295 bool FullDependence::isPeelFirst(unsigned Level) const {
296   assert(0 < Level && Level <= Levels && "Level out of range");
297   return DV[Level - 1].PeelFirst;
298 }
299 
300 
301 // Returns true if peeling the last iteration from this loop
302 // will break this dependence.
303 bool FullDependence::isPeelLast(unsigned Level) const {
304   assert(0 < Level && Level <= Levels && "Level out of range");
305   return DV[Level - 1].PeelLast;
306 }
307 
308 
309 // Returns true if splitting this loop will break the dependence.
310 bool FullDependence::isSplitable(unsigned Level) const {
311   assert(0 < Level && Level <= Levels && "Level out of range");
312   return DV[Level - 1].Splitable;
313 }
314 
315 
316 //===----------------------------------------------------------------------===//
317 // DependenceInfo::Constraint methods
318 
319 // If constraint is a point <X, Y>, returns X.
320 // Otherwise assert.
321 const SCEV *DependenceInfo::Constraint::getX() const {
322   assert(Kind == Point && "Kind should be Point");
323   return A;
324 }
325 
326 
327 // If constraint is a point <X, Y>, returns Y.
328 // Otherwise assert.
329 const SCEV *DependenceInfo::Constraint::getY() const {
330   assert(Kind == Point && "Kind should be Point");
331   return B;
332 }
333 
334 
335 // If constraint is a line AX + BY = C, returns A.
336 // Otherwise assert.
337 const SCEV *DependenceInfo::Constraint::getA() const {
338   assert((Kind == Line || Kind == Distance) &&
339          "Kind should be Line (or Distance)");
340   return A;
341 }
342 
343 
344 // If constraint is a line AX + BY = C, returns B.
345 // Otherwise assert.
346 const SCEV *DependenceInfo::Constraint::getB() const {
347   assert((Kind == Line || Kind == Distance) &&
348          "Kind should be Line (or Distance)");
349   return B;
350 }
351 
352 
353 // If constraint is a line AX + BY = C, returns C.
354 // Otherwise assert.
355 const SCEV *DependenceInfo::Constraint::getC() const {
356   assert((Kind == Line || Kind == Distance) &&
357          "Kind should be Line (or Distance)");
358   return C;
359 }
360 
361 
362 // If constraint is a distance, returns D.
363 // Otherwise assert.
364 const SCEV *DependenceInfo::Constraint::getD() const {
365   assert(Kind == Distance && "Kind should be Distance");
366   return SE->getNegativeSCEV(C);
367 }
368 
369 
370 // Returns the loop associated with this constraint.
371 const Loop *DependenceInfo::Constraint::getAssociatedLoop() const {
372   assert((Kind == Distance || Kind == Line || Kind == Point) &&
373          "Kind should be Distance, Line, or Point");
374   return AssociatedLoop;
375 }
376 
377 void DependenceInfo::Constraint::setPoint(const SCEV *X, const SCEV *Y,
378                                           const Loop *CurLoop) {
379   Kind = Point;
380   A = X;
381   B = Y;
382   AssociatedLoop = CurLoop;
383 }
384 
385 void DependenceInfo::Constraint::setLine(const SCEV *AA, const SCEV *BB,
386                                          const SCEV *CC, const Loop *CurLoop) {
387   Kind = Line;
388   A = AA;
389   B = BB;
390   C = CC;
391   AssociatedLoop = CurLoop;
392 }
393 
394 void DependenceInfo::Constraint::setDistance(const SCEV *D,
395                                              const Loop *CurLoop) {
396   Kind = Distance;
397   A = SE->getOne(D->getType());
398   B = SE->getNegativeSCEV(A);
399   C = SE->getNegativeSCEV(D);
400   AssociatedLoop = CurLoop;
401 }
402 
403 void DependenceInfo::Constraint::setEmpty() { Kind = Empty; }
404 
405 void DependenceInfo::Constraint::setAny(ScalarEvolution *NewSE) {
406   SE = NewSE;
407   Kind = Any;
408 }
409 
410 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
411 // For debugging purposes. Dumps the constraint out to OS.
412 LLVM_DUMP_METHOD void DependenceInfo::Constraint::dump(raw_ostream &OS) const {
413   if (isEmpty())
414     OS << " Empty\n";
415   else if (isAny())
416     OS << " Any\n";
417   else if (isPoint())
418     OS << " Point is <" << *getX() << ", " << *getY() << ">\n";
419   else if (isDistance())
420     OS << " Distance is " << *getD() <<
421       " (" << *getA() << "*X + " << *getB() << "*Y = " << *getC() << ")\n";
422   else if (isLine())
423     OS << " Line is " << *getA() << "*X + " <<
424       *getB() << "*Y = " << *getC() << "\n";
425   else
426     llvm_unreachable("unknown constraint type in Constraint::dump");
427 }
428 #endif
429 
430 
431 // Updates X with the intersection
432 // of the Constraints X and Y. Returns true if X has changed.
433 // Corresponds to Figure 4 from the paper
434 //
435 //            Practical Dependence Testing
436 //            Goff, Kennedy, Tseng
437 //            PLDI 1991
438 bool DependenceInfo::intersectConstraints(Constraint *X, const Constraint *Y) {
439   ++DeltaApplications;
440   LLVM_DEBUG(dbgs() << "\tintersect constraints\n");
441   LLVM_DEBUG(dbgs() << "\t    X ="; X->dump(dbgs()));
442   LLVM_DEBUG(dbgs() << "\t    Y ="; Y->dump(dbgs()));
443   assert(!Y->isPoint() && "Y must not be a Point");
444   if (X->isAny()) {
445     if (Y->isAny())
446       return false;
447     *X = *Y;
448     return true;
449   }
450   if (X->isEmpty())
451     return false;
452   if (Y->isEmpty()) {
453     X->setEmpty();
454     return true;
455   }
456 
457   if (X->isDistance() && Y->isDistance()) {
458     LLVM_DEBUG(dbgs() << "\t    intersect 2 distances\n");
459     if (isKnownPredicate(CmpInst::ICMP_EQ, X->getD(), Y->getD()))
460       return false;
461     if (isKnownPredicate(CmpInst::ICMP_NE, X->getD(), Y->getD())) {
462       X->setEmpty();
463       ++DeltaSuccesses;
464       return true;
465     }
466     // Hmmm, interesting situation.
467     // I guess if either is constant, keep it and ignore the other.
468     if (isa<SCEVConstant>(Y->getD())) {
469       *X = *Y;
470       return true;
471     }
472     return false;
473   }
474 
475   // At this point, the pseudo-code in Figure 4 of the paper
476   // checks if (X->isPoint() && Y->isPoint()).
477   // This case can't occur in our implementation,
478   // since a Point can only arise as the result of intersecting
479   // two Line constraints, and the right-hand value, Y, is never
480   // the result of an intersection.
481   assert(!(X->isPoint() && Y->isPoint()) &&
482          "We shouldn't ever see X->isPoint() && Y->isPoint()");
483 
484   if (X->isLine() && Y->isLine()) {
485     LLVM_DEBUG(dbgs() << "\t    intersect 2 lines\n");
486     const SCEV *Prod1 = SE->getMulExpr(X->getA(), Y->getB());
487     const SCEV *Prod2 = SE->getMulExpr(X->getB(), Y->getA());
488     if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2)) {
489       // slopes are equal, so lines are parallel
490       LLVM_DEBUG(dbgs() << "\t\tsame slope\n");
491       Prod1 = SE->getMulExpr(X->getC(), Y->getB());
492       Prod2 = SE->getMulExpr(X->getB(), Y->getC());
493       if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2))
494         return false;
495       if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) {
496         X->setEmpty();
497         ++DeltaSuccesses;
498         return true;
499       }
500       return false;
501     }
502     if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) {
503       // slopes differ, so lines intersect
504       LLVM_DEBUG(dbgs() << "\t\tdifferent slopes\n");
505       const SCEV *C1B2 = SE->getMulExpr(X->getC(), Y->getB());
506       const SCEV *C1A2 = SE->getMulExpr(X->getC(), Y->getA());
507       const SCEV *C2B1 = SE->getMulExpr(Y->getC(), X->getB());
508       const SCEV *C2A1 = SE->getMulExpr(Y->getC(), X->getA());
509       const SCEV *A1B2 = SE->getMulExpr(X->getA(), Y->getB());
510       const SCEV *A2B1 = SE->getMulExpr(Y->getA(), X->getB());
511       const SCEVConstant *C1A2_C2A1 =
512         dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1A2, C2A1));
513       const SCEVConstant *C1B2_C2B1 =
514         dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1B2, C2B1));
515       const SCEVConstant *A1B2_A2B1 =
516         dyn_cast<SCEVConstant>(SE->getMinusSCEV(A1B2, A2B1));
517       const SCEVConstant *A2B1_A1B2 =
518         dyn_cast<SCEVConstant>(SE->getMinusSCEV(A2B1, A1B2));
519       if (!C1B2_C2B1 || !C1A2_C2A1 ||
520           !A1B2_A2B1 || !A2B1_A1B2)
521         return false;
522       APInt Xtop = C1B2_C2B1->getAPInt();
523       APInt Xbot = A1B2_A2B1->getAPInt();
524       APInt Ytop = C1A2_C2A1->getAPInt();
525       APInt Ybot = A2B1_A1B2->getAPInt();
526       LLVM_DEBUG(dbgs() << "\t\tXtop = " << Xtop << "\n");
527       LLVM_DEBUG(dbgs() << "\t\tXbot = " << Xbot << "\n");
528       LLVM_DEBUG(dbgs() << "\t\tYtop = " << Ytop << "\n");
529       LLVM_DEBUG(dbgs() << "\t\tYbot = " << Ybot << "\n");
530       APInt Xq = Xtop; // these need to be initialized, even
531       APInt Xr = Xtop; // though they're just going to be overwritten
532       APInt::sdivrem(Xtop, Xbot, Xq, Xr);
533       APInt Yq = Ytop;
534       APInt Yr = Ytop;
535       APInt::sdivrem(Ytop, Ybot, Yq, Yr);
536       if (Xr != 0 || Yr != 0) {
537         X->setEmpty();
538         ++DeltaSuccesses;
539         return true;
540       }
541       LLVM_DEBUG(dbgs() << "\t\tX = " << Xq << ", Y = " << Yq << "\n");
542       if (Xq.slt(0) || Yq.slt(0)) {
543         X->setEmpty();
544         ++DeltaSuccesses;
545         return true;
546       }
547       if (const SCEVConstant *CUB =
548           collectConstantUpperBound(X->getAssociatedLoop(), Prod1->getType())) {
549         const APInt &UpperBound = CUB->getAPInt();
550         LLVM_DEBUG(dbgs() << "\t\tupper bound = " << UpperBound << "\n");
551         if (Xq.sgt(UpperBound) || Yq.sgt(UpperBound)) {
552           X->setEmpty();
553           ++DeltaSuccesses;
554           return true;
555         }
556       }
557       X->setPoint(SE->getConstant(Xq),
558                   SE->getConstant(Yq),
559                   X->getAssociatedLoop());
560       ++DeltaSuccesses;
561       return true;
562     }
563     return false;
564   }
565 
566   // if (X->isLine() && Y->isPoint()) This case can't occur.
567   assert(!(X->isLine() && Y->isPoint()) && "This case should never occur");
568 
569   if (X->isPoint() && Y->isLine()) {
570     LLVM_DEBUG(dbgs() << "\t    intersect Point and Line\n");
571     const SCEV *A1X1 = SE->getMulExpr(Y->getA(), X->getX());
572     const SCEV *B1Y1 = SE->getMulExpr(Y->getB(), X->getY());
573     const SCEV *Sum = SE->getAddExpr(A1X1, B1Y1);
574     if (isKnownPredicate(CmpInst::ICMP_EQ, Sum, Y->getC()))
575       return false;
576     if (isKnownPredicate(CmpInst::ICMP_NE, Sum, Y->getC())) {
577       X->setEmpty();
578       ++DeltaSuccesses;
579       return true;
580     }
581     return false;
582   }
583 
584   llvm_unreachable("shouldn't reach the end of Constraint intersection");
585   return false;
586 }
587 
588 
589 //===----------------------------------------------------------------------===//
590 // DependenceInfo methods
591 
592 // For debugging purposes. Dumps a dependence to OS.
593 void Dependence::dump(raw_ostream &OS) const {
594   bool Splitable = false;
595   if (isConfused())
596     OS << "confused";
597   else {
598     if (isConsistent())
599       OS << "consistent ";
600     if (isFlow())
601       OS << "flow";
602     else if (isOutput())
603       OS << "output";
604     else if (isAnti())
605       OS << "anti";
606     else if (isInput())
607       OS << "input";
608     unsigned Levels = getLevels();
609     OS << " [";
610     for (unsigned II = 1; II <= Levels; ++II) {
611       if (isSplitable(II))
612         Splitable = true;
613       if (isPeelFirst(II))
614         OS << 'p';
615       const SCEV *Distance = getDistance(II);
616       if (Distance)
617         OS << *Distance;
618       else if (isScalar(II))
619         OS << "S";
620       else {
621         unsigned Direction = getDirection(II);
622         if (Direction == DVEntry::ALL)
623           OS << "*";
624         else {
625           if (Direction & DVEntry::LT)
626             OS << "<";
627           if (Direction & DVEntry::EQ)
628             OS << "=";
629           if (Direction & DVEntry::GT)
630             OS << ">";
631         }
632       }
633       if (isPeelLast(II))
634         OS << 'p';
635       if (II < Levels)
636         OS << " ";
637     }
638     if (isLoopIndependent())
639       OS << "|<";
640     OS << "]";
641     if (Splitable)
642       OS << " splitable";
643   }
644   OS << "!\n";
645 }
646 
647 // Returns NoAlias/MayAliass/MustAlias for two memory locations based upon their
648 // underlaying objects. If LocA and LocB are known to not alias (for any reason:
649 // tbaa, non-overlapping regions etc), then it is known there is no dependecy.
650 // Otherwise the underlying objects are checked to see if they point to
651 // different identifiable objects.
652 static AliasResult underlyingObjectsAlias(AAResults *AA,
653                                           const DataLayout &DL,
654                                           const MemoryLocation &LocA,
655                                           const MemoryLocation &LocB) {
656   // Check the original locations (minus size) for noalias, which can happen for
657   // tbaa, incompatible underlying object locations, etc.
658   MemoryLocation LocAS =
659       MemoryLocation::getBeforeOrAfter(LocA.Ptr, LocA.AATags);
660   MemoryLocation LocBS =
661       MemoryLocation::getBeforeOrAfter(LocB.Ptr, LocB.AATags);
662   if (AA->isNoAlias(LocAS, LocBS))
663     return AliasResult::NoAlias;
664 
665   // Check the underlying objects are the same
666   const Value *AObj = getUnderlyingObject(LocA.Ptr);
667   const Value *BObj = getUnderlyingObject(LocB.Ptr);
668 
669   // If the underlying objects are the same, they must alias
670   if (AObj == BObj)
671     return AliasResult::MustAlias;
672 
673   // We may have hit the recursion limit for underlying objects, or have
674   // underlying objects where we don't know they will alias.
675   if (!isIdentifiedObject(AObj) || !isIdentifiedObject(BObj))
676     return AliasResult::MayAlias;
677 
678   // Otherwise we know the objects are different and both identified objects so
679   // must not alias.
680   return AliasResult::NoAlias;
681 }
682 
683 
684 // Returns true if the load or store can be analyzed. Atomic and volatile
685 // operations have properties which this analysis does not understand.
686 static
687 bool isLoadOrStore(const Instruction *I) {
688   if (const LoadInst *LI = dyn_cast<LoadInst>(I))
689     return LI->isUnordered();
690   else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
691     return SI->isUnordered();
692   return false;
693 }
694 
695 
696 // Examines the loop nesting of the Src and Dst
697 // instructions and establishes their shared loops. Sets the variables
698 // CommonLevels, SrcLevels, and MaxLevels.
699 // The source and destination instructions needn't be contained in the same
700 // loop. The routine establishNestingLevels finds the level of most deeply
701 // nested loop that contains them both, CommonLevels. An instruction that's
702 // not contained in a loop is at level = 0. MaxLevels is equal to the level
703 // of the source plus the level of the destination, minus CommonLevels.
704 // This lets us allocate vectors MaxLevels in length, with room for every
705 // distinct loop referenced in both the source and destination subscripts.
706 // The variable SrcLevels is the nesting depth of the source instruction.
707 // It's used to help calculate distinct loops referenced by the destination.
708 // Here's the map from loops to levels:
709 //            0 - unused
710 //            1 - outermost common loop
711 //          ... - other common loops
712 // CommonLevels - innermost common loop
713 //          ... - loops containing Src but not Dst
714 //    SrcLevels - innermost loop containing Src but not Dst
715 //          ... - loops containing Dst but not Src
716 //    MaxLevels - innermost loops containing Dst but not Src
717 // Consider the follow code fragment:
718 //   for (a = ...) {
719 //     for (b = ...) {
720 //       for (c = ...) {
721 //         for (d = ...) {
722 //           A[] = ...;
723 //         }
724 //       }
725 //       for (e = ...) {
726 //         for (f = ...) {
727 //           for (g = ...) {
728 //             ... = A[];
729 //           }
730 //         }
731 //       }
732 //     }
733 //   }
734 // If we're looking at the possibility of a dependence between the store
735 // to A (the Src) and the load from A (the Dst), we'll note that they
736 // have 2 loops in common, so CommonLevels will equal 2 and the direction
737 // vector for Result will have 2 entries. SrcLevels = 4 and MaxLevels = 7.
738 // A map from loop names to loop numbers would look like
739 //     a - 1
740 //     b - 2 = CommonLevels
741 //     c - 3
742 //     d - 4 = SrcLevels
743 //     e - 5
744 //     f - 6
745 //     g - 7 = MaxLevels
746 void DependenceInfo::establishNestingLevels(const Instruction *Src,
747                                             const Instruction *Dst) {
748   const BasicBlock *SrcBlock = Src->getParent();
749   const BasicBlock *DstBlock = Dst->getParent();
750   unsigned SrcLevel = LI->getLoopDepth(SrcBlock);
751   unsigned DstLevel = LI->getLoopDepth(DstBlock);
752   const Loop *SrcLoop = LI->getLoopFor(SrcBlock);
753   const Loop *DstLoop = LI->getLoopFor(DstBlock);
754   SrcLevels = SrcLevel;
755   MaxLevels = SrcLevel + DstLevel;
756   while (SrcLevel > DstLevel) {
757     SrcLoop = SrcLoop->getParentLoop();
758     SrcLevel--;
759   }
760   while (DstLevel > SrcLevel) {
761     DstLoop = DstLoop->getParentLoop();
762     DstLevel--;
763   }
764   while (SrcLoop != DstLoop) {
765     SrcLoop = SrcLoop->getParentLoop();
766     DstLoop = DstLoop->getParentLoop();
767     SrcLevel--;
768   }
769   CommonLevels = SrcLevel;
770   MaxLevels -= CommonLevels;
771 }
772 
773 
774 // Given one of the loops containing the source, return
775 // its level index in our numbering scheme.
776 unsigned DependenceInfo::mapSrcLoop(const Loop *SrcLoop) const {
777   return SrcLoop->getLoopDepth();
778 }
779 
780 
781 // Given one of the loops containing the destination,
782 // return its level index in our numbering scheme.
783 unsigned DependenceInfo::mapDstLoop(const Loop *DstLoop) const {
784   unsigned D = DstLoop->getLoopDepth();
785   if (D > CommonLevels)
786     // This tries to make sure that we assign unique numbers to src and dst when
787     // the memory accesses reside in different loops that have the same depth.
788     return D - CommonLevels + SrcLevels;
789   else
790     return D;
791 }
792 
793 
794 // Returns true if Expression is loop invariant in LoopNest.
795 bool DependenceInfo::isLoopInvariant(const SCEV *Expression,
796                                      const Loop *LoopNest) const {
797   // Unlike ScalarEvolution::isLoopInvariant() we consider an access outside of
798   // any loop as invariant, because we only consier expression evaluation at a
799   // specific position (where the array access takes place), and not across the
800   // entire function.
801   if (!LoopNest)
802     return true;
803 
804   // If the expression is invariant in the outermost loop of the loop nest, it
805   // is invariant anywhere in the loop nest.
806   return SE->isLoopInvariant(Expression, LoopNest->getOutermostLoop());
807 }
808 
809 
810 
811 // Finds the set of loops from the LoopNest that
812 // have a level <= CommonLevels and are referred to by the SCEV Expression.
813 void DependenceInfo::collectCommonLoops(const SCEV *Expression,
814                                         const Loop *LoopNest,
815                                         SmallBitVector &Loops) const {
816   while (LoopNest) {
817     unsigned Level = LoopNest->getLoopDepth();
818     if (Level <= CommonLevels && !SE->isLoopInvariant(Expression, LoopNest))
819       Loops.set(Level);
820     LoopNest = LoopNest->getParentLoop();
821   }
822 }
823 
824 void DependenceInfo::unifySubscriptType(ArrayRef<Subscript *> Pairs) {
825 
826   unsigned widestWidthSeen = 0;
827   Type *widestType;
828 
829   // Go through each pair and find the widest bit to which we need
830   // to extend all of them.
831   for (Subscript *Pair : Pairs) {
832     const SCEV *Src = Pair->Src;
833     const SCEV *Dst = Pair->Dst;
834     IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
835     IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
836     if (SrcTy == nullptr || DstTy == nullptr) {
837       assert(SrcTy == DstTy && "This function only unify integer types and "
838              "expect Src and Dst share the same type "
839              "otherwise.");
840       continue;
841     }
842     if (SrcTy->getBitWidth() > widestWidthSeen) {
843       widestWidthSeen = SrcTy->getBitWidth();
844       widestType = SrcTy;
845     }
846     if (DstTy->getBitWidth() > widestWidthSeen) {
847       widestWidthSeen = DstTy->getBitWidth();
848       widestType = DstTy;
849     }
850   }
851 
852 
853   assert(widestWidthSeen > 0);
854 
855   // Now extend each pair to the widest seen.
