1 //===-- DependenceAnalysis.cpp - DA Implementation --------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // DependenceAnalysis is an LLVM pass that analyses dependences between memory
10 // accesses. Currently, it is an (incomplete) implementation of the approach
11 // described in
12 //
13 // Practical Dependence Testing
14 // Goff, Kennedy, Tseng
15 // PLDI 1991
16 //
17 // There's a single entry point that analyzes the dependence between a pair
18 // of memory references in a function, returning either NULL, for no dependence,
19 // or a more-or-less detailed description of the dependence between them.
20 //
21 // Currently, the implementation cannot propagate constraints between
22 // coupled RDIV subscripts and lacks a multi-subscript MIV test.
23 // Both of these are conservative weaknesses;
24 // that is, not a source of correctness problems.
25 //
26 // Since Clang linearizes some array subscripts, the dependence
27 // analysis is using SCEV->delinearize to recover the representation of multiple
28 // subscripts, and thus avoid the more expensive and less precise MIV tests. The
29 // delinearization is controlled by the flag -da-delinearize.
30 //
31 // We should pay some careful attention to the possibility of integer overflow
32 // in the implementation of the various tests. This could happen with Add,
33 // Subtract, or Multiply, with both APInt's and SCEV's.
34 //
35 // Some non-linear subscript pairs can be handled by the GCD test
36 // (and perhaps other tests).
37 // Should explore how often these things occur.
38 //
39 // Finally, it seems like certain test cases expose weaknesses in the SCEV
40 // simplification, especially in the handling of sign and zero extensions.
41 // It could be useful to spend time exploring these.
42 //
43 // Please note that this is work in progress and the interface is subject to
44 // change.
45 //
46 //===----------------------------------------------------------------------===//
47 // //
48 // In memory of Ken Kennedy, 1945 - 2007 //
49 // //
50 //===----------------------------------------------------------------------===//
51
52 #include "llvm/Analysis/DependenceAnalysis.h"
53 #include "llvm/ADT/STLExtras.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/LoopInfo.h"
57 #include "llvm/Analysis/ScalarEvolution.h"
58 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
59 #include "llvm/Analysis/ValueTracking.h"
60 #include "llvm/Config/llvm-config.h"
61 #include "llvm/IR/InstIterator.h"
62 #include "llvm/IR/Module.h"
63 #include "llvm/IR/Operator.h"
64 #include "llvm/InitializePasses.h"
65 #include "llvm/Support/CommandLine.h"
66 #include "llvm/Support/Debug.h"
67 #include "llvm/Support/ErrorHandling.h"
68 #include "llvm/Support/raw_ostream.h"
69
70 using namespace llvm;
71
72 #define DEBUG_TYPE "da"
73
74 //===----------------------------------------------------------------------===//
75 // statistics
76
77 STATISTIC(TotalArrayPairs, "Array pairs tested");
78 STATISTIC(SeparableSubscriptPairs, "Separable subscript pairs");
79 STATISTIC(CoupledSubscriptPairs, "Coupled subscript pairs");
80 STATISTIC(NonlinearSubscriptPairs, "Nonlinear subscript pairs");
81 STATISTIC(ZIVapplications, "ZIV applications");
82 STATISTIC(ZIVindependence, "ZIV independence");
83 STATISTIC(StrongSIVapplications, "Strong SIV applications");
84 STATISTIC(StrongSIVsuccesses, "Strong SIV successes");
85 STATISTIC(StrongSIVindependence, "Strong SIV independence");
86 STATISTIC(WeakCrossingSIVapplications, "Weak-Crossing SIV applications");
87 STATISTIC(WeakCrossingSIVsuccesses, "Weak-Crossing SIV successes");
88 STATISTIC(WeakCrossingSIVindependence, "Weak-Crossing SIV independence");
89 STATISTIC(ExactSIVapplications, "Exact SIV applications");
90 STATISTIC(ExactSIVsuccesses, "Exact SIV successes");
91 STATISTIC(ExactSIVindependence, "Exact SIV independence");
92 STATISTIC(WeakZeroSIVapplications, "Weak-Zero SIV applications");
93 STATISTIC(WeakZeroSIVsuccesses, "Weak-Zero SIV successes");
94 STATISTIC(WeakZeroSIVindependence, "Weak-Zero SIV independence");
95 STATISTIC(ExactRDIVapplications, "Exact RDIV applications");
96 STATISTIC(ExactRDIVindependence, "Exact RDIV independence");
97 STATISTIC(SymbolicRDIVapplications, "Symbolic RDIV applications");
98 STATISTIC(SymbolicRDIVindependence, "Symbolic RDIV independence");
99 STATISTIC(DeltaApplications, "Delta applications");
100 STATISTIC(DeltaSuccesses, "Delta successes");
101 STATISTIC(DeltaIndependence, "Delta independence");
102 STATISTIC(DeltaPropagations, "Delta propagations");
103 STATISTIC(GCDapplications, "GCD applications");
104 STATISTIC(GCDsuccesses, "GCD successes");
105 STATISTIC(GCDindependence, "GCD independence");
106 STATISTIC(BanerjeeApplications, "Banerjee applications");
107 STATISTIC(BanerjeeIndependence, "Banerjee independence");
108 STATISTIC(BanerjeeSuccesses, "Banerjee successes");
109
110 static cl::opt<bool>
111 Delinearize("da-delinearize", cl::init(true), cl::Hidden, cl::ZeroOrMore,
112 cl::desc("Try to delinearize array references."));
113 static cl::opt<bool> DisableDelinearizationChecks(
114 "da-disable-delinearization-checks", cl::init(false), cl::Hidden,
115 cl::ZeroOrMore,
116 cl::desc(
117 "Disable checks that try to statically verify validity of "
118 "delinearized subscripts. Enabling this option may result in incorrect "
119 "dependence vectors for languages that allow the subscript of one "
120 "dimension to underflow or overflow into another dimension."));
121
122 //===----------------------------------------------------------------------===//
123 // basics
124
125 DependenceAnalysis::Result
run(Function & F,FunctionAnalysisManager & FAM)126 DependenceAnalysis::run(Function &F, FunctionAnalysisManager &FAM) {
127 auto &AA = FAM.getResult<AAManager>(F);
128 auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F);
129 auto &LI = FAM.getResult<LoopAnalysis>(F);
130 return DependenceInfo(&F, &AA, &SE, &LI);
131 }
132
133 AnalysisKey DependenceAnalysis::Key;
134
135 INITIALIZE_PASS_BEGIN(DependenceAnalysisWrapperPass, "da",
136 "Dependence Analysis", true, true)
137 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
138 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
139 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
140 INITIALIZE_PASS_END(DependenceAnalysisWrapperPass, "da", "Dependence Analysis",
141 true, true)
142
143 char DependenceAnalysisWrapperPass::ID = 0;
144
DependenceAnalysisWrapperPass()145 DependenceAnalysisWrapperPass::DependenceAnalysisWrapperPass()
146 : FunctionPass(ID) {
147 initializeDependenceAnalysisWrapperPassPass(*PassRegistry::getPassRegistry());
148 }
149
createDependenceAnalysisWrapperPass()150 FunctionPass *llvm::createDependenceAnalysisWrapperPass() {
151 return new DependenceAnalysisWrapperPass();
152 }
153
runOnFunction(Function & F)154 bool DependenceAnalysisWrapperPass::runOnFunction(Function &F) {
155 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
156 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
157 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
158 info.reset(new DependenceInfo(&F, &AA, &SE, &LI));
159 return false;
160 }
161
getDI() const162 DependenceInfo &DependenceAnalysisWrapperPass::getDI() const { return *info; }
163
releaseMemory()164 void DependenceAnalysisWrapperPass::releaseMemory() { info.reset(); }
165
getAnalysisUsage(AnalysisUsage & AU) const166 void DependenceAnalysisWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
167 AU.setPreservesAll();
168 AU.addRequiredTransitive<AAResultsWrapperPass>();
169 AU.addRequiredTransitive<ScalarEvolutionWrapperPass>();
170 AU.addRequiredTransitive<LoopInfoWrapperPass>();
171 }
172
173 // Used to test the dependence analyzer.
174 // Looks through the function, noting instructions that may access memory.
175 // Calls depends() on every possible pair and prints out the result.
176 // Ignores all other instructions.
dumpExampleDependence(raw_ostream & OS,DependenceInfo * DA)177 static void dumpExampleDependence(raw_ostream &OS, DependenceInfo *DA) {
178 auto *F = DA->getFunction();
179 for (inst_iterator SrcI = inst_begin(F), SrcE = inst_end(F); SrcI != SrcE;
180 ++SrcI) {
181 if (SrcI->mayReadOrWriteMemory()) {
182 for (inst_iterator DstI = SrcI, DstE = inst_end(F);
183 DstI != DstE; ++DstI) {
184 if (DstI->mayReadOrWriteMemory()) {
185 OS << "Src:" << *SrcI << " --> Dst:" << *DstI << "\n";
186 OS << " da analyze - ";
187 if (auto D = DA->depends(&*SrcI, &*DstI, true)) {
188 D->dump(OS);
189 for (unsigned Level = 1; Level <= D->getLevels(); Level++) {
190 if (D->isSplitable(Level)) {
191 OS << " da analyze - split level = " << Level;
192 OS << ", iteration = " << *DA->getSplitIteration(*D, Level);
193 OS << "!\n";
194 }
195 }
196 }
197 else
198 OS << "none!\n";
199 }
200 }
201 }
202 }
203 }
204
print(raw_ostream & OS,const Module *) const205 void DependenceAnalysisWrapperPass::print(raw_ostream &OS,
206 const Module *) const {
207 dumpExampleDependence(OS, info.get());
208 }
209
210 PreservedAnalyses
run(Function & F,FunctionAnalysisManager & FAM)211 DependenceAnalysisPrinterPass::run(Function &F, FunctionAnalysisManager &FAM) {
212 OS << "'Dependence Analysis' for function '" << F.getName() << "':\n";
213 dumpExampleDependence(OS, &FAM.getResult<DependenceAnalysis>(F));
214 return PreservedAnalyses::all();
215 }
216
217 //===----------------------------------------------------------------------===//
218 // Dependence methods
219
220 // Returns true if this is an input dependence.
isInput() const221 bool Dependence::isInput() const {
222 return Src->mayReadFromMemory() && Dst->mayReadFromMemory();
223 }
224
225
226 // Returns true if this is an output dependence.
isOutput() const227 bool Dependence::isOutput() const {
228 return Src->mayWriteToMemory() && Dst->mayWriteToMemory();
229 }
230
231
232 // Returns true if this is an flow (aka true) dependence.
isFlow() const233 bool Dependence::isFlow() const {
234 return Src->mayWriteToMemory() && Dst->mayReadFromMemory();
235 }
236
237
238 // Returns true if this is an anti dependence.
isAnti() const239 bool Dependence::isAnti() const {
240 return Src->mayReadFromMemory() && Dst->mayWriteToMemory();
241 }
242
243
244 // Returns true if a particular level is scalar; that is,
245 // if no subscript in the source or destination mention the induction
246 // variable associated with the loop at this level.
247 // Leave this out of line, so it will serve as a virtual method anchor
isScalar(unsigned level) const248 bool Dependence::isScalar(unsigned level) const {
249 return false;
250 }
251
252
253 //===----------------------------------------------------------------------===//
254 // FullDependence methods
255
FullDependence(Instruction * Source,Instruction * Destination,bool PossiblyLoopIndependent,unsigned CommonLevels)256 FullDependence::FullDependence(Instruction *Source, Instruction *Destination,
257 bool PossiblyLoopIndependent,
258 unsigned CommonLevels)
259 : Dependence(Source, Destination), Levels(CommonLevels),
260 LoopIndependent(PossiblyLoopIndependent) {
261 Consistent = true;
262 if (CommonLevels)
263 DV = std::make_unique<DVEntry[]>(CommonLevels);
264 }
265
266 // The rest are simple getters that hide the implementation.
267
268 // getDirection - Returns the direction associated with a particular level.
getDirection(unsigned Level) const269 unsigned FullDependence::getDirection(unsigned Level) const {
270 assert(0 < Level && Level <= Levels && "Level out of range");
271 return DV[Level - 1].Direction;
272 }
273
274
275 // Returns the distance (or NULL) associated with a particular level.
getDistance(unsigned Level) const276 const SCEV *FullDependence::getDistance(unsigned Level) const {
277 assert(0 < Level && Level <= Levels && "Level out of range");
278 return DV[Level - 1].Distance;
279 }
280
281
282 // Returns true if a particular level is scalar; that is,
283 // if no subscript in the source or destination mention the induction
284 // variable associated with the loop at this level.
isScalar(unsigned Level) const285 bool FullDependence::isScalar(unsigned Level) const {
286 assert(0 < Level && Level <= Levels && "Level out of range");
287 return DV[Level - 1].Scalar;
288 }
289
290
291 // Returns true if peeling the first iteration from this loop
292 // will break this dependence.
isPeelFirst(unsigned Level) const293 bool FullDependence::isPeelFirst(unsigned Level) const {
294 assert(0 < Level && Level <= Levels && "Level out of range");
295 return DV[Level - 1].PeelFirst;
296 }
297
298
299 // Returns true if peeling the last iteration from this loop
300 // will break this dependence.
isPeelLast(unsigned Level) const301 bool FullDependence::isPeelLast(unsigned Level) const {
302 assert(0 < Level && Level <= Levels && "Level out of range");
303 return DV[Level - 1].PeelLast;
304 }
305
306
307 // Returns true if splitting this loop will break the dependence.
isSplitable(unsigned Level) const308 bool FullDependence::isSplitable(unsigned Level) const {
309 assert(0 < Level && Level <= Levels && "Level out of range");
310 return DV[Level - 1].Splitable;
311 }
312
313
314 //===----------------------------------------------------------------------===//
315 // DependenceInfo::Constraint methods
316
317 // If constraint is a point <X, Y>, returns X.
318 // Otherwise assert.
getX() const319 const SCEV *DependenceInfo::Constraint::getX() const {
320 assert(Kind == Point && "Kind should be Point");
321 return A;
322 }
323
324
325 // If constraint is a point <X, Y>, returns Y.
326 // Otherwise assert.
getY() const327 const SCEV *DependenceInfo::Constraint::getY() const {
328 assert(Kind == Point && "Kind should be Point");
329 return B;
330 }
331
332
333 // If constraint is a line AX + BY = C, returns A.
334 // Otherwise assert.
getA() const335 const SCEV *DependenceInfo::Constraint::getA() const {
336 assert((Kind == Line || Kind == Distance) &&
337 "Kind should be Line (or Distance)");
338 return A;
339 }
340
341
342 // If constraint is a line AX + BY = C, returns B.
343 // Otherwise assert.
getB() const344 const SCEV *DependenceInfo::Constraint::getB() const {
345 assert((Kind == Line || Kind == Distance) &&
346 "Kind should be Line (or Distance)");
347 return B;
348 }
349
350
351 // If constraint is a line AX + BY = C, returns C.
352 // Otherwise assert.
getC() const353 const SCEV *DependenceInfo::Constraint::getC() const {
354 assert((Kind == Line || Kind == Distance) &&
355 "Kind should be Line (or Distance)");
356 return C;
357 }
358
359
360 // If constraint is a distance, returns D.
361 // Otherwise assert.
getD() const362 const SCEV *DependenceInfo::Constraint::getD() const {
363 assert(Kind == Distance && "Kind should be Distance");
364 return SE->getNegativeSCEV(C);
365 }
366
367
368 // Returns the loop associated with this constraint.
getAssociatedLoop() const369 const Loop *DependenceInfo::Constraint::getAssociatedLoop() const {
370 assert((Kind == Distance || Kind == Line || Kind == Point) &&
371 "Kind should be Distance, Line, or Point");
372 return AssociatedLoop;
373 }
374
setPoint(const SCEV * X,const SCEV * Y,const Loop * CurLoop)375 void DependenceInfo::Constraint::setPoint(const SCEV *X, const SCEV *Y,
376 const Loop *CurLoop) {
377 Kind = Point;
378 A = X;
379 B = Y;
380 AssociatedLoop = CurLoop;
381 }
382
setLine(const SCEV * AA,const SCEV * BB,const SCEV * CC,const Loop * CurLoop)383 void DependenceInfo::Constraint::setLine(const SCEV *AA, const SCEV *BB,
384 const SCEV *CC, const Loop *CurLoop) {
385 Kind = Line;
386 A = AA;
387 B = BB;
388 C = CC;
389 AssociatedLoop = CurLoop;
390 }
391
setDistance(const SCEV * D,const Loop * CurLoop)392 void DependenceInfo::Constraint::setDistance(const SCEV *D,
393 const Loop *CurLoop) {
394 Kind = Distance;
395 A = SE->getOne(D->getType());
396 B = SE->getNegativeSCEV(A);
397 C = SE->getNegativeSCEV(D);
398 AssociatedLoop = CurLoop;
399 }
400
setEmpty()401 void DependenceInfo::Constraint::setEmpty() { Kind = Empty; }
402
setAny(ScalarEvolution * NewSE)403 void DependenceInfo::Constraint::setAny(ScalarEvolution *NewSE) {
404 SE = NewSE;
405 Kind = Any;
406 }
407
408 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
409 // For debugging purposes. Dumps the constraint out to OS.
dump(raw_ostream & OS) const410 LLVM_DUMP_METHOD void DependenceInfo::Constraint::dump(raw_ostream &OS) const {
411 if (isEmpty())
412 OS << " Empty\n";
413 else if (isAny())
414 OS << " Any\n";
415 else if (isPoint())
416 OS << " Point is <" << *getX() << ", " << *getY() << ">\n";
417 else if (isDistance())
418 OS << " Distance is " << *getD() <<
419 " (" << *getA() << "*X + " << *getB() << "*Y = " << *getC() << ")\n";
420 else if (isLine())
421 OS << " Line is " << *getA() << "*X + " <<
422 *getB() << "*Y = " << *getC() << "\n";
423 else
424 llvm_unreachable("unknown constraint type in Constraint::dump");
425 }
426 #endif
427
428
429 // Updates X with the intersection
430 // of the Constraints X and Y. Returns true if X has changed.