856   for (Subscript *Pair : Pairs) {
857     const SCEV *Src = Pair->Src;
858     const SCEV *Dst = Pair->Dst;
859     IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
860     IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
861     if (SrcTy == nullptr || DstTy == nullptr) {
862       assert(SrcTy == DstTy && "This function only unify integer types and "
863              "expect Src and Dst share the same type "
864              "otherwise.");
865       continue;
866     }
867     if (SrcTy->getBitWidth() < widestWidthSeen)
868       // Sign-extend Src to widestType
869       Pair->Src = SE->getSignExtendExpr(Src, widestType);
870     if (DstTy->getBitWidth() < widestWidthSeen) {
871       // Sign-extend Dst to widestType
872       Pair->Dst = SE->getSignExtendExpr(Dst, widestType);
873     }
874   }
875 }
876 
877 // removeMatchingExtensions - Examines a subscript pair.
878 // If the source and destination are identically sign (or zero)
879 // extended, it strips off the extension in an effect to simplify
880 // the actual analysis.
881 void DependenceInfo::removeMatchingExtensions(Subscript *Pair) {
882   const SCEV *Src = Pair->Src;
883   const SCEV *Dst = Pair->Dst;
884   if ((isa<SCEVZeroExtendExpr>(Src) && isa<SCEVZeroExtendExpr>(Dst)) ||
885       (isa<SCEVSignExtendExpr>(Src) && isa<SCEVSignExtendExpr>(Dst))) {
886     const SCEVIntegralCastExpr *SrcCast = cast<SCEVIntegralCastExpr>(Src);
887     const SCEVIntegralCastExpr *DstCast = cast<SCEVIntegralCastExpr>(Dst);
888     const SCEV *SrcCastOp = SrcCast->getOperand();
889     const SCEV *DstCastOp = DstCast->getOperand();
890     if (SrcCastOp->getType() == DstCastOp->getType()) {
891       Pair->Src = SrcCastOp;
892       Pair->Dst = DstCastOp;
893     }
894   }
895 }
896 
897 // Examine the scev and return true iff it's affine.
898 // Collect any loops mentioned in the set of "Loops".
899 bool DependenceInfo::checkSubscript(const SCEV *Expr, const Loop *LoopNest,
900                                     SmallBitVector &Loops, bool IsSrc) {
901   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
902   if (!AddRec)
903     return isLoopInvariant(Expr, LoopNest);
904 
905   // The AddRec must depend on one of the containing loops. Otherwise,
906   // mapSrcLoop and mapDstLoop return indices outside the intended range. This
907   // can happen when a subscript in one loop references an IV from a sibling
908   // loop that could not be replaced with a concrete exit value by
909   // getSCEVAtScope.
910   const Loop *L = LoopNest;
911   while (L && AddRec->getLoop() != L)
912     L = L->getParentLoop();
913   if (!L)
914     return false;
915 
916   const SCEV *Start = AddRec->getStart();
917   const SCEV *Step = AddRec->getStepRecurrence(*SE);
918   const SCEV *UB = SE->getBackedgeTakenCount(AddRec->getLoop());
919   if (!isa<SCEVCouldNotCompute>(UB)) {
920     if (SE->getTypeSizeInBits(Start->getType()) <
921         SE->getTypeSizeInBits(UB->getType())) {
922       if (!AddRec->getNoWrapFlags())
923         return false;
924     }
925   }
926   if (!isLoopInvariant(Step, LoopNest))
927     return false;
928   if (IsSrc)
929     Loops.set(mapSrcLoop(AddRec->getLoop()));
930   else
931     Loops.set(mapDstLoop(AddRec->getLoop()));
932   return checkSubscript(Start, LoopNest, Loops, IsSrc);
933 }
934 
935 // Examine the scev and return true iff it's linear.
936 // Collect any loops mentioned in the set of "Loops".
937 bool DependenceInfo::checkSrcSubscript(const SCEV *Src, const Loop *LoopNest,
938                                        SmallBitVector &Loops) {
939   return checkSubscript(Src, LoopNest, Loops, true);
940 }
941 
942 // Examine the scev and return true iff it's linear.
943 // Collect any loops mentioned in the set of "Loops".
944 bool DependenceInfo::checkDstSubscript(const SCEV *Dst, const Loop *LoopNest,
945                                        SmallBitVector &Loops) {
946   return checkSubscript(Dst, LoopNest, Loops, false);
947 }
948 
949 
950 // Examines the subscript pair (the Src and Dst SCEVs)
951 // and classifies it as either ZIV, SIV, RDIV, MIV, or Nonlinear.
952 // Collects the associated loops in a set.
953 DependenceInfo::Subscript::ClassificationKind
954 DependenceInfo::classifyPair(const SCEV *Src, const Loop *SrcLoopNest,
955                              const SCEV *Dst, const Loop *DstLoopNest,
956                              SmallBitVector &Loops) {
957   SmallBitVector SrcLoops(MaxLevels + 1);
958   SmallBitVector DstLoops(MaxLevels + 1);
959   if (!checkSrcSubscript(Src, SrcLoopNest, SrcLoops))
960     return Subscript::NonLinear;
961   if (!checkDstSubscript(Dst, DstLoopNest, DstLoops))
962     return Subscript::NonLinear;
963   Loops = SrcLoops;
964   Loops |= DstLoops;
965   unsigned N = Loops.count();
966   if (N == 0)
967     return Subscript::ZIV;
968   if (N == 1)
969     return Subscript::SIV;
970   if (N == 2 && (SrcLoops.count() == 0 ||
971                  DstLoops.count() == 0 ||
972                  (SrcLoops.count() == 1 && DstLoops.count() == 1)))
973     return Subscript::RDIV;
974   return Subscript::MIV;
975 }
976 
977 
978 // A wrapper around SCEV::isKnownPredicate.
979 // Looks for cases where we're interested in comparing for equality.
980 // If both X and Y have been identically sign or zero extended,
981 // it strips off the (confusing) extensions before invoking
982 // SCEV::isKnownPredicate. Perhaps, someday, the ScalarEvolution package
983 // will be similarly updated.
984 //
985 // If SCEV::isKnownPredicate can't prove the predicate,
986 // we try simple subtraction, which seems to help in some cases
987 // involving symbolics.
988 bool DependenceInfo::isKnownPredicate(ICmpInst::Predicate Pred, const SCEV *X,
989                                       const SCEV *Y) const {
990   if (Pred == CmpInst::ICMP_EQ ||
991       Pred == CmpInst::ICMP_NE) {
992     if ((isa<SCEVSignExtendExpr>(X) &&
993          isa<SCEVSignExtendExpr>(Y)) ||
994         (isa<SCEVZeroExtendExpr>(X) &&
995          isa<SCEVZeroExtendExpr>(Y))) {
996       const SCEVIntegralCastExpr *CX = cast<SCEVIntegralCastExpr>(X);
997       const SCEVIntegralCastExpr *CY = cast<SCEVIntegralCastExpr>(Y);
998       const SCEV *Xop = CX->getOperand();
999       const SCEV *Yop = CY->getOperand();
1000       if (Xop->getType() == Yop->getType()) {
1001         X = Xop;
1002         Y = Yop;
1003       }
1004     }
1005   }
1006   if (SE->isKnownPredicate(Pred, X, Y))
1007     return true;
1008   // If SE->isKnownPredicate can't prove the condition,
1009   // we try the brute-force approach of subtracting
1010   // and testing the difference.
1011   // By testing with SE->isKnownPredicate first, we avoid
1012   // the possibility of overflow when the arguments are constants.
1013   const SCEV *Delta = SE->getMinusSCEV(X, Y);
1014   switch (Pred) {
1015   case CmpInst::ICMP_EQ:
1016     return Delta->isZero();
1017   case CmpInst::ICMP_NE:
1018     return SE->isKnownNonZero(Delta);
1019   case CmpInst::ICMP_SGE:
1020     return SE->isKnownNonNegative(Delta);
1021   case CmpInst::ICMP_SLE:
1022     return SE->isKnownNonPositive(Delta);
1023   case CmpInst::ICMP_SGT:
1024     return SE->isKnownPositive(Delta);
1025   case CmpInst::ICMP_SLT:
1026     return SE->isKnownNegative(Delta);
1027   default:
1028     llvm_unreachable("unexpected predicate in isKnownPredicate");
1029   }
1030 }
1031 
1032 /// Compare to see if S is less than Size, using isKnownNegative(S - max(Size, 1))
1033 /// with some extra checking if S is an AddRec and we can prove less-than using
1034 /// the loop bounds.
1035 bool DependenceInfo::isKnownLessThan(const SCEV *S, const SCEV *Size) const {
1036   // First unify to the same type
1037   auto *SType = dyn_cast<IntegerType>(S->getType());
1038   auto *SizeType = dyn_cast<IntegerType>(Size->getType());
1039   if (!SType || !SizeType)
1040     return false;
1041   Type *MaxType =
1042       (SType->getBitWidth() >= SizeType->getBitWidth()) ? SType : SizeType;
1043   S = SE->getTruncateOrZeroExtend(S, MaxType);
1044   Size = SE->getTruncateOrZeroExtend(Size, MaxType);
1045 
1046   // Special check for addrecs using BE taken count
1047   const SCEV *Bound = SE->getMinusSCEV(S, Size);
1048   if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Bound)) {
1049     if (AddRec->isAffine()) {
1050       const SCEV *BECount = SE->getBackedgeTakenCount(AddRec->getLoop());
1051       if (!isa<SCEVCouldNotCompute>(BECount)) {
1052         const SCEV *Limit = AddRec->evaluateAtIteration(BECount, *SE);
1053         if (SE->isKnownNegative(Limit))
1054           return true;
1055       }
1056     }
1057   }
1058 
1059   // Check using normal isKnownNegative
1060   const SCEV *LimitedBound =
1061       SE->getMinusSCEV(S, SE->getSMaxExpr(Size, SE->getOne(Size->getType())));
1062   return SE->isKnownNegative(LimitedBound);
1063 }
1064 
1065 bool DependenceInfo::isKnownNonNegative(const SCEV *S, const Value *Ptr) const {
1066   bool Inbounds = false;
1067   if (auto *SrcGEP = dyn_cast<GetElementPtrInst>(Ptr))
1068     Inbounds = SrcGEP->isInBounds();
1069   if (Inbounds) {
1070     if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
1071       if (AddRec->isAffine()) {
1072         // We know S is for Ptr, the operand on a load/store, so doesn't wrap.
1073         // If both parts are NonNegative, the end result will be NonNegative
1074         if (SE->isKnownNonNegative(AddRec->getStart()) &&
1075             SE->isKnownNonNegative(AddRec->getOperand(1)))
1076           return true;
1077       }
1078     }
1079   }
1080 
1081   return SE->isKnownNonNegative(S);
1082 }
1083 
1084 // All subscripts are all the same type.
1085 // Loop bound may be smaller (e.g., a char).
1086 // Should zero extend loop bound, since it's always >= 0.
1087 // This routine collects upper bound and extends or truncates if needed.
1088 // Truncating is safe when subscripts are known not to wrap. Cases without
1089 // nowrap flags should have been rejected earlier.
1090 // Return null if no bound available.
1091 const SCEV *DependenceInfo::collectUpperBound(const Loop *L, Type *T) const {
1092   if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
1093     const SCEV *UB = SE->getBackedgeTakenCount(L);
1094     return SE->getTruncateOrZeroExtend(UB, T);
1095   }
1096   return nullptr;
1097 }
1098 
1099 
1100 // Calls collectUpperBound(), then attempts to cast it to SCEVConstant.
1101 // If the cast fails, returns NULL.
1102 const SCEVConstant *DependenceInfo::collectConstantUpperBound(const Loop *L,
1103                                                               Type *T) const {
1104   if (const SCEV *UB = collectUpperBound(L, T))
1105     return dyn_cast<SCEVConstant>(UB);
1106   return nullptr;
1107 }
1108 
1109 
1110 // testZIV -
1111 // When we have a pair of subscripts of the form [c1] and [c2],
1112 // where c1 and c2 are both loop invariant, we attack it using
1113 // the ZIV test. Basically, we test by comparing the two values,
1114 // but there are actually three possible results:
1115 // 1) the values are equal, so there's a dependence
1116 // 2) the values are different, so there's no dependence
1117 // 3) the values might be equal, so we have to assume a dependence.
1118 //
1119 // Return true if dependence disproved.
1120 bool DependenceInfo::testZIV(const SCEV *Src, const SCEV *Dst,
1121                              FullDependence &Result) const {
1122   LLVM_DEBUG(dbgs() << "    src = " << *Src << "\n");
1123   LLVM_DEBUG(dbgs() << "    dst = " << *Dst << "\n");
1124   ++ZIVapplications;
1125   if (isKnownPredicate(CmpInst::ICMP_EQ, Src, Dst)) {
1126     LLVM_DEBUG(dbgs() << "    provably dependent\n");
1127     return false; // provably dependent
1128   }
1129   if (isKnownPredicate(CmpInst::ICMP_NE, Src, Dst)) {
1130     LLVM_DEBUG(dbgs() << "    provably independent\n");
1131     ++ZIVindependence;
1132     return true; // provably independent
1133   }
1134   LLVM_DEBUG(dbgs() << "    possibly dependent\n");
1135   Result.Consistent = false;
1136   return false; // possibly dependent
1137 }
1138 
1139 
1140 // strongSIVtest -
1141 // From the paper, Practical Dependence Testing, Section 4.2.1
1142 //
1143 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 + a*i],
1144 // where i is an induction variable, c1 and c2 are loop invariant,
1145 //  and a is a constant, we can solve it exactly using the Strong SIV test.
1146 //
1147 // Can prove independence. Failing that, can compute distance (and direction).
1148 // In the presence of symbolic terms, we can sometimes make progress.
1149 //
1150 // If there's a dependence,
1151 //
1152 //    c1 + a*i = c2 + a*i'
1153 //
1154 // The dependence distance is
1155 //
1156 //    d = i' - i = (c1 - c2)/a
1157 //
1158 // A dependence only exists if d is an integer and abs(d) <= U, where U is the
1159 // loop's upper bound. If a dependence exists, the dependence direction is
1160 // defined as
1161 //
1162 //                { < if d > 0
1163 //    direction = { = if d = 0
1164 //                { > if d < 0
1165 //
1166 // Return true if dependence disproved.
1167 bool DependenceInfo::strongSIVtest(const SCEV *Coeff, const SCEV *SrcConst,
1168                                    const SCEV *DstConst, const Loop *CurLoop,
1169                                    unsigned Level, FullDependence &Result,
1170                                    Constraint &NewConstraint) const {
1171   LLVM_DEBUG(dbgs() << "\tStrong SIV test\n");
1172   LLVM_DEBUG(dbgs() << "\t    Coeff = " << *Coeff);
1173   LLVM_DEBUG(dbgs() << ", " << *Coeff->getType() << "\n");
1174   LLVM_DEBUG(dbgs() << "\t    SrcConst = " << *SrcConst);
1175   LLVM_DEBUG(dbgs() << ", " << *SrcConst->getType() << "\n");
1176   LLVM_DEBUG(dbgs() << "\t    DstConst = " << *DstConst);
1177   LLVM_DEBUG(dbgs() << ", " << *DstConst->getType() << "\n");
1178   ++StrongSIVapplications;
1179   assert(0 < Level && Level <= CommonLevels && "level out of range");
1180   Level--;
1181 
1182   const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst);
1183   LLVM_DEBUG(dbgs() << "\t    Delta = " << *Delta);
1184   LLVM_DEBUG(dbgs() << ", " << *Delta->getType() << "\n");
1185 
1186   // check that |Delta| < iteration count
1187   if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1188     LLVM_DEBUG(dbgs() << "\t    UpperBound = " << *UpperBound);
1189     LLVM_DEBUG(dbgs() << ", " << *UpperBound->getType() << "\n");
1190     const SCEV *AbsDelta =
1191       SE->isKnownNonNegative(Delta) ? Delta : SE->getNegativeSCEV(Delta);
1192     const SCEV *AbsCoeff =
1193       SE->isKnownNonNegative(Coeff) ? Coeff : SE->getNegativeSCEV(Coeff);
1194     const SCEV *Product = SE->getMulExpr(UpperBound, AbsCoeff);
1195     if (isKnownPredicate(CmpInst::ICMP_SGT, AbsDelta, Product)) {
1196       // Distance greater than trip count - no dependence
1197       ++StrongSIVindependence;
1198       ++StrongSIVsuccesses;
1199       return true;
1200     }
1201   }
1202 
1203   // Can we compute distance?
1204   if (isa<SCEVConstant>(Delta) && isa<SCEVConstant>(Coeff)) {
1205     APInt ConstDelta = cast<SCEVConstant>(Delta)->getAPInt();
1206     APInt ConstCoeff = cast<SCEVConstant>(Coeff)->getAPInt();
1207     APInt Distance  = ConstDelta; // these need to be initialized
1208     APInt Remainder = ConstDelta;
1209     APInt::sdivrem(ConstDelta, ConstCoeff, Distance, Remainder);
1210     LLVM_DEBUG(dbgs() << "\t    Distance = " << Distance << "\n");
1211     LLVM_DEBUG(dbgs() << "\t    Remainder = " << Remainder << "\n");
1212     // Make sure Coeff divides Delta exactly
1213     if (Remainder != 0) {
1214       // Coeff doesn't divide Distance, no dependence
1215       ++StrongSIVindependence;
1216       ++StrongSIVsuccesses;
1217       return true;
1218     }
1219     Result.DV[Level].Distance = SE->getConstant(Distance);
1220     NewConstraint.setDistance(SE->getConstant(Distance), CurLoop);
1221     if (Distance.sgt(0))
1222       Result.DV[Level].Direction &= Dependence::DVEntry::LT;
1223     else if (Distance.slt(0))
1224       Result.DV[Level].Direction &= Dependence::DVEntry::GT;
1225     else
1226       Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
1227     ++StrongSIVsuccesses;
1228   }
1229   else if (Delta->isZero()) {
1230     // since 0/X == 0
1231     Result.DV[Level].Distance = Delta;
1232     NewConstraint.setDistance(Delta, CurLoop);
1233     Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
1234     ++StrongSIVsuccesses;
1235   }
1236   else {
1237     if (Coeff->isOne()) {
1238       LLVM_DEBUG(dbgs() << "\t    Distance = " << *Delta << "\n");
1239       Result.DV[Level].Distance = Delta; // since X/1 == X
1240       NewConstraint.setDistance(Delta, CurLoop);
1241     }
1242     else {
1243       Result.Consistent = false;
1244       NewConstraint.setLine(Coeff,
1245                             SE->getNegativeSCEV(Coeff),
1246                             SE->getNegativeSCEV(Delta), CurLoop);
1247     }
1248 
1249     // maybe we can get a useful direction
1250     bool DeltaMaybeZero     = !SE->isKnownNonZero(Delta);
1251     bool DeltaMaybePositive = !SE->isKnownNonPositive(Delta);
1252     bool DeltaMaybeNegative = !SE->isKnownNonNegative(Delta);
1253     bool CoeffMaybePositive = !SE->isKnownNonPositive(Coeff);
1254     bool CoeffMaybeNegative = !SE->isKnownNonNegative(Coeff);
1255     // The double negatives above are confusing.
1256     // It helps to read !SE->isKnownNonZero(Delta)
1257     // as "Delta might be Zero"
1258     unsigned NewDirection = Dependence::DVEntry::NONE;
1259     if ((DeltaMaybePositive && CoeffMaybePositive) ||
1260         (DeltaMaybeNegative && CoeffMaybeNegative))
1261       NewDirection = Dependence::DVEntry::LT;
1262     if (DeltaMaybeZero)
1263       NewDirection |= Dependence::DVEntry::EQ;
1264     if ((DeltaMaybeNegative && CoeffMaybePositive) ||
1265         (DeltaMaybePositive && CoeffMaybeNegative))
1266       NewDirection |= Dependence::DVEntry::GT;
1267     if (NewDirection < Result.DV[Level].Direction)
1268       ++StrongSIVsuccesses;
1269     Result.DV[Level].Direction &= NewDirection;
1270   }
1271   return false;
1272 }
1273 
1274 
1275 // weakCrossingSIVtest -
1276 // From the paper, Practical Dependence Testing, Section 4.2.2
1277 //
1278 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 - a*i],
1279 // where i is an induction variable, c1 and c2 are loop invariant,
1280 // and a is a constant, we can solve it exactly using the
1281 // Weak-Crossing SIV test.
1282 //
1283 // Given c1 + a*i = c2 - a*i', we can look for the intersection of
1284 // the two lines, where i = i', yielding
1285 //
1286 //    c1 + a*i = c2 - a*i
1287 //    2a*i = c2 - c1
1288 //    i = (c2 - c1)/2a
1289 //
1290 // If i < 0, there is no dependence.
1291 // If i > upperbound, there is no dependence.
1292 // If i = 0 (i.e., if c1 = c2), there's a dependence with distance = 0.
1293 // If i = upperbound, there's a dependence with distance = 0.
1294 // If i is integral, there's a dependence (all directions).
1295 // If the non-integer part = 1/2, there's a dependence (<> directions).
1296 // Otherwise, there's no dependence.
1297 //
1298 // Can prove independence. Failing that,
1299 // can sometimes refine the directions.
1300 // Can determine iteration for splitting.
1301 //
1302 // Return true if dependence disproved.
1303 bool DependenceInfo::weakCrossingSIVtest(
1304     const SCEV *Coeff, const SCEV *SrcConst, const SCEV *DstConst,
1305     const Loop *CurLoop, unsigned Level, FullDependence &Result,
1306     Constraint &NewConstraint, const SCEV *&SplitIter) const {
1307   LLVM_DEBUG(dbgs() << "\tWeak-Crossing SIV test\n");
1308   LLVM_DEBUG(dbgs() << "\t    Coeff = " << *Coeff << "\n");
1309   LLVM_DEBUG(dbgs() << "\t    SrcConst = " << *SrcConst << "\n");
1310   LLVM_DEBUG(dbgs() << "\t    DstConst = " << *DstConst << "\n");
1311   ++WeakCrossingSIVapplications;
1312   assert(0 < Level && Level <= CommonLevels && "Level out of range");
1313   Level--;
1314   Result.Consistent = false;
1315   const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1316   LLVM_DEBUG(dbgs() << "\t    Delta = " << *Delta << "\n");
1317   NewConstraint.setLine(Coeff, Coeff, Delta, CurLoop);
1318   if (Delta->isZero()) {
1319     Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT);
1320     Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT);
1321     ++WeakCrossingSIVsuccesses;
1322     if (!Result.DV[Level].Direction) {
1323       ++WeakCrossingSIVindependence;
1324       return true;
1325     }
1326     Result.DV[Level].Distance = Delta; // = 0
1327     return false;
1328   }
1329   const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(Coeff);
1330   if (!ConstCoeff)
1331     return false;
1332 
1333   Result.DV[Level].Splitable = true;
1334   if (SE->isKnownNegative(ConstCoeff)) {
1335     ConstCoeff = dyn_cast<SCEVConstant>(SE->getNegativeSCEV(ConstCoeff));
1336     assert(ConstCoeff &&
1337            "dynamic cast of negative of ConstCoeff should yield constant");
1338     Delta = SE->getNegativeSCEV(Delta);
1339   }
1340   assert(SE->isKnownPositive(ConstCoeff) && "ConstCoeff should be positive");
1341 
1342   // compute SplitIter for use by DependenceInfo::getSplitIteration()
1343   SplitIter = SE->getUDivExpr(
1344       SE->getSMaxExpr(SE->getZero(Delta->getType()), Delta),
1345       SE->getMulExpr(SE->getConstant(Delta->getType(), 2), ConstCoeff));
1346   LLVM_DEBUG(dbgs() << "\t    Split iter = " << *SplitIter << "\n");
1347 
1348   const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1349   if (!ConstDelta)
1350     return false;
1351 
1352   // We're certain that ConstCoeff > 0; therefore,
1353   // if Delta < 0, then no dependence.
1354   LLVM_DEBUG(dbgs() << "\t    Delta = " << *Delta << "\n");
1355   LLVM_DEBUG(dbgs() << "\t    ConstCoeff = " << *ConstCoeff << "\n");
1356   if (SE->isKnownNegative(Delta)) {
1357     // No dependence, Delta < 0
1358     ++WeakCrossingSIVindependence;
1359     ++WeakCrossingSIVsuccesses;
1360     return true;
1361   }
1362 
1363   // We're certain that Delta > 0 and ConstCoeff > 0.
1364   // Check Delta/(2*ConstCoeff) against upper loop bound
1365   if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1366     LLVM_DEBUG(dbgs() << "\t    UpperBound = " << *UpperBound << "\n");
1367     const SCEV *ConstantTwo = SE->getConstant(UpperBound->getType(), 2);
1368     const SCEV *ML = SE->getMulExpr(SE->getMulExpr(ConstCoeff, UpperBound),
1369                                     ConstantTwo);
1370     LLVM_DEBUG(dbgs() << "\t    ML = " << *ML << "\n");
1371     if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, ML)) {
1372       // Delta too big, no dependence
1373       ++WeakCrossingSIVindependence;
1374       ++WeakCrossingSIVsuccesses;
1375       return true;
1376     }
1377     if (isKnownPredicate(CmpInst::ICMP_EQ, Delta, ML)) {
1378       // i = i' = UB
1379       Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT);
1380       Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT);
1381       ++WeakCrossingSIVsuccesses;
1382       if (!Result.DV[Level].Direction) {
1383         ++WeakCrossingSIVindependence;
1384         return true;
1385       }
1386       Result.DV[Level].Splitable = false;
1387       Result.DV[Level].Distance = SE->getZero(Delta->getType());
1388       return false;
1389     }
1390   }
1391 
1392   // check that Coeff divides Delta
1393   APInt APDelta = ConstDelta->getAPInt();
1394   APInt APCoeff = ConstCoeff->getAPInt();
1395   APInt Distance = APDelta; // these need to be initialzed
1396   APInt Remainder = APDelta;
1397   APInt::sdivrem(APDelta, APCoeff, Distance, Remainder);
1398   LLVM_DEBUG(dbgs() << "\t    Remainder = " << Remainder << "\n");
1399   if (Remainder != 0) {
1400     // Coeff doesn't divide Delta, no dependence
1401     ++WeakCrossingSIVindependence;
1402     ++WeakCrossingSIVsuccesses;
1403     return true;
1404   }
1405   LLVM_DEBUG(dbgs() << "\t    Distance = " << Distance << "\n");
1406 
1407   // if 2*Coeff doesn't divide Delta, then the equal direction isn't possible
1408   APInt Two = APInt(Distance.getBitWidth(), 2, true);
1409   Remainder = Distance.srem(Two);
1410   LLVM_DEBUG(dbgs() << "\t    Remainder = " << Remainder << "\n");
1411   if (Remainder != 0) {
1412     // Equal direction isn't possible
1413     Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::EQ);
1414     ++WeakCrossingSIVsuccesses;
1415   }
1416   return false;
1417 }
1418 
1419 
1420 // Kirch's algorithm, from
1421 //
1422 //        Optimizing Supercompilers for Supercomputers
1423 //        Michael Wolfe
1424 //        MIT Press, 1989
1425 //
1426 // Program 2.1, page 29.
1427 // Computes the GCD of AM and BM.
1428 // Also finds a solution to the equation ax - by = gcd(a, b).
1429 // Returns true if dependence disproved; i.e., gcd does not divide Delta.
1430 static bool findGCD(unsigned Bits, const APInt &AM, const APInt &BM,
1431                     const APInt &Delta, APInt &G, APInt &X, APInt &Y) {
1432   APInt A0(Bits, 1, true), A1(Bits, 0, true);
1433   APInt B0(Bits, 0, true), B1(Bits, 1, true);
1434   APInt G0 = AM.abs();
1435   APInt G1 = BM.abs();
1436   APInt Q = G0; // these need to be initialized
1437   APInt R = G0;
1438   APInt::sdivrem(G0, G1, Q, R);
1439   while (R != 0) {
1440     APInt A2 = A0 - Q*A1; A0 = A1; A1 = A2;
1441     APInt B2 = B0 - Q*B1; B0 = B1; B1 = B2;
1442     G0 = G1; G1 = R;
1443     APInt::sdivrem(G0, G1, Q, R);
1444   }
1445   G = G1;
1446   LLVM_DEBUG(dbgs() << "\t    GCD = " << G << "\n");
1447   X = AM.slt(0) ? -A1 : A1;
1448   Y = BM.slt(0) ? B1 : -B1;
1449 
1450   // make sure gcd divides Delta
1451   R = Delta.srem(G);
1452   if (R != 0)
1453     return true; // gcd doesn't divide Delta, no dependence
1454   Q = Delta.sdiv(G);
1455   return false;
1456 }
1457 
1458 static APInt floorOfQuotient(const APInt &A, const APInt &B) {
1459   APInt Q = A; // these need to be initialized
1460   APInt R = A;
1461   APInt::sdivrem(A, B, Q, R);
1462   if (R == 0)
1463     return Q;
1464   if ((A.sgt(0) && B.sgt(0)) ||
1465       (A.slt(0) && B.slt(0)))
1466     return Q;
1467   else
1468     return Q - 1;
1469 }
1470 
1471 static APInt ceilingOfQuotient(const APInt &A, const APInt &B) {
1472   APInt Q = A; // these need to be initialized
1473   APInt R = A;
1474   APInt::sdivrem(A, B, Q, R);
1475   if (R == 0)
1476     return Q;
1477   if ((A.sgt(0) && B.sgt(0)) ||
1478       (A.slt(0) && B.slt(0)))
1479     return Q + 1;
1480   else
1481     return Q;
1482 }
1483 
1484 // exactSIVtest -
1485 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*i],
1486 // where i is an induction variable, c1 and c2 are loop invariant, and a1
1487 // and a2 are constant, we can solve it exactly using an algorithm developed
1488 // by Banerjee and Wolfe. See Algorithm 6.2.1 (case 2.5) in:
1489 //
1490 //        Dependence Analysis for Supercomputing
1491 //        Utpal Banerjee
1492 //        Kluwer Academic Publishers, 1988
1493 //
1494 // It's slower than the specialized tests (strong SIV, weak-zero SIV, etc),
1495 // so use them if possible. They're also a bit better with symbolics and,
1496 // in the case of the strong SIV test, can compute Distances.
1497 //
1498 // Return true if dependence disproved.
1499 //
1500 // This is a modified version of the original Banerjee algorithm. The original
1501 // only tested whether Dst depends on Src. This algorithm extends that and
1502 // returns all the dependencies that exist between Dst and Src.
1503 bool DependenceInfo::exactSIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff,
1504                                   const SCEV *SrcConst, const SCEV *DstConst,
1505                                   const Loop *CurLoop, unsigned Level,
1506                                   FullDependence &Result,
1507                                   Constraint &NewConstraint) const {
1508   LLVM_DEBUG(dbgs() << "\tExact SIV test\n");
1509   LLVM_DEBUG(dbgs() << "\t    SrcCoeff = " << *SrcCoeff << " = AM\n");
1510   LLVM_DEBUG(dbgs() << "\t    DstCoeff = " << *DstCoeff << " = BM\n");
1511   LLVM_DEBUG(dbgs() << "\t    SrcConst = " << *SrcConst << "\n");
1512   LLVM_DEBUG(dbgs() << "\t    DstConst = " << *DstConst << "\n");
1513   ++ExactSIVapplications;
1514   assert(0 < Level && Level <= CommonLevels && "Level out of range");
1515   Level--;
1516   Result.Consistent = false;
1517   const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1518   LLVM_DEBUG(dbgs() << "\t    Delta = " << *Delta << "\n");
1519   NewConstraint.setLine(SrcCoeff, SE->getNegativeSCEV(DstCoeff), Delta,
1520                         CurLoop);
1521   const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1522   const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1523   const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1524   if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff)
1525     return false;
1526 
1527   // find gcd
1528   APInt G, X, Y;
1529   APInt AM = ConstSrcCoeff->getAPInt();
1530   APInt BM = ConstDstCoeff->getAPInt();
1531   APInt CM = ConstDelta->getAPInt();
1532   unsigned Bits = AM.getBitWidth();
1533   if (findGCD(Bits, AM, BM, CM, G, X, Y)) {
1534     // gcd doesn't divide Delta, no dependence
1535     ++ExactSIVindependence;
1536     ++ExactSIVsuccesses;
1537     return true;
1538   }
1539 
1540   LLVM_DEBUG(dbgs() << "\t    X = " << X << ", Y = " << Y << "\n");
1541 
1542   // since SCEV construction normalizes, LM = 0
1543   APInt UM(Bits, 1, true);
1544   bool UMValid = false;
1545   // UM is perhaps unavailable, let's check
1546   if (const SCEVConstant *CUB =
1547           collectConstantUpperBound(CurLoop, Delta->getType())) {
1548     UM = CUB->getAPInt();
1549     LLVM_DEBUG(dbgs() << "\t    UM = " << UM << "\n");
1550     UMValid = true;
1551   }
1552 
1553   APInt TU(APInt::getSignedMaxValue(Bits));
1554   APInt TL(APInt::getSignedMinValue(Bits));
1555   APInt TC = CM.sdiv(G);
1556   APInt TX = X * TC;
1557   APInt TY = Y * TC;
1558   LLVM_DEBUG(dbgs() << "\t    TC = " << TC << "\n");
1559   LLVM_DEBUG(dbgs() << "\t    TX = " << TX << "\n");
1560   LLVM_DEBUG(dbgs() << "\t    TY = " << TY << "\n");
1561 
1562   SmallVector<APInt, 2> TLVec, TUVec;
1563   APInt TB = BM.sdiv(G);
1564   if (TB.sgt(0)) {
1565     TLVec.push_back(ceilingOfQuotient(-TX, TB));
1566     LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1567     // New bound check - modification to Banerjee's e3 check
1568     if (UMValid) {
1569       TUVec.push_back(floorOfQuotient(UM - TX, TB));
1570       LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1571     }
1572   } else {
1573     TUVec.push_back(floorOfQuotient(-TX, TB));
1574     LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1575     // New bound check - modification to Banerjee's e3 check
1576     if (UMValid) {
1577       TLVec.push_back(ceilingOfQuotient(UM - TX, TB));
1578       LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1579     }
1580   }
1581 
1582   APInt TA = AM.sdiv(G);
1583   if (TA.sgt(0)) {
1584     if (UMValid) {
1585       TUVec.push_back(floorOfQuotient(UM - TY, TA));
1586       LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1587     }
1588     // New bound check - modification to Banerjee's e3 check
1589     TLVec.push_back(ceilingOfQuotient(-TY, TA));
1590     LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1591   } else {
1592     if (UMValid) {
1593       TLVec.push_back(ceilingOfQuotient(UM - TY, TA));
1594       LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1595     }
1596     // New bound check - modification to Banerjee's e3 check
1597     TUVec.push_back(floorOfQuotient(-TY, TA));
1598     LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1599   }
1600 
1601   LLVM_DEBUG(dbgs() << "\t    TA = " << TA << "\n");
1602   LLVM_DEBUG(dbgs() << "\t    TB = " << TB << "\n");
1603 
1604   if (TLVec.empty() || TUVec.empty())
1605     return false;
1606   TL = APIntOps::smax(TLVec.front(), TLVec.back());
1607   TU = APIntOps::smin(TUVec.front(), TUVec.back());
1608   LLVM_DEBUG(dbgs() << "\t    TL = " << TL << "\n");
1609   LLVM_DEBUG(dbgs() << "\t    TU = " << TU << "\n");
1610 
1611   if (TL.sgt(TU)) {
1612     ++ExactSIVindependence;
1613     ++ExactSIVsuccesses;
1614     return true;
1615   }
1616 
1617   // explore directions
1618   unsigned NewDirection = Dependence::DVEntry::NONE;
1619   APInt LowerDistance, UpperDistance;
1620   if (TA.sgt(TB)) {
1621     LowerDistance = (TY - TX) + (TA - TB) * TL;
1622     UpperDistance = (TY - TX) + (TA - TB) * TU;
1623   } else {
1624     LowerDistance = (TY - TX) + (TA - TB) * TU;
1625     UpperDistance = (TY - TX) + (TA - TB) * TL;
1626   }
1627 
1628   LLVM_DEBUG(dbgs() << "\t    LowerDistance = " << LowerDistance << "\n");
1629   LLVM_DEBUG(dbgs() << "\t    UpperDistance = " << UpperDistance << "\n");
1630 
1631   APInt Zero(Bits, 0, true);
1632   if (LowerDistance.sle(Zero) && UpperDistance.sge(Zero)) {
1633     NewDirection |= Dependence::DVEntry::EQ;
1634     ++ExactSIVsuccesses;
1635   }
1636   if (LowerDistance.slt(0)) {
1637     NewDirection |= Dependence::DVEntry::GT;
1638     ++ExactSIVsuccesses;
1639   }
1640   if (UpperDistance.sgt(0)) {
1641     NewDirection |= Dependence::DVEntry::LT;
1642     ++ExactSIVsuccesses;
1643   }
1644 
1645   // finished
1646   Result.DV[Level].Direction &= NewDirection;
1647   if (Result.DV[Level].Direction == Dependence::DVEntry::NONE)
1648     ++ExactSIVindependence;
1649   LLVM_DEBUG(dbgs() << "\t    Result = ");
1650   LLVM_DEBUG(Result.dump(dbgs()));
1651   return Result.DV[Level].Direction == Dependence::DVEntry::NONE;
1652 }
1653 
1654 
1655 // Return true if the divisor evenly divides the dividend.
1656 static
1657 bool isRemainderZero(const SCEVConstant *Dividend,
1658                      const SCEVConstant *Divisor) {
1659   const APInt &ConstDividend = Dividend->getAPInt();
1660   const APInt &ConstDivisor = Divisor->getAPInt();
1661   return ConstDividend.srem(ConstDivisor) == 0;
1662 }
1663 
1664 
1665 // weakZeroSrcSIVtest -
1666 // From the paper, Practical Dependence Testing, Section 4.2.2
1667 //
1668 // When we have a pair of subscripts of the form [c1] and [c2 + a*i],
1669 // where i is an induction variable, c1 and c2 are loop invariant,
1670 // and a is a constant, we can solve it exactly using the
1671 // Weak-Zero SIV test.
1672 //
1673 // Given
1674 //
1675 //    c1 = c2 + a*i
1676 //
1677 // we get
1678 //
1679 //    (c1 - c2)/a = i
1680 //
1681 // If i is not an integer, there's no dependence.
1682 // If i < 0 or > UB, there's no dependence.
1683 // If i = 0, the direction is >= and peeling the
1684 // 1st iteration will break the dependence.
1685 // If i = UB, the direction is <= and peeling the
1686 // last iteration will break the dependence.
1687 // Otherwise, the direction is *.
1688 //
1689 // Can prove independence. Failing that, we can sometimes refine
1690 // the directions. Can sometimes show that first or last
1691 // iteration carries all the dependences (so worth peeling).
1692 //
1693 // (see also weakZeroDstSIVtest)
1694 //
1695 // Return true if dependence disproved.
1696 bool DependenceInfo::weakZeroSrcSIVtest(const SCEV *DstCoeff,
1697                                         const SCEV *SrcConst,
1698                                         const SCEV *DstConst,
1699                                         const Loop *CurLoop, unsigned Level,
1700                                         FullDependence &Result,
1701                                         Constraint &NewConstraint) const {
1702   // For the WeakSIV test, it's possible the loop isn't common to
1703   // the Src and Dst loops. If it isn't, then there's no need to
1704   // record a direction.
1705   LLVM_DEBUG(dbgs() << "\tWeak-Zero (src) SIV test\n");
1706   LLVM_DEBUG(dbgs() << "\t    DstCoeff = " << *DstCoeff << "\n");
1707   LLVM_DEBUG(dbgs() << "\t    SrcConst = " << *SrcConst << "\n");
1708   LLVM_DEBUG(dbgs() << "\t    DstConst = " << *DstConst << "\n");
1709   ++WeakZeroSIVapplications;
1710   assert(0 < Level && Level <= MaxLevels && "Level out of range");
1711   Level--;
1712   Result.Consistent = false;
1713   const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst);
1714   NewConstraint.setLine(SE->getZero(Delta->getType()), DstCoeff, Delta,
1715                         CurLoop);
1716   LLVM_DEBUG(dbgs() << "\t    Delta = " << *Delta << "\n");
1717   if (isKnownPredicate(CmpInst::ICMP_EQ, SrcConst, DstConst)) {
1718     if (Level < CommonLevels) {
1719       Result.DV[Level].Direction &= Dependence::DVEntry::GE;
1720       Result.DV[Level].PeelFirst = true;
1721       ++WeakZeroSIVsuccesses;
1722     }
1723     return false; // dependences caused by first iteration
1724   }
1725   const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1726   if (!ConstCoeff)
1727     return false;
1728   const SCEV *AbsCoeff =
1729     SE->isKnownNegative(ConstCoeff) ?
1730     SE->getNegativeSCEV(ConstCoeff) : ConstCoeff;
1731   const SCEV *NewDelta =
1732     SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta;
1733 
1734   // check that Delta/SrcCoeff < iteration count
1735   // really check NewDelta < count*AbsCoeff
1736   if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1737     LLVM_DEBUG(dbgs() << "\t    UpperBound = " << *UpperBound << "\n");
1738     const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound);
1739     if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) {
1740       ++WeakZeroSIVindependence;
1741       ++WeakZeroSIVsuccesses;
1742       return true;
1743     }
1744     if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) {
1745       // dependences caused by last iteration
1746       if (Level < CommonLevels) {
1747         Result.DV[Level].Direction &= Dependence::DVEntry::LE;
1748         Result.DV[Level].PeelLast = true;
1749         ++WeakZeroSIVsuccesses;
1750       }
1751       return false;
1752     }
1753   }
1754 
1755   // check that Delta/SrcCoeff >= 0
1756   // really check that NewDelta >= 0
1757   if (SE->isKnownNegative(NewDelta)) {
1758     // No dependence, newDelta < 0
1759     ++WeakZeroSIVindependence;
1760     ++WeakZeroSIVsuccesses;
1761     return true;
1762   }
1763 
1764   // if SrcCoeff doesn't divide Delta, then no dependence
1765   if (isa<SCEVConstant>(Delta) &&
1766       !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) {
1767     ++WeakZeroSIVindependence;
1768     ++WeakZeroSIVsuccesses;
1769     return true;
1770   }
1771   return false;
1772 }
1773 
1774 
1775 // weakZeroDstSIVtest -
1776 // From the paper, Practical Dependence Testing, Section 4.2.2
1777 //
1778 // When we have a pair of subscripts of the form [c1 + a*i] and [c2],
1779 // where i is an induction variable, c1 and c2 are loop invariant,
1780 // and a is a constant, we can solve it exactly using the
1781 // Weak-Zero SIV test.
1782 //
1783 // Given
1784 //
1785 //    c1 + a*i = c2
1786 //
1787 // we get
1788 //
1789 //    i = (c2 - c1)/a
1790 //
1791 // If i is not an integer, there's no dependence.