431 // Corresponds to Figure 4 from the paper
432 //
433 // Practical Dependence Testing
434 // Goff, Kennedy, Tseng
435 // PLDI 1991
intersectConstraints(Constraint * X,const Constraint * Y)436 bool DependenceInfo::intersectConstraints(Constraint *X, const Constraint *Y) {
437 ++DeltaApplications;
438 LLVM_DEBUG(dbgs() << "\tintersect constraints\n");
439 LLVM_DEBUG(dbgs() << "\t X ="; X->dump(dbgs()));
440 LLVM_DEBUG(dbgs() << "\t Y ="; Y->dump(dbgs()));
441 assert(!Y->isPoint() && "Y must not be a Point");
442 if (X->isAny()) {
443 if (Y->isAny())
444 return false;
445 *X = *Y;
446 return true;
447 }
448 if (X->isEmpty())
449 return false;
450 if (Y->isEmpty()) {
451 X->setEmpty();
452 return true;
453 }
454
455 if (X->isDistance() && Y->isDistance()) {
456 LLVM_DEBUG(dbgs() << "\t intersect 2 distances\n");
457 if (isKnownPredicate(CmpInst::ICMP_EQ, X->getD(), Y->getD()))
458 return false;
459 if (isKnownPredicate(CmpInst::ICMP_NE, X->getD(), Y->getD())) {
460 X->setEmpty();
461 ++DeltaSuccesses;
462 return true;
463 }
464 // Hmmm, interesting situation.
465 // I guess if either is constant, keep it and ignore the other.
466 if (isa<SCEVConstant>(Y->getD())) {
467 *X = *Y;
468 return true;
469 }
470 return false;
471 }
472
473 // At this point, the pseudo-code in Figure 4 of the paper
474 // checks if (X->isPoint() && Y->isPoint()).
475 // This case can't occur in our implementation,
476 // since a Point can only arise as the result of intersecting
477 // two Line constraints, and the right-hand value, Y, is never
478 // the result of an intersection.
479 assert(!(X->isPoint() && Y->isPoint()) &&
480 "We shouldn't ever see X->isPoint() && Y->isPoint()");
481
482 if (X->isLine() && Y->isLine()) {
483 LLVM_DEBUG(dbgs() << "\t intersect 2 lines\n");
484 const SCEV *Prod1 = SE->getMulExpr(X->getA(), Y->getB());
485 const SCEV *Prod2 = SE->getMulExpr(X->getB(), Y->getA());
486 if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2)) {
487 // slopes are equal, so lines are parallel
488 LLVM_DEBUG(dbgs() << "\t\tsame slope\n");
489 Prod1 = SE->getMulExpr(X->getC(), Y->getB());
490 Prod2 = SE->getMulExpr(X->getB(), Y->getC());
491 if (isKnownPredicate(CmpInst::ICMP_EQ, Prod1, Prod2))
492 return false;
493 if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) {
494 X->setEmpty();
495 ++DeltaSuccesses;
496 return true;
497 }
498 return false;
499 }
500 if (isKnownPredicate(CmpInst::ICMP_NE, Prod1, Prod2)) {
501 // slopes differ, so lines intersect
502 LLVM_DEBUG(dbgs() << "\t\tdifferent slopes\n");
503 const SCEV *C1B2 = SE->getMulExpr(X->getC(), Y->getB());
504 const SCEV *C1A2 = SE->getMulExpr(X->getC(), Y->getA());
505 const SCEV *C2B1 = SE->getMulExpr(Y->getC(), X->getB());
506 const SCEV *C2A1 = SE->getMulExpr(Y->getC(), X->getA());
507 const SCEV *A1B2 = SE->getMulExpr(X->getA(), Y->getB());
508 const SCEV *A2B1 = SE->getMulExpr(Y->getA(), X->getB());
509 const SCEVConstant *C1A2_C2A1 =
510 dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1A2, C2A1));
511 const SCEVConstant *C1B2_C2B1 =
512 dyn_cast<SCEVConstant>(SE->getMinusSCEV(C1B2, C2B1));
513 const SCEVConstant *A1B2_A2B1 =
514 dyn_cast<SCEVConstant>(SE->getMinusSCEV(A1B2, A2B1));
515 const SCEVConstant *A2B1_A1B2 =
516 dyn_cast<SCEVConstant>(SE->getMinusSCEV(A2B1, A1B2));
517 if (!C1B2_C2B1 || !C1A2_C2A1 ||
518 !A1B2_A2B1 || !A2B1_A1B2)
519 return false;
520 APInt Xtop = C1B2_C2B1->getAPInt();
521 APInt Xbot = A1B2_A2B1->getAPInt();
522 APInt Ytop = C1A2_C2A1->getAPInt();
523 APInt Ybot = A2B1_A1B2->getAPInt();
524 LLVM_DEBUG(dbgs() << "\t\tXtop = " << Xtop << "\n");
525 LLVM_DEBUG(dbgs() << "\t\tXbot = " << Xbot << "\n");
526 LLVM_DEBUG(dbgs() << "\t\tYtop = " << Ytop << "\n");
527 LLVM_DEBUG(dbgs() << "\t\tYbot = " << Ybot << "\n");
528 APInt Xq = Xtop; // these need to be initialized, even
529 APInt Xr = Xtop; // though they're just going to be overwritten
530 APInt::sdivrem(Xtop, Xbot, Xq, Xr);
531 APInt Yq = Ytop;
532 APInt Yr = Ytop;
533 APInt::sdivrem(Ytop, Ybot, Yq, Yr);
534 if (Xr != 0 || Yr != 0) {
535 X->setEmpty();
536 ++DeltaSuccesses;
537 return true;
538 }
539 LLVM_DEBUG(dbgs() << "\t\tX = " << Xq << ", Y = " << Yq << "\n");
540 if (Xq.slt(0) || Yq.slt(0)) {
541 X->setEmpty();
542 ++DeltaSuccesses;
543 return true;
544 }
545 if (const SCEVConstant *CUB =
546 collectConstantUpperBound(X->getAssociatedLoop(), Prod1->getType())) {
547 const APInt &UpperBound = CUB->getAPInt();
548 LLVM_DEBUG(dbgs() << "\t\tupper bound = " << UpperBound << "\n");
549 if (Xq.sgt(UpperBound) || Yq.sgt(UpperBound)) {
550 X->setEmpty();
551 ++DeltaSuccesses;
552 return true;
553 }
554 }
555 X->setPoint(SE->getConstant(Xq),
556 SE->getConstant(Yq),
557 X->getAssociatedLoop());
558 ++DeltaSuccesses;
559 return true;
560 }
561 return false;
562 }
563
564 // if (X->isLine() && Y->isPoint()) This case can't occur.
565 assert(!(X->isLine() && Y->isPoint()) && "This case should never occur");
566
567 if (X->isPoint() && Y->isLine()) {
568 LLVM_DEBUG(dbgs() << "\t intersect Point and Line\n");
569 const SCEV *A1X1 = SE->getMulExpr(Y->getA(), X->getX());
570 const SCEV *B1Y1 = SE->getMulExpr(Y->getB(), X->getY());
571 const SCEV *Sum = SE->getAddExpr(A1X1, B1Y1);
572 if (isKnownPredicate(CmpInst::ICMP_EQ, Sum, Y->getC()))
573 return false;
574 if (isKnownPredicate(CmpInst::ICMP_NE, Sum, Y->getC())) {
575 X->setEmpty();
576 ++DeltaSuccesses;
577 return true;
578 }
579 return false;
580 }
581
582 llvm_unreachable("shouldn't reach the end of Constraint intersection");
583 return false;
584 }
585
586
587 //===----------------------------------------------------------------------===//
588 // DependenceInfo methods
589
590 // For debugging purposes. Dumps a dependence to OS.
dump(raw_ostream & OS) const591 void Dependence::dump(raw_ostream &OS) const {
592 bool Splitable = false;
593 if (isConfused())
594 OS << "confused";
595 else {
596 if (isConsistent())
597 OS << "consistent ";
598 if (isFlow())
599 OS << "flow";
600 else if (isOutput())
601 OS << "output";
602 else if (isAnti())
603 OS << "anti";
604 else if (isInput())
605 OS << "input";
606 unsigned Levels = getLevels();
607 OS << " [";
608 for (unsigned II = 1; II <= Levels; ++II) {
609 if (isSplitable(II))
610 Splitable = true;
611 if (isPeelFirst(II))
612 OS << 'p';
613 const SCEV *Distance = getDistance(II);
614 if (Distance)
615 OS << *Distance;
616 else if (isScalar(II))
617 OS << "S";
618 else {
619 unsigned Direction = getDirection(II);
620 if (Direction == DVEntry::ALL)
621 OS << "*";
622 else {
623 if (Direction & DVEntry::LT)
624 OS << "<";
625 if (Direction & DVEntry::EQ)
626 OS << "=";
627 if (Direction & DVEntry::GT)
628 OS << ">";
629 }
630 }
631 if (isPeelLast(II))
632 OS << 'p';
633 if (II < Levels)
634 OS << " ";
635 }
636 if (isLoopIndependent())
637 OS << "|<";
638 OS << "]";
639 if (Splitable)
640 OS << " splitable";
641 }
642 OS << "!\n";
643 }
644
645 // Returns NoAlias/MayAliass/MustAlias for two memory locations based upon their
646 // underlaying objects. If LocA and LocB are known to not alias (for any reason:
647 // tbaa, non-overlapping regions etc), then it is known there is no dependecy.
648 // Otherwise the underlying objects are checked to see if they point to
649 // different identifiable objects.
underlyingObjectsAlias(AAResults * AA,const DataLayout & DL,const MemoryLocation & LocA,const MemoryLocation & LocB)650 static AliasResult underlyingObjectsAlias(AAResults *AA,
651 const DataLayout &DL,
652 const MemoryLocation &LocA,
653 const MemoryLocation &LocB) {
654 // Check the original locations (minus size) for noalias, which can happen for
655 // tbaa, incompatible underlying object locations, etc.
656 MemoryLocation LocAS =
657 MemoryLocation::getBeforeOrAfter(LocA.Ptr, LocA.AATags);
658 MemoryLocation LocBS =
659 MemoryLocation::getBeforeOrAfter(LocB.Ptr, LocB.AATags);
660 if (AA->isNoAlias(LocAS, LocBS))
661 return AliasResult::NoAlias;
662
663 // Check the underlying objects are the same
664 const Value *AObj = getUnderlyingObject(LocA.Ptr);
665 const Value *BObj = getUnderlyingObject(LocB.Ptr);
666
667 // If the underlying objects are the same, they must alias
668 if (AObj == BObj)
669 return AliasResult::MustAlias;
670
671 // We may have hit the recursion limit for underlying objects, or have
672 // underlying objects where we don't know they will alias.
673 if (!isIdentifiedObject(AObj) || !isIdentifiedObject(BObj))
674 return AliasResult::MayAlias;
675
676 // Otherwise we know the objects are different and both identified objects so
677 // must not alias.
678 return AliasResult::NoAlias;
679 }
680
681
682 // Returns true if the load or store can be analyzed. Atomic and volatile
683 // operations have properties which this analysis does not understand.
684 static
isLoadOrStore(const Instruction * I)685 bool isLoadOrStore(const Instruction *I) {
686 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
687 return LI->isUnordered();
688 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
689 return SI->isUnordered();
690 return false;
691 }
692
693
694 // Examines the loop nesting of the Src and Dst
695 // instructions and establishes their shared loops. Sets the variables
696 // CommonLevels, SrcLevels, and MaxLevels.
697 // The source and destination instructions needn't be contained in the same
698 // loop. The routine establishNestingLevels finds the level of most deeply
699 // nested loop that contains them both, CommonLevels. An instruction that's
700 // not contained in a loop is at level = 0. MaxLevels is equal to the level
701 // of the source plus the level of the destination, minus CommonLevels.
702 // This lets us allocate vectors MaxLevels in length, with room for every
703 // distinct loop referenced in both the source and destination subscripts.
704 // The variable SrcLevels is the nesting depth of the source instruction.
705 // It's used to help calculate distinct loops referenced by the destination.
706 // Here's the map from loops to levels:
707 // 0 - unused
708 // 1 - outermost common loop
709 // ... - other common loops
710 // CommonLevels - innermost common loop
711 // ... - loops containing Src but not Dst
712 // SrcLevels - innermost loop containing Src but not Dst
713 // ... - loops containing Dst but not Src
714 // MaxLevels - innermost loops containing Dst but not Src
715 // Consider the follow code fragment:
716 // for (a = ...) {
717 // for (b = ...) {
718 // for (c = ...) {
719 // for (d = ...) {
720 // A[] = ...;
721 // }
722 // }
723 // for (e = ...) {
724 // for (f = ...) {
725 // for (g = ...) {
726 // ... = A[];
727 // }
728 // }
729 // }
730 // }
731 // }
732 // If we're looking at the possibility of a dependence between the store
733 // to A (the Src) and the load from A (the Dst), we'll note that they
734 // have 2 loops in common, so CommonLevels will equal 2 and the direction
735 // vector for Result will have 2 entries. SrcLevels = 4 and MaxLevels = 7.
736 // A map from loop names to loop numbers would look like
737 // a - 1
738 // b - 2 = CommonLevels
739 // c - 3
740 // d - 4 = SrcLevels
741 // e - 5
742 // f - 6
743 // g - 7 = MaxLevels
establishNestingLevels(const Instruction * Src,const Instruction * Dst)744 void DependenceInfo::establishNestingLevels(const Instruction *Src,
745 const Instruction *Dst) {
746 const BasicBlock *SrcBlock = Src->getParent();
747 const BasicBlock *DstBlock = Dst->getParent();
748 unsigned SrcLevel = LI->getLoopDepth(SrcBlock);
749 unsigned DstLevel = LI->getLoopDepth(DstBlock);
750 const Loop *SrcLoop = LI->getLoopFor(SrcBlock);
751 const Loop *DstLoop = LI->getLoopFor(DstBlock);
752 SrcLevels = SrcLevel;
753 MaxLevels = SrcLevel + DstLevel;
754 while (SrcLevel > DstLevel) {
755 SrcLoop = SrcLoop->getParentLoop();
756 SrcLevel--;
757 }
758 while (DstLevel > SrcLevel) {
759 DstLoop = DstLoop->getParentLoop();
760 DstLevel--;
761 }
762 while (SrcLoop != DstLoop) {
763 SrcLoop = SrcLoop->getParentLoop();
764 DstLoop = DstLoop->getParentLoop();
765 SrcLevel--;
766 }
767 CommonLevels = SrcLevel;
768 MaxLevels -= CommonLevels;
769 }
770
771
772 // Given one of the loops containing the source, return
773 // its level index in our numbering scheme.
mapSrcLoop(const Loop * SrcLoop) const774 unsigned DependenceInfo::mapSrcLoop(const Loop *SrcLoop) const {
775 return SrcLoop->getLoopDepth();
776 }
777
778
779 // Given one of the loops containing the destination,
780 // return its level index in our numbering scheme.
mapDstLoop(const Loop * DstLoop) const781 unsigned DependenceInfo::mapDstLoop(const Loop *DstLoop) const {
782 unsigned D = DstLoop->getLoopDepth();
783 if (D > CommonLevels)
784 return D - CommonLevels + SrcLevels;
785 else
786 return D;
787 }
788
789
790 // Returns true if Expression is loop invariant in LoopNest.
isLoopInvariant(const SCEV * Expression,const Loop * LoopNest) const791 bool DependenceInfo::isLoopInvariant(const SCEV *Expression,
792 const Loop *LoopNest) const {
793 if (!LoopNest)
794 return true;
795 return SE->isLoopInvariant(Expression, LoopNest) &&
796 isLoopInvariant(Expression, LoopNest->getParentLoop());
797 }
798
799
800
801 // Finds the set of loops from the LoopNest that
802 // have a level <= CommonLevels and are referred to by the SCEV Expression.
collectCommonLoops(const SCEV * Expression,const Loop * LoopNest,SmallBitVector & Loops) const803 void DependenceInfo::collectCommonLoops(const SCEV *Expression,
804 const Loop *LoopNest,
805 SmallBitVector &Loops) const {
806 while (LoopNest) {
807 unsigned Level = LoopNest->getLoopDepth();
808 if (Level <= CommonLevels && !SE->isLoopInvariant(Expression, LoopNest))
809 Loops.set(Level);
810 LoopNest = LoopNest->getParentLoop();
811 }
812 }
813
unifySubscriptType(ArrayRef<Subscript * > Pairs)814 void DependenceInfo::unifySubscriptType(ArrayRef<Subscript *> Pairs) {
815
816 unsigned widestWidthSeen = 0;
817 Type *widestType;
818
819 // Go through each pair and find the widest bit to which we need
820 // to extend all of them.
821 for (Subscript *Pair : Pairs) {
822 const SCEV *Src = Pair->Src;
823 const SCEV *Dst = Pair->Dst;
824 IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
825 IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
826 if (SrcTy == nullptr || DstTy == nullptr) {
827 assert(SrcTy == DstTy && "This function only unify integer types and "
828 "expect Src and Dst share the same type "
829 "otherwise.");
830 continue;
831 }
832 if (SrcTy->getBitWidth() > widestWidthSeen) {
833 widestWidthSeen = SrcTy->getBitWidth();
834 widestType = SrcTy;
835 }
836 if (DstTy->getBitWidth() > widestWidthSeen) {
837 widestWidthSeen = DstTy->getBitWidth();
838 widestType = DstTy;
839 }
840 }
841
842
843 assert(widestWidthSeen > 0);
844
845 // Now extend each pair to the widest seen.
846 for (Subscript *Pair : Pairs) {
847 const SCEV *Src = Pair->Src;
848 const SCEV *Dst = Pair->Dst;
849 IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
850 IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
851 if (SrcTy == nullptr || DstTy == nullptr) {
852 assert(SrcTy == DstTy && "This function only unify integer types and "
853 "expect Src and Dst share the same type "
854 "otherwise.");
855 continue;
856 }
857 if (SrcTy->getBitWidth() < widestWidthSeen)
858 // Sign-extend Src to widestType
859 Pair->Src = SE->getSignExtendExpr(Src, widestType);
860 if (DstTy->getBitWidth() < widestWidthSeen) {
861 // Sign-extend Dst to widestType
862 Pair->Dst = SE->getSignExtendExpr(Dst, widestType);
863 }
864 }
865 }
866
867 // removeMatchingExtensions - Examines a subscript pair.