1792 // If i < 0 or > UB, there's no dependence.
1793 // If i = 0, the direction is <= and peeling the
1794 // 1st iteration will break the dependence.
1795 // If i = UB, the direction is >= and peeling the
1796 // last iteration will break the dependence.
1797 // Otherwise, the direction is *.
1798 //
1799 // Can prove independence. Failing that, we can sometimes refine
1800 // the directions. Can sometimes show that first or last
1801 // iteration carries all the dependences (so worth peeling).
1802 //
1803 // (see also weakZeroSrcSIVtest)
1804 //
1805 // Return true if dependence disproved.
1806 bool DependenceInfo::weakZeroDstSIVtest(const SCEV *SrcCoeff,
1807                                         const SCEV *SrcConst,
1808                                         const SCEV *DstConst,
1809                                         const Loop *CurLoop, unsigned Level,
1810                                         FullDependence &Result,
1811                                         Constraint &NewConstraint) const {
1812   // For the WeakSIV test, it's possible the loop isn't common to the
1813   // Src and Dst loops. If it isn't, then there's no need to record a direction.
1814   LLVM_DEBUG(dbgs() << "\tWeak-Zero (dst) SIV test\n");
1815   LLVM_DEBUG(dbgs() << "\t    SrcCoeff = " << *SrcCoeff << "\n");
1816   LLVM_DEBUG(dbgs() << "\t    SrcConst = " << *SrcConst << "\n");
1817   LLVM_DEBUG(dbgs() << "\t    DstConst = " << *DstConst << "\n");
1818   ++WeakZeroSIVapplications;
1819   assert(0 < Level && Level <= SrcLevels && "Level out of range");
1820   Level--;
1821   Result.Consistent = false;
1822   const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1823   NewConstraint.setLine(SrcCoeff, SE->getZero(Delta->getType()), Delta,
1824                         CurLoop);
1825   LLVM_DEBUG(dbgs() << "\t    Delta = " << *Delta << "\n");
1826   if (isKnownPredicate(CmpInst::ICMP_EQ, DstConst, SrcConst)) {
1827     if (Level < CommonLevels) {
1828       Result.DV[Level].Direction &= Dependence::DVEntry::LE;
1829       Result.DV[Level].PeelFirst = true;
1830       ++WeakZeroSIVsuccesses;
1831     }
1832     return false; // dependences caused by first iteration
1833   }
1834   const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1835   if (!ConstCoeff)
1836     return false;
1837   const SCEV *AbsCoeff =
1838     SE->isKnownNegative(ConstCoeff) ?
1839     SE->getNegativeSCEV(ConstCoeff) : ConstCoeff;
1840   const SCEV *NewDelta =
1841     SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta;
1842 
1843   // check that Delta/SrcCoeff < iteration count
1844   // really check NewDelta < count*AbsCoeff
1845   if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1846     LLVM_DEBUG(dbgs() << "\t    UpperBound = " << *UpperBound << "\n");
1847     const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound);
1848     if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) {
1849       ++WeakZeroSIVindependence;
1850       ++WeakZeroSIVsuccesses;
1851       return true;
1852     }
1853     if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) {
1854       // dependences caused by last iteration
1855       if (Level < CommonLevels) {
1856         Result.DV[Level].Direction &= Dependence::DVEntry::GE;
1857         Result.DV[Level].PeelLast = true;
1858         ++WeakZeroSIVsuccesses;
1859       }
1860       return false;
1861     }
1862   }
1863 
1864   // check that Delta/SrcCoeff >= 0
1865   // really check that NewDelta >= 0
1866   if (SE->isKnownNegative(NewDelta)) {
1867     // No dependence, newDelta < 0
1868     ++WeakZeroSIVindependence;
1869     ++WeakZeroSIVsuccesses;
1870     return true;
1871   }
1872 
1873   // if SrcCoeff doesn't divide Delta, then no dependence
1874   if (isa<SCEVConstant>(Delta) &&
1875       !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) {
1876     ++WeakZeroSIVindependence;
1877     ++WeakZeroSIVsuccesses;
1878     return true;
1879   }
1880   return false;
1881 }
1882 
1883 
1884 // exactRDIVtest - Tests the RDIV subscript pair for dependence.
1885 // Things of the form [c1 + a*i] and [c2 + b*j],
1886 // where i and j are induction variable, c1 and c2 are loop invariant,
1887 // and a and b are constants.
1888 // Returns true if any possible dependence is disproved.
1889 // Marks the result as inconsistent.
1890 // Works in some cases that symbolicRDIVtest doesn't, and vice versa.
1891 bool DependenceInfo::exactRDIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff,
1892                                    const SCEV *SrcConst, const SCEV *DstConst,
1893                                    const Loop *SrcLoop, const Loop *DstLoop,
1894                                    FullDependence &Result) const {
1895   LLVM_DEBUG(dbgs() << "\tExact RDIV test\n");
1896   LLVM_DEBUG(dbgs() << "\t    SrcCoeff = " << *SrcCoeff << " = AM\n");
1897   LLVM_DEBUG(dbgs() << "\t    DstCoeff = " << *DstCoeff << " = BM\n");
1898   LLVM_DEBUG(dbgs() << "\t    SrcConst = " << *SrcConst << "\n");
1899   LLVM_DEBUG(dbgs() << "\t    DstConst = " << *DstConst << "\n");
1900   ++ExactRDIVapplications;
1901   Result.Consistent = false;
1902   const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1903   LLVM_DEBUG(dbgs() << "\t    Delta = " << *Delta << "\n");
1904   const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1905   const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1906   const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1907   if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff)
1908     return false;
1909 
1910   // find gcd
1911   APInt G, X, Y;
1912   APInt AM = ConstSrcCoeff->getAPInt();
1913   APInt BM = ConstDstCoeff->getAPInt();
1914   APInt CM = ConstDelta->getAPInt();
1915   unsigned Bits = AM.getBitWidth();
1916   if (findGCD(Bits, AM, BM, CM, G, X, Y)) {
1917     // gcd doesn't divide Delta, no dependence
1918     ++ExactRDIVindependence;
1919     return true;
1920   }
1921 
1922   LLVM_DEBUG(dbgs() << "\t    X = " << X << ", Y = " << Y << "\n");
1923 
1924   // since SCEV construction seems to normalize, LM = 0
1925   APInt SrcUM(Bits, 1, true);
1926   bool SrcUMvalid = false;
1927   // SrcUM is perhaps unavailable, let's check
1928   if (const SCEVConstant *UpperBound =
1929           collectConstantUpperBound(SrcLoop, Delta->getType())) {
1930     SrcUM = UpperBound->getAPInt();
1931     LLVM_DEBUG(dbgs() << "\t    SrcUM = " << SrcUM << "\n");
1932     SrcUMvalid = true;
1933   }
1934 
1935   APInt DstUM(Bits, 1, true);
1936   bool DstUMvalid = false;
1937   // UM is perhaps unavailable, let's check
1938   if (const SCEVConstant *UpperBound =
1939           collectConstantUpperBound(DstLoop, Delta->getType())) {
1940     DstUM = UpperBound->getAPInt();
1941     LLVM_DEBUG(dbgs() << "\t    DstUM = " << DstUM << "\n");
1942     DstUMvalid = true;
1943   }
1944 
1945   APInt TU(APInt::getSignedMaxValue(Bits));
1946   APInt TL(APInt::getSignedMinValue(Bits));
1947   APInt TC = CM.sdiv(G);
1948   APInt TX = X * TC;
1949   APInt TY = Y * TC;
1950   LLVM_DEBUG(dbgs() << "\t    TC = " << TC << "\n");
1951   LLVM_DEBUG(dbgs() << "\t    TX = " << TX << "\n");
1952   LLVM_DEBUG(dbgs() << "\t    TY = " << TY << "\n");
1953 
1954   SmallVector<APInt, 2> TLVec, TUVec;
1955   APInt TB = BM.sdiv(G);
1956   if (TB.sgt(0)) {
1957     TLVec.push_back(ceilingOfQuotient(-TX, TB));
1958     LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1959     if (SrcUMvalid) {
1960       TUVec.push_back(floorOfQuotient(SrcUM - TX, TB));
1961       LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1962     }
1963   } else {
1964     TUVec.push_back(floorOfQuotient(-TX, TB));
1965     LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1966     if (SrcUMvalid) {
1967       TLVec.push_back(ceilingOfQuotient(SrcUM - TX, TB));
1968       LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1969     }
1970   }
1971 
1972   APInt TA = AM.sdiv(G);
1973   if (TA.sgt(0)) {
1974     TLVec.push_back(ceilingOfQuotient(-TY, TA));
1975     LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1976     if (DstUMvalid) {
1977       TUVec.push_back(floorOfQuotient(DstUM - TY, TA));
1978       LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1979     }
1980   } else {
1981     TUVec.push_back(floorOfQuotient(-TY, TA));
1982     LLVM_DEBUG(dbgs() << "\t    Possible TU = " << TUVec.back() << "\n");
1983     if (DstUMvalid) {
1984       TLVec.push_back(ceilingOfQuotient(DstUM - TY, TA));
1985       LLVM_DEBUG(dbgs() << "\t    Possible TL = " << TLVec.back() << "\n");
1986     }
1987   }
1988 
1989   if (TLVec.empty() || TUVec.empty())
1990     return false;
1991 
1992   LLVM_DEBUG(dbgs() << "\t    TA = " << TA << "\n");
1993   LLVM_DEBUG(dbgs() << "\t    TB = " << TB << "\n");
1994 
1995   TL = APIntOps::smax(TLVec.front(), TLVec.back());
1996   TU = APIntOps::smin(TUVec.front(), TUVec.back());
1997   LLVM_DEBUG(dbgs() << "\t    TL = " << TL << "\n");
1998   LLVM_DEBUG(dbgs() << "\t    TU = " << TU << "\n");
1999 
2000   if (TL.sgt(TU))
2001     ++ExactRDIVindependence;
2002   return TL.sgt(TU);
2003 }
2004 
2005 
2006 // symbolicRDIVtest -
2007 // In Section 4.5 of the Practical Dependence Testing paper,the authors
2008 // introduce a special case of Banerjee's Inequalities (also called the
2009 // Extreme-Value Test) that can handle some of the SIV and RDIV cases,
2010 // particularly cases with symbolics. Since it's only able to disprove
2011 // dependence (not compute distances or directions), we'll use it as a
2012 // fall back for the other tests.
2013 //
2014 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j]
2015 // where i and j are induction variables and c1 and c2 are loop invariants,
2016 // we can use the symbolic tests to disprove some dependences, serving as a
2017 // backup for the RDIV test. Note that i and j can be the same variable,
2018 // letting this test serve as a backup for the various SIV tests.
2019 //
2020 // For a dependence to exist, c1 + a1*i must equal c2 + a2*j for some
2021 //  0 <= i <= N1 and some 0 <= j <= N2, where N1 and N2 are the (normalized)
2022 // loop bounds for the i and j loops, respectively. So, ...
2023 //
2024 // c1 + a1*i = c2 + a2*j
2025 // a1*i - a2*j = c2 - c1
2026 //
2027 // To test for a dependence, we compute c2 - c1 and make sure it's in the
2028 // range of the maximum and minimum possible values of a1*i - a2*j.
2029 // Considering the signs of a1 and a2, we have 4 possible cases:
2030 //
2031 // 1) If a1 >= 0 and a2 >= 0, then
2032 //        a1*0 - a2*N2 <= c2 - c1 <= a1*N1 - a2*0
2033 //              -a2*N2 <= c2 - c1 <= a1*N1
2034 //
2035 // 2) If a1 >= 0 and a2 <= 0, then
2036 //        a1*0 - a2*0 <= c2 - c1 <= a1*N1 - a2*N2
2037 //                  0 <= c2 - c1 <= a1*N1 - a2*N2
2038 //
2039 // 3) If a1 <= 0 and a2 >= 0, then
2040 //        a1*N1 - a2*N2 <= c2 - c1 <= a1*0 - a2*0
2041 //        a1*N1 - a2*N2 <= c2 - c1 <= 0
2042 //
2043 // 4) If a1 <= 0 and a2 <= 0, then
2044 //        a1*N1 - a2*0  <= c2 - c1 <= a1*0 - a2*N2
2045 //        a1*N1         <= c2 - c1 <=       -a2*N2
2046 //
2047 // return true if dependence disproved
2048 bool DependenceInfo::symbolicRDIVtest(const SCEV *A1, const SCEV *A2,
2049                                       const SCEV *C1, const SCEV *C2,
2050                                       const Loop *Loop1,
2051                                       const Loop *Loop2) const {
2052   ++SymbolicRDIVapplications;
2053   LLVM_DEBUG(dbgs() << "\ttry symbolic RDIV test\n");
2054   LLVM_DEBUG(dbgs() << "\t    A1 = " << *A1);
2055   LLVM_DEBUG(dbgs() << ", type = " << *A1->getType() << "\n");
2056   LLVM_DEBUG(dbgs() << "\t    A2 = " << *A2 << "\n");
2057   LLVM_DEBUG(dbgs() << "\t    C1 = " << *C1 << "\n");
2058   LLVM_DEBUG(dbgs() << "\t    C2 = " << *C2 << "\n");
2059   const SCEV *N1 = collectUpperBound(Loop1, A1->getType());
2060   const SCEV *N2 = collectUpperBound(Loop2, A1->getType());
2061   LLVM_DEBUG(if (N1) dbgs() << "\t    N1 = " << *N1 << "\n");
2062   LLVM_DEBUG(if (N2) dbgs() << "\t    N2 = " << *N2 << "\n");
2063   const SCEV *C2_C1 = SE->getMinusSCEV(C2, C1);
2064   const SCEV *C1_C2 = SE->getMinusSCEV(C1, C2);
2065   LLVM_DEBUG(dbgs() << "\t    C2 - C1 = " << *C2_C1 << "\n");
2066   LLVM_DEBUG(dbgs() << "\t    C1 - C2 = " << *C1_C2 << "\n");
2067   if (SE->isKnownNonNegative(A1)) {
2068     if (SE->isKnownNonNegative(A2)) {
2069       // A1 >= 0 && A2 >= 0
2070       if (N1) {
2071         // make sure that c2 - c1 <= a1*N1
2072         const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2073         LLVM_DEBUG(dbgs() << "\t    A1*N1 = " << *A1N1 << "\n");
2074         if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1)) {
2075           ++SymbolicRDIVindependence;
2076           return true;
2077         }
2078       }
2079       if (N2) {
2080         // make sure that -a2*N2 <= c2 - c1, or a2*N2 >= c1 - c2
2081         const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2082         LLVM_DEBUG(dbgs() << "\t    A2*N2 = " << *A2N2 << "\n");
2083         if (isKnownPredicate(CmpInst::ICMP_SLT, A2N2, C1_C2)) {
2084           ++SymbolicRDIVindependence;
2085           return true;
2086         }
2087       }
2088     }
2089     else if (SE->isKnownNonPositive(A2)) {
2090       // a1 >= 0 && a2 <= 0
2091       if (N1 && N2) {
2092         // make sure that c2 - c1 <= a1*N1 - a2*N2
2093         const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2094         const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2095         const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2);
2096         LLVM_DEBUG(dbgs() << "\t    A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n");
2097         if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1_A2N2)) {
2098           ++SymbolicRDIVindependence;
2099           return true;
2100         }
2101       }
2102       // make sure that 0 <= c2 - c1
2103       if (SE->isKnownNegative(C2_C1)) {
2104         ++SymbolicRDIVindependence;
2105         return true;
2106       }
2107     }
2108   }
2109   else if (SE->isKnownNonPositive(A1)) {
2110     if (SE->isKnownNonNegative(A2)) {
2111       // a1 <= 0 && a2 >= 0
2112       if (N1 && N2) {
2113         // make sure that a1*N1 - a2*N2 <= c2 - c1
2114         const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2115         const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2116         const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2);
2117         LLVM_DEBUG(dbgs() << "\t    A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n");
2118         if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1_A2N2, C2_C1)) {
2119           ++SymbolicRDIVindependence;
2120           return true;
2121         }
2122       }
2123       // make sure that c2 - c1 <= 0
2124       if (SE->isKnownPositive(C2_C1)) {
2125         ++SymbolicRDIVindependence;
2126         return true;
2127       }
2128     }
2129     else if (SE->isKnownNonPositive(A2)) {
2130       // a1 <= 0 && a2 <= 0
2131       if (N1) {
2132         // make sure that a1*N1 <= c2 - c1
2133         const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2134         LLVM_DEBUG(dbgs() << "\t    A1*N1 = " << *A1N1 << "\n");
2135         if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1, C2_C1)) {
2136           ++SymbolicRDIVindependence;
2137           return true;
2138         }
2139       }
2140       if (N2) {
2141         // make sure that c2 - c1 <= -a2*N2, or c1 - c2 >= a2*N2
2142         const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2143         LLVM_DEBUG(dbgs() << "\t    A2*N2 = " << *A2N2 << "\n");
2144         if (isKnownPredicate(CmpInst::ICMP_SLT, C1_C2, A2N2)) {
2145           ++SymbolicRDIVindependence;
2146           return true;
2147         }
2148       }
2149     }
2150   }
2151   return false;
2152 }
2153 
2154 
2155 // testSIV -
2156 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 - a2*i]
2157 // where i is an induction variable, c1 and c2 are loop invariant, and a1 and
2158 // a2 are constant, we attack it with an SIV test. While they can all be
2159 // solved with the Exact SIV test, it's worthwhile to use simpler tests when
2160 // they apply; they're cheaper and sometimes more precise.
2161 //
2162 // Return true if dependence disproved.
2163 bool DependenceInfo::testSIV(const SCEV *Src, const SCEV *Dst, unsigned &Level,
2164                              FullDependence &Result, Constraint &NewConstraint,
2165                              const SCEV *&SplitIter) const {
2166   LLVM_DEBUG(dbgs() << "    src = " << *Src << "\n");
2167   LLVM_DEBUG(dbgs() << "    dst = " << *Dst << "\n");
2168   const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src);
2169   const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst);
2170   if (SrcAddRec && DstAddRec) {
2171     const SCEV *SrcConst = SrcAddRec->getStart();
2172     const SCEV *DstConst = DstAddRec->getStart();
2173     const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2174     const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE);
2175     const Loop *CurLoop = SrcAddRec->getLoop();
2176     assert(CurLoop == DstAddRec->getLoop() &&
2177            "both loops in SIV should be same");
2178     Level = mapSrcLoop(CurLoop);
2179     bool disproven;
2180     if (SrcCoeff == DstCoeff)
2181       disproven = strongSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2182                                 Level, Result, NewConstraint);
2183     else if (SrcCoeff == SE->getNegativeSCEV(DstCoeff))
2184       disproven = weakCrossingSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2185                                       Level, Result, NewConstraint, SplitIter);
2186     else
2187       disproven = exactSIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop,
2188                                Level, Result, NewConstraint);
2189     return disproven ||
2190       gcdMIVtest(Src, Dst, Result) ||
2191       symbolicRDIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop, CurLoop);
2192   }
2193   if (SrcAddRec) {
2194     const SCEV *SrcConst = SrcAddRec->getStart();
2195     const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2196     const SCEV *DstConst = Dst;
2197     const Loop *CurLoop = SrcAddRec->getLoop();
2198     Level = mapSrcLoop(CurLoop);
2199     return weakZeroDstSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2200                               Level, Result, NewConstraint) ||
2201       gcdMIVtest(Src, Dst, Result);
2202   }
2203   if (DstAddRec) {
2204     const SCEV *DstConst = DstAddRec->getStart();
2205     const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE);
2206     const SCEV *SrcConst = Src;
2207     const Loop *CurLoop = DstAddRec->getLoop();
2208     Level = mapDstLoop(CurLoop);
2209     return weakZeroSrcSIVtest(DstCoeff, SrcConst, DstConst,
2210                               CurLoop, Level, Result, NewConstraint) ||
2211       gcdMIVtest(Src, Dst, Result);
2212   }
2213   llvm_unreachable("SIV test expected at least one AddRec");
2214   return false;
2215 }
2216 
2217 
2218 // testRDIV -
2219 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j]
2220 // where i and j are induction variables, c1 and c2 are loop invariant,
2221 // and a1 and a2 are constant, we can solve it exactly with an easy adaptation
2222 // of the Exact SIV test, the Restricted Double Index Variable (RDIV) test.
2223 // It doesn't make sense to talk about distance or direction in this case,
2224 // so there's no point in making special versions of the Strong SIV test or
2225 // the Weak-crossing SIV test.
2226 //
2227 // With minor algebra, this test can also be used for things like
2228 // [c1 + a1*i + a2*j][c2].
2229 //
2230 // Return true if dependence disproved.