868 // If the source and destination are identically sign (or zero)
869 // extended, it strips off the extension in an effect to simplify
870 // the actual analysis.
removeMatchingExtensions(Subscript * Pair)871 void DependenceInfo::removeMatchingExtensions(Subscript *Pair) {
872 const SCEV *Src = Pair->Src;
873 const SCEV *Dst = Pair->Dst;
874 if ((isa<SCEVZeroExtendExpr>(Src) && isa<SCEVZeroExtendExpr>(Dst)) ||
875 (isa<SCEVSignExtendExpr>(Src) && isa<SCEVSignExtendExpr>(Dst))) {
876 const SCEVIntegralCastExpr *SrcCast = cast<SCEVIntegralCastExpr>(Src);
877 const SCEVIntegralCastExpr *DstCast = cast<SCEVIntegralCastExpr>(Dst);
878 const SCEV *SrcCastOp = SrcCast->getOperand();
879 const SCEV *DstCastOp = DstCast->getOperand();
880 if (SrcCastOp->getType() == DstCastOp->getType()) {
881 Pair->Src = SrcCastOp;
882 Pair->Dst = DstCastOp;
883 }
884 }
885 }
886
887 // Examine the scev and return true iff it's linear.
888 // Collect any loops mentioned in the set of "Loops".
checkSubscript(const SCEV * Expr,const Loop * LoopNest,SmallBitVector & Loops,bool IsSrc)889 bool DependenceInfo::checkSubscript(const SCEV *Expr, const Loop *LoopNest,
890 SmallBitVector &Loops, bool IsSrc) {
891 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
892 if (!AddRec)
893 return isLoopInvariant(Expr, LoopNest);
894 const SCEV *Start = AddRec->getStart();
895 const SCEV *Step = AddRec->getStepRecurrence(*SE);
896 const SCEV *UB = SE->getBackedgeTakenCount(AddRec->getLoop());
897 if (!isa<SCEVCouldNotCompute>(UB)) {
898 if (SE->getTypeSizeInBits(Start->getType()) <
899 SE->getTypeSizeInBits(UB->getType())) {
900 if (!AddRec->getNoWrapFlags())
901 return false;
902 }
903 }
904 if (!isLoopInvariant(Step, LoopNest))
905 return false;
906 if (IsSrc)
907 Loops.set(mapSrcLoop(AddRec->getLoop()));
908 else
909 Loops.set(mapDstLoop(AddRec->getLoop()));
910 return checkSubscript(Start, LoopNest, Loops, IsSrc);
911 }
912
913 // Examine the scev and return true iff it's linear.
914 // Collect any loops mentioned in the set of "Loops".
checkSrcSubscript(const SCEV * Src,const Loop * LoopNest,SmallBitVector & Loops)915 bool DependenceInfo::checkSrcSubscript(const SCEV *Src, const Loop *LoopNest,
916 SmallBitVector &Loops) {
917 return checkSubscript(Src, LoopNest, Loops, true);
918 }
919
920 // Examine the scev and return true iff it's linear.
921 // Collect any loops mentioned in the set of "Loops".
checkDstSubscript(const SCEV * Dst,const Loop * LoopNest,SmallBitVector & Loops)922 bool DependenceInfo::checkDstSubscript(const SCEV *Dst, const Loop *LoopNest,
923 SmallBitVector &Loops) {
924 return checkSubscript(Dst, LoopNest, Loops, false);
925 }
926
927
928 // Examines the subscript pair (the Src and Dst SCEVs)
929 // and classifies it as either ZIV, SIV, RDIV, MIV, or Nonlinear.
930 // Collects the associated loops in a set.
931 DependenceInfo::Subscript::ClassificationKind
classifyPair(const SCEV * Src,const Loop * SrcLoopNest,const SCEV * Dst,const Loop * DstLoopNest,SmallBitVector & Loops)932 DependenceInfo::classifyPair(const SCEV *Src, const Loop *SrcLoopNest,
933 const SCEV *Dst, const Loop *DstLoopNest,
934 SmallBitVector &Loops) {
935 SmallBitVector SrcLoops(MaxLevels + 1);
936 SmallBitVector DstLoops(MaxLevels + 1);
937 if (!checkSrcSubscript(Src, SrcLoopNest, SrcLoops))
938 return Subscript::NonLinear;
939 if (!checkDstSubscript(Dst, DstLoopNest, DstLoops))
940 return Subscript::NonLinear;
941 Loops = SrcLoops;
942 Loops |= DstLoops;
943 unsigned N = Loops.count();
944 if (N == 0)
945 return Subscript::ZIV;
946 if (N == 1)
947 return Subscript::SIV;
948 if (N == 2 && (SrcLoops.count() == 0 ||
949 DstLoops.count() == 0 ||
950 (SrcLoops.count() == 1 && DstLoops.count() == 1)))
951 return Subscript::RDIV;
952 return Subscript::MIV;
953 }
954
955
956 // A wrapper around SCEV::isKnownPredicate.
957 // Looks for cases where we're interested in comparing for equality.
958 // If both X and Y have been identically sign or zero extended,
959 // it strips off the (confusing) extensions before invoking
960 // SCEV::isKnownPredicate. Perhaps, someday, the ScalarEvolution package
961 // will be similarly updated.
962 //
963 // If SCEV::isKnownPredicate can't prove the predicate,
964 // we try simple subtraction, which seems to help in some cases
965 // involving symbolics.
isKnownPredicate(ICmpInst::Predicate Pred,const SCEV * X,const SCEV * Y) const966 bool DependenceInfo::isKnownPredicate(ICmpInst::Predicate Pred, const SCEV *X,
967 const SCEV *Y) const {
968 if (Pred == CmpInst::ICMP_EQ ||
969 Pred == CmpInst::ICMP_NE) {
970 if ((isa<SCEVSignExtendExpr>(X) &&
971 isa<SCEVSignExtendExpr>(Y)) ||
972 (isa<SCEVZeroExtendExpr>(X) &&
973 isa<SCEVZeroExtendExpr>(Y))) {
974 const SCEVIntegralCastExpr *CX = cast<SCEVIntegralCastExpr>(X);
975 const SCEVIntegralCastExpr *CY = cast<SCEVIntegralCastExpr>(Y);
976 const SCEV *Xop = CX->getOperand();
977 const SCEV *Yop = CY->getOperand();
978 if (Xop->getType() == Yop->getType()) {
979 X = Xop;
980 Y = Yop;
981 }
982 }
983 }
984 if (SE->isKnownPredicate(Pred, X, Y))
985 return true;
986 // If SE->isKnownPredicate can't prove the condition,
987 // we try the brute-force approach of subtracting
988 // and testing the difference.
989 // By testing with SE->isKnownPredicate first, we avoid
990 // the possibility of overflow when the arguments are constants.
991 const SCEV *Delta = SE->getMinusSCEV(X, Y);
992 switch (Pred) {
993 case CmpInst::ICMP_EQ:
994 return Delta->isZero();
995 case CmpInst::ICMP_NE:
996 return SE->isKnownNonZero(Delta);
997 case CmpInst::ICMP_SGE:
998 return SE->isKnownNonNegative(Delta);
999 case CmpInst::ICMP_SLE:
1000 return SE->isKnownNonPositive(Delta);
1001 case CmpInst::ICMP_SGT:
1002 return SE->isKnownPositive(Delta);
1003 case CmpInst::ICMP_SLT:
1004 return SE->isKnownNegative(Delta);
1005 default:
1006 llvm_unreachable("unexpected predicate in isKnownPredicate");
1007 }
1008 }
1009
1010 /// Compare to see if S is less than Size, using isKnownNegative(S - max(Size, 1))
1011 /// with some extra checking if S is an AddRec and we can prove less-than using
1012 /// the loop bounds.
isKnownLessThan(const SCEV * S,const SCEV * Size) const1013 bool DependenceInfo::isKnownLessThan(const SCEV *S, const SCEV *Size) const {
1014 // First unify to the same type
1015 auto *SType = dyn_cast<IntegerType>(S->getType());
1016 auto *SizeType = dyn_cast<IntegerType>(Size->getType());
1017 if (!SType || !SizeType)
1018 return false;
1019 Type *MaxType =
1020 (SType->getBitWidth() >= SizeType->getBitWidth()) ? SType : SizeType;
1021 S = SE->getTruncateOrZeroExtend(S, MaxType);
1022 Size = SE->getTruncateOrZeroExtend(Size, MaxType);
1023
1024 // Special check for addrecs using BE taken count
1025 const SCEV *Bound = SE->getMinusSCEV(S, Size);
1026 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Bound)) {
1027 if (AddRec->isAffine()) {
1028 const SCEV *BECount = SE->getBackedgeTakenCount(AddRec->getLoop());
1029 if (!isa<SCEVCouldNotCompute>(BECount)) {
1030 const SCEV *Limit = AddRec->evaluateAtIteration(BECount, *SE);
1031 if (SE->isKnownNegative(Limit))
1032 return true;
1033 }
1034 }
1035 }
1036
1037 // Check using normal isKnownNegative
1038 const SCEV *LimitedBound =
1039 SE->getMinusSCEV(S, SE->getSMaxExpr(Size, SE->getOne(Size->getType())));
1040 return SE->isKnownNegative(LimitedBound);
1041 }
1042
isKnownNonNegative(const SCEV * S,const Value * Ptr) const1043 bool DependenceInfo::isKnownNonNegative(const SCEV *S, const Value *Ptr) const {
1044 bool Inbounds = false;
1045 if (auto *SrcGEP = dyn_cast<GetElementPtrInst>(Ptr))
1046 Inbounds = SrcGEP->isInBounds();
1047 if (Inbounds) {
1048 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
1049 if (AddRec->isAffine()) {
1050 // We know S is for Ptr, the operand on a load/store, so doesn't wrap.
1051 // If both parts are NonNegative, the end result will be NonNegative
1052 if (SE->isKnownNonNegative(AddRec->getStart()) &&
1053 SE->isKnownNonNegative(AddRec->getOperand(1)))
1054 return true;
1055 }
1056 }
1057 }
1058
1059 return SE->isKnownNonNegative(S);
1060 }
1061
1062 // All subscripts are all the same type.
1063 // Loop bound may be smaller (e.g., a char).
1064 // Should zero extend loop bound, since it's always >= 0.
1065 // This routine collects upper bound and extends or truncates if needed.
1066 // Truncating is safe when subscripts are known not to wrap. Cases without
1067 // nowrap flags should have been rejected earlier.
1068 // Return null if no bound available.
collectUpperBound(const Loop * L,Type * T) const1069 const SCEV *DependenceInfo::collectUpperBound(const Loop *L, Type *T) const {
1070 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
1071 const SCEV *UB = SE->getBackedgeTakenCount(L);
1072 return SE->getTruncateOrZeroExtend(UB, T);
1073 }
1074 return nullptr;
1075 }
1076
1077
1078 // Calls collectUpperBound(), then attempts to cast it to SCEVConstant.
1079 // If the cast fails, returns NULL.
collectConstantUpperBound(const Loop * L,Type * T) const1080 const SCEVConstant *DependenceInfo::collectConstantUpperBound(const Loop *L,
1081 Type *T) const {
1082 if (const SCEV *UB = collectUpperBound(L, T))
1083 return dyn_cast<SCEVConstant>(UB);
1084 return nullptr;
1085 }
1086
1087
1088 // testZIV -
1089 // When we have a pair of subscripts of the form [c1] and [c2],
1090 // where c1 and c2 are both loop invariant, we attack it using
1091 // the ZIV test. Basically, we test by comparing the two values,
1092 // but there are actually three possible results:
1093 // 1) the values are equal, so there's a dependence
1094 // 2) the values are different, so there's no dependence
1095 // 3) the values might be equal, so we have to assume a dependence.
1096 //
1097 // Return true if dependence disproved.
testZIV(const SCEV * Src,const SCEV * Dst,FullDependence & Result) const1098 bool DependenceInfo::testZIV(const SCEV *Src, const SCEV *Dst,
1099 FullDependence &Result) const {
1100 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
1101 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
1102 ++ZIVapplications;
1103 if (isKnownPredicate(CmpInst::ICMP_EQ, Src, Dst)) {
1104 LLVM_DEBUG(dbgs() << " provably dependent\n");
1105 return false; // provably dependent
1106 }
1107 if (isKnownPredicate(CmpInst::ICMP_NE, Src, Dst)) {
1108 LLVM_DEBUG(dbgs() << " provably independent\n");
1109 ++ZIVindependence;
1110 return true; // provably independent
1111 }
1112 LLVM_DEBUG(dbgs() << " possibly dependent\n");
1113 Result.Consistent = false;
1114 return false; // possibly dependent
1115 }
1116
1117
1118 // strongSIVtest -
1119 // From the paper, Practical Dependence Testing, Section 4.2.1
1120 //
1121 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 + a*i],
1122 // where i is an induction variable, c1 and c2 are loop invariant,
1123 // and a is a constant, we can solve it exactly using the Strong SIV test.
1124 //
1125 // Can prove independence. Failing that, can compute distance (and direction).
1126 // In the presence of symbolic terms, we can sometimes make progress.
1127 //
1128 // If there's a dependence,
1129 //
1130 // c1 + a*i = c2 + a*i'
1131 //
1132 // The dependence distance is
1133 //
1134 // d = i' - i = (c1 - c2)/a
1135 //
1136 // A dependence only exists if d is an integer and abs(d) <= U, where U is the
1137 // loop's upper bound. If a dependence exists, the dependence direction is
1138 // defined as
1139 //
1140 // { < if d > 0
1141 // direction = { = if d = 0
1142 // { > if d < 0
1143 //
1144 // Return true if dependence disproved.
strongSIVtest(const SCEV * Coeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1145 bool DependenceInfo::strongSIVtest(const SCEV *Coeff, const SCEV *SrcConst,
1146 const SCEV *DstConst, const Loop *CurLoop,
1147 unsigned Level, FullDependence &Result,
1148 Constraint &NewConstraint) const {
1149 LLVM_DEBUG(dbgs() << "\tStrong SIV test\n");
1150 LLVM_DEBUG(dbgs() << "\t Coeff = " << *Coeff);
1151 LLVM_DEBUG(dbgs() << ", " << *Coeff->getType() << "\n");
1152 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst);
1153 LLVM_DEBUG(dbgs() << ", " << *SrcConst->getType() << "\n");
1154 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst);
1155 LLVM_DEBUG(dbgs() << ", " << *DstConst->getType() << "\n");
1156 ++StrongSIVapplications;
1157 assert(0 < Level && Level <= CommonLevels && "level out of range");
1158 Level--;
1159
1160 const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst);
1161 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta);
1162 LLVM_DEBUG(dbgs() << ", " << *Delta->getType() << "\n");
1163
1164 // check that |Delta| < iteration count
1165 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1166 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound);
1167 LLVM_DEBUG(dbgs() << ", " << *UpperBound->getType() << "\n");
1168 const SCEV *AbsDelta =
1169 SE->isKnownNonNegative(Delta) ? Delta : SE->getNegativeSCEV(Delta);
1170 const SCEV *AbsCoeff =
1171 SE->isKnownNonNegative(Coeff) ? Coeff : SE->getNegativeSCEV(Coeff);
1172 const SCEV *Product = SE->getMulExpr(UpperBound, AbsCoeff);
1173 if (isKnownPredicate(CmpInst::ICMP_SGT, AbsDelta, Product)) {
1174 // Distance greater than trip count - no dependence
1175 ++StrongSIVindependence;
1176 ++StrongSIVsuccesses;
1177 return true;
1178 }
1179 }
1180
1181 // Can we compute distance?
1182 if (isa<SCEVConstant>(Delta) && isa<SCEVConstant>(Coeff)) {
1183 APInt ConstDelta = cast<SCEVConstant>(Delta)->getAPInt();
1184 APInt ConstCoeff = cast<SCEVConstant>(Coeff)->getAPInt();
1185 APInt Distance = ConstDelta; // these need to be initialized
1186 APInt Remainder = ConstDelta;
1187 APInt::sdivrem(ConstDelta, ConstCoeff, Distance, Remainder);
1188 LLVM_DEBUG(dbgs() << "\t Distance = " << Distance << "\n");
1189 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n");
1190 // Make sure Coeff divides Delta exactly
1191 if (Remainder != 0) {
1192 // Coeff doesn't divide Distance, no dependence
1193 ++StrongSIVindependence;
1194 ++StrongSIVsuccesses;
1195 return true;
1196 }
1197 Result.DV[Level].Distance = SE->getConstant(Distance);
1198 NewConstraint.setDistance(SE->getConstant(Distance), CurLoop);
1199 if (Distance.sgt(0))
1200 Result.DV[Level].Direction &= Dependence::DVEntry::LT;
1201 else if (Distance.slt(0))
1202 Result.DV[Level].Direction &= Dependence::DVEntry::GT;
1203 else
1204 Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
1205 ++StrongSIVsuccesses;
1206 }
1207 else if (Delta->isZero()) {
1208 // since 0/X == 0
1209 Result.DV[Level].Distance = Delta;
1210 NewConstraint.setDistance(Delta, CurLoop);
1211 Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
1212 ++StrongSIVsuccesses;
1213 }
1214 else {
1215 if (Coeff->isOne()) {
1216 LLVM_DEBUG(dbgs() << "\t Distance = " << *Delta << "\n");
1217 Result.DV[Level].Distance = Delta; // since X/1 == X
1218 NewConstraint.setDistance(Delta, CurLoop);
1219 }
1220 else {
1221 Result.Consistent = false;
1222 NewConstraint.setLine(Coeff,
1223 SE->getNegativeSCEV(Coeff),
1224 SE->getNegativeSCEV(Delta), CurLoop);
1225 }
1226
1227 // maybe we can get a useful direction
1228 bool DeltaMaybeZero = !SE->isKnownNonZero(Delta);
1229 bool DeltaMaybePositive = !SE->isKnownNonPositive(Delta);
1230 bool DeltaMaybeNegative = !SE->isKnownNonNegative(Delta);
1231 bool CoeffMaybePositive = !SE->isKnownNonPositive(Coeff);
1232 bool CoeffMaybeNegative = !SE->isKnownNonNegative(Coeff);
1233 // The double negatives above are confusing.