2231 bool DependenceInfo::testRDIV(const SCEV *Src, const SCEV *Dst,
2232                               FullDependence &Result) const {
2233   // we have 3 possible situations here:
2234   //   1) [a*i + b] and [c*j + d]
2235   //   2) [a*i + c*j + b] and [d]
2236   //   3) [b] and [a*i + c*j + d]
2237   // We need to find what we've got and get organized
2238 
2239   const SCEV *SrcConst, *DstConst;
2240   const SCEV *SrcCoeff, *DstCoeff;
2241   const Loop *SrcLoop, *DstLoop;
2242 
2243   LLVM_DEBUG(dbgs() << "    src = " << *Src << "\n");
2244   LLVM_DEBUG(dbgs() << "    dst = " << *Dst << "\n");
2245   const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src);
2246   const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst);
2247   if (SrcAddRec && DstAddRec) {
2248     SrcConst = SrcAddRec->getStart();
2249     SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2250     SrcLoop = SrcAddRec->getLoop();
2251     DstConst = DstAddRec->getStart();
2252     DstCoeff = DstAddRec->getStepRecurrence(*SE);
2253     DstLoop = DstAddRec->getLoop();
2254   }
2255   else if (SrcAddRec) {
2256     if (const SCEVAddRecExpr *tmpAddRec =
2257         dyn_cast<SCEVAddRecExpr>(SrcAddRec->getStart())) {
2258       SrcConst = tmpAddRec->getStart();
2259       SrcCoeff = tmpAddRec->getStepRecurrence(*SE);
2260       SrcLoop = tmpAddRec->getLoop();
2261       DstConst = Dst;
2262       DstCoeff = SE->getNegativeSCEV(SrcAddRec->getStepRecurrence(*SE));
2263       DstLoop = SrcAddRec->getLoop();
2264     }
2265     else
2266       llvm_unreachable("RDIV reached by surprising SCEVs");
2267   }
2268   else if (DstAddRec) {
2269     if (const SCEVAddRecExpr *tmpAddRec =
2270         dyn_cast<SCEVAddRecExpr>(DstAddRec->getStart())) {
2271       DstConst = tmpAddRec->getStart();
2272       DstCoeff = tmpAddRec->getStepRecurrence(*SE);
2273       DstLoop = tmpAddRec->getLoop();
2274       SrcConst = Src;
2275       SrcCoeff = SE->getNegativeSCEV(DstAddRec->getStepRecurrence(*SE));
2276       SrcLoop = DstAddRec->getLoop();
2277     }
2278     else
2279       llvm_unreachable("RDIV reached by surprising SCEVs");
2280   }
2281   else
2282     llvm_unreachable("RDIV expected at least one AddRec");
2283   return exactRDIVtest(SrcCoeff, DstCoeff,
2284                        SrcConst, DstConst,
2285                        SrcLoop, DstLoop,
2286                        Result) ||
2287     gcdMIVtest(Src, Dst, Result) ||
2288     symbolicRDIVtest(SrcCoeff, DstCoeff,
2289                      SrcConst, DstConst,
2290                      SrcLoop, DstLoop);
2291 }
2292 
2293 
2294 // Tests the single-subscript MIV pair (Src and Dst) for dependence.
2295 // Return true if dependence disproved.
2296 // Can sometimes refine direction vectors.
2297 bool DependenceInfo::testMIV(const SCEV *Src, const SCEV *Dst,
2298                              const SmallBitVector &Loops,
2299                              FullDependence &Result) const {
2300   LLVM_DEBUG(dbgs() << "    src = " << *Src << "\n");
2301   LLVM_DEBUG(dbgs() << "    dst = " << *Dst << "\n");
2302   Result.Consistent = false;
2303   return gcdMIVtest(Src, Dst, Result) ||
2304     banerjeeMIVtest(Src, Dst, Loops, Result);
2305 }
2306 
2307 
2308 // Given a product, e.g., 10*X*Y, returns the first constant operand,
2309 // in this case 10. If there is no constant part, returns NULL.
2310 static
2311 const SCEVConstant *getConstantPart(const SCEV *Expr) {
2312   if (const auto *Constant = dyn_cast<SCEVConstant>(Expr))
2313     return Constant;
2314   else if (const auto *Product = dyn_cast<SCEVMulExpr>(Expr))
2315     if (const auto *Constant = dyn_cast<SCEVConstant>(Product->getOperand(0)))
2316       return Constant;
2317   return nullptr;
2318 }
2319 
2320 
2321 //===----------------------------------------------------------------------===//
2322 // gcdMIVtest -
2323 // Tests an MIV subscript pair for dependence.
2324 // Returns true if any possible dependence is disproved.
2325 // Marks the result as inconsistent.
2326 // Can sometimes disprove the equal direction for 1 or more loops,
2327 // as discussed in Michael Wolfe's book,
2328 // High Performance Compilers for Parallel Computing, page 235.
2329 //
2330 // We spend some effort (code!) to handle cases like
2331 // [10*i + 5*N*j + 15*M + 6], where i and j are induction variables,
2332 // but M and N are just loop-invariant variables.
2333 // This should help us handle linearized subscripts;
2334 // also makes this test a useful backup to the various SIV tests.
2335 //
2336 // It occurs to me that the presence of loop-invariant variables
2337 // changes the nature of the test from "greatest common divisor"
2338 // to "a common divisor".
2339 bool DependenceInfo::gcdMIVtest(const SCEV *Src, const SCEV *Dst,
2340                                 FullDependence &Result) const {
2341   LLVM_DEBUG(dbgs() << "starting gcd\n");
2342   ++GCDapplications;
2343   unsigned BitWidth = SE->getTypeSizeInBits(Src->getType());
2344   APInt RunningGCD = APInt::getZero(BitWidth);
2345 
2346   // Examine Src coefficients.
2347   // Compute running GCD and record source constant.
2348   // Because we're looking for the constant at the end of the chain,
2349   // we can't quit the loop just because the GCD == 1.
2350   const SCEV *Coefficients = Src;
2351   while (const SCEVAddRecExpr *AddRec =
2352          dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2353     const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2354     // If the coefficient is the product of a constant and other stuff,
2355     // we can use the constant in the GCD computation.
2356     const auto *Constant = getConstantPart(Coeff);
2357     if (!Constant)
2358       return false;
2359     APInt ConstCoeff = Constant->getAPInt();
2360     RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2361     Coefficients = AddRec->getStart();
2362   }
2363   const SCEV *SrcConst = Coefficients;
2364 
2365   // Examine Dst coefficients.
2366   // Compute running GCD and record destination constant.
2367   // Because we're looking for the constant at the end of the chain,
2368   // we can't quit the loop just because the GCD == 1.
2369   Coefficients = Dst;
2370   while (const SCEVAddRecExpr *AddRec =
2371          dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2372     const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2373     // If the coefficient is the product of a constant and other stuff,
2374     // we can use the constant in the GCD computation.
2375     const auto *Constant = getConstantPart(Coeff);
2376     if (!Constant)
2377       return false;
2378     APInt ConstCoeff = Constant->getAPInt();
2379     RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2380     Coefficients = AddRec->getStart();
2381   }
2382   const SCEV *DstConst = Coefficients;
2383 
2384   APInt ExtraGCD = APInt::getZero(BitWidth);
2385   const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
2386   LLVM_DEBUG(dbgs() << "    Delta = " << *Delta << "\n");
2387   const SCEVConstant *Constant = dyn_cast<SCEVConstant>(Delta);
2388   if (const SCEVAddExpr *Sum = dyn_cast<SCEVAddExpr>(Delta)) {
2389     // If Delta is a sum of products, we may be able to make further progress.
2390     for (unsigned Op = 0, Ops = Sum->getNumOperands(); Op < Ops; Op++) {
2391       const SCEV *Operand = Sum->getOperand(Op);
2392       if (isa<SCEVConstant>(Operand)) {
2393         assert(!Constant && "Surprised to find multiple constants");
2394         Constant = cast<SCEVConstant>(Operand);
2395       }
2396       else if (const SCEVMulExpr *Product = dyn_cast<SCEVMulExpr>(Operand)) {
2397         // Search for constant operand to participate in GCD;
2398         // If none found; return false.
2399         const SCEVConstant *ConstOp = getConstantPart(Product);
2400         if (!ConstOp)
2401           return false;
2402         APInt ConstOpValue = ConstOp->getAPInt();
2403         ExtraGCD = APIntOps::GreatestCommonDivisor(ExtraGCD,
2404                                                    ConstOpValue.abs());
2405       }
2406       else
2407         return false;
2408     }
2409   }
2410   if (!Constant)
2411     return false;
2412   APInt ConstDelta = cast<SCEVConstant>(Constant)->getAPInt();
2413   LLVM_DEBUG(dbgs() << "    ConstDelta = " << ConstDelta << "\n");
2414   if (ConstDelta == 0)
2415     return false;
2416   RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ExtraGCD);
2417   LLVM_DEBUG(dbgs() << "    RunningGCD = " << RunningGCD << "\n");
2418   APInt Remainder = ConstDelta.srem(RunningGCD);
2419   if (Remainder != 0) {
2420     ++GCDindependence;
2421     return true;
2422   }
2423 
2424   // Try to disprove equal directions.
2425   // For example, given a subscript pair [3*i + 2*j] and [i' + 2*j' - 1],
2426   // the code above can't disprove the dependence because the GCD = 1.
2427   // So we consider what happen if i = i' and what happens if j = j'.
2428   // If i = i', we can simplify the subscript to [2*i + 2*j] and [2*j' - 1],
2429   // which is infeasible, so we can disallow the = direction for the i level.
2430   // Setting j = j' doesn't help matters, so we end up with a direction vector
2431   // of [<>, *]
2432   //
2433   // Given A[5*i + 10*j*M + 9*M*N] and A[15*i + 20*j*M - 21*N*M + 5],
2434   // we need to remember that the constant part is 5 and the RunningGCD should
2435   // be initialized to ExtraGCD = 30.
2436   LLVM_DEBUG(dbgs() << "    ExtraGCD = " << ExtraGCD << '\n');
2437 
2438   bool Improved = false;
2439   Coefficients = Src;
2440   while (const SCEVAddRecExpr *AddRec =
2441          dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2442     Coefficients = AddRec->getStart();
2443     const Loop *CurLoop = AddRec->getLoop();
2444     RunningGCD = ExtraGCD;
2445     const SCEV *SrcCoeff = AddRec->getStepRecurrence(*SE);
2446     const SCEV *DstCoeff = SE->getMinusSCEV(SrcCoeff, SrcCoeff);
2447     const SCEV *Inner = Src;
2448     while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) {
2449       AddRec = cast<SCEVAddRecExpr>(Inner);
2450       const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2451       if (CurLoop == AddRec->getLoop())
2452         ; // SrcCoeff == Coeff
2453       else {
2454         // If the coefficient is the product of a constant and other stuff,
2455         // we can use the constant in the GCD computation.
2456         Constant = getConstantPart(Coeff);
2457         if (!Constant)
2458           return false;
2459         APInt ConstCoeff = Constant->getAPInt();
2460         RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2461       }
2462       Inner = AddRec->getStart();
2463     }
2464     Inner = Dst;
2465     while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) {
2466       AddRec = cast<SCEVAddRecExpr>(Inner);
2467       const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2468       if (CurLoop == AddRec->getLoop())
2469         DstCoeff = Coeff;
2470       else {
2471         // If the coefficient is the product of a constant and other stuff,
2472         // we can use the constant in the GCD computation.
2473         Constant = getConstantPart(Coeff);
2474         if (!Constant)
2475           return false;
2476         APInt ConstCoeff = Constant->getAPInt();
2477         RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2478       }
2479       Inner = AddRec->getStart();
2480     }
2481     Delta = SE->getMinusSCEV(SrcCoeff, DstCoeff);
2482     // If the coefficient is the product of a constant and other stuff,
2483     // we can use the constant in the GCD computation.
2484     Constant = getConstantPart(Delta);
2485     if (!Constant)
2486       // The difference of the two coefficients might not be a product
2487       // or constant, in which case we give up on this direction.
2488       continue;
2489     APInt ConstCoeff = Constant->getAPInt();
2490     RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2491     LLVM_DEBUG(dbgs() << "\tRunningGCD = " << RunningGCD << "\n");
2492     if (RunningGCD != 0) {
2493       Remainder = ConstDelta.srem(RunningGCD);
2494       LLVM_DEBUG(dbgs() << "\tRemainder = " << Remainder << "\n");
2495       if (Remainder != 0) {
2496         unsigned Level = mapSrcLoop(CurLoop);
2497         Result.DV[Level - 1].Direction &= unsigned(~Dependence::DVEntry::EQ);
2498         Improved = true;
2499       }
2500     }
2501   }
2502   if (Improved)
2503     ++GCDsuccesses;
2504   LLVM_DEBUG(dbgs() << "all done\n");
2505   return false;
2506 }
2507 
2508 
2509 //===----------------------------------------------------------------------===//
2510 // banerjeeMIVtest -
2511 // Use Banerjee's Inequalities to test an MIV subscript pair.
2512 // (Wolfe, in the race-car book, calls this the Extreme Value Test.)
2513 // Generally follows the discussion in Section 2.5.2 of
2514 //
2515 //    Optimizing Supercompilers for Supercomputers
2516 //    Michael Wolfe
2517 //
2518 // The inequalities given on page 25 are simplified in that loops are
2519 // normalized so that the lower bound is always 0 and the stride is always 1.
2520 // For example, Wolfe gives
2521 //
2522 //     LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2523 //
2524 // where A_k is the coefficient of the kth index in the source subscript,
2525 // B_k is the coefficient of the kth index in the destination subscript,
2526 // U_k is the upper bound of the kth index, L_k is the lower bound of the Kth
2527 // index, and N_k is the stride of the kth index. Since all loops are normalized
2528 // by the SCEV package, N_k = 1 and L_k = 0, allowing us to simplify the
2529 // equation to
2530 //
2531 //     LB^<_k = (A^-_k - B_k)^- (U_k - 0 - 1) + (A_k - B_k)0 - B_k 1
2532 //            = (A^-_k - B_k)^- (U_k - 1)  - B_k
2533 //
2534 // Similar simplifications are possible for the other equations.
2535 //
2536 // When we can't determine the number of iterations for a loop,
2537 // we use NULL as an indicator for the worst case, infinity.
2538 // When computing the upper bound, NULL denotes +inf;
2539 // for the lower bound, NULL denotes -inf.
2540 //
2541 // Return true if dependence disproved.
2542 bool DependenceInfo::banerjeeMIVtest(const SCEV *Src, const SCEV *Dst,
2543                                      const SmallBitVector &Loops,
2544                                      FullDependence &Result) const {
2545   LLVM_DEBUG(dbgs() << "starting Banerjee\n");
2546   ++BanerjeeApplications;
2547   LLVM_DEBUG(dbgs() << "    Src = " << *Src << '\n');
2548   const SCEV *A0;
2549   CoefficientInfo *A = collectCoeffInfo(Src, true, A0);
2550   LLVM_DEBUG(dbgs() << "    Dst = " << *Dst << '\n');
2551   const SCEV *B0;
2552   CoefficientInfo *B = collectCoeffInfo(Dst, false, B0);
2553   BoundInfo *Bound = new BoundInfo[MaxLevels + 1];
2554   const SCEV *Delta = SE->getMinusSCEV(B0, A0);
2555   LLVM_DEBUG(dbgs() << "\tDelta = " << *Delta << '\n');
2556 
2557   // Compute bounds for all the * directions.
2558   LLVM_DEBUG(dbgs() << "\tBounds[*]\n");
2559   for (unsigned K = 1; K <= MaxLevels; ++K) {
2560     Bound[K].Iterations = A[K].Iterations ? A[K].Iterations : B[K].Iterations;
2561     Bound[K].Direction = Dependence::DVEntry::ALL;
2562     Bound[K].DirSet = Dependence::DVEntry::NONE;
2563     findBoundsALL(A, B, Bound, K);
2564 #ifndef NDEBUG
2565     LLVM_DEBUG(dbgs() << "\t    " << K << '\t');
2566     if (Bound[K].Lower[Dependence::DVEntry::ALL])
2567       LLVM_DEBUG(dbgs() << *Bound[K].Lower[Dependence::DVEntry::ALL] << '\t');
2568     else
2569       LLVM_DEBUG(dbgs() << "-inf\t");
2570     if (Bound[K].Upper[Dependence::DVEntry::ALL])
2571       LLVM_DEBUG(dbgs() << *Bound[K].Upper[Dependence::DVEntry::ALL] << '\n');
2572     else
2573       LLVM_DEBUG(dbgs() << "+inf\n");
2574 #endif
2575   }
2576 
2577   // Test the *, *, *, ... case.
2578   bool Disproved = false;
2579   if (testBounds(Dependence::DVEntry::ALL, 0, Bound, Delta)) {
2580     // Explore the direction vector hierarchy.
2581     unsigned DepthExpanded = 0;
2582     unsigned NewDeps = exploreDirections(1, A, B, Bound,
2583                                          Loops, DepthExpanded, Delta);
2584     if (NewDeps > 0) {
2585       bool Improved = false;
2586       for (unsigned K = 1; K <= CommonLevels; ++K) {
2587         if (Loops[K]) {
2588           unsigned Old = Result.DV[K - 1].Direction;
2589           Result.DV[K - 1].Direction = Old & Bound[K].DirSet;
2590           Improved |= Old != Result.DV[K - 1].Direction;
2591           if (!Result.DV[K - 1].Direction) {
2592             Improved = false;
2593             Disproved = true;
2594             break;
2595           }
2596         }
2597       }
2598       if (Improved)
2599         ++BanerjeeSuccesses;
2600     }
2601     else {
2602       ++BanerjeeIndependence;
2603       Disproved = true;
2604     }
2605   }
2606   else {
2607     ++BanerjeeIndependence;
2608     Disproved = true;
2609   }
2610   delete [] Bound;
2611   delete [] A;
2612   delete [] B;
2613   return Disproved;
2614 }
2615 
2616 
2617 // Hierarchically expands the direction vector
2618 // search space, combining the directions of discovered dependences
2619 // in the DirSet field of Bound. Returns the number of distinct
2620 // dependences discovered. If the dependence is disproved,
2621 // it will return 0.
2622 unsigned DependenceInfo::exploreDirections(unsigned Level, CoefficientInfo *A,
2623                                            CoefficientInfo *B, BoundInfo *Bound,
2624                                            const SmallBitVector &Loops,
2625                                            unsigned &DepthExpanded,
2626                                            const SCEV *Delta) const {
2627   // This algorithm has worst case complexity of O(3^n), where 'n' is the number
2628   // of common loop levels. To avoid excessive compile-time, pessimize all the
2629   // results and immediately return when the number of common levels is beyond
2630   // the given threshold.
2631   if (CommonLevels > MIVMaxLevelThreshold) {
2632     LLVM_DEBUG(dbgs() << "Number of common levels exceeded the threshold. MIV "
2633                          "direction exploration is terminated.\n");
2634     for (unsigned K = 1; K <= CommonLevels; ++K)
2635       if (Loops[K])
2636         Bound[K].DirSet = Dependence::DVEntry::ALL;
2637     return 1;
2638   }
2639 
2640   if (Level > CommonLevels) {
2641     // record result
2642     LLVM_DEBUG(dbgs() << "\t[");
2643     for (unsigned K = 1; K <= CommonLevels; ++K) {
2644       if (Loops[K]) {
2645         Bound[K].DirSet |= Bound[K].Direction;
2646 #ifndef NDEBUG
2647         switch (Bound[K].Direction) {
2648         case Dependence::DVEntry::LT:
2649           LLVM_DEBUG(dbgs() << " <");
2650           break;
2651         case Dependence::DVEntry::EQ:
2652           LLVM_DEBUG(dbgs() << " =");
2653           break;
2654         case Dependence::DVEntry::GT:
2655           LLVM_DEBUG(dbgs() << " >");
2656           break;
2657         case Dependence::DVEntry::ALL:
2658           LLVM_DEBUG(dbgs() << " *");
2659           break;
2660         default:
2661           llvm_unreachable("unexpected Bound[K].Direction");
2662         }
2663 #endif
2664       }
2665     }
2666     LLVM_DEBUG(dbgs() << " ]\n");
2667     return 1;
2668   }
2669   if (Loops[Level]) {
2670     if (Level > DepthExpanded) {
2671       DepthExpanded = Level;
2672       // compute bounds for <, =, > at current level
2673       findBoundsLT(A, B, Bound, Level);
2674       findBoundsGT(A, B, Bound, Level);
2675       findBoundsEQ(A, B, Bound, Level);
2676 #ifndef NDEBUG
2677       LLVM_DEBUG(dbgs() << "\tBound for level = " << Level << '\n');
2678       LLVM_DEBUG(dbgs() << "\t    <\t");
2679       if (Bound[Level].Lower[Dependence::DVEntry::LT])
2680         LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::LT]
2681                           << '\t');
2682       else
2683         LLVM_DEBUG(dbgs() << "-inf\t");
2684       if (Bound[Level].Upper[Dependence::DVEntry::LT])
2685         LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::LT]
2686                           << '\n');
2687       else
2688         LLVM_DEBUG(dbgs() << "+inf\n");
2689       LLVM_DEBUG(dbgs() << "\t    =\t");
2690       if (Bound[Level].Lower[Dependence::DVEntry::EQ])
2691         LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::EQ]
2692                           << '\t');
2693       else
2694         LLVM_DEBUG(dbgs() << "-inf\t");
2695       if (Bound[Level].Upper[Dependence::DVEntry::EQ])
2696         LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::EQ]
2697                           << '\n');
2698       else
2699         LLVM_DEBUG(dbgs() << "+inf\n");
2700       LLVM_DEBUG(dbgs() << "\t    >\t");
2701       if (Bound[Level].Lower[Dependence::DVEntry::GT])
2702         LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::GT]
2703                           << '\t');
2704       else
2705         LLVM_DEBUG(dbgs() << "-inf\t");
2706       if (Bound[Level].Upper[Dependence::DVEntry::GT])
2707         LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::GT]
2708                           << '\n');
2709       else
2710         LLVM_DEBUG(dbgs() << "+inf\n");
2711 #endif
2712     }
2713 
2714     unsigned NewDeps = 0;
2715 
2716     // test bounds for <, *, *, ...
2717     if (testBounds(Dependence::DVEntry::LT, Level, Bound, Delta))
2718       NewDeps += exploreDirections(Level + 1, A, B, Bound,
2719                                    Loops, DepthExpanded, Delta);
2720 
2721     // Test bounds for =, *, *, ...
2722     if (testBounds(Dependence::DVEntry::EQ, Level, Bound, Delta))
2723       NewDeps += exploreDirections(Level + 1, A, B, Bound,
2724                                    Loops, DepthExpanded, Delta);
2725 
2726     // test bounds for >, *, *, ...
2727     if (testBounds(Dependence::DVEntry::GT, Level, Bound, Delta))
2728       NewDeps += exploreDirections(Level + 1, A, B, Bound,
2729                                    Loops, DepthExpanded, Delta);
2730 
2731     Bound[Level].Direction = Dependence::DVEntry::ALL;
2732     return NewDeps;
2733   }
2734   else
2735     return exploreDirections(Level + 1, A, B, Bound, Loops, DepthExpanded, Delta);
2736 }
2737 
2738 
2739 // Returns true iff the current bounds are plausible.
2740 bool DependenceInfo::testBounds(unsigned char DirKind, unsigned Level,
2741                                 BoundInfo *Bound, const SCEV *Delta) const {
2742   Bound[Level].Direction = DirKind;
2743   if (const SCEV *LowerBound = getLowerBound(Bound))
2744     if (isKnownPredicate(CmpInst::ICMP_SGT, LowerBound, Delta))
2745       return false;
2746   if (const SCEV *UpperBound = getUpperBound(Bound))
2747     if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, UpperBound))
2748       return false;
2749   return true;
2750 }
2751 
2752 
2753 // Computes the upper and lower bounds for level K
2754 // using the * direction. Records them in Bound.
2755 // Wolfe gives the equations
2756 //
2757 //    LB^*_k = (A^-_k - B^+_k)(U_k - L_k) + (A_k - B_k)L_k
2758 //    UB^*_k = (A^+_k - B^-_k)(U_k - L_k) + (A_k - B_k)L_k
2759 //
2760 // Since we normalize loops, we can simplify these equations to
2761 //
2762 //    LB^*_k = (A^-_k - B^+_k)U_k
2763 //    UB^*_k = (A^+_k - B^-_k)U_k
2764 //
2765 // We must be careful to handle the case where the upper bound is unknown.
2766 // Note that the lower bound is always <= 0
2767 // and the upper bound is always >= 0.
2768 void DependenceInfo::findBoundsALL(CoefficientInfo *A, CoefficientInfo *B,
2769                                    BoundInfo *Bound, unsigned K) const {
2770   Bound[K].Lower[Dependence::DVEntry::ALL] = nullptr; // Default value = -infinity.
2771   Bound[K].Upper[Dependence::DVEntry::ALL] = nullptr; // Default value = +infinity.
2772   if (Bound[K].Iterations) {
2773     Bound[K].Lower[Dependence::DVEntry::ALL] =
2774       SE->getMulExpr(SE->getMinusSCEV(A[K].NegPart, B[K].PosPart),
2775                      Bound[K].Iterations);
2776     Bound[K].Upper[Dependence::DVEntry::ALL] =
2777       SE->getMulExpr(SE->getMinusSCEV(A[K].PosPart, B[K].NegPart),
2778                      Bound[K].Iterations);
2779   }
2780   else {
2781     // If the difference is 0, we won't need to know the number of iterations.
2782     if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].NegPart, B[K].PosPart))
2783       Bound[K].Lower[Dependence::DVEntry::ALL] =
2784           SE->getZero(A[K].Coeff->getType());
2785     if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].PosPart, B[K].NegPart))
2786       Bound[K].Upper[Dependence::DVEntry::ALL] =
2787           SE->getZero(A[K].Coeff->getType());
2788   }
2789 }
2790 
2791 
2792 // Computes the upper and lower bounds for level K
2793 // using the = direction. Records them in Bound.
2794 // Wolfe gives the equations
2795 //
2796 //    LB^=_k = (A_k - B_k)^- (U_k - L_k) + (A_k - B_k)L_k
2797 //    UB^=_k = (A_k - B_k)^+ (U_k - L_k) + (A_k - B_k)L_k
2798 //
2799 // Since we normalize loops, we can simplify these equations to
2800 //
2801 //    LB^=_k = (A_k - B_k)^- U_k
2802 //    UB^=_k = (A_k - B_k)^+ U_k
2803 //
2804 // We must be careful to handle the case where the upper bound is unknown.
2805 // Note that the lower bound is always <= 0
2806 // and the upper bound is always >= 0.
2807 void DependenceInfo::findBoundsEQ(CoefficientInfo *A, CoefficientInfo *B,
2808                                   BoundInfo *Bound, unsigned K) const {
2809   Bound[K].Lower[Dependence::DVEntry::EQ] = nullptr; // Default value = -infinity.
2810   Bound[K].Upper[Dependence::DVEntry::EQ] = nullptr; // Default value = +infinity.
2811   if (Bound[K].Iterations) {
2812     const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff);
2813     const SCEV *NegativePart = getNegativePart(Delta);
2814     Bound[K].Lower[Dependence::DVEntry::EQ] =
2815       SE->getMulExpr(NegativePart, Bound[K].Iterations);
2816     const SCEV *PositivePart = getPositivePart(Delta);
2817     Bound[K].Upper[Dependence::DVEntry::EQ] =
2818       SE->getMulExpr(PositivePart, Bound[K].Iterations);
2819   }
2820   else {
2821     // If the positive/negative part of the difference is 0,
2822     // we won't need to know the number of iterations.
2823     const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff);
2824     const SCEV *NegativePart = getNegativePart(Delta);
2825     if (NegativePart->isZero())
2826       Bound[K].Lower[Dependence::DVEntry::EQ] = NegativePart; // Zero
2827     const SCEV *PositivePart = getPositivePart(Delta);
2828     if (PositivePart->isZero())
2829       Bound[K].Upper[Dependence::DVEntry::EQ] = PositivePart; // Zero
2830   }
2831 }
2832 
2833 
2834 // Computes the upper and lower bounds for level K
2835 // using the < direction. Records them in Bound.
2836 // Wolfe gives the equations
2837 //
2838 //    LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2839 //    UB^<_k = (A^+_k - B_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2840 //
2841 // Since we normalize loops, we can simplify these equations to
2842 //
2843 //    LB^<_k = (A^-_k - B_k)^- (U_k - 1) - B_k
2844 //    UB^<_k = (A^+_k - B_k)^+ (U_k - 1) - B_k
2845 //
2846 // We must be careful to handle the case where the upper bound is unknown.
2847 void DependenceInfo::findBoundsLT(CoefficientInfo *A, CoefficientInfo *B,
2848                                   BoundInfo *Bound, unsigned K) const {
2849   Bound[K].Lower[Dependence::DVEntry::LT] = nullptr; // Default value = -infinity.
2850   Bound[K].Upper[Dependence::DVEntry::LT] = nullptr; // Default value = +infinity.
2851   if (Bound[K].Iterations) {
2852     const SCEV *Iter_1 = SE->getMinusSCEV(
2853         Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType()));
2854     const SCEV *NegPart =
2855       getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff));
2856     Bound[K].Lower[Dependence::DVEntry::LT] =
2857       SE->getMinusSCEV(SE->getMulExpr(NegPart, Iter_1), B[K].Coeff);
2858     const SCEV *PosPart =
2859       getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff));
2860     Bound[K].Upper[Dependence::DVEntry::LT] =
2861       SE->getMinusSCEV(SE->getMulExpr(PosPart, Iter_1), B[K].Coeff);
2862   }
2863   else {
2864     // If the positive/negative part of the difference is 0,
2865     // we won't need to know the number of iterations.
2866     const SCEV *NegPart =
2867       getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff));
2868     if (NegPart->isZero())
2869       Bound[K].Lower[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff);
2870     const SCEV *PosPart =
2871       getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff));
2872     if (PosPart->isZero())
2873       Bound[K].Upper[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff);
2874   }
2875 }
2876 
2877 
2878 // Computes the upper and lower bounds for level K
2879 // using the > direction. Records them in Bound.
2880 // Wolfe gives the equations
2881 //
2882 //    LB^>_k = (A_k - B^+_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k
2883 //    UB^>_k = (A_k - B^-_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k
2884 //
2885 // Since we normalize loops, we can simplify these equations to
2886 //
2887 //    LB^>_k = (A_k - B^+_k)^- (U_k - 1) + A_k
2888 //    UB^>_k = (A_k - B^-_k)^+ (U_k - 1) + A_k
2889 //
2890 // We must be careful to handle the case where the upper bound is unknown.
2891 void DependenceInfo::findBoundsGT(CoefficientInfo *A, CoefficientInfo *B,
2892                                   BoundInfo *Bound, unsigned K) const {
2893   Bound[K].Lower[Dependence::DVEntry::GT] = nullptr; // Default value = -infinity.
2894   Bound[K].Upper[Dependence::DVEntry::GT] = nullptr; // Default value = +infinity.
2895   if (Bound[K].Iterations) {
2896     const SCEV *Iter_1 = SE->getMinusSCEV(
2897         Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType()));
2898     const SCEV *NegPart =
2899       getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart));
2900     Bound[K].Lower[Dependence::DVEntry::GT] =
2901       SE->getAddExpr(SE->getMulExpr(NegPart, Iter_1), A[K].Coeff);
2902     const SCEV *PosPart =
2903       getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart));
2904     Bound[K].Upper[Dependence::DVEntry::GT] =
2905       SE->getAddExpr(SE->getMulExpr(PosPart, Iter_1), A[K].Coeff);
2906   }
2907   else {
2908     // If the positive/negative part of the difference is 0,
2909     // we won't need to know the number of iterations.
2910     const SCEV *NegPart = getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart));
2911     if (NegPart->isZero())
2912       Bound[K].Lower[Dependence::DVEntry::GT] = A[K].Coeff;
2913     const SCEV *PosPart = getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart));
2914     if (PosPart->isZero())
2915       Bound[K].Upper[Dependence::DVEntry::GT] = A[K].Coeff;
2916   }
2917 }
2918 
2919 
2920 // X^+ = max(X, 0)
2921 const SCEV *DependenceInfo::getPositivePart(const SCEV *X) const {
2922   return SE->getSMaxExpr(X, SE->getZero(X->getType()));
2923 }
2924 
2925 
2926 // X^- = min(X, 0)
2927 const SCEV *DependenceInfo::getNegativePart(const SCEV *X) const {
2928   return SE->getSMinExpr(X, SE->getZero(X->getType()));
2929 }
2930 
2931 
2932 // Walks through the subscript,
2933 // collecting each coefficient, the associated loop bounds,
2934 // and recording its positive and negative parts for later use.
2935 DependenceInfo::CoefficientInfo *
2936 DependenceInfo::collectCoeffInfo(const SCEV *Subscript, bool SrcFlag,
2937                                  const SCEV *&Constant) const {
2938   const SCEV *Zero = SE->getZero(Subscript->getType());
2939   CoefficientInfo *CI = new CoefficientInfo[MaxLevels + 1];
2940   for (unsigned K = 1; K <= MaxLevels; ++K) {
2941     CI[K].Coeff = Zero;
2942     CI[K].PosPart = Zero;
2943     CI[K].NegPart = Zero;
2944     CI[K].Iterations = nullptr;
2945   }
2946   while (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Subscript)) {
2947     const Loop *L = AddRec->getLoop();
2948     unsigned K = SrcFlag ? mapSrcLoop(L) : mapDstLoop(L);
2949     CI[K].Coeff = AddRec->getStepRecurrence(*SE);
2950     CI[K].PosPart = getPositivePart(CI[K].Coeff);
2951     CI[K].NegPart = getNegativePart(CI[K].Coeff);
2952     CI[K].Iterations = collectUpperBound(L, Subscript->getType());
2953     Subscript = AddRec->getStart();
2954   }
2955   Constant = Subscript;
2956 #ifndef NDEBUG
2957   LLVM_DEBUG(dbgs() << "\tCoefficient Info\n");
2958   for (unsigned K = 1; K <= MaxLevels; ++K) {
2959     LLVM_DEBUG(dbgs() << "\t    " << K << "\t" << *CI[K].Coeff);
2960     LLVM_DEBUG(dbgs() << "\tPos Part = ");
2961     LLVM_DEBUG(dbgs() << *CI[K].PosPart);
2962     LLVM_DEBUG(dbgs() << "\tNeg Part = ");
2963     LLVM_DEBUG(dbgs() << *CI[K].NegPart);
2964     LLVM_DEBUG(dbgs() << "\tUpper Bound = ");
2965     if (CI[K].Iterations)
2966       LLVM_DEBUG(dbgs() << *CI[K].Iterations);
2967     else
2968       LLVM_DEBUG(dbgs() << "+inf");
2969     LLVM_DEBUG(dbgs() << '\n');
2970   }
2971   LLVM_DEBUG(dbgs() << "\t    Constant = " << *Subscript << '\n');
2972 #endif
2973   return CI;
2974 }
2975 
2976 
2977 // Looks through all the bounds info and
2978 // computes the lower bound given the current direction settings
2979 // at each level. If the lower bound for any level is -inf,
2980 // the result is -inf.
2981 const SCEV *DependenceInfo::getLowerBound(BoundInfo *Bound) const {
2982   const SCEV *Sum = Bound[1].Lower[Bound[1].Direction];
2983   for (unsigned K = 2; Sum && K <= MaxLevels; ++K) {
2984     if (Bound[K].Lower[Bound[K].Direction])
2985       Sum = SE->getAddExpr(Sum, Bound[K].Lower[Bound[K].Direction]);
2986     else
2987       Sum = nullptr;
2988   }
2989   return Sum;
2990 }
2991 
2992 
2993 // Looks through all the bounds info and
2994 // computes the upper bound given the current direction settings
2995 // at each level. If the upper bound at any level is +inf,
2996 // the result is +inf.
2997 const SCEV *DependenceInfo::getUpperBound(BoundInfo *Bound) const {
2998   const SCEV *Sum = Bound[1].Upper[Bound[1].Direction];
2999   for (unsigned K = 2; Sum && K <= MaxLevels; ++K) {
3000     if (Bound[K].Upper[Bound[K].Direction])
3001       Sum = SE->getAddExpr(Sum, Bound[K].Upper[Bound[K].Direction]);
3002     else
3003       Sum = nullptr;
3004   }
3005   return Sum;
3006 }
3007 
3008 
3009 //===----------------------------------------------------------------------===//
3010 // Constraint manipulation for Delta test.
3011 
3012 // Given a linear SCEV,
3013 // return the coefficient (the step)
3014 // corresponding to the specified loop.
3015 // If there isn't one, return 0.
3016 // For example, given a*i + b*j + c*k, finding the coefficient
3017 // corresponding to the j loop would yield b.
3018 const SCEV *DependenceInfo::findCoefficient(const SCEV *Expr,
3019                                             const Loop *TargetLoop) const {
3020   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
3021   if (!AddRec)
3022     return SE->getZero(Expr->getType());
3023   if (AddRec->getLoop() == TargetLoop)
3024     return AddRec->getStepRecurrence(*SE);
3025   return findCoefficient(AddRec->getStart(), TargetLoop);
3026 }
3027 
3028 
3029 // Given a linear SCEV,
3030 // return the SCEV given by zeroing out the coefficient
3031 // corresponding to the specified loop.
3032 // For example, given a*i + b*j + c*k, zeroing the coefficient
3033 // corresponding to the j loop would yield a*i + c*k.
3034 const SCEV *DependenceInfo::zeroCoefficient(const SCEV *Expr,
3035                                             const Loop *TargetLoop) const {
3036   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
3037   if (!AddRec)
3038     return Expr; // ignore
3039   if (AddRec->getLoop() == TargetLoop)
3040     return AddRec->getStart();
3041   return SE->getAddRecExpr(zeroCoefficient(AddRec->getStart(), TargetLoop),
3042                            AddRec->getStepRecurrence(*SE),
3043                            AddRec->getLoop(),
3044                            AddRec->getNoWrapFlags());
3045 }
3046 
3047 
3048 // Given a linear SCEV Expr,
3049 // return the SCEV given by adding some Value to the
3050 // coefficient corresponding to the specified TargetLoop.
3051 // For example, given a*i + b*j + c*k, adding 1 to the coefficient
3052 // corresponding to the j loop would yield a*i + (b+1)*j + c*k.
3053 const SCEV *DependenceInfo::addToCoefficient(const SCEV *Expr,
3054                                              const Loop *TargetLoop,
3055                                              const SCEV *Value) const {
3056   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
3057   if (!AddRec) // create a new addRec
3058     return SE->getAddRecExpr(Expr,
3059                              Value,
3060                              TargetLoop,
3061                              SCEV::FlagAnyWrap); // Worst case, with no info.
3062   if (AddRec->getLoop() == TargetLoop) {
3063     const SCEV *Sum = SE->getAddExpr(AddRec->getStepRecurrence(*SE), Value);
3064     if (Sum->isZero())
3065       return AddRec->getStart();
3066     return SE->getAddRecExpr(AddRec->getStart(),
3067                              Sum,
3068                              AddRec->getLoop(),
3069                              AddRec->getNoWrapFlags());
3070   }
3071   if (SE->isLoopInvariant(AddRec, TargetLoop))
3072     return SE->getAddRecExpr(AddRec, Value, TargetLoop, SCEV::FlagAnyWrap);
3073   return SE->getAddRecExpr(
3074       addToCoefficient(AddRec->getStart(), TargetLoop, Value),
3075       AddRec->getStepRecurrence(*SE), AddRec->getLoop(),
3076       AddRec->getNoWrapFlags());
3077 }
3078 
3079 
3080 // Review the constraints, looking for opportunities
3081 // to simplify a subscript pair (Src and Dst).
3082 // Return true if some simplification occurs.
3083 // If the simplification isn't exact (that is, if it is conservative
3084 // in terms of dependence), set consistent to false.
3085 // Corresponds to Figure 5 from the paper
3086 //
3087 //            Practical Dependence Testing
3088 //            Goff, Kennedy, Tseng
3089 //            PLDI 1991
3090 bool DependenceInfo::propagate(const SCEV *&Src, const SCEV *&Dst,
3091                                SmallBitVector &Loops,
3092                                SmallVectorImpl<Constraint> &Constraints,
3093                                bool &Consistent) {
3094   bool Result = false;
3095   for (unsigned LI : Loops.set_bits()) {
3096     LLVM_DEBUG(dbgs() << "\t    Constraint[" << LI << "] is");
3097     LLVM_DEBUG(Constraints[LI].dump(dbgs()));
3098     if (Constraints[LI].isDistance())
3099       Result |= propagateDistance(Src, Dst, Constraints[LI], Consistent);
3100     else if (Constraints[LI].isLine())
3101       Result |= propagateLine(Src, Dst, Constraints[LI], Consistent);
3102     else if (Constraints[LI].isPoint())
3103       Result |= propagatePoint(Src, Dst, Constraints[LI]);
3104   }
3105   return Result;
3106 }
3107 
3108 
3109 // Attempt to propagate a distance
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.
3114 bool DependenceInfo::propagateDistance(const SCEV *&Src, const SCEV *&Dst,
3115                                        Constraint &CurConstraint,
3116                                        bool &Consistent) {
3117   const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3118   LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n");
3119   const SCEV *A_K = findCoefficient(Src, CurLoop);
3120   if (A_K->isZero())
3121     return false;
3122   const SCEV *DA_K = SE->getMulExpr(A_K, CurConstraint.getD());
3123   Src = SE->getMinusSCEV(Src, DA_K);
3124   Src = zeroCoefficient(Src, CurLoop);
3125   LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n");
3126   LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n");
3127   Dst = addToCoefficient(Dst, CurLoop, SE->getNegativeSCEV(A_K));
3128   LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n");
3129   if (!findCoefficient(Dst, CurLoop)->isZero())
3130     Consistent = false;
3131   return true;
3132 }
3133 
3134 
3135 // Attempt to propagate a line
3136 // constraint into a subscript pair (Src and Dst).
3137 // Return true if some simplification occurs.
3138 // If the simplification isn't exact (that is, if it is conservative
3139 // in terms of dependence), set consistent to false.
3140 bool DependenceInfo::propagateLine(const SCEV *&Src, const SCEV *&Dst,
3141                                    Constraint &CurConstraint,
3142                                    bool &Consistent) {
3143   const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3144   const SCEV *A = CurConstraint.getA();
3145   const SCEV *B = CurConstraint.getB();
3146   const SCEV *C = CurConstraint.getC();
3147   LLVM_DEBUG(dbgs() << "\t\tA = " << *A << ", B = " << *B << ", C = " << *C
3148                     << "\n");
3149   LLVM_DEBUG(dbgs() << "\t\tSrc = " << *Src << "\n");
3150   LLVM_DEBUG(dbgs() << "\t\tDst = " << *Dst << "\n");
3151   if (A->isZero()) {
3152     const SCEVConstant *Bconst = dyn_cast<SCEVConstant>(B);
3153     const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3154     if (!Bconst || !Cconst) return false;
3155     APInt Beta = Bconst->getAPInt();
3156     APInt Charlie = Cconst->getAPInt();
3157     APInt CdivB = Charlie.sdiv(Beta);
3158     assert(Charlie.srem(Beta) == 0 && "C should be evenly divisible by B");
3159     const SCEV *AP_K = findCoefficient(Dst, CurLoop);
3160     //    Src = SE->getAddExpr(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB)));
3161     Src = SE->getMinusSCEV(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB)));
3162     Dst = zeroCoefficient(Dst, CurLoop);
3163     if (!findCoefficient(Src, CurLoop)->isZero())
3164       Consistent = false;
3165   }
3166   else if (B->isZero()) {
3167     const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A);
3168     const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3169     if (!Aconst || !Cconst) return false;
3170     APInt Alpha = Aconst->getAPInt();
3171     APInt Charlie = Cconst->getAPInt();
3172     APInt CdivA = Charlie.sdiv(Alpha);
3173     assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A");
3174     const SCEV *A_K = findCoefficient(Src, CurLoop);
3175     Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA)));
3176     Src = zeroCoefficient(Src, CurLoop);
3177     if (!findCoefficient(Dst, CurLoop)->isZero())
3178       Consistent = false;
3179   }
3180   else if (isKnownPredicate(CmpInst::ICMP_EQ, A, B)) {
3181     const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A);
3182     const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3183     if (!Aconst || !Cconst) return false;
3184     APInt Alpha = Aconst->getAPInt();
3185     APInt Charlie = Cconst->getAPInt();
3186     APInt CdivA = Charlie.sdiv(Alpha);
3187     assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A");
3188     const SCEV *A_K = findCoefficient(Src, CurLoop);
3189     Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA)));
3190     Src = zeroCoefficient(Src, CurLoop);
3191     Dst = addToCoefficient(Dst, CurLoop, A_K);
3192     if (!findCoefficient(Dst, CurLoop)->isZero())
3193       Consistent = false;
3194   }
3195   else {
3196     // paper is incorrect here, or perhaps just misleading
3197     const SCEV *A_K = findCoefficient(Src, CurLoop);
3198     Src = SE->getMulExpr(Src, A);
3199     Dst = SE->getMulExpr(Dst, A);
3200     Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, C));
3201     Src = zeroCoefficient(Src, CurLoop);
3202     Dst = addToCoefficient(Dst, CurLoop, SE->getMulExpr(A_K, B));
3203     if (!findCoefficient(Dst, CurLoop)->isZero())
3204       Consistent = false;
3205   }
3206   LLVM_DEBUG(dbgs() << "\t\tnew Src = " << *Src << "\n");
3207   LLVM_DEBUG(dbgs() << "\t\tnew Dst = " << *Dst << "\n");
3208   return true;
3209 }
3210 
3211 
3212 // Attempt to propagate a point
3213 // constraint into a subscript pair (Src and Dst).