1234 // It helps to read !SE->isKnownNonZero(Delta)
1235 // as "Delta might be Zero"
1236 unsigned NewDirection = Dependence::DVEntry::NONE;
1237 if ((DeltaMaybePositive && CoeffMaybePositive) ||
1238 (DeltaMaybeNegative && CoeffMaybeNegative))
1239 NewDirection = Dependence::DVEntry::LT;
1240 if (DeltaMaybeZero)
1241 NewDirection |= Dependence::DVEntry::EQ;
1242 if ((DeltaMaybeNegative && CoeffMaybePositive) ||
1243 (DeltaMaybePositive && CoeffMaybeNegative))
1244 NewDirection |= Dependence::DVEntry::GT;
1245 if (NewDirection < Result.DV[Level].Direction)
1246 ++StrongSIVsuccesses;
1247 Result.DV[Level].Direction &= NewDirection;
1248 }
1249 return false;
1250 }
1251
1252
1253 // weakCrossingSIVtest -
1254 // From the paper, Practical Dependence Testing, Section 4.2.2
1255 //
1256 // When we have a pair of subscripts of the form [c1 + a*i] and [c2 - a*i],
1257 // where i is an induction variable, c1 and c2 are loop invariant,
1258 // and a is a constant, we can solve it exactly using the
1259 // Weak-Crossing SIV test.
1260 //
1261 // Given c1 + a*i = c2 - a*i', we can look for the intersection of
1262 // the two lines, where i = i', yielding
1263 //
1264 // c1 + a*i = c2 - a*i
1265 // 2a*i = c2 - c1
1266 // i = (c2 - c1)/2a
1267 //
1268 // If i < 0, there is no dependence.
1269 // If i > upperbound, there is no dependence.
1270 // If i = 0 (i.e., if c1 = c2), there's a dependence with distance = 0.
1271 // If i = upperbound, there's a dependence with distance = 0.
1272 // If i is integral, there's a dependence (all directions).
1273 // If the non-integer part = 1/2, there's a dependence (<> directions).
1274 // Otherwise, there's no dependence.
1275 //
1276 // Can prove independence. Failing that,
1277 // can sometimes refine the directions.
1278 // Can determine iteration for splitting.
1279 //
1280 // Return true if dependence disproved.
weakCrossingSIVtest(const SCEV * Coeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint,const SCEV * & SplitIter) const1281 bool DependenceInfo::weakCrossingSIVtest(
1282 const SCEV *Coeff, const SCEV *SrcConst, const SCEV *DstConst,
1283 const Loop *CurLoop, unsigned Level, FullDependence &Result,
1284 Constraint &NewConstraint, const SCEV *&SplitIter) const {
1285 LLVM_DEBUG(dbgs() << "\tWeak-Crossing SIV test\n");
1286 LLVM_DEBUG(dbgs() << "\t Coeff = " << *Coeff << "\n");
1287 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1288 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1289 ++WeakCrossingSIVapplications;
1290 assert(0 < Level && Level <= CommonLevels && "Level out of range");
1291 Level--;
1292 Result.Consistent = false;
1293 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1294 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1295 NewConstraint.setLine(Coeff, Coeff, Delta, CurLoop);
1296 if (Delta->isZero()) {
1297 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT);
1298 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT);
1299 ++WeakCrossingSIVsuccesses;
1300 if (!Result.DV[Level].Direction) {
1301 ++WeakCrossingSIVindependence;
1302 return true;
1303 }
1304 Result.DV[Level].Distance = Delta; // = 0
1305 return false;
1306 }
1307 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(Coeff);
1308 if (!ConstCoeff)
1309 return false;
1310
1311 Result.DV[Level].Splitable = true;
1312 if (SE->isKnownNegative(ConstCoeff)) {
1313 ConstCoeff = dyn_cast<SCEVConstant>(SE->getNegativeSCEV(ConstCoeff));
1314 assert(ConstCoeff &&
1315 "dynamic cast of negative of ConstCoeff should yield constant");
1316 Delta = SE->getNegativeSCEV(Delta);
1317 }
1318 assert(SE->isKnownPositive(ConstCoeff) && "ConstCoeff should be positive");
1319
1320 // compute SplitIter for use by DependenceInfo::getSplitIteration()
1321 SplitIter = SE->getUDivExpr(
1322 SE->getSMaxExpr(SE->getZero(Delta->getType()), Delta),
1323 SE->getMulExpr(SE->getConstant(Delta->getType(), 2), ConstCoeff));
1324 LLVM_DEBUG(dbgs() << "\t Split iter = " << *SplitIter << "\n");
1325
1326 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1327 if (!ConstDelta)
1328 return false;
1329
1330 // We're certain that ConstCoeff > 0; therefore,
1331 // if Delta < 0, then no dependence.
1332 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1333 LLVM_DEBUG(dbgs() << "\t ConstCoeff = " << *ConstCoeff << "\n");
1334 if (SE->isKnownNegative(Delta)) {
1335 // No dependence, Delta < 0
1336 ++WeakCrossingSIVindependence;
1337 ++WeakCrossingSIVsuccesses;
1338 return true;
1339 }
1340
1341 // We're certain that Delta > 0 and ConstCoeff > 0.
1342 // Check Delta/(2*ConstCoeff) against upper loop bound
1343 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1344 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n");
1345 const SCEV *ConstantTwo = SE->getConstant(UpperBound->getType(), 2);
1346 const SCEV *ML = SE->getMulExpr(SE->getMulExpr(ConstCoeff, UpperBound),
1347 ConstantTwo);
1348 LLVM_DEBUG(dbgs() << "\t ML = " << *ML << "\n");
1349 if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, ML)) {
1350 // Delta too big, no dependence
1351 ++WeakCrossingSIVindependence;
1352 ++WeakCrossingSIVsuccesses;
1353 return true;
1354 }
1355 if (isKnownPredicate(CmpInst::ICMP_EQ, Delta, ML)) {
1356 // i = i' = UB
1357 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::LT);
1358 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::GT);
1359 ++WeakCrossingSIVsuccesses;
1360 if (!Result.DV[Level].Direction) {
1361 ++WeakCrossingSIVindependence;
1362 return true;
1363 }
1364 Result.DV[Level].Splitable = false;
1365 Result.DV[Level].Distance = SE->getZero(Delta->getType());
1366 return false;
1367 }
1368 }
1369
1370 // check that Coeff divides Delta
1371 APInt APDelta = ConstDelta->getAPInt();
1372 APInt APCoeff = ConstCoeff->getAPInt();
1373 APInt Distance = APDelta; // these need to be initialzed
1374 APInt Remainder = APDelta;
1375 APInt::sdivrem(APDelta, APCoeff, Distance, Remainder);
1376 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n");
1377 if (Remainder != 0) {
1378 // Coeff doesn't divide Delta, no dependence
1379 ++WeakCrossingSIVindependence;
1380 ++WeakCrossingSIVsuccesses;
1381 return true;
1382 }
1383 LLVM_DEBUG(dbgs() << "\t Distance = " << Distance << "\n");
1384
1385 // if 2*Coeff doesn't divide Delta, then the equal direction isn't possible
1386 APInt Two = APInt(Distance.getBitWidth(), 2, true);
1387 Remainder = Distance.srem(Two);
1388 LLVM_DEBUG(dbgs() << "\t Remainder = " << Remainder << "\n");
1389 if (Remainder != 0) {
1390 // Equal direction isn't possible
1391 Result.DV[Level].Direction &= unsigned(~Dependence::DVEntry::EQ);
1392 ++WeakCrossingSIVsuccesses;
1393 }
1394 return false;
1395 }
1396
1397
1398 // Kirch's algorithm, from
1399 //
1400 // Optimizing Supercompilers for Supercomputers
1401 // Michael Wolfe
1402 // MIT Press, 1989
1403 //
1404 // Program 2.1, page 29.
1405 // Computes the GCD of AM and BM.
1406 // Also finds a solution to the equation ax - by = gcd(a, b).
1407 // Returns true if dependence disproved; i.e., gcd does not divide Delta.
findGCD(unsigned Bits,const APInt & AM,const APInt & BM,const APInt & Delta,APInt & G,APInt & X,APInt & Y)1408 static bool findGCD(unsigned Bits, const APInt &AM, const APInt &BM,
1409 const APInt &Delta, APInt &G, APInt &X, APInt &Y) {
1410 APInt A0(Bits, 1, true), A1(Bits, 0, true);
1411 APInt B0(Bits, 0, true), B1(Bits, 1, true);
1412 APInt G0 = AM.abs();
1413 APInt G1 = BM.abs();
1414 APInt Q = G0; // these need to be initialized
1415 APInt R = G0;
1416 APInt::sdivrem(G0, G1, Q, R);
1417 while (R != 0) {
1418 APInt A2 = A0 - Q*A1; A0 = A1; A1 = A2;
1419 APInt B2 = B0 - Q*B1; B0 = B1; B1 = B2;
1420 G0 = G1; G1 = R;
1421 APInt::sdivrem(G0, G1, Q, R);
1422 }
1423 G = G1;
1424 LLVM_DEBUG(dbgs() << "\t GCD = " << G << "\n");
1425 X = AM.slt(0) ? -A1 : A1;
1426 Y = BM.slt(0) ? B1 : -B1;
1427
1428 // make sure gcd divides Delta
1429 R = Delta.srem(G);
1430 if (R != 0)
1431 return true; // gcd doesn't divide Delta, no dependence
1432 Q = Delta.sdiv(G);
1433 return false;
1434 }
1435
floorOfQuotient(const APInt & A,const APInt & B)1436 static APInt floorOfQuotient(const APInt &A, const APInt &B) {
1437 APInt Q = A; // these need to be initialized
1438 APInt R = A;
1439 APInt::sdivrem(A, B, Q, R);
1440 if (R == 0)
1441 return Q;
1442 if ((A.sgt(0) && B.sgt(0)) ||
1443 (A.slt(0) && B.slt(0)))
1444 return Q;
1445 else
1446 return Q - 1;
1447 }
1448
ceilingOfQuotient(const APInt & A,const APInt & B)1449 static APInt ceilingOfQuotient(const APInt &A, const APInt &B) {
1450 APInt Q = A; // these need to be initialized
1451 APInt R = A;
1452 APInt::sdivrem(A, B, Q, R);
1453 if (R == 0)
1454 return Q;
1455 if ((A.sgt(0) && B.sgt(0)) ||
1456 (A.slt(0) && B.slt(0)))
1457 return Q + 1;
1458 else
1459 return Q;
1460 }
1461
1462 // exactSIVtest -
1463 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*i],
1464 // where i is an induction variable, c1 and c2 are loop invariant, and a1
1465 // and a2 are constant, we can solve it exactly using an algorithm developed
1466 // by Banerjee and Wolfe. See Algorithm 6.2.1 (case 2.5) in:
1467 //
1468 // Dependence Analysis for Supercomputing
1469 // Utpal Banerjee
1470 // Kluwer Academic Publishers, 1988
1471 //
1472 // It's slower than the specialized tests (strong SIV, weak-zero SIV, etc),
1473 // so use them if possible. They're also a bit better with symbolics and,
1474 // in the case of the strong SIV test, can compute Distances.
1475 //
1476 // Return true if dependence disproved.
1477 //
1478 // This is a modified version of the original Banerjee algorithm. The original
1479 // only tested whether Dst depends on Src. This algorithm extends that and
1480 // returns all the dependencies that exist between Dst and Src.
exactSIVtest(const SCEV * SrcCoeff,const SCEV * DstCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1481 bool DependenceInfo::exactSIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff,
1482 const SCEV *SrcConst, const SCEV *DstConst,
1483 const Loop *CurLoop, unsigned Level,
1484 FullDependence &Result,
1485 Constraint &NewConstraint) const {
1486 LLVM_DEBUG(dbgs() << "\tExact SIV test\n");
1487 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << " = AM\n");
1488 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << " = BM\n");
1489 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1490 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1491 ++ExactSIVapplications;
1492 assert(0 < Level && Level <= CommonLevels && "Level out of range");
1493 Level--;
1494 Result.Consistent = false;
1495 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1496 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1497 NewConstraint.setLine(SrcCoeff, SE->getNegativeSCEV(DstCoeff), Delta,
1498 CurLoop);
1499 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1500 const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1501 const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1502 if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff)
1503 return false;
1504
1505 // find gcd
1506 APInt G, X, Y;
1507 APInt AM = ConstSrcCoeff->getAPInt();
1508 APInt BM = ConstDstCoeff->getAPInt();
1509 APInt CM = ConstDelta->getAPInt();
1510 unsigned Bits = AM.getBitWidth();
1511 if (findGCD(Bits, AM, BM, CM, G, X, Y)) {
1512 // gcd doesn't divide Delta, no dependence
1513 ++ExactSIVindependence;
1514 ++ExactSIVsuccesses;
1515 return true;
1516 }
1517
1518 LLVM_DEBUG(dbgs() << "\t X = " << X << ", Y = " << Y << "\n");
1519
1520 // since SCEV construction normalizes, LM = 0
1521 APInt UM(Bits, 1, true);
1522 bool UMValid = false;
1523 // UM is perhaps unavailable, let's check
1524 if (const SCEVConstant *CUB =
1525 collectConstantUpperBound(CurLoop, Delta->getType())) {
1526 UM = CUB->getAPInt();
1527 LLVM_DEBUG(dbgs() << "\t UM = " << UM << "\n");
1528 UMValid = true;
1529 }
1530
1531 APInt TU(APInt::getSignedMaxValue(Bits));
1532 APInt TL(APInt::getSignedMinValue(Bits));
1533 APInt TC = CM.sdiv(G);
1534 APInt TX = X * TC;
1535 APInt TY = Y * TC;
1536 LLVM_DEBUG(dbgs() << "\t TC = " << TC << "\n");
1537 LLVM_DEBUG(dbgs() << "\t TX = " << TX << "\n");
1538 LLVM_DEBUG(dbgs() << "\t TY = " << TY << "\n");
1539
1540 SmallVector<APInt, 2> TLVec, TUVec;
1541 APInt TB = BM.sdiv(G);
1542 if (TB.sgt(0)) {
1543 TLVec.push_back(ceilingOfQuotient(-TX, TB));
1544 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1545 // New bound check - modification to Banerjee's e3 check
1546 if (UMValid) {
1547 TUVec.push_back(floorOfQuotient(UM - TX, TB));
1548 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1549 }
1550 } else {
1551 TUVec.push_back(floorOfQuotient(-TX, TB));
1552 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1553 // New bound check - modification to Banerjee's e3 check
1554 if (UMValid) {
1555 TLVec.push_back(ceilingOfQuotient(UM - TX, TB));
1556 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1557 }
1558 }
1559
1560 APInt TA = AM.sdiv(G);
1561 if (TA.sgt(0)) {
1562 if (UMValid) {
1563 TUVec.push_back(floorOfQuotient(UM - TY, TA));
1564 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1565 }
1566 // New bound check - modification to Banerjee's e3 check
1567 TLVec.push_back(ceilingOfQuotient(-TY, TA));
1568 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1569 } else {
1570 if (UMValid) {
1571 TLVec.push_back(ceilingOfQuotient(UM - TY, TA));
1572 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1573 }
1574 // New bound check - modification to Banerjee's e3 check
1575 TUVec.push_back(floorOfQuotient(-TY, TA));
1576 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1577 }
1578
1579 LLVM_DEBUG(dbgs() << "\t TA = " << TA << "\n");
1580 LLVM_DEBUG(dbgs() << "\t TB = " << TB << "\n");
1581
1582 if (TLVec.empty() || TUVec.empty())
1583 return false;
1584 TL = APIntOps::smax(TLVec.front(), TLVec.back());
1585 TU = APIntOps::smin(TUVec.front(), TUVec.back());
1586 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1587 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1588
1589 if (TL.sgt(TU)) {
1590 ++ExactSIVindependence;
1591 ++ExactSIVsuccesses;
1592 return true;
1593 }
1594
1595 // explore directions
1596 unsigned NewDirection = Dependence::DVEntry::NONE;
1597 APInt LowerDistance, UpperDistance;
1598 if (TA.sgt(TB)) {
1599 LowerDistance = (TY - TX) + (TA - TB) * TL;
1600 UpperDistance = (TY - TX) + (TA - TB) * TU;
1601 } else {
1602 LowerDistance = (TY - TX) + (TA - TB) * TU;
1603 UpperDistance = (TY - TX) + (TA - TB) * TL;
1604 }
1605
1606 LLVM_DEBUG(dbgs() << "\t LowerDistance = " << LowerDistance << "\n");
1607 LLVM_DEBUG(dbgs() << "\t UpperDistance = " << UpperDistance << "\n");
1608
1609 APInt Zero(Bits, 0, true);
1610 if (LowerDistance.sle(Zero) && UpperDistance.sge(Zero)) {
1611 NewDirection |= Dependence::DVEntry::EQ;
1612 ++ExactSIVsuccesses;
1613 }
1614 if (LowerDistance.slt(0)) {
1615 NewDirection |= Dependence::DVEntry::GT;
1616 ++ExactSIVsuccesses;
1617 }
1618 if (UpperDistance.sgt(0)) {
1619 NewDirection |= Dependence::DVEntry::LT;
1620 ++ExactSIVsuccesses;
1621 }
1622
1623 // finished
1624 Result.DV[Level].Direction &= NewDirection;
1625 if (Result.DV[Level].Direction == Dependence::DVEntry::NONE)
1626 ++ExactSIVindependence;
1627 LLVM_DEBUG(dbgs() << "\t Result = ");
1628 LLVM_DEBUG(Result.dump(dbgs()));
1629 return Result.DV[Level].Direction == Dependence::DVEntry::NONE;
1630 }
1631
1632
1633 // Return true if the divisor evenly divides the dividend.
1634 static
isRemainderZero(const SCEVConstant * Dividend,const SCEVConstant * Divisor)1635 bool isRemainderZero(const SCEVConstant *Dividend,
1636 const SCEVConstant *Divisor) {
1637 const APInt &ConstDividend = Dividend->getAPInt();
1638 const APInt &ConstDivisor = Divisor->getAPInt();
1639 return ConstDividend.srem(ConstDivisor) == 0;
1640 }
1641
1642
1643 // weakZeroSrcSIVtest -
1644 // From the paper, Practical Dependence Testing, Section 4.2.2
1645 //
1646 // When we have a pair of subscripts of the form [c1] and [c2 + a*i],
1647 // where i is an induction variable, c1 and c2 are loop invariant,
1648 // and a is a constant, we can solve it exactly using the
1649 // Weak-Zero SIV test.
1650 //
1651 // Given
1652 //
1653 // c1 = c2 + a*i
1654 //
1655 // we get
1656 //
1657 // (c1 - c2)/a = i
1658 //
1659 // If i is not an integer, there's no dependence.
1660 // If i < 0 or > UB, there's no dependence.
1661 // If i = 0, the direction is >= and peeling the
1662 // 1st iteration will break the dependence.
1663 // If i = UB, the direction is <= and peeling the
1664 // last iteration will break the dependence.
1665 // Otherwise, the direction is *.
1666 //
1667 // Can prove independence. Failing that, we can sometimes refine
1668 // the directions. Can sometimes show that first or last
1669 // iteration carries all the dependences (so worth peeling).
1670 //
1671 // (see also weakZeroDstSIVtest)
1672 //
1673 // Return true if dependence disproved.
weakZeroSrcSIVtest(const SCEV * DstCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1674 bool DependenceInfo::weakZeroSrcSIVtest(const SCEV *DstCoeff,
1675 const SCEV *SrcConst,
1676 const SCEV *DstConst,
1677 const Loop *CurLoop, unsigned Level,
1678 FullDependence &Result,
1679 Constraint &NewConstraint) const {
1680 // For the WeakSIV test, it's possible the loop isn't common to
1681 // the Src and Dst loops. If it isn't, then there's no need to
1682 // record a direction.
1683 LLVM_DEBUG(dbgs() << "\tWeak-Zero (src) SIV test\n");
1684 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << "\n");
1685 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1686 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1687 ++WeakZeroSIVapplications;
1688 assert(0 < Level && Level <= MaxLevels && "Level out of range");
1689 Level--;
1690 Result.Consistent = false;
1691 const SCEV *Delta = SE->getMinusSCEV(SrcConst, DstConst);
1692 NewConstraint.setLine(SE->getZero(Delta->getType()), DstCoeff, Delta,
1693 CurLoop);
1694 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1695 if (isKnownPredicate(CmpInst::ICMP_EQ, SrcConst, DstConst)) {
1696 if (Level < CommonLevels) {
1697 Result.DV[Level].Direction &= Dependence::DVEntry::GE;
1698 Result.DV[Level].PeelFirst = true;
1699 ++WeakZeroSIVsuccesses;
1700 }
1701 return false; // dependences caused by first iteration
1702 }
1703 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1704 if (!ConstCoeff)
1705 return false;
1706 const SCEV *AbsCoeff =
1707 SE->isKnownNegative(ConstCoeff) ?
1708 SE->getNegativeSCEV(ConstCoeff) : ConstCoeff;
1709 const SCEV *NewDelta =
1710 SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta;
1711
1712 // check that Delta/SrcCoeff < iteration count
1713 // really check NewDelta < count*AbsCoeff
1714 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1715 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n");
1716 const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound);
1717 if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) {
1718 ++WeakZeroSIVindependence;
1719 ++WeakZeroSIVsuccesses;
1720 return true;
1721 }
1722 if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) {
1723 // dependences caused by last iteration
1724 if (Level < CommonLevels) {
1725 Result.DV[Level].Direction &= Dependence::DVEntry::LE;
1726 Result.DV[Level].PeelLast = true;
1727 ++WeakZeroSIVsuccesses;
1728 }
1729 return false;
1730 }
1731 }
1732
1733 // check that Delta/SrcCoeff >= 0
1734 // really check that NewDelta >= 0
1735 if (SE->isKnownNegative(NewDelta)) {
1736 // No dependence, newDelta < 0
1737 ++WeakZeroSIVindependence;
1738 ++WeakZeroSIVsuccesses;
1739 return true;
1740 }
1741
1742 // if SrcCoeff doesn't divide Delta, then no dependence
1743 if (isa<SCEVConstant>(Delta) &&
1744 !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) {
1745 ++WeakZeroSIVindependence;
1746 ++WeakZeroSIVsuccesses;
1747 return true;
1748 }
1749 return false;
1750 }
1751
1752
1753 // weakZeroDstSIVtest -
1754 // From the paper, Practical Dependence Testing, Section 4.2.2
1755 //
1756 // When we have a pair of subscripts of the form [c1 + a*i] and [c2],
1757 // where i is an induction variable, c1 and c2 are loop invariant,
1758 // and a is a constant, we can solve it exactly using the
1759 // Weak-Zero SIV test.
1760 //
1761 // Given
1762 //
1763 // c1 + a*i = c2
1764 //
1765 // we get
1766 //
1767 // i = (c2 - c1)/a
1768 //
1769 // If i is not an integer, there's no dependence.
1770 // If i < 0 or > UB, there's no dependence.
1771 // If i = 0, the direction is <= and peeling the
1772 // 1st iteration will break the dependence.
1773 // If i = UB, the direction is >= and peeling the
1774 // last iteration will break the dependence.
1775 // Otherwise, the direction is *.
1776 //
1777 // Can prove independence. Failing that, we can sometimes refine
1778 // the directions. Can sometimes show that first or last
1779 // iteration carries all the dependences (so worth peeling).
1780 //
1781 // (see also weakZeroSrcSIVtest)
1782 //
1783 // Return true if dependence disproved.
weakZeroDstSIVtest(const SCEV * SrcCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * CurLoop,unsigned Level,FullDependence & Result,Constraint & NewConstraint) const1784 bool DependenceInfo::weakZeroDstSIVtest(const SCEV *SrcCoeff,
1785 const SCEV *SrcConst,
1786 const SCEV *DstConst,
1787 const Loop *CurLoop, unsigned Level,
1788 FullDependence &Result,
1789 Constraint &NewConstraint) const {
1790 // For the WeakSIV test, it's possible the loop isn't common to the
1791 // Src and Dst loops. If it isn't, then there's no need to record a direction.
1792 LLVM_DEBUG(dbgs() << "\tWeak-Zero (dst) SIV test\n");
1793 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << "\n");
1794 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1795 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1796 ++WeakZeroSIVapplications;
1797 assert(0 < Level && Level <= SrcLevels && "Level out of range");
1798 Level--;
1799 Result.Consistent = false;
1800 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1801 NewConstraint.setLine(SrcCoeff, SE->getZero(Delta->getType()), Delta,
1802 CurLoop);
1803 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1804 if (isKnownPredicate(CmpInst::ICMP_EQ, DstConst, SrcConst)) {
1805 if (Level < CommonLevels) {
1806 Result.DV[Level].Direction &= Dependence::DVEntry::LE;
1807 Result.DV[Level].PeelFirst = true;
1808 ++WeakZeroSIVsuccesses;
1809 }
1810 return false; // dependences caused by first iteration
1811 }
1812 const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1813 if (!ConstCoeff)
1814 return false;
1815 const SCEV *AbsCoeff =
1816 SE->isKnownNegative(ConstCoeff) ?
1817 SE->getNegativeSCEV(ConstCoeff) : ConstCoeff;
1818 const SCEV *NewDelta =
1819 SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(Delta) : Delta;
1820
1821 // check that Delta/SrcCoeff < iteration count
1822 // really check NewDelta < count*AbsCoeff
1823 if (const SCEV *UpperBound = collectUpperBound(CurLoop, Delta->getType())) {
1824 LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound << "\n");
1825 const SCEV *Product = SE->getMulExpr(AbsCoeff, UpperBound);
1826 if (isKnownPredicate(CmpInst::ICMP_SGT, NewDelta, Product)) {
1827 ++WeakZeroSIVindependence;
1828 ++WeakZeroSIVsuccesses;
1829 return true;
1830 }
1831 if (isKnownPredicate(CmpInst::ICMP_EQ, NewDelta, Product)) {
1832 // dependences caused by last iteration
1833 if (Level < CommonLevels) {
1834 Result.DV[Level].Direction &= Dependence::DVEntry::GE;
1835 Result.DV[Level].PeelLast = true;
1836 ++WeakZeroSIVsuccesses;
1837 }
1838 return false;
1839 }
1840 }
1841
1842 // check that Delta/SrcCoeff >= 0
1843 // really check that NewDelta >= 0
1844 if (SE->isKnownNegative(NewDelta)) {
1845 // No dependence, newDelta < 0
1846 ++WeakZeroSIVindependence;
1847 ++WeakZeroSIVsuccesses;
1848 return true;
1849 }
1850
1851 // if SrcCoeff doesn't divide Delta, then no dependence
1852 if (isa<SCEVConstant>(Delta) &&
1853 !isRemainderZero(cast<SCEVConstant>(Delta), ConstCoeff)) {
1854 ++WeakZeroSIVindependence;
1855 ++WeakZeroSIVsuccesses;
1856 return true;
1857 }
1858 return false;
1859 }
1860
1861
1862 // exactRDIVtest - Tests the RDIV subscript pair for dependence.
1863 // Things of the form [c1 + a*i] and [c2 + b*j],
1864 // where i and j are induction variable, c1 and c2 are loop invariant,
1865 // and a and b are constants.
1866 // Returns true if any possible dependence is disproved.
1867 // Marks the result as inconsistent.
1868 // Works in some cases that symbolicRDIVtest doesn't, and vice versa.
exactRDIVtest(const SCEV * SrcCoeff,const SCEV * DstCoeff,const SCEV * SrcConst,const SCEV * DstConst,const Loop * SrcLoop,const Loop * DstLoop,FullDependence & Result) const1869 bool DependenceInfo::exactRDIVtest(const SCEV *SrcCoeff, const SCEV *DstCoeff,
1870 const SCEV *SrcConst, const SCEV *DstConst,
1871 const Loop *SrcLoop, const Loop *DstLoop,
1872 FullDependence &Result) const {
1873 LLVM_DEBUG(dbgs() << "\tExact RDIV test\n");
1874 LLVM_DEBUG(dbgs() << "\t SrcCoeff = " << *SrcCoeff << " = AM\n");
1875 LLVM_DEBUG(dbgs() << "\t DstCoeff = " << *DstCoeff << " = BM\n");
1876 LLVM_DEBUG(dbgs() << "\t SrcConst = " << *SrcConst << "\n");
1877 LLVM_DEBUG(dbgs() << "\t DstConst = " << *DstConst << "\n");
1878 ++ExactRDIVapplications;
1879 Result.Consistent = false;
1880 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
1881 LLVM_DEBUG(dbgs() << "\t Delta = " << *Delta << "\n");
1882 const SCEVConstant *ConstDelta = dyn_cast<SCEVConstant>(Delta);
1883 const SCEVConstant *ConstSrcCoeff = dyn_cast<SCEVConstant>(SrcCoeff);
1884 const SCEVConstant *ConstDstCoeff = dyn_cast<SCEVConstant>(DstCoeff);
1885 if (!ConstDelta || !ConstSrcCoeff || !ConstDstCoeff)
1886 return false;
1887
1888 // find gcd
1889 APInt G, X, Y;
1890 APInt AM = ConstSrcCoeff->getAPInt();
1891 APInt BM = ConstDstCoeff->getAPInt();
1892 APInt CM = ConstDelta->getAPInt();
1893 unsigned Bits = AM.getBitWidth();
1894 if (findGCD(Bits, AM, BM, CM, G, X, Y)) {
1895 // gcd doesn't divide Delta, no dependence
1896 ++ExactRDIVindependence;
1897 return true;
1898 }
1899
1900 LLVM_DEBUG(dbgs() << "\t X = " << X << ", Y = " << Y << "\n");
1901
1902 // since SCEV construction seems to normalize, LM = 0
1903 APInt SrcUM(Bits, 1, true);
1904 bool SrcUMvalid = false;
1905 // SrcUM is perhaps unavailable, let's check
1906 if (const SCEVConstant *UpperBound =
1907 collectConstantUpperBound(SrcLoop, Delta->getType())) {
1908 SrcUM = UpperBound->getAPInt();
1909 LLVM_DEBUG(dbgs() << "\t SrcUM = " << SrcUM << "\n");
1910 SrcUMvalid = true;
1911 }
1912
1913 APInt DstUM(Bits, 1, true);
1914 bool DstUMvalid = false;
1915 // UM is perhaps unavailable, let's check
1916 if (const SCEVConstant *UpperBound =
1917 collectConstantUpperBound(DstLoop, Delta->getType())) {
1918 DstUM = UpperBound->getAPInt();
1919 LLVM_DEBUG(dbgs() << "\t DstUM = " << DstUM << "\n");
1920 DstUMvalid = true;
1921 }
1922
1923 APInt TU(APInt::getSignedMaxValue(Bits));
1924 APInt TL(APInt::getSignedMinValue(Bits));
1925 APInt TC = CM.sdiv(G);
1926 APInt TX = X * TC;
1927 APInt TY = Y * TC;
1928 LLVM_DEBUG(dbgs() << "\t TC = " << TC << "\n");
1929 LLVM_DEBUG(dbgs() << "\t TX = " << TX << "\n");
1930 LLVM_DEBUG(dbgs() << "\t TY = " << TY << "\n");
1931
1932 SmallVector<APInt, 2> TLVec, TUVec;
1933 APInt TB = BM.sdiv(G);
1934 if (TB.sgt(0)) {
1935 TLVec.push_back(ceilingOfQuotient(-TX, TB));
1936 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1937 if (SrcUMvalid) {
1938 TUVec.push_back(floorOfQuotient(SrcUM - TX, TB));
1939 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1940 }
1941 } else {
1942 TUVec.push_back(floorOfQuotient(-TX, TB));
1943 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1944 if (SrcUMvalid) {
1945 TLVec.push_back(ceilingOfQuotient(SrcUM - TX, TB));
1946 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1947 }
1948 }
1949
1950 APInt TA = AM.sdiv(G);
1951 if (TA.sgt(0)) {
1952 TLVec.push_back(ceilingOfQuotient(-TY, TA));
1953 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1954 if (DstUMvalid) {
1955 TUVec.push_back(floorOfQuotient(DstUM - TY, TA));
1956 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1957 }
1958 } else {
1959 TUVec.push_back(floorOfQuotient(-TY, TA));
1960 LLVM_DEBUG(dbgs() << "\t Possible TU = " << TUVec.back() << "\n");
1961 if (DstUMvalid) {
1962 TLVec.push_back(ceilingOfQuotient(DstUM - TY, TA));
1963 LLVM_DEBUG(dbgs() << "\t Possible TL = " << TLVec.back() << "\n");
1964 }
1965 }
1966
1967 if (TLVec.empty() || TUVec.empty())
1968 return false;
1969
1970 LLVM_DEBUG(dbgs() << "\t TA = " << TA << "\n");
1971 LLVM_DEBUG(dbgs() << "\t TB = " << TB << "\n");
1972
1973 TL = APIntOps::smax(TLVec.front(), TLVec.back());
1974 TU = APIntOps::smin(TUVec.front(), TUVec.back());
1975 LLVM_DEBUG(dbgs() << "\t TL = " << TL << "\n");
1976 LLVM_DEBUG(dbgs() << "\t TU = " << TU << "\n");
1977
1978 if (TL.sgt(TU))
1979 ++ExactRDIVindependence;
1980 return TL.sgt(TU);
1981 }
1982
1983
1984 // symbolicRDIVtest -
1985 // In Section 4.5 of the Practical Dependence Testing paper,the authors
1986 // introduce a special case of Banerjee's Inequalities (also called the
1987 // Extreme-Value Test) that can handle some of the SIV and RDIV cases,
1988 // particularly cases with symbolics. Since it's only able to disprove
1989 // dependence (not compute distances or directions), we'll use it as a
1990 // fall back for the other tests.
1991 //
1992 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j]
1993 // where i and j are induction variables and c1 and c2 are loop invariants,
1994 // we can use the symbolic tests to disprove some dependences, serving as a
1995 // backup for the RDIV test. Note that i and j can be the same variable,
1996 // letting this test serve as a backup for the various SIV tests.
1997 //
1998 // For a dependence to exist, c1 + a1*i must equal c2 + a2*j for some
1999 // 0 <= i <= N1 and some 0 <= j <= N2, where N1 and N2 are the (normalized)
2000 // loop bounds for the i and j loops, respectively. So, ...
2001 //
2002 // c1 + a1*i = c2 + a2*j
2003 // a1*i - a2*j = c2 - c1
2004 //
2005 // To test for a dependence, we compute c2 - c1 and make sure it's in the
2006 // range of the maximum and minimum possible values of a1*i - a2*j.