3214 // Return true if some simplification occurs.
3215 bool DependenceInfo::propagatePoint(const SCEV *&Src, const SCEV *&Dst,
3216                                     Constraint &CurConstraint) {
3217   const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3218   const SCEV *A_K = findCoefficient(Src, CurLoop);
3219   const SCEV *AP_K = findCoefficient(Dst, CurLoop);
3220   const SCEV *XA_K = SE->getMulExpr(A_K, CurConstraint.getX());
3221   const SCEV *YAP_K = SE->getMulExpr(AP_K, CurConstraint.getY());
3222   LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n");
3223   Src = SE->getAddExpr(Src, SE->getMinusSCEV(XA_K, YAP_K));
3224   Src = zeroCoefficient(Src, CurLoop);
3225   LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n");
3226   LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n");
3227   Dst = zeroCoefficient(Dst, CurLoop);
3228   LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n");
3229   return true;
3230 }
3231 
3232 
3233 // Update direction vector entry based on the current constraint.
3234 void DependenceInfo::updateDirection(Dependence::DVEntry &Level,
3235                                      const Constraint &CurConstraint) const {
3236   LLVM_DEBUG(dbgs() << "\tUpdate direction, constraint =");
3237   LLVM_DEBUG(CurConstraint.dump(dbgs()));
3238   if (CurConstraint.isAny())
3239     ; // use defaults
3240   else if (CurConstraint.isDistance()) {
3241     // this one is consistent, the others aren't
3242     Level.Scalar = false;
3243     Level.Distance = CurConstraint.getD();
3244     unsigned NewDirection = Dependence::DVEntry::NONE;
3245     if (!SE->isKnownNonZero(Level.Distance)) // if may be zero
3246       NewDirection = Dependence::DVEntry::EQ;
3247     if (!SE->isKnownNonPositive(Level.Distance)) // if may be positive
3248       NewDirection |= Dependence::DVEntry::LT;
3249     if (!SE->isKnownNonNegative(Level.Distance)) // if may be negative
3250       NewDirection |= Dependence::DVEntry::GT;
3251     Level.Direction &= NewDirection;
3252   }
3253   else if (CurConstraint.isLine()) {
3254     Level.Scalar = false;
3255     Level.Distance = nullptr;
3256     // direction should be accurate
3257   }
3258   else if (CurConstraint.isPoint()) {
3259     Level.Scalar = false;
3260     Level.Distance = nullptr;
3261     unsigned NewDirection = Dependence::DVEntry::NONE;
3262     if (!isKnownPredicate(CmpInst::ICMP_NE,
3263                           CurConstraint.getY(),
3264                           CurConstraint.getX()))
3265       // if X may be = Y
3266       NewDirection |= Dependence::DVEntry::EQ;
3267     if (!isKnownPredicate(CmpInst::ICMP_SLE,
3268                           CurConstraint.getY(),
3269                           CurConstraint.getX()))
3270       // if Y may be > X
3271       NewDirection |= Dependence::DVEntry::LT;
3272     if (!isKnownPredicate(CmpInst::ICMP_SGE,
3273                           CurConstraint.getY(),
3274                           CurConstraint.getX()))
3275       // if Y may be < X
3276       NewDirection |= Dependence::DVEntry::GT;
3277     Level.Direction &= NewDirection;
3278   }
3279   else
3280     llvm_unreachable("constraint has unexpected kind");
3281 }
3282 
3283 /// Check if we can delinearize the subscripts. If the SCEVs representing the
3284 /// source and destination array references are recurrences on a nested loop,
3285 /// this function flattens the nested recurrences into separate recurrences
3286 /// for each loop level.
3287 bool DependenceInfo::tryDelinearize(Instruction *Src, Instruction *Dst,
3288                                     SmallVectorImpl<Subscript> &Pair) {
3289   assert(isLoadOrStore(Src) && "instruction is not load or store");
3290   assert(isLoadOrStore(Dst) && "instruction is not load or store");
3291   Value *SrcPtr = getLoadStorePointerOperand(Src);
3292   Value *DstPtr = getLoadStorePointerOperand(Dst);
3293   Loop *SrcLoop = LI->getLoopFor(Src->getParent());
3294   Loop *DstLoop = LI->getLoopFor(Dst->getParent());
3295   const SCEV *SrcAccessFn = SE->getSCEVAtScope(SrcPtr, SrcLoop);
3296   const SCEV *DstAccessFn = SE->getSCEVAtScope(DstPtr, DstLoop);
3297   const SCEVUnknown *SrcBase =
3298       dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
3299   const SCEVUnknown *DstBase =
3300       dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
3301 
3302   if (!SrcBase || !DstBase || SrcBase != DstBase)
3303     return false;
3304 
3305   SmallVector<const SCEV *, 4> SrcSubscripts, DstSubscripts;
3306 
3307   if (!tryDelinearizeFixedSize(Src, Dst, SrcAccessFn, DstAccessFn,
3308                                SrcSubscripts, DstSubscripts) &&
3309       !tryDelinearizeParametricSize(Src, Dst, SrcAccessFn, DstAccessFn,
3310                                     SrcSubscripts, DstSubscripts))
3311     return false;
3312 
3313   int Size = SrcSubscripts.size();
3314   LLVM_DEBUG({
3315     dbgs() << "\nSrcSubscripts: ";
3316     for (int I = 0; I < Size; I++)
3317       dbgs() << *SrcSubscripts[I];
3318     dbgs() << "\nDstSubscripts: ";
3319     for (int I = 0; I < Size; I++)
3320       dbgs() << *DstSubscripts[I];
3321   });
3322 
3323   // The delinearization transforms a single-subscript MIV dependence test into
3324   // a multi-subscript SIV dependence test that is easier to compute. So we
3325   // resize Pair to contain as many pairs of subscripts as the delinearization
3326   // has found, and then initialize the pairs following the delinearization.
3327   Pair.resize(Size);
3328   for (int I = 0; I < Size; ++I) {
3329     Pair[I].Src = SrcSubscripts[I];
3330     Pair[I].Dst = DstSubscripts[I];
3331     unifySubscriptType(&Pair[I]);
3332   }
3333 
3334   return true;
3335 }
3336 
3337 /// Try to delinearize \p SrcAccessFn and \p DstAccessFn if the underlying
3338 /// arrays accessed are fixed-size arrays. Return true if delinearization was
3339 /// successful.
3340 bool DependenceInfo::tryDelinearizeFixedSize(
3341     Instruction *Src, Instruction *Dst, const SCEV *SrcAccessFn,
3342     const SCEV *DstAccessFn, SmallVectorImpl<const SCEV *> &SrcSubscripts,
3343     SmallVectorImpl<const SCEV *> &DstSubscripts) {
3344   LLVM_DEBUG({
3345     const SCEVUnknown *SrcBase =
3346         dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
3347     const SCEVUnknown *DstBase =
3348         dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
3349     assert(SrcBase && DstBase && SrcBase == DstBase &&
3350            "expected src and dst scev unknowns to be equal");
3351     });
3352 
3353   SmallVector<int, 4> SrcSizes;
3354   SmallVector<int, 4> DstSizes;
3355   if (!tryDelinearizeFixedSizeImpl(SE, Src, SrcAccessFn, SrcSubscripts,
3356                                    SrcSizes) ||
3357       !tryDelinearizeFixedSizeImpl(SE, Dst, DstAccessFn, DstSubscripts,
3358                                    DstSizes))
3359     return false;
3360 
3361   // Check that the two size arrays are non-empty and equal in length and
3362   // value.
3363   if (SrcSizes.size() != DstSizes.size() ||
3364       !std::equal(SrcSizes.begin(), SrcSizes.end(), DstSizes.begin())) {
3365     SrcSubscripts.clear();
3366     DstSubscripts.clear();
3367     return false;
3368   }
3369 
3370   assert(SrcSubscripts.size() == DstSubscripts.size() &&
3371          "Expected equal number of entries in the list of SrcSubscripts and "
3372          "DstSubscripts.");
3373 
3374   Value *SrcPtr = getLoadStorePointerOperand(Src);
3375   Value *DstPtr = getLoadStorePointerOperand(Dst);
3376 
3377   // In general we cannot safely assume that the subscripts recovered from GEPs
3378   // are in the range of values defined for their corresponding array
3379   // dimensions. For example some C language usage/interpretation make it
3380   // impossible to verify this at compile-time. As such we can only delinearize
3381   // iff the subscripts are positive and are less than the range of the
3382   // dimension.
3383   if (!DisableDelinearizationChecks) {
3384     auto AllIndiciesInRange = [&](SmallVector<int, 4> &DimensionSizes,
3385                                   SmallVectorImpl<const SCEV *> &Subscripts,
3386                                   Value *Ptr) {
3387       size_t SSize = Subscripts.size();
3388       for (size_t I = 1; I < SSize; ++I) {
3389         const SCEV *S = Subscripts[I];
3390         if (!isKnownNonNegative(S, Ptr))
3391           return false;
3392         if (auto *SType = dyn_cast<IntegerType>(S->getType())) {
3393           const SCEV *Range = SE->getConstant(
3394               ConstantInt::get(SType, DimensionSizes[I - 1], false));
3395           if (!isKnownLessThan(S, Range))
3396             return false;
3397         }
3398       }
3399       return true;
3400     };
3401 
3402     if (!AllIndiciesInRange(SrcSizes, SrcSubscripts, SrcPtr) ||
3403         !AllIndiciesInRange(DstSizes, DstSubscripts, DstPtr)) {
3404       SrcSubscripts.clear();
3405       DstSubscripts.clear();
3406       return false;
3407     }
3408   }
3409   LLVM_DEBUG({
3410     dbgs() << "Delinearized subscripts of fixed-size array\n"
3411            << "SrcGEP:" << *SrcPtr << "\n"
3412            << "DstGEP:" << *DstPtr << "\n";
3413   });
3414   return true;
3415 }
3416 
3417 bool DependenceInfo::tryDelinearizeParametricSize(
3418     Instruction *Src, Instruction *Dst, const SCEV *SrcAccessFn,
3419     const SCEV *DstAccessFn, SmallVectorImpl<const SCEV *> &SrcSubscripts,
3420     SmallVectorImpl<const SCEV *> &DstSubscripts) {
3421 
3422   Value *SrcPtr = getLoadStorePointerOperand(Src);
3423   Value *DstPtr = getLoadStorePointerOperand(Dst);
3424   const SCEVUnknown *SrcBase =
3425       dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
3426   const SCEVUnknown *DstBase =
3427       dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
3428   assert(SrcBase && DstBase && SrcBase == DstBase &&
3429          "expected src and dst scev unknowns to be equal");
3430 
3431   const SCEV *ElementSize = SE->getElementSize(Src);
3432   if (ElementSize != SE->getElementSize(Dst))
3433     return false;
3434 
3435   const SCEV *SrcSCEV = SE->getMinusSCEV(SrcAccessFn, SrcBase);
3436   const SCEV *DstSCEV = SE->getMinusSCEV(DstAccessFn, DstBase);
3437 
3438   const SCEVAddRecExpr *SrcAR = dyn_cast<SCEVAddRecExpr>(SrcSCEV);
3439   const SCEVAddRecExpr *DstAR = dyn_cast<SCEVAddRecExpr>(DstSCEV);
3440   if (!SrcAR || !DstAR || !SrcAR->isAffine() || !DstAR->isAffine())
3441     return false;
3442 
3443   // First step: collect parametric terms in both array references.
3444   SmallVector<const SCEV *, 4> Terms;
3445   collectParametricTerms(*SE, SrcAR, Terms);
3446   collectParametricTerms(*SE, DstAR, Terms);
3447 
3448   // Second step: find subscript sizes.
3449   SmallVector<const SCEV *, 4> Sizes;
3450   findArrayDimensions(*SE, Terms, Sizes, ElementSize);
3451 
3452   // Third step: compute the access functions for each subscript.
3453   computeAccessFunctions(*SE, SrcAR, SrcSubscripts, Sizes);
3454   computeAccessFunctions(*SE, DstAR, DstSubscripts, Sizes);
3455 
3456   // Fail when there is only a subscript: that's a linearized access function.
3457   if (SrcSubscripts.size() < 2 || DstSubscripts.size() < 2 ||
3458       SrcSubscripts.size() != DstSubscripts.size())
3459     return false;
3460 
3461   size_t Size = SrcSubscripts.size();
3462 
3463   // Statically check that the array bounds are in-range. The first subscript we
3464   // don't have a size for and it cannot overflow into another subscript, so is
3465   // always safe. The others need to be 0 <= subscript[i] < bound, for both src
3466   // and dst.
3467   // FIXME: It may be better to record these sizes and add them as constraints
3468   // to the dependency checks.
3469   if (!DisableDelinearizationChecks)
3470     for (size_t I = 1; I < Size; ++I) {
3471       if (!isKnownNonNegative(SrcSubscripts[I], SrcPtr))
3472         return false;
3473 
3474       if (!isKnownLessThan(SrcSubscripts[I], Sizes[I - 1]))
3475         return false;
3476 
3477       if (!isKnownNonNegative(DstSubscripts[I], DstPtr))
3478         return false;
3479 
3480       if (!isKnownLessThan(DstSubscripts[I], Sizes[I - 1]))
3481         return false;
3482     }
3483 
3484   return true;
3485 }
3486 
3487 //===----------------------------------------------------------------------===//
3488 
3489 #ifndef NDEBUG
3490 // For debugging purposes, dump a small bit vector to dbgs().
3491 static void dumpSmallBitVector(SmallBitVector &BV) {
3492   dbgs() << "{";
3493   for (unsigned VI : BV.set_bits()) {
3494     dbgs() << VI;
3495     if (BV.find_next(VI) >= 0)
3496       dbgs() << ' ';
3497   }
3498   dbgs() << "}\n";
3499 }
3500 #endif
3501 
3502 bool DependenceInfo::invalidate(Function &F, const PreservedAnalyses &PA,
3503                                 FunctionAnalysisManager::Invalidator &Inv) {
3504   // Check if the analysis itself has been invalidated.
3505   auto PAC = PA.getChecker<DependenceAnalysis>();
3506   if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
3507     return true;
3508 
3509   // Check transitive dependencies.
3510   return Inv.invalidate<AAManager>(F, PA) ||
3511          Inv.invalidate<ScalarEvolutionAnalysis>(F, PA) ||
3512          Inv.invalidate<LoopAnalysis>(F, PA);
3513 }
3514 
3515 // depends -
3516 // Returns NULL if there is no dependence.
3517 // Otherwise, return a Dependence with as many details as possible.
3518 // Corresponds to Section 3.1 in the paper
3519 //
3520 //            Practical Dependence Testing
3521 //            Goff, Kennedy, Tseng
3522 //            PLDI 1991
3523 //
3524 // Care is required to keep the routine below, getSplitIteration(),
3525 // up to date with respect to this routine.
3526 std::unique_ptr<Dependence>
3527 DependenceInfo::depends(Instruction *Src, Instruction *Dst,
3528                         bool PossiblyLoopIndependent) {
3529   if (Src == Dst)
3530     PossiblyLoopIndependent = false;
3531 
3532   if (!(Src->mayReadOrWriteMemory() && Dst->mayReadOrWriteMemory()))
3533     // if both instructions don't reference memory, there's no dependence
3534     return nullptr;
3535 
3536   if (!isLoadOrStore(Src) || !isLoadOrStore(Dst)) {
3537     // can only analyze simple loads and stores, i.e., no calls, invokes, etc.
3538     LLVM_DEBUG(dbgs() << "can only handle simple loads and stores\n");
3539     return std::make_unique<Dependence>(Src, Dst);
3540   }
3541 
3542   assert(isLoadOrStore(Src) && "instruction is not load or store");
3543   assert(isLoadOrStore(Dst) && "instruction is not load or store");
3544   Value *SrcPtr = getLoadStorePointerOperand(Src);
3545   Value *DstPtr = getLoadStorePointerOperand(Dst);
3546 
3547   switch (underlyingObjectsAlias(AA, F->getParent()->getDataLayout(),
3548                                  MemoryLocation::get(Dst),
3549                                  MemoryLocation::get(Src))) {
3550   case AliasResult::MayAlias:
3551   case AliasResult::PartialAlias:
3552     // cannot analyse objects if we don't understand their aliasing.
3553     LLVM_DEBUG(dbgs() << "can't analyze may or partial alias\n");
3554     return std::make_unique<Dependence>(Src, Dst);
3555   case AliasResult::NoAlias:
3556     // If the objects noalias, they are distinct, accesses are independent.
3557     LLVM_DEBUG(dbgs() << "no alias\n");
3558     return nullptr;
3559   case AliasResult::MustAlias:
3560     break; // The underlying objects alias; test accesses for dependence.
3561   }
3562 
3563   // establish loop nesting levels
3564   establishNestingLevels(Src, Dst);
3565   LLVM_DEBUG(dbgs() << "    common nesting levels = " << CommonLevels << "\n");
3566   LLVM_DEBUG(dbgs() << "    maximum nesting levels = " << MaxLevels << "\n");
3567 
3568   FullDependence Result(Src, Dst, PossiblyLoopIndependent, CommonLevels);
3569   ++TotalArrayPairs;
3570 
3571   unsigned Pairs = 1;
3572   SmallVector<Subscript, 2> Pair(Pairs);
3573   const SCEV *SrcSCEV = SE->getSCEV(SrcPtr);
3574   const SCEV *DstSCEV = SE->getSCEV(DstPtr);
3575   LLVM_DEBUG(dbgs() << "    SrcSCEV = " << *SrcSCEV << "\n");
3576   LLVM_DEBUG(dbgs() << "    DstSCEV = " << *DstSCEV << "\n");
3577   if (SE->getPointerBase(SrcSCEV) != SE->getPointerBase(DstSCEV)) {
3578     // If two pointers have different bases, trying to analyze indexes won't
3579     // work; we can't compare them to each other. This can happen, for example,
3580     // if one is produced by an LCSSA PHI node.
3581     //
3582     // We check this upfront so we don't crash in cases where getMinusSCEV()
3583     // returns a SCEVCouldNotCompute.
3584     LLVM_DEBUG(dbgs() << "can't analyze SCEV with different pointer base\n");
3585     return std::make_unique<Dependence>(Src, Dst);
3586   }
3587   Pair[0].Src = SrcSCEV;
3588   Pair[0].Dst = DstSCEV;
3589 
3590   if (Delinearize) {
3591     if (tryDelinearize(Src, Dst, Pair)) {
3592       LLVM_DEBUG(dbgs() << "    delinearized\n");
3593       Pairs = Pair.size();
3594     }
3595   }
3596 
3597   for (unsigned P = 0; P < Pairs; ++P) {
3598     Pair[P].Loops.resize(MaxLevels + 1);
3599     Pair[P].GroupLoops.resize(MaxLevels + 1);
3600     Pair[P].Group.resize(Pairs);
3601     removeMatchingExtensions(&Pair[P]);
3602     Pair[P].Classification =
3603       classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()),
3604                    Pair[P].Dst, LI->getLoopFor(Dst->getParent()),
3605                    Pair[P].Loops);
3606     Pair[P].GroupLoops = Pair[P].Loops;
3607     Pair[P].Group.set(P);
3608     LLVM_DEBUG(dbgs() << "    subscript " << P << "\n");
3609     LLVM_DEBUG(dbgs() << "\tsrc = " << *Pair[P].Src << "\n");
3610     LLVM_DEBUG(dbgs() << "\tdst = " << *Pair[P].Dst << "\n");
3611     LLVM_DEBUG(dbgs() << "\tclass = " << Pair[P].Classification << "\n");
3612     LLVM_DEBUG(dbgs() << "\tloops = ");
3613     LLVM_DEBUG(dumpSmallBitVector(Pair[P].Loops));
3614   }
3615 
3616   SmallBitVector Separable(Pairs);
3617   SmallBitVector Coupled(Pairs);
3618 
3619   // Partition subscripts into separable and minimally-coupled groups
3620   // Algorithm in paper is algorithmically better;
3621   // this may be faster in practice. Check someday.
3622   //
3623   // Here's an example of how it works. Consider this code:
3624   //
3625   //   for (i = ...) {
3626   //     for (j = ...) {
3627   //       for (k = ...) {
3628   //         for (l = ...) {
3629   //           for (m = ...) {
3630   //             A[i][j][k][m] = ...;
3631   //             ... = A[0][j][l][i + j];
3632   //           }
3633   //         }
3634   //       }
3635   //     }
3636   //   }
3637   //
3638   // There are 4 subscripts here:
3639   //    0 [i] and [0]
3640   //    1 [j] and [j]
3641   //    2 [k] and [l]
3642   //    3 [m] and [i + j]
3643   //
3644   // We've already classified each subscript pair as ZIV, SIV, etc.,
3645   // and collected all the loops mentioned by pair P in Pair[P].Loops.
3646   // In addition, we've initialized Pair[P].GroupLoops to Pair[P].Loops
3647   // and set Pair[P].Group = {P}.
3648   //
3649   //      Src Dst    Classification Loops  GroupLoops Group
3650   //    0 [i] [0]         SIV       {1}      {1}        {0}
3651   //    1 [j] [j]         SIV       {2}      {2}        {1}
3652   //    2 [k] [l]         RDIV      {3,4}    {3,4}      {2}
3653   //    3 [m] [i + j]     MIV       {1,2,5}  {1,2,5}    {3}
3654   //
3655   // For each subscript SI 0 .. 3, we consider each remaining subscript, SJ.
3656   // So, 0 is compared against 1, 2, and 3; 1 is compared against 2 and 3, etc.
3657   //
3658   // We begin by comparing 0 and 1. The intersection of the GroupLoops is empty.
3659   // Next, 0 and 2. Again, the intersection of their GroupLoops is empty.
3660   // Next 0 and 3. The intersection of their GroupLoop = {1}, not empty,
3661   // so Pair[3].Group = {0,3} and Done = false (that is, 0 will not be added
3662   // to either Separable or Coupled).
3663   //
3664   // Next, we consider 1 and 2. The intersection of the GroupLoops is empty.
3665   // Next, 1 and 3. The intersection of their GroupLoops = {2}, not empty,
3666   // so Pair[3].Group = {0, 1, 3} and Done = false.
3667   //
3668   // Next, we compare 2 against 3. The intersection of the GroupLoops is empty.
3669   // Since Done remains true, we add 2 to the set of Separable pairs.
3670   //
3671   // Finally, we consider 3. There's nothing to compare it with,
3672   // so Done remains true and we add it to the Coupled set.
3673   // Pair[3].Group = {0, 1, 3} and GroupLoops = {1, 2, 5}.
3674   //
3675   // In the end, we've got 1 separable subscript and 1 coupled group.
3676   for (unsigned SI = 0; SI < Pairs; ++SI) {
3677     if (Pair[SI].Classification == Subscript::NonLinear) {
3678       // ignore these, but collect loops for later
3679       ++NonlinearSubscriptPairs;
3680       collectCommonLoops(Pair[SI].Src,
3681                          LI->getLoopFor(Src->getParent()),
3682                          Pair[SI].Loops);
3683       collectCommonLoops(Pair[SI].Dst,
3684                          LI->getLoopFor(Dst->getParent()),
3685                          Pair[SI].Loops);
3686       Result.Consistent = false;
3687     } else if (Pair[SI].Classification == Subscript::ZIV) {
3688       // always separable
3689       Separable.set(SI);
3690     }
3691     else {
3692       // SIV, RDIV, or MIV, so check for coupled group
3693       bool Done = true;
3694       for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) {
3695         SmallBitVector Intersection = Pair[SI].GroupLoops;
3696         Intersection &= Pair[SJ].GroupLoops;
3697         if (Intersection.any()) {
3698           // accumulate set of all the loops in group
3699           Pair[SJ].GroupLoops |= Pair[SI].GroupLoops;
3700           // accumulate set of all subscripts in group
3701           Pair[SJ].Group |= Pair[SI].Group;
3702           Done = false;
3703         }
3704       }
3705       if (Done) {
3706         if (Pair[SI].Group.count() == 1) {
3707           Separable.set(SI);
3708           ++SeparableSubscriptPairs;
3709         }
3710         else {
3711           Coupled.set(SI);
3712           ++CoupledSubscriptPairs;
3713         }
3714       }
3715     }
3716   }
3717 
3718   LLVM_DEBUG(dbgs() << "    Separable = ");
3719   LLVM_DEBUG(dumpSmallBitVector(Separable));
3720   LLVM_DEBUG(dbgs() << "    Coupled = ");
3721   LLVM_DEBUG(dumpSmallBitVector(Coupled));
3722 
3723   Constraint NewConstraint;
3724   NewConstraint.setAny(SE);
3725 
3726   // test separable subscripts
3727   for (unsigned SI : Separable.set_bits()) {
3728     LLVM_DEBUG(dbgs() << "testing subscript " << SI);
3729     switch (Pair[SI].Classification) {
3730     case Subscript::ZIV:
3731       LLVM_DEBUG(dbgs() << ", ZIV\n");
3732       if (testZIV(Pair[SI].Src, Pair[SI].Dst, Result))
3733         return nullptr;
3734       break;
3735     case Subscript::SIV: {
3736       LLVM_DEBUG(dbgs() << ", SIV\n");
3737       unsigned Level;
3738       const SCEV *SplitIter = nullptr;
3739       if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level, Result, NewConstraint,
3740                   SplitIter))
3741         return nullptr;
3742       break;
3743     }
3744     case Subscript::RDIV:
3745       LLVM_DEBUG(dbgs() << ", RDIV\n");
3746       if (testRDIV(Pair[SI].Src, Pair[SI].Dst, Result))
3747         return nullptr;
3748       break;
3749     case Subscript::MIV:
3750       LLVM_DEBUG(dbgs() << ", MIV\n");
3751       if (testMIV(Pair[SI].Src, Pair[SI].Dst, Pair[SI].Loops, Result))
3752         return nullptr;
3753       break;
3754     default:
3755       llvm_unreachable("subscript has unexpected classification");
3756     }
3757   }
3758 
3759   if (Coupled.count()) {
3760     // test coupled subscript groups
3761     LLVM_DEBUG(dbgs() << "starting on coupled subscripts\n");
3762     LLVM_DEBUG(dbgs() << "MaxLevels + 1 = " << MaxLevels + 1 << "\n");
3763     SmallVector<Constraint, 4> Constraints(MaxLevels + 1);
3764     for (unsigned II = 0; II <= MaxLevels; ++II)
3765       Constraints[II].setAny(SE);
3766     for (unsigned SI : Coupled.set_bits()) {
3767       LLVM_DEBUG(dbgs() << "testing subscript group " << SI << " { ");
3768       SmallBitVector Group(Pair[SI].Group);
3769       SmallBitVector Sivs(Pairs);
3770       SmallBitVector Mivs(Pairs);
3771       SmallBitVector ConstrainedLevels(MaxLevels + 1);
3772       SmallVector<Subscript *, 4> PairsInGroup;
3773       for (unsigned SJ : Group.set_bits()) {
3774         LLVM_DEBUG(dbgs() << SJ << " ");
3775         if (Pair[SJ].Classification == Subscript::SIV)
3776           Sivs.set(SJ);
3777         else
3778           Mivs.set(SJ);
3779         PairsInGroup.push_back(&Pair[SJ]);
3780       }
3781       unifySubscriptType(PairsInGroup);
3782       LLVM_DEBUG(dbgs() << "}\n");
3783       while (Sivs.any()) {
3784         bool Changed = false;
3785         for (unsigned SJ : Sivs.set_bits()) {
3786           LLVM_DEBUG(dbgs() << "testing subscript " << SJ << ", SIV\n");
3787           // SJ is an SIV subscript that's part of the current coupled group
3788           unsigned Level;
3789           const SCEV *SplitIter = nullptr;
3790           LLVM_DEBUG(dbgs() << "SIV\n");
3791           if (testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level, Result, NewConstraint,
3792                       SplitIter))
3793             return nullptr;
3794           ConstrainedLevels.set(Level);
3795           if (intersectConstraints(&Constraints[Level], &NewConstraint)) {
3796             if (Constraints[Level].isEmpty()) {
3797               ++DeltaIndependence;
3798               return nullptr;
3799             }
3800             Changed = true;
3801           }
3802           Sivs.reset(SJ);
3803         }
3804         if (Changed) {
3805           // propagate, possibly creating new SIVs and ZIVs
3806           LLVM_DEBUG(dbgs() << "    propagating\n");
3807           LLVM_DEBUG(dbgs() << "\tMivs = ");
3808           LLVM_DEBUG(dumpSmallBitVector(Mivs));
3809           for (unsigned SJ : Mivs.set_bits()) {
3810             // SJ is an MIV subscript that's part of the current coupled group
3811             LLVM_DEBUG(dbgs() << "\tSJ = " << SJ << "\n");
3812             if (propagate(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops,
3813                           Constraints, Result.Consistent)) {
3814               LLVM_DEBUG(dbgs() << "\t    Changed\n");
3815               ++DeltaPropagations;
3816               Pair[SJ].Classification =
3817                 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()),
3818                              Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()),
3819                              Pair[SJ].Loops);
3820               switch (Pair[SJ].Classification) {
3821               case Subscript::ZIV:
3822                 LLVM_DEBUG(dbgs() << "ZIV\n");
3823                 if (testZIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
3824                   return nullptr;
3825                 Mivs.reset(SJ);
3826                 break;
3827               case Subscript::SIV:
3828                 Sivs.set(SJ);
3829                 Mivs.reset(SJ);
3830                 break;
3831               case Subscript::RDIV:
3832               case Subscript::MIV:
3833                 break;
3834               default:
3835                 llvm_unreachable("bad subscript classification");
3836               }
3837             }
3838           }
3839         }
3840       }
3841 
3842       // test & propagate remaining RDIVs
3843       for (unsigned SJ : Mivs.set_bits()) {
3844         if (Pair[SJ].Classification == Subscript::RDIV) {
3845           LLVM_DEBUG(dbgs() << "RDIV test\n");
3846           if (testRDIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
3847             return nullptr;
3848           // I don't yet understand how to propagate RDIV results
3849           Mivs.reset(SJ);
3850         }
3851       }
3852 
3853       // test remaining MIVs
3854       // This code is temporary.
3855       // Better to somehow test all remaining subscripts simultaneously.
3856       for (unsigned SJ : Mivs.set_bits()) {
3857         if (Pair[SJ].Classification == Subscript::MIV) {
3858           LLVM_DEBUG(dbgs() << "MIV test\n");
3859           if (testMIV(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops, Result))
3860             return nullptr;
3861         }
3862         else
3863           llvm_unreachable("expected only MIV subscripts at this point");
3864       }
3865 
3866       // update Result.DV from constraint vector
3867       LLVM_DEBUG(dbgs() << "    updating\n");
3868       for (unsigned SJ : ConstrainedLevels.set_bits()) {
3869         if (SJ > CommonLevels)
3870           break;
3871         updateDirection(Result.DV[SJ - 1], Constraints[SJ]);
3872         if (Result.DV[SJ - 1].Direction == Dependence::DVEntry::NONE)
3873           return nullptr;
3874       }
3875     }
3876   }
3877 
3878   // Make sure the Scalar flags are set correctly.
3879   SmallBitVector CompleteLoops(MaxLevels + 1);
3880   for (unsigned SI = 0; SI < Pairs; ++SI)
3881     CompleteLoops |= Pair[SI].Loops;
3882   for (unsigned II = 1; II <= CommonLevels; ++II)
3883     if (CompleteLoops[II])
3884       Result.DV[II - 1].Scalar = false;
3885 
3886   if (PossiblyLoopIndependent) {
3887     // Make sure the LoopIndependent flag is set correctly.
3888     // All directions must include equal, otherwise no
3889     // loop-independent dependence is possible.
3890     for (unsigned II = 1; II <= CommonLevels; ++II) {
3891       if (!(Result.getDirection(II) & Dependence::DVEntry::EQ)) {
3892         Result.LoopIndependent = false;
3893         break;
3894       }
3895     }
3896   }
3897   else {
3898     // On the other hand, if all directions are equal and there's no
3899     // loop-independent dependence possible, then no dependence exists.
3900     bool AllEqual = true;
3901     for (unsigned II = 1; II <= CommonLevels; ++II) {
3902       if (Result.getDirection(II) != Dependence::DVEntry::EQ) {
3903         AllEqual = false;
3904         break;
3905       }
3906     }
3907     if (AllEqual)
3908       return nullptr;
3909   }
3910 
3911   return std::make_unique<FullDependence>(std::move(Result));
3912 }
3913 
3914 //===----------------------------------------------------------------------===//
3915 // getSplitIteration -
3916 // Rather than spend rarely-used space recording the splitting iteration
3917 // during the Weak-Crossing SIV test, we re-compute it on demand.
3918 // The re-computation is basically a repeat of the entire dependence test,
3919 // though simplified since we know that the dependence exists.
3920 // It's tedious, since we must go through all propagations, etc.
3921 //
3922 // Care is required to keep this code up to date with respect to the routine
3923 // above, depends().
3924 //
3925 // Generally, the dependence analyzer will be used to build
3926 // a dependence graph for a function (basically a map from instructions
3927 // to dependences). Looking for cycles in the graph shows us loops
3928 // that cannot be trivially vectorized/parallelized.
3929 //
3930 // We can try to improve the situation by examining all the dependences
3931 // that make up the cycle, looking for ones we can break.
3932 // Sometimes, peeling the first or last iteration of a loop will break
3933 // dependences, and we've got flags for those possibilities.
3934 // Sometimes, splitting a loop at some other iteration will do the trick,
3935 // and we've got a flag for that case. Rather than waste the space to
3936 // record the exact iteration (since we rarely know), we provide
3937 // a method that calculates the iteration. It's a drag that it must work
3938 // from scratch, but wonderful in that it's possible.
3939 //
3940 // Here's an example:
3941 //
3942 //    for (i = 0; i < 10; i++)
3943 //        A[i] = ...
3944 //        ... = A[11 - i]
3945 //
3946 // There's a loop-carried flow dependence from the store to the load,
3947 // found by the weak-crossing SIV test. The dependence will have a flag,
3948 // indicating that the dependence can be broken by splitting the loop.
3949 // Calling getSplitIteration will return 5.
3950 // Splitting the loop breaks the dependence, like so:
3951 //
3952 //    for (i = 0; i <= 5; i++)
3953 //        A[i] = ...
3954 //        ... = A[11 - i]
3955 //    for (i = 6; i < 10; i++)
3956 //        A[i] = ...
3957 //        ... = A[11 - i]
3958 //
3959 // breaks the dependence and allows us to vectorize/parallelize
3960 // both loops.
3961 const SCEV *DependenceInfo::getSplitIteration(const Dependence &Dep,
3962                                               unsigned SplitLevel) {
3963   assert(Dep.isSplitable(SplitLevel) &&
3964          "Dep should be splitable at SplitLevel");
3965   Instruction *Src = Dep.getSrc();
3966   Instruction *Dst = Dep.getDst();
3967   assert(Src->mayReadFromMemory() || Src->mayWriteToMemory());
3968   assert(Dst->mayReadFromMemory() || Dst->mayWriteToMemory());
3969   assert(isLoadOrStore(Src));
3970   assert(isLoadOrStore(Dst));
3971   Value *SrcPtr = getLoadStorePointerOperand(Src);
3972   Value *DstPtr = getLoadStorePointerOperand(Dst);
3973   assert(underlyingObjectsAlias(
3974              AA, F->getParent()->getDataLayout(), MemoryLocation::get(Dst),
3975              MemoryLocation::get(Src)) == AliasResult::MustAlias);
3976 
3977   // establish loop nesting levels
3978   establishNestingLevels(Src, Dst);
3979 
3980   FullDependence Result(Src, Dst, false, CommonLevels);
3981 
3982   unsigned Pairs = 1;
3983   SmallVector<Subscript, 2> Pair(Pairs);
3984   const SCEV *SrcSCEV = SE->getSCEV(SrcPtr);
3985   const SCEV *DstSCEV = SE->getSCEV(DstPtr);
3986   Pair[0].Src = SrcSCEV;
3987   Pair[0].Dst = DstSCEV;
3988 
3989   if (Delinearize) {
3990     if (tryDelinearize(Src, Dst, Pair)) {
3991       LLVM_DEBUG(dbgs() << "    delinearized\n");
3992       Pairs = Pair.size();
3993     }
3994   }
3995 
3996   for (unsigned P = 0; P < Pairs; ++P) {
3997     Pair[P].Loops.resize(MaxLevels + 1);
3998     Pair[P].GroupLoops.resize(MaxLevels + 1);
3999     Pair[P].Group.resize(Pairs);
4000     removeMatchingExtensions(&Pair[P]);
4001     Pair[P].Classification =
4002       classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()),
4003                    Pair[P].Dst, LI->getLoopFor(Dst->getParent()),
4004                    Pair[P].Loops);
4005     Pair[P].GroupLoops = Pair[P].Loops;
4006     Pair[P].Group.set(P);
4007   }
4008 
4009   SmallBitVector Separable(Pairs);
4010   SmallBitVector Coupled(Pairs);
4011 
4012   // partition subscripts into separable and minimally-coupled groups
4013   for (unsigned SI = 0; SI < Pairs; ++SI) {
4014     if (Pair[SI].Classification == Subscript::NonLinear) {
4015       // ignore these, but collect loops for later
4016       collectCommonLoops(Pair[SI].Src,
4017                          LI->getLoopFor(Src->getParent()),
4018                          Pair[SI].Loops);
4019       collectCommonLoops(Pair[SI].Dst,
4020                          LI->getLoopFor(Dst->getParent()),
4021                          Pair[SI].Loops);
4022       Result.Consistent = false;
4023     }
4024     else if (Pair[SI].Classification == Subscript::ZIV)
4025       Separable.set(SI);
4026     else {
4027       // SIV, RDIV, or MIV, so check for coupled group
4028       bool Done = true;
4029       for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) {
4030         SmallBitVector Intersection = Pair[SI].GroupLoops;
4031         Intersection &= Pair[SJ].GroupLoops;
4032         if (Intersection.any()) {
4033           // accumulate set of all the loops in group
4034           Pair[SJ].GroupLoops |= Pair[SI].GroupLoops;
4035           // accumulate set of all subscripts in group
4036           Pair[SJ].Group |= Pair[SI].Group;
4037           Done = false;
4038         }
4039       }
4040       if (Done) {
4041         if (Pair[SI].Group.count() == 1)
4042           Separable.set(SI);
4043         else
4044           Coupled.set(SI);
4045       }
4046     }
4047   }
4048 
4049   Constraint NewConstraint;
4050   NewConstraint.setAny(SE);
4051 
4052   // test separable subscripts
4053   for (unsigned SI : Separable.set_bits()) {
4054     switch (Pair[SI].Classification) {
4055     case Subscript::SIV: {
4056       unsigned Level;
4057       const SCEV *SplitIter = nullptr;
4058       (void) testSIV(Pair[SI].Src, Pair[SI].Dst, Level,
4059                      Result, NewConstraint, SplitIter);
4060       if (Level == SplitLevel) {
4061         assert(SplitIter != nullptr);
4062         return SplitIter;
4063       }
4064       break;
4065     }
4066     case Subscript::ZIV:
4067     case Subscript::RDIV:
4068     case Subscript::MIV:
4069       break;
4070     default:
4071       llvm_unreachable("subscript has unexpected classification");
4072     }
4073   }
4074 
4075   if (Coupled.count()) {
4076     // test coupled subscript groups
4077     SmallVector<Constraint, 4> Constraints(MaxLevels + 1);
4078     for (unsigned II = 0; II <= MaxLevels; ++II)
4079       Constraints[II].setAny(SE);
4080     for (unsigned SI : Coupled.set_bits()) {
4081       SmallBitVector Group(Pair[SI].Group);
4082       SmallBitVector Sivs(Pairs);
4083       SmallBitVector Mivs(Pairs);
4084       SmallBitVector ConstrainedLevels(MaxLevels + 1);
4085       for (unsigned SJ : Group.set_bits()) {
4086         if (Pair[SJ].Classification == Subscript::SIV)
4087           Sivs.set(SJ);
4088         else
4089           Mivs.set(SJ);
4090       }
4091       while (Sivs.any()) {
4092         bool Changed = false;
4093         for (unsigned SJ : Sivs.set_bits()) {
4094           // SJ is an SIV subscript that's part of the current coupled group
4095           unsigned Level;
4096           const SCEV *SplitIter = nullptr;
4097           (void) testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level,
4098                          Result, NewConstraint, SplitIter);
4099           if (Level == SplitLevel && SplitIter)
4100             return SplitIter;
4101           ConstrainedLevels.set(Level);
4102           if (intersectConstraints(&Constraints[Level], &NewConstraint))
4103             Changed = true;
4104           Sivs.reset(SJ);
4105         }
4106         if (Changed) {
4107           // propagate, possibly creating new SIVs and ZIVs
4108           for (unsigned SJ : Mivs.set_bits()) {
4109             // SJ is an MIV subscript that's part of the current coupled group
4110             if (propagate(Pair[SJ].Src, Pair[SJ].Dst,
4111                           Pair[SJ].Loops, Constraints, Result.Consistent)) {
4112               Pair[SJ].Classification =
4113                 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()),
4114                              Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()),
4115                              Pair[SJ].Loops);
4116               switch (Pair[SJ].Classification) {
4117               case Subscript::ZIV:
4118                 Mivs.reset(SJ);
4119                 break;
4120               case Subscript::SIV:
4121                 Sivs.set(SJ);
4122                 Mivs.reset(SJ);
4123                 break;
4124               case Subscript::RDIV:
4125               case Subscript::MIV:
4126                 break;
4127               default:
4128                 llvm_unreachable("bad subscript classification");
4129               }
4130             }
4131           }
4132         }
4133       }
4134     }
4135   }
4136   llvm_unreachable("somehow reached end of routine");
4137   return nullptr;
4138 }
4139