2007 // Considering the signs of a1 and a2, we have 4 possible cases:
2008 //
2009 // 1) If a1 >= 0 and a2 >= 0, then
2010 // a1*0 - a2*N2 <= c2 - c1 <= a1*N1 - a2*0
2011 // -a2*N2 <= c2 - c1 <= a1*N1
2012 //
2013 // 2) If a1 >= 0 and a2 <= 0, then
2014 // a1*0 - a2*0 <= c2 - c1 <= a1*N1 - a2*N2
2015 // 0 <= c2 - c1 <= a1*N1 - a2*N2
2016 //
2017 // 3) If a1 <= 0 and a2 >= 0, then
2018 // a1*N1 - a2*N2 <= c2 - c1 <= a1*0 - a2*0
2019 // a1*N1 - a2*N2 <= c2 - c1 <= 0
2020 //
2021 // 4) If a1 <= 0 and a2 <= 0, then
2022 // a1*N1 - a2*0 <= c2 - c1 <= a1*0 - a2*N2
2023 // a1*N1 <= c2 - c1 <= -a2*N2
2024 //
2025 // return true if dependence disproved
symbolicRDIVtest(const SCEV * A1,const SCEV * A2,const SCEV * C1,const SCEV * C2,const Loop * Loop1,const Loop * Loop2) const2026 bool DependenceInfo::symbolicRDIVtest(const SCEV *A1, const SCEV *A2,
2027 const SCEV *C1, const SCEV *C2,
2028 const Loop *Loop1,
2029 const Loop *Loop2) const {
2030 ++SymbolicRDIVapplications;
2031 LLVM_DEBUG(dbgs() << "\ttry symbolic RDIV test\n");
2032 LLVM_DEBUG(dbgs() << "\t A1 = " << *A1);
2033 LLVM_DEBUG(dbgs() << ", type = " << *A1->getType() << "\n");
2034 LLVM_DEBUG(dbgs() << "\t A2 = " << *A2 << "\n");
2035 LLVM_DEBUG(dbgs() << "\t C1 = " << *C1 << "\n");
2036 LLVM_DEBUG(dbgs() << "\t C2 = " << *C2 << "\n");
2037 const SCEV *N1 = collectUpperBound(Loop1, A1->getType());
2038 const SCEV *N2 = collectUpperBound(Loop2, A1->getType());
2039 LLVM_DEBUG(if (N1) dbgs() << "\t N1 = " << *N1 << "\n");
2040 LLVM_DEBUG(if (N2) dbgs() << "\t N2 = " << *N2 << "\n");
2041 const SCEV *C2_C1 = SE->getMinusSCEV(C2, C1);
2042 const SCEV *C1_C2 = SE->getMinusSCEV(C1, C2);
2043 LLVM_DEBUG(dbgs() << "\t C2 - C1 = " << *C2_C1 << "\n");
2044 LLVM_DEBUG(dbgs() << "\t C1 - C2 = " << *C1_C2 << "\n");
2045 if (SE->isKnownNonNegative(A1)) {
2046 if (SE->isKnownNonNegative(A2)) {
2047 // A1 >= 0 && A2 >= 0
2048 if (N1) {
2049 // make sure that c2 - c1 <= a1*N1
2050 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2051 LLVM_DEBUG(dbgs() << "\t A1*N1 = " << *A1N1 << "\n");
2052 if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1)) {
2053 ++SymbolicRDIVindependence;
2054 return true;
2055 }
2056 }
2057 if (N2) {
2058 // make sure that -a2*N2 <= c2 - c1, or a2*N2 >= c1 - c2
2059 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2060 LLVM_DEBUG(dbgs() << "\t A2*N2 = " << *A2N2 << "\n");
2061 if (isKnownPredicate(CmpInst::ICMP_SLT, A2N2, C1_C2)) {
2062 ++SymbolicRDIVindependence;
2063 return true;
2064 }
2065 }
2066 }
2067 else if (SE->isKnownNonPositive(A2)) {
2068 // a1 >= 0 && a2 <= 0
2069 if (N1 && N2) {
2070 // make sure that c2 - c1 <= a1*N1 - a2*N2
2071 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2072 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2073 const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2);
2074 LLVM_DEBUG(dbgs() << "\t A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n");
2075 if (isKnownPredicate(CmpInst::ICMP_SGT, C2_C1, A1N1_A2N2)) {
2076 ++SymbolicRDIVindependence;
2077 return true;
2078 }
2079 }
2080 // make sure that 0 <= c2 - c1
2081 if (SE->isKnownNegative(C2_C1)) {
2082 ++SymbolicRDIVindependence;
2083 return true;
2084 }
2085 }
2086 }
2087 else if (SE->isKnownNonPositive(A1)) {
2088 if (SE->isKnownNonNegative(A2)) {
2089 // a1 <= 0 && a2 >= 0
2090 if (N1 && N2) {
2091 // make sure that a1*N1 - a2*N2 <= c2 - c1
2092 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2093 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2094 const SCEV *A1N1_A2N2 = SE->getMinusSCEV(A1N1, A2N2);
2095 LLVM_DEBUG(dbgs() << "\t A1*N1 - A2*N2 = " << *A1N1_A2N2 << "\n");
2096 if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1_A2N2, C2_C1)) {
2097 ++SymbolicRDIVindependence;
2098 return true;
2099 }
2100 }
2101 // make sure that c2 - c1 <= 0
2102 if (SE->isKnownPositive(C2_C1)) {
2103 ++SymbolicRDIVindependence;
2104 return true;
2105 }
2106 }
2107 else if (SE->isKnownNonPositive(A2)) {
2108 // a1 <= 0 && a2 <= 0
2109 if (N1) {
2110 // make sure that a1*N1 <= c2 - c1
2111 const SCEV *A1N1 = SE->getMulExpr(A1, N1);
2112 LLVM_DEBUG(dbgs() << "\t A1*N1 = " << *A1N1 << "\n");
2113 if (isKnownPredicate(CmpInst::ICMP_SGT, A1N1, C2_C1)) {
2114 ++SymbolicRDIVindependence;
2115 return true;
2116 }
2117 }
2118 if (N2) {
2119 // make sure that c2 - c1 <= -a2*N2, or c1 - c2 >= a2*N2
2120 const SCEV *A2N2 = SE->getMulExpr(A2, N2);
2121 LLVM_DEBUG(dbgs() << "\t A2*N2 = " << *A2N2 << "\n");
2122 if (isKnownPredicate(CmpInst::ICMP_SLT, C1_C2, A2N2)) {
2123 ++SymbolicRDIVindependence;
2124 return true;
2125 }
2126 }
2127 }
2128 }
2129 return false;
2130 }
2131
2132
2133 // testSIV -
2134 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 - a2*i]
2135 // where i is an induction variable, c1 and c2 are loop invariant, and a1 and
2136 // a2 are constant, we attack it with an SIV test. While they can all be
2137 // solved with the Exact SIV test, it's worthwhile to use simpler tests when
2138 // they apply; they're cheaper and sometimes more precise.
2139 //
2140 // Return true if dependence disproved.
testSIV(const SCEV * Src,const SCEV * Dst,unsigned & Level,FullDependence & Result,Constraint & NewConstraint,const SCEV * & SplitIter) const2141 bool DependenceInfo::testSIV(const SCEV *Src, const SCEV *Dst, unsigned &Level,
2142 FullDependence &Result, Constraint &NewConstraint,
2143 const SCEV *&SplitIter) const {
2144 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
2145 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
2146 const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src);
2147 const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst);
2148 if (SrcAddRec && DstAddRec) {
2149 const SCEV *SrcConst = SrcAddRec->getStart();
2150 const SCEV *DstConst = DstAddRec->getStart();
2151 const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2152 const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE);
2153 const Loop *CurLoop = SrcAddRec->getLoop();
2154 assert(CurLoop == DstAddRec->getLoop() &&
2155 "both loops in SIV should be same");
2156 Level = mapSrcLoop(CurLoop);
2157 bool disproven;
2158 if (SrcCoeff == DstCoeff)
2159 disproven = strongSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2160 Level, Result, NewConstraint);
2161 else if (SrcCoeff == SE->getNegativeSCEV(DstCoeff))
2162 disproven = weakCrossingSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2163 Level, Result, NewConstraint, SplitIter);
2164 else
2165 disproven = exactSIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop,
2166 Level, Result, NewConstraint);
2167 return disproven ||
2168 gcdMIVtest(Src, Dst, Result) ||
2169 symbolicRDIVtest(SrcCoeff, DstCoeff, SrcConst, DstConst, CurLoop, CurLoop);
2170 }
2171 if (SrcAddRec) {
2172 const SCEV *SrcConst = SrcAddRec->getStart();
2173 const SCEV *SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2174 const SCEV *DstConst = Dst;
2175 const Loop *CurLoop = SrcAddRec->getLoop();
2176 Level = mapSrcLoop(CurLoop);
2177 return weakZeroDstSIVtest(SrcCoeff, SrcConst, DstConst, CurLoop,
2178 Level, Result, NewConstraint) ||
2179 gcdMIVtest(Src, Dst, Result);
2180 }
2181 if (DstAddRec) {
2182 const SCEV *DstConst = DstAddRec->getStart();
2183 const SCEV *DstCoeff = DstAddRec->getStepRecurrence(*SE);
2184 const SCEV *SrcConst = Src;
2185 const Loop *CurLoop = DstAddRec->getLoop();
2186 Level = mapDstLoop(CurLoop);
2187 return weakZeroSrcSIVtest(DstCoeff, SrcConst, DstConst,
2188 CurLoop, Level, Result, NewConstraint) ||
2189 gcdMIVtest(Src, Dst, Result);
2190 }
2191 llvm_unreachable("SIV test expected at least one AddRec");
2192 return false;
2193 }
2194
2195
2196 // testRDIV -
2197 // When we have a pair of subscripts of the form [c1 + a1*i] and [c2 + a2*j]
2198 // where i and j are induction variables, c1 and c2 are loop invariant,
2199 // and a1 and a2 are constant, we can solve it exactly with an easy adaptation
2200 // of the Exact SIV test, the Restricted Double Index Variable (RDIV) test.
2201 // It doesn't make sense to talk about distance or direction in this case,
2202 // so there's no point in making special versions of the Strong SIV test or
2203 // the Weak-crossing SIV test.
2204 //
2205 // With minor algebra, this test can also be used for things like
2206 // [c1 + a1*i + a2*j][c2].
2207 //
2208 // Return true if dependence disproved.
testRDIV(const SCEV * Src,const SCEV * Dst,FullDependence & Result) const2209 bool DependenceInfo::testRDIV(const SCEV *Src, const SCEV *Dst,
2210 FullDependence &Result) const {
2211 // we have 3 possible situations here:
2212 // 1) [a*i + b] and [c*j + d]
2213 // 2) [a*i + c*j + b] and [d]
2214 // 3) [b] and [a*i + c*j + d]
2215 // We need to find what we've got and get organized
2216
2217 const SCEV *SrcConst, *DstConst;
2218 const SCEV *SrcCoeff, *DstCoeff;
2219 const Loop *SrcLoop, *DstLoop;
2220
2221 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
2222 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
2223 const SCEVAddRecExpr *SrcAddRec = dyn_cast<SCEVAddRecExpr>(Src);
2224 const SCEVAddRecExpr *DstAddRec = dyn_cast<SCEVAddRecExpr>(Dst);
2225 if (SrcAddRec && DstAddRec) {
2226 SrcConst = SrcAddRec->getStart();
2227 SrcCoeff = SrcAddRec->getStepRecurrence(*SE);
2228 SrcLoop = SrcAddRec->getLoop();
2229 DstConst = DstAddRec->getStart();
2230 DstCoeff = DstAddRec->getStepRecurrence(*SE);
2231 DstLoop = DstAddRec->getLoop();
2232 }
2233 else if (SrcAddRec) {
2234 if (const SCEVAddRecExpr *tmpAddRec =
2235 dyn_cast<SCEVAddRecExpr>(SrcAddRec->getStart())) {
2236 SrcConst = tmpAddRec->getStart();
2237 SrcCoeff = tmpAddRec->getStepRecurrence(*SE);
2238 SrcLoop = tmpAddRec->getLoop();
2239 DstConst = Dst;
2240 DstCoeff = SE->getNegativeSCEV(SrcAddRec->getStepRecurrence(*SE));
2241 DstLoop = SrcAddRec->getLoop();
2242 }
2243 else
2244 llvm_unreachable("RDIV reached by surprising SCEVs");
2245 }
2246 else if (DstAddRec) {
2247 if (const SCEVAddRecExpr *tmpAddRec =
2248 dyn_cast<SCEVAddRecExpr>(DstAddRec->getStart())) {
2249 DstConst = tmpAddRec->getStart();
2250 DstCoeff = tmpAddRec->getStepRecurrence(*SE);
2251 DstLoop = tmpAddRec->getLoop();
2252 SrcConst = Src;
2253 SrcCoeff = SE->getNegativeSCEV(DstAddRec->getStepRecurrence(*SE));
2254 SrcLoop = DstAddRec->getLoop();
2255 }
2256 else
2257 llvm_unreachable("RDIV reached by surprising SCEVs");
2258 }
2259 else
2260 llvm_unreachable("RDIV expected at least one AddRec");
2261 return exactRDIVtest(SrcCoeff, DstCoeff,
2262 SrcConst, DstConst,
2263 SrcLoop, DstLoop,
2264 Result) ||
2265 gcdMIVtest(Src, Dst, Result) ||
2266 symbolicRDIVtest(SrcCoeff, DstCoeff,
2267 SrcConst, DstConst,
2268 SrcLoop, DstLoop);
2269 }
2270
2271
2272 // Tests the single-subscript MIV pair (Src and Dst) for dependence.
2273 // Return true if dependence disproved.
2274 // Can sometimes refine direction vectors.
testMIV(const SCEV * Src,const SCEV * Dst,const SmallBitVector & Loops,FullDependence & Result) const2275 bool DependenceInfo::testMIV(const SCEV *Src, const SCEV *Dst,
2276 const SmallBitVector &Loops,
2277 FullDependence &Result) const {
2278 LLVM_DEBUG(dbgs() << " src = " << *Src << "\n");
2279 LLVM_DEBUG(dbgs() << " dst = " << *Dst << "\n");
2280 Result.Consistent = false;
2281 return gcdMIVtest(Src, Dst, Result) ||
2282 banerjeeMIVtest(Src, Dst, Loops, Result);
2283 }
2284
2285
2286 // Given a product, e.g., 10*X*Y, returns the first constant operand,
2287 // in this case 10. If there is no constant part, returns NULL.
2288 static
getConstantPart(const SCEV * Expr)2289 const SCEVConstant *getConstantPart(const SCEV *Expr) {
2290 if (const auto *Constant = dyn_cast<SCEVConstant>(Expr))
2291 return Constant;
2292 else if (const auto *Product = dyn_cast<SCEVMulExpr>(Expr))
2293 if (const auto *Constant = dyn_cast<SCEVConstant>(Product->getOperand(0)))
2294 return Constant;
2295 return nullptr;
2296 }
2297
2298
2299 //===----------------------------------------------------------------------===//
2300 // gcdMIVtest -
2301 // Tests an MIV subscript pair for dependence.
2302 // Returns true if any possible dependence is disproved.
2303 // Marks the result as inconsistent.
2304 // Can sometimes disprove the equal direction for 1 or more loops,
2305 // as discussed in Michael Wolfe's book,
2306 // High Performance Compilers for Parallel Computing, page 235.
2307 //
2308 // We spend some effort (code!) to handle cases like
2309 // [10*i + 5*N*j + 15*M + 6], where i and j are induction variables,
2310 // but M and N are just loop-invariant variables.
2311 // This should help us handle linearized subscripts;
2312 // also makes this test a useful backup to the various SIV tests.
2313 //
2314 // It occurs to me that the presence of loop-invariant variables
2315 // changes the nature of the test from "greatest common divisor"
2316 // to "a common divisor".
gcdMIVtest(const SCEV * Src,const SCEV * Dst,FullDependence & Result) const2317 bool DependenceInfo::gcdMIVtest(const SCEV *Src, const SCEV *Dst,
2318 FullDependence &Result) const {
2319 LLVM_DEBUG(dbgs() << "starting gcd\n");
2320 ++GCDapplications;
2321 unsigned BitWidth = SE->getTypeSizeInBits(Src->getType());
2322 APInt RunningGCD = APInt::getNullValue(BitWidth);
2323
2324 // Examine Src coefficients.
2325 // Compute running GCD and record source constant.
2326 // Because we're looking for the constant at the end of the chain,
2327 // we can't quit the loop just because the GCD == 1.
2328 const SCEV *Coefficients = Src;
2329 while (const SCEVAddRecExpr *AddRec =
2330 dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2331 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2332 // If the coefficient is the product of a constant and other stuff,
2333 // we can use the constant in the GCD computation.
2334 const auto *Constant = getConstantPart(Coeff);
2335 if (!Constant)
2336 return false;
2337 APInt ConstCoeff = Constant->getAPInt();
2338 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2339 Coefficients = AddRec->getStart();
2340 }
2341 const SCEV *SrcConst = Coefficients;
2342
2343 // Examine Dst coefficients.
2344 // Compute running GCD and record destination constant.
2345 // Because we're looking for the constant at the end of the chain,
2346 // we can't quit the loop just because the GCD == 1.
2347 Coefficients = Dst;
2348 while (const SCEVAddRecExpr *AddRec =
2349 dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2350 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2351 // If the coefficient is the product of a constant and other stuff,
2352 // we can use the constant in the GCD computation.
2353 const auto *Constant = getConstantPart(Coeff);
2354 if (!Constant)
2355 return false;
2356 APInt ConstCoeff = Constant->getAPInt();
2357 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2358 Coefficients = AddRec->getStart();
2359 }
2360 const SCEV *DstConst = Coefficients;
2361
2362 APInt ExtraGCD = APInt::getNullValue(BitWidth);
2363 const SCEV *Delta = SE->getMinusSCEV(DstConst, SrcConst);
2364 LLVM_DEBUG(dbgs() << " Delta = " << *Delta << "\n");
2365 const SCEVConstant *Constant = dyn_cast<SCEVConstant>(Delta);
2366 if (const SCEVAddExpr *Sum = dyn_cast<SCEVAddExpr>(Delta)) {
2367 // If Delta is a sum of products, we may be able to make further progress.
2368 for (unsigned Op = 0, Ops = Sum->getNumOperands(); Op < Ops; Op++) {
2369 const SCEV *Operand = Sum->getOperand(Op);
2370 if (isa<SCEVConstant>(Operand)) {
2371 assert(!Constant && "Surprised to find multiple constants");
2372 Constant = cast<SCEVConstant>(Operand);
2373 }
2374 else if (const SCEVMulExpr *Product = dyn_cast<SCEVMulExpr>(Operand)) {
2375 // Search for constant operand to participate in GCD;
2376 // If none found; return false.
2377 const SCEVConstant *ConstOp = getConstantPart(Product);
2378 if (!ConstOp)
2379 return false;
2380 APInt ConstOpValue = ConstOp->getAPInt();
2381 ExtraGCD = APIntOps::GreatestCommonDivisor(ExtraGCD,
2382 ConstOpValue.abs());
2383 }
2384 else
2385 return false;
2386 }
2387 }
2388 if (!Constant)
2389 return false;
2390 APInt ConstDelta = cast<SCEVConstant>(Constant)->getAPInt();
2391 LLVM_DEBUG(dbgs() << " ConstDelta = " << ConstDelta << "\n");
2392 if (ConstDelta == 0)
2393 return false;
2394 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ExtraGCD);
2395 LLVM_DEBUG(dbgs() << " RunningGCD = " << RunningGCD << "\n");
2396 APInt Remainder = ConstDelta.srem(RunningGCD);
2397 if (Remainder != 0) {
2398 ++GCDindependence;
2399 return true;
2400 }
2401
2402 // Try to disprove equal directions.
2403 // For example, given a subscript pair [3*i + 2*j] and [i' + 2*j' - 1],
2404 // the code above can't disprove the dependence because the GCD = 1.
2405 // So we consider what happen if i = i' and what happens if j = j'.
2406 // If i = i', we can simplify the subscript to [2*i + 2*j] and [2*j' - 1],
2407 // which is infeasible, so we can disallow the = direction for the i level.
2408 // Setting j = j' doesn't help matters, so we end up with a direction vector
2409 // of [<>, *]
2410 //
2411 // Given A[5*i + 10*j*M + 9*M*N] and A[15*i + 20*j*M - 21*N*M + 5],
2412 // we need to remember that the constant part is 5 and the RunningGCD should
2413 // be initialized to ExtraGCD = 30.
2414 LLVM_DEBUG(dbgs() << " ExtraGCD = " << ExtraGCD << '\n');
2415
2416 bool Improved = false;
2417 Coefficients = Src;
2418 while (const SCEVAddRecExpr *AddRec =
2419 dyn_cast<SCEVAddRecExpr>(Coefficients)) {
2420 Coefficients = AddRec->getStart();
2421 const Loop *CurLoop = AddRec->getLoop();
2422 RunningGCD = ExtraGCD;
2423 const SCEV *SrcCoeff = AddRec->getStepRecurrence(*SE);
2424 const SCEV *DstCoeff = SE->getMinusSCEV(SrcCoeff, SrcCoeff);
2425 const SCEV *Inner = Src;
2426 while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) {
2427 AddRec = cast<SCEVAddRecExpr>(Inner);
2428 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2429 if (CurLoop == AddRec->getLoop())
2430 ; // SrcCoeff == Coeff
2431 else {
2432 // If the coefficient is the product of a constant and other stuff,
2433 // we can use the constant in the GCD computation.
2434 Constant = getConstantPart(Coeff);
2435 if (!Constant)
2436 return false;
2437 APInt ConstCoeff = Constant->getAPInt();
2438 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2439 }
2440 Inner = AddRec->getStart();
2441 }
2442 Inner = Dst;
2443 while (RunningGCD != 1 && isa<SCEVAddRecExpr>(Inner)) {
2444 AddRec = cast<SCEVAddRecExpr>(Inner);
2445 const SCEV *Coeff = AddRec->getStepRecurrence(*SE);
2446 if (CurLoop == AddRec->getLoop())
2447 DstCoeff = Coeff;
2448 else {
2449 // If the coefficient is the product of a constant and other stuff,
2450 // we can use the constant in the GCD computation.
2451 Constant = getConstantPart(Coeff);
2452 if (!Constant)
2453 return false;
2454 APInt ConstCoeff = Constant->getAPInt();
2455 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2456 }
2457 Inner = AddRec->getStart();
2458 }
2459 Delta = SE->getMinusSCEV(SrcCoeff, DstCoeff);
2460 // If the coefficient is the product of a constant and other stuff,
2461 // we can use the constant in the GCD computation.
2462 Constant = getConstantPart(Delta);
2463 if (!Constant)
2464 // The difference of the two coefficients might not be a product
2465 // or constant, in which case we give up on this direction.
2466 continue;
2467 APInt ConstCoeff = Constant->getAPInt();
2468 RunningGCD = APIntOps::GreatestCommonDivisor(RunningGCD, ConstCoeff.abs());
2469 LLVM_DEBUG(dbgs() << "\tRunningGCD = " << RunningGCD << "\n");
2470 if (RunningGCD != 0) {
2471 Remainder = ConstDelta.srem(RunningGCD);
2472 LLVM_DEBUG(dbgs() << "\tRemainder = " << Remainder << "\n");
2473 if (Remainder != 0) {
2474 unsigned Level = mapSrcLoop(CurLoop);
2475 Result.DV[Level - 1].Direction &= unsigned(~Dependence::DVEntry::EQ);
2476 Improved = true;
2477 }
2478 }
2479 }
2480 if (Improved)
2481 ++GCDsuccesses;
2482 LLVM_DEBUG(dbgs() << "all done\n");
2483 return false;
2484 }
2485
2486
2487 //===----------------------------------------------------------------------===//
2488 // banerjeeMIVtest -
2489 // Use Banerjee's Inequalities to test an MIV subscript pair.
2490 // (Wolfe, in the race-car book, calls this the Extreme Value Test.)
2491 // Generally follows the discussion in Section 2.5.2 of
2492 //
2493 // Optimizing Supercompilers for Supercomputers
2494 // Michael Wolfe
2495 //
2496 // The inequalities given on page 25 are simplified in that loops are
2497 // normalized so that the lower bound is always 0 and the stride is always 1.
2498 // For example, Wolfe gives
2499 //
2500 // LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2501 //
2502 // where A_k is the coefficient of the kth index in the source subscript,
2503 // B_k is the coefficient of the kth index in the destination subscript,
2504 // U_k is the upper bound of the kth index, L_k is the lower bound of the Kth
2505 // index, and N_k is the stride of the kth index. Since all loops are normalized
2506 // by the SCEV package, N_k = 1 and L_k = 0, allowing us to simplify the
2507 // equation to
2508 //
2509 // LB^<_k = (A^-_k - B_k)^- (U_k - 0 - 1) + (A_k - B_k)0 - B_k 1
2510 // = (A^-_k - B_k)^- (U_k - 1) - B_k
2511 //
2512 // Similar simplifications are possible for the other equations.
2513 //
2514 // When we can't determine the number of iterations for a loop,
2515 // we use NULL as an indicator for the worst case, infinity.
2516 // When computing the upper bound, NULL denotes +inf;
2517 // for the lower bound, NULL denotes -inf.
2518 //
2519 // Return true if dependence disproved.
banerjeeMIVtest(const SCEV * Src,const SCEV * Dst,const SmallBitVector & Loops,FullDependence & Result) const2520 bool DependenceInfo::banerjeeMIVtest(const SCEV *Src, const SCEV *Dst,
2521 const SmallBitVector &Loops,
2522 FullDependence &Result) const {
2523 LLVM_DEBUG(dbgs() << "starting Banerjee\n");
2524 ++BanerjeeApplications;
2525 LLVM_DEBUG(dbgs() << " Src = " << *Src << '\n');
2526 const SCEV *A0;
2527 CoefficientInfo *A = collectCoeffInfo(Src, true, A0);
2528 LLVM_DEBUG(dbgs() << " Dst = " << *Dst << '\n');
2529 const SCEV *B0;
2530 CoefficientInfo *B = collectCoeffInfo(Dst, false, B0);
2531 BoundInfo *Bound = new BoundInfo[MaxLevels + 1];
2532 const SCEV *Delta = SE->getMinusSCEV(B0, A0);
2533 LLVM_DEBUG(dbgs() << "\tDelta = " << *Delta << '\n');
2534
2535 // Compute bounds for all the * directions.
2536 LLVM_DEBUG(dbgs() << "\tBounds[*]\n");
2537 for (unsigned K = 1; K <= MaxLevels; ++K) {
2538 Bound[K].Iterations = A[K].Iterations ? A[K].Iterations : B[K].Iterations;
2539 Bound[K].Direction = Dependence::DVEntry::ALL;
2540 Bound[K].DirSet = Dependence::DVEntry::NONE;
2541 findBoundsALL(A, B, Bound, K);
2542 #ifndef NDEBUG
2543 LLVM_DEBUG(dbgs() << "\t " << K << '\t');
2544 if (Bound[K].Lower[Dependence::DVEntry::ALL])
2545 LLVM_DEBUG(dbgs() << *Bound[K].Lower[Dependence::DVEntry::ALL] << '\t');
2546 else
2547 LLVM_DEBUG(dbgs() << "-inf\t");
2548 if (Bound[K].Upper[Dependence::DVEntry::ALL])
2549 LLVM_DEBUG(dbgs() << *Bound[K].Upper[Dependence::DVEntry::ALL] << '\n');
2550 else
2551 LLVM_DEBUG(dbgs() << "+inf\n");
2552 #endif
2553 }
2554
2555 // Test the *, *, *, ... case.
2556 bool Disproved = false;
2557 if (testBounds(Dependence::DVEntry::ALL, 0, Bound, Delta)) {
2558 // Explore the direction vector hierarchy.
2559 unsigned DepthExpanded = 0;
2560 unsigned NewDeps = exploreDirections(1, A, B, Bound,
2561 Loops, DepthExpanded, Delta);
2562 if (NewDeps > 0) {
2563 bool Improved = false;
2564 for (unsigned K = 1; K <= CommonLevels; ++K) {
2565 if (Loops[K]) {
2566 unsigned Old = Result.DV[K - 1].Direction;
2567 Result.DV[K - 1].Direction = Old & Bound[K].DirSet;
2568 Improved |= Old != Result.DV[K - 1].Direction;
2569 if (!Result.DV[K - 1].Direction) {
2570 Improved = false;
2571 Disproved = true;
2572 break;
2573 }
2574 }
2575 }
2576 if (Improved)
2577 ++BanerjeeSuccesses;
2578 }
2579 else {
2580 ++BanerjeeIndependence;
2581 Disproved = true;
2582 }
2583 }
2584 else {
2585 ++BanerjeeIndependence;
2586 Disproved = true;
2587 }
2588 delete [] Bound;
2589 delete [] A;
2590 delete [] B;
2591 return Disproved;
2592 }
2593
2594
2595 // Hierarchically expands the direction vector
2596 // search space, combining the directions of discovered dependences
2597 // in the DirSet field of Bound. Returns the number of distinct
2598 // dependences discovered. If the dependence is disproved,
2599 // it will return 0.
exploreDirections(unsigned Level,CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,const SmallBitVector & Loops,unsigned & DepthExpanded,const SCEV * Delta) const2600 unsigned DependenceInfo::exploreDirections(unsigned Level, CoefficientInfo *A,
2601 CoefficientInfo *B, BoundInfo *Bound,
2602 const SmallBitVector &Loops,
2603 unsigned &DepthExpanded,
2604 const SCEV *Delta) const {
2605 if (Level > CommonLevels) {
2606 // record result
2607 LLVM_DEBUG(dbgs() << "\t[");
2608 for (unsigned K = 1; K <= CommonLevels; ++K) {
2609 if (Loops[K]) {
2610 Bound[K].DirSet |= Bound[K].Direction;
2611 #ifndef NDEBUG
2612 switch (Bound[K].Direction) {
2613 case Dependence::DVEntry::LT:
2614 LLVM_DEBUG(dbgs() << " <");
2615 break;
2616 case Dependence::DVEntry::EQ:
2617 LLVM_DEBUG(dbgs() << " =");
2618 break;
2619 case Dependence::DVEntry::GT:
2620 LLVM_DEBUG(dbgs() << " >");
2621 break;
2622 case Dependence::DVEntry::ALL:
2623 LLVM_DEBUG(dbgs() << " *");
2624 break;
2625 default:
2626 llvm_unreachable("unexpected Bound[K].Direction");
2627 }
2628 #endif
2629 }
2630 }
2631 LLVM_DEBUG(dbgs() << " ]\n");
2632 return 1;
2633 }
2634 if (Loops[Level]) {
2635 if (Level > DepthExpanded) {
2636 DepthExpanded = Level;
2637 // compute bounds for <, =, > at current level
2638 findBoundsLT(A, B, Bound, Level);
2639 findBoundsGT(A, B, Bound, Level);
2640 findBoundsEQ(A, B, Bound, Level);
2641 #ifndef NDEBUG
2642 LLVM_DEBUG(dbgs() << "\tBound for level = " << Level << '\n');
2643 LLVM_DEBUG(dbgs() << "\t <\t");
2644 if (Bound[Level].Lower[Dependence::DVEntry::LT])
2645 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::LT]
2646 << '\t');
2647 else
2648 LLVM_DEBUG(dbgs() << "-inf\t");
2649 if (Bound[Level].Upper[Dependence::DVEntry::LT])
2650 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::LT]
2651 << '\n');
2652 else
2653 LLVM_DEBUG(dbgs() << "+inf\n");
2654 LLVM_DEBUG(dbgs() << "\t =\t");
2655 if (Bound[Level].Lower[Dependence::DVEntry::EQ])
2656 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::EQ]
2657 << '\t');
2658 else
2659 LLVM_DEBUG(dbgs() << "-inf\t");
2660 if (Bound[Level].Upper[Dependence::DVEntry::EQ])
2661 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::EQ]
2662 << '\n');
2663 else
2664 LLVM_DEBUG(dbgs() << "+inf\n");
2665 LLVM_DEBUG(dbgs() << "\t >\t");
2666 if (Bound[Level].Lower[Dependence::DVEntry::GT])
2667 LLVM_DEBUG(dbgs() << *Bound[Level].Lower[Dependence::DVEntry::GT]
2668 << '\t');
2669 else
2670 LLVM_DEBUG(dbgs() << "-inf\t");
2671 if (Bound[Level].Upper[Dependence::DVEntry::GT])
2672 LLVM_DEBUG(dbgs() << *Bound[Level].Upper[Dependence::DVEntry::GT]
2673 << '\n');
2674 else
2675 LLVM_DEBUG(dbgs() << "+inf\n");
2676 #endif
2677 }
2678
2679 unsigned NewDeps = 0;
2680
2681 // test bounds for <, *, *, ...
2682 if (testBounds(Dependence::DVEntry::LT, Level, Bound, Delta))
2683 NewDeps += exploreDirections(Level + 1, A, B, Bound,
2684 Loops, DepthExpanded, Delta);
2685
2686 // Test bounds for =, *, *, ...
2687 if (testBounds(Dependence::DVEntry::EQ, Level, Bound, Delta))
2688 NewDeps += exploreDirections(Level + 1, A, B, Bound,
2689 Loops, DepthExpanded, Delta);
2690
2691 // test bounds for >, *, *, ...
2692 if (testBounds(Dependence::DVEntry::GT, Level, Bound, Delta))
2693 NewDeps += exploreDirections(Level + 1, A, B, Bound,
2694 Loops, DepthExpanded, Delta);
2695
2696 Bound[Level].Direction = Dependence::DVEntry::ALL;
2697 return NewDeps;
2698 }
2699 else
2700 return exploreDirections(Level + 1, A, B, Bound, Loops, DepthExpanded, Delta);
2701 }
2702
2703
2704 // Returns true iff the current bounds are plausible.
testBounds(unsigned char DirKind,unsigned Level,BoundInfo * Bound,const SCEV * Delta) const2705 bool DependenceInfo::testBounds(unsigned char DirKind, unsigned Level,
2706 BoundInfo *Bound, const SCEV *Delta) const {
2707 Bound[Level].Direction = DirKind;
2708 if (const SCEV *LowerBound = getLowerBound(Bound))
2709 if (isKnownPredicate(CmpInst::ICMP_SGT, LowerBound, Delta))
2710 return false;
2711 if (const SCEV *UpperBound = getUpperBound(Bound))
2712 if (isKnownPredicate(CmpInst::ICMP_SGT, Delta, UpperBound))
2713 return false;
2714 return true;
2715 }
2716
2717
2718 // Computes the upper and lower bounds for level K
2719 // using the * direction. Records them in Bound.
2720 // Wolfe gives the equations
2721 //
2722 // LB^*_k = (A^-_k - B^+_k)(U_k - L_k) + (A_k - B_k)L_k
2723 // UB^*_k = (A^+_k - B^-_k)(U_k - L_k) + (A_k - B_k)L_k
2724 //
2725 // Since we normalize loops, we can simplify these equations to
2726 //
2727 // LB^*_k = (A^-_k - B^+_k)U_k
2728 // UB^*_k = (A^+_k - B^-_k)U_k
2729 //
2730 // We must be careful to handle the case where the upper bound is unknown.
2731 // Note that the lower bound is always <= 0
2732 // and the upper bound is always >= 0.
findBoundsALL(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2733 void DependenceInfo::findBoundsALL(CoefficientInfo *A, CoefficientInfo *B,
2734 BoundInfo *Bound, unsigned K) const {
2735 Bound[K].Lower[Dependence::DVEntry::ALL] = nullptr; // Default value = -infinity.
2736 Bound[K].Upper[Dependence::DVEntry::ALL] = nullptr; // Default value = +infinity.
2737 if (Bound[K].Iterations) {
2738 Bound[K].Lower[Dependence::DVEntry::ALL] =
2739 SE->getMulExpr(SE->getMinusSCEV(A[K].NegPart, B[K].PosPart),
2740 Bound[K].Iterations);
2741 Bound[K].Upper[Dependence::DVEntry::ALL] =
2742 SE->getMulExpr(SE->getMinusSCEV(A[K].PosPart, B[K].NegPart),
2743 Bound[K].Iterations);
2744 }
2745 else {
2746 // If the difference is 0, we won't need to know the number of iterations.
2747 if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].NegPart, B[K].PosPart))
2748 Bound[K].Lower[Dependence::DVEntry::ALL] =
2749 SE->getZero(A[K].Coeff->getType());
2750 if (isKnownPredicate(CmpInst::ICMP_EQ, A[K].PosPart, B[K].NegPart))
2751 Bound[K].Upper[Dependence::DVEntry::ALL] =
2752 SE->getZero(A[K].Coeff->getType());
2753 }
2754 }
2755
2756
2757 // Computes the upper and lower bounds for level K
2758 // using the = direction. Records them in Bound.
2759 // Wolfe gives the equations
2760 //
2761 // LB^=_k = (A_k - B_k)^- (U_k - L_k) + (A_k - B_k)L_k
2762 // UB^=_k = (A_k - B_k)^+ (U_k - L_k) + (A_k - B_k)L_k
2763 //
2764 // Since we normalize loops, we can simplify these equations to
2765 //
2766 // LB^=_k = (A_k - B_k)^- U_k
2767 // UB^=_k = (A_k - B_k)^+ U_k
2768 //
2769 // We must be careful to handle the case where the upper bound is unknown.
2770 // Note that the lower bound is always <= 0
2771 // and the upper bound is always >= 0.
findBoundsEQ(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2772 void DependenceInfo::findBoundsEQ(CoefficientInfo *A, CoefficientInfo *B,
2773 BoundInfo *Bound, unsigned K) const {
2774 Bound[K].Lower[Dependence::DVEntry::EQ] = nullptr; // Default value = -infinity.
2775 Bound[K].Upper[Dependence::DVEntry::EQ] = nullptr; // Default value = +infinity.
2776 if (Bound[K].Iterations) {
2777 const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff);
2778 const SCEV *NegativePart = getNegativePart(Delta);
2779 Bound[K].Lower[Dependence::DVEntry::EQ] =
2780 SE->getMulExpr(NegativePart, Bound[K].Iterations);
2781 const SCEV *PositivePart = getPositivePart(Delta);
2782 Bound[K].Upper[Dependence::DVEntry::EQ] =
2783 SE->getMulExpr(PositivePart, Bound[K].Iterations);
2784 }
2785 else {
2786 // If the positive/negative part of the difference is 0,
2787 // we won't need to know the number of iterations.
2788 const SCEV *Delta = SE->getMinusSCEV(A[K].Coeff, B[K].Coeff);
2789 const SCEV *NegativePart = getNegativePart(Delta);
2790 if (NegativePart->isZero())
2791 Bound[K].Lower[Dependence::DVEntry::EQ] = NegativePart; // Zero
2792 const SCEV *PositivePart = getPositivePart(Delta);
2793 if (PositivePart->isZero())
2794 Bound[K].Upper[Dependence::DVEntry::EQ] = PositivePart; // Zero
2795 }
2796 }
2797
2798
2799 // Computes the upper and lower bounds for level K
2800 // using the < direction. Records them in Bound.
2801 // Wolfe gives the equations
2802 //
2803 // LB^<_k = (A^-_k - B_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2804 // UB^<_k = (A^+_k - B_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k - B_k N_k
2805 //
2806 // Since we normalize loops, we can simplify these equations to
2807 //
2808 // LB^<_k = (A^-_k - B_k)^- (U_k - 1) - B_k
2809 // UB^<_k = (A^+_k - B_k)^+ (U_k - 1) - B_k
2810 //
2811 // We must be careful to handle the case where the upper bound is unknown.
findBoundsLT(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2812 void DependenceInfo::findBoundsLT(CoefficientInfo *A, CoefficientInfo *B,
2813 BoundInfo *Bound, unsigned K) const {
2814 Bound[K].Lower[Dependence::DVEntry::LT] = nullptr; // Default value = -infinity.
2815 Bound[K].Upper[Dependence::DVEntry::LT] = nullptr; // Default value = +infinity.
2816 if (Bound[K].Iterations) {
2817 const SCEV *Iter_1 = SE->getMinusSCEV(
2818 Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType()));
2819 const SCEV *NegPart =
2820 getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff));
2821 Bound[K].Lower[Dependence::DVEntry::LT] =
2822 SE->getMinusSCEV(SE->getMulExpr(NegPart, Iter_1), B[K].Coeff);
2823 const SCEV *PosPart =
2824 getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff));
2825 Bound[K].Upper[Dependence::DVEntry::LT] =
2826 SE->getMinusSCEV(SE->getMulExpr(PosPart, Iter_1), B[K].Coeff);
2827 }
2828 else {
2829 // If the positive/negative part of the difference is 0,
2830 // we won't need to know the number of iterations.
2831 const SCEV *NegPart =
2832 getNegativePart(SE->getMinusSCEV(A[K].NegPart, B[K].Coeff));
2833 if (NegPart->isZero())
2834 Bound[K].Lower[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff);
2835 const SCEV *PosPart =
2836 getPositivePart(SE->getMinusSCEV(A[K].PosPart, B[K].Coeff));
2837 if (PosPart->isZero())
2838 Bound[K].Upper[Dependence::DVEntry::LT] = SE->getNegativeSCEV(B[K].Coeff);
2839 }
2840 }
2841
2842
2843 // Computes the upper and lower bounds for level K
2844 // using the > direction. Records them in Bound.
2845 // Wolfe gives the equations
2846 //
2847 // LB^>_k = (A_k - B^+_k)^- (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k
2848 // UB^>_k = (A_k - B^-_k)^+ (U_k - L_k - N_k) + (A_k - B_k)L_k + A_k N_k
2849 //
2850 // Since we normalize loops, we can simplify these equations to
2851 //
2852 // LB^>_k = (A_k - B^+_k)^- (U_k - 1) + A_k
2853 // UB^>_k = (A_k - B^-_k)^+ (U_k - 1) + A_k
2854 //
2855 // We must be careful to handle the case where the upper bound is unknown.
findBoundsGT(CoefficientInfo * A,CoefficientInfo * B,BoundInfo * Bound,unsigned K) const2856 void DependenceInfo::findBoundsGT(CoefficientInfo *A, CoefficientInfo *B,
2857 BoundInfo *Bound, unsigned K) const {
2858 Bound[K].Lower[Dependence::DVEntry::GT] = nullptr; // Default value = -infinity.
2859 Bound[K].Upper[Dependence::DVEntry::GT] = nullptr; // Default value = +infinity.
2860 if (Bound[K].Iterations) {
2861 const SCEV *Iter_1 = SE->getMinusSCEV(
2862 Bound[K].Iterations, SE->getOne(Bound[K].Iterations->getType()));
2863 const SCEV *NegPart =
2864 getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart));
2865 Bound[K].Lower[Dependence::DVEntry::GT] =
2866 SE->getAddExpr(SE->getMulExpr(NegPart, Iter_1), A[K].Coeff);
2867 const SCEV *PosPart =
2868 getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart));
2869 Bound[K].Upper[Dependence::DVEntry::GT] =
2870 SE->getAddExpr(SE->getMulExpr(PosPart, Iter_1), A[K].Coeff);
2871 }
2872 else {
2873 // If the positive/negative part of the difference is 0,
2874 // we won't need to know the number of iterations.
2875 const SCEV *NegPart = getNegativePart(SE->getMinusSCEV(A[K].Coeff, B[K].PosPart));
2876 if (NegPart->isZero())
2877 Bound[K].Lower[Dependence::DVEntry::GT] = A[K].Coeff;
2878 const SCEV *PosPart = getPositivePart(SE->getMinusSCEV(A[K].Coeff, B[K].NegPart));
2879 if (PosPart->isZero())
2880 Bound[K].Upper[Dependence::DVEntry::GT] = A[K].Coeff;
2881 }
2882 }
2883
2884
2885 // X^+ = max(X, 0)
getPositivePart(const SCEV * X) const2886 const SCEV *DependenceInfo::getPositivePart(const SCEV *X) const {
2887 return SE->getSMaxExpr(X, SE->getZero(X->getType()));
2888 }
2889
2890
2891 // X^- = min(X, 0)
getNegativePart(const SCEV * X) const2892 const SCEV *DependenceInfo::getNegativePart(const SCEV *X) const {
2893 return SE->getSMinExpr(X, SE->getZero(X->getType()));
2894 }
2895
2896
2897 // Walks through the subscript,
2898 // collecting each coefficient, the associated loop bounds,
2899 // and recording its positive and negative parts for later use.
2900 DependenceInfo::CoefficientInfo *
collectCoeffInfo(const SCEV * Subscript,bool SrcFlag,const SCEV * & Constant) const2901 DependenceInfo::collectCoeffInfo(const SCEV *Subscript, bool SrcFlag,
2902 const SCEV *&Constant) const {
2903 const SCEV *Zero = SE->getZero(Subscript->getType());
2904 CoefficientInfo *CI = new CoefficientInfo[MaxLevels + 1];
2905 for (unsigned K = 1; K <= MaxLevels; ++K) {
2906 CI[K].Coeff = Zero;
2907 CI[K].PosPart = Zero;
2908 CI[K].NegPart = Zero;
2909 CI[K].Iterations = nullptr;
2910 }
2911 while (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Subscript)) {
2912 const Loop *L = AddRec->getLoop();
2913 unsigned K = SrcFlag ? mapSrcLoop(L) : mapDstLoop(L);
2914 CI[K].Coeff = AddRec->getStepRecurrence(*SE);
2915 CI[K].PosPart = getPositivePart(CI[K].Coeff);
2916 CI[K].NegPart = getNegativePart(CI[K].Coeff);
2917 CI[K].Iterations = collectUpperBound(L, Subscript->getType());
2918 Subscript = AddRec->getStart();
2919 }
2920 Constant = Subscript;
2921 #ifndef NDEBUG
2922 LLVM_DEBUG(dbgs() << "\tCoefficient Info\n");
2923 for (unsigned K = 1; K <= MaxLevels; ++K) {
2924 LLVM_DEBUG(dbgs() << "\t " << K << "\t" << *CI[K].Coeff);
2925 LLVM_DEBUG(dbgs() << "\tPos Part = ");
2926 LLVM_DEBUG(dbgs() << *CI[K].PosPart);
2927 LLVM_DEBUG(dbgs() << "\tNeg Part = ");
2928 LLVM_DEBUG(dbgs() << *CI[K].NegPart);
2929 LLVM_DEBUG(dbgs() << "\tUpper Bound = ");
2930 if (CI[K].Iterations)
2931 LLVM_DEBUG(dbgs() << *CI[K].Iterations);
2932 else
2933 LLVM_DEBUG(dbgs() << "+inf");
2934 LLVM_DEBUG(dbgs() << '\n');
2935 }
2936 LLVM_DEBUG(dbgs() << "\t Constant = " << *Subscript << '\n');
2937 #endif
2938 return CI;
2939 }
2940
2941
2942 // Looks through all the bounds info and
2943 // computes the lower bound given the current direction settings
2944 // at each level. If the lower bound for any level is -inf,
2945 // the result is -inf.
getLowerBound(BoundInfo * Bound) const2946 const SCEV *DependenceInfo::getLowerBound(BoundInfo *Bound) const {
2947 const SCEV *Sum = Bound[1].Lower[Bound[1].Direction];
2948 for (unsigned K = 2; Sum && K <= MaxLevels; ++K) {
2949 if (Bound[K].Lower[Bound[K].Direction])
2950 Sum = SE->getAddExpr(Sum, Bound[K].Lower[Bound[K].Direction]);
2951 else
2952 Sum = nullptr;
2953 }
2954 return Sum;
2955 }
2956
2957
2958 // Looks through all the bounds info and
2959 // computes the upper bound given the current direction settings
2960 // at each level. If the upper bound at any level is +inf,
2961 // the result is +inf.
getUpperBound(BoundInfo * Bound) const2962 const SCEV *DependenceInfo::getUpperBound(BoundInfo *Bound) const {
2963 const SCEV *Sum = Bound[1].Upper[Bound[1].Direction];
2964 for (unsigned K = 2; Sum && K <= MaxLevels; ++K) {
2965 if (Bound[K].Upper[Bound[K].Direction])
2966 Sum = SE->getAddExpr(Sum, Bound[K].Upper[Bound[K].Direction]);
2967 else
2968 Sum = nullptr;
2969 }
2970 return Sum;
2971 }
2972
2973
2974 //===----------------------------------------------------------------------===//
2975 // Constraint manipulation for Delta test.
2976
2977 // Given a linear SCEV,
2978 // return the coefficient (the step)
2979 // corresponding to the specified loop.
2980 // If there isn't one, return 0.
2981 // For example, given a*i + b*j + c*k, finding the coefficient
2982 // corresponding to the j loop would yield b.
findCoefficient(const SCEV * Expr,const Loop * TargetLoop) const2983 const SCEV *DependenceInfo::findCoefficient(const SCEV *Expr,
2984 const Loop *TargetLoop) const {
2985 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
2986 if (!AddRec)
2987 return SE->getZero(Expr->getType());
2988 if (AddRec->getLoop() == TargetLoop)
2989 return AddRec->getStepRecurrence(*SE);
2990 return findCoefficient(AddRec->getStart(), TargetLoop);
2991 }
2992
2993
2994 // Given a linear SCEV,
2995 // return the SCEV given by zeroing out the coefficient
2996 // corresponding to the specified loop.
2997 // For example, given a*i + b*j + c*k, zeroing the coefficient
2998 // corresponding to the j loop would yield a*i + c*k.
zeroCoefficient(const SCEV * Expr,const Loop * TargetLoop) const2999 const SCEV *DependenceInfo::zeroCoefficient(const SCEV *Expr,
3000 const Loop *TargetLoop) const {
3001 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
3002 if (!AddRec)
3003 return Expr; // ignore
3004 if (AddRec->getLoop() == TargetLoop)
3005 return AddRec->getStart();
3006 return SE->getAddRecExpr(zeroCoefficient(AddRec->getStart(), TargetLoop),
3007 AddRec->getStepRecurrence(*SE),
3008 AddRec->getLoop(),
3009 AddRec->getNoWrapFlags());
3010 }
3011
3012
3013 // Given a linear SCEV Expr,
3014 // return the SCEV given by adding some Value to the
3015 // coefficient corresponding to the specified TargetLoop.
3016 // For example, given a*i + b*j + c*k, adding 1 to the coefficient
3017 // corresponding to the j loop would yield a*i + (b+1)*j + c*k.
addToCoefficient(const SCEV * Expr,const Loop * TargetLoop,const SCEV * Value) const3018 const SCEV *DependenceInfo::addToCoefficient(const SCEV *Expr,
3019 const Loop *TargetLoop,
3020 const SCEV *Value) const {
3021 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Expr);
3022 if (!AddRec) // create a new addRec
3023 return SE->getAddRecExpr(Expr,
3024 Value,
3025 TargetLoop,
3026 SCEV::FlagAnyWrap); // Worst case, with no info.
3027 if (AddRec->getLoop() == TargetLoop) {
3028 const SCEV *Sum = SE->getAddExpr(AddRec->getStepRecurrence(*SE), Value);
3029 if (Sum->isZero())
3030 return AddRec->getStart();
3031 return SE->getAddRecExpr(AddRec->getStart(),
3032 Sum,
3033 AddRec->getLoop(),
3034 AddRec->getNoWrapFlags());
3035 }
3036 if (SE->isLoopInvariant(AddRec, TargetLoop))
3037 return SE->getAddRecExpr(AddRec, Value, TargetLoop, SCEV::FlagAnyWrap);
3038 return SE->getAddRecExpr(
3039 addToCoefficient(AddRec->getStart(), TargetLoop, Value),
3040 AddRec->getStepRecurrence(*SE), AddRec->getLoop(),
3041 AddRec->getNoWrapFlags());
3042 }
3043
3044
3045 // Review the constraints, looking for opportunities
3046 // to simplify a subscript pair (Src and Dst).
3047 // Return true if some simplification occurs.
3048 // If the simplification isn't exact (that is, if it is conservative
3049 // in terms of dependence), set consistent to false.
3050 // Corresponds to Figure 5 from the paper
3051 //
3052 // Practical Dependence Testing
3053 // Goff, Kennedy, Tseng
3054 // PLDI 1991
propagate(const SCEV * & Src,const SCEV * & Dst,SmallBitVector & Loops,SmallVectorImpl<Constraint> & Constraints,bool & Consistent)3055 bool DependenceInfo::propagate(const SCEV *&Src, const SCEV *&Dst,
3056 SmallBitVector &Loops,
3057 SmallVectorImpl<Constraint> &Constraints,
3058 bool &Consistent) {
3059 bool Result = false;
3060 for (unsigned LI : Loops.set_bits()) {
3061 LLVM_DEBUG(dbgs() << "\t Constraint[" << LI << "] is");
3062 LLVM_DEBUG(Constraints[LI].dump(dbgs()));
3063 if (Constraints[LI].isDistance())
3064 Result |= propagateDistance(Src, Dst, Constraints[LI], Consistent);
3065 else if (Constraints[LI].isLine())
3066 Result |= propagateLine(Src, Dst, Constraints[LI], Consistent);
3067 else if (Constraints[LI].isPoint())
3068 Result |= propagatePoint(Src, Dst, Constraints[LI]);
3069 }
3070 return Result;
3071 }
3072
3073
3074 // Attempt to propagate a distance
3075 // constraint into a subscript pair (Src and Dst).
3076 // Return true if some simplification occurs.
3077 // If the simplification isn't exact (that is, if it is conservative
3078 // in terms of dependence), set consistent to false.
propagateDistance(const SCEV * & Src,const SCEV * & Dst,Constraint & CurConstraint,bool & Consistent)3079 bool DependenceInfo::propagateDistance(const SCEV *&Src, const SCEV *&Dst,
3080 Constraint &CurConstraint,
3081 bool &Consistent) {
3082 const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3083 LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n");
3084 const SCEV *A_K = findCoefficient(Src, CurLoop);
3085 if (A_K->isZero())
3086 return false;
3087 const SCEV *DA_K = SE->getMulExpr(A_K, CurConstraint.getD());
3088 Src = SE->getMinusSCEV(Src, DA_K);
3089 Src = zeroCoefficient(Src, CurLoop);
3090 LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n");
3091 LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n");
3092 Dst = addToCoefficient(Dst, CurLoop, SE->getNegativeSCEV(A_K));
3093 LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n");
3094 if (!findCoefficient(Dst, CurLoop)->isZero())
3095 Consistent = false;
3096 return true;
3097 }
3098
3099
3100 // Attempt to propagate a line
3101 // constraint into a subscript pair (Src and Dst).
3102 // Return true if some simplification occurs.
3103 // If the simplification isn't exact (that is, if it is conservative
3104 // in terms of dependence), set consistent to false.
propagateLine(const SCEV * & Src,const SCEV * & Dst,Constraint & CurConstraint,bool & Consistent)3105 bool DependenceInfo::propagateLine(const SCEV *&Src, const SCEV *&Dst,
3106 Constraint &CurConstraint,
3107 bool &Consistent) {
3108 const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3109 const SCEV *A = CurConstraint.getA();
3110 const SCEV *B = CurConstraint.getB();
3111 const SCEV *C = CurConstraint.getC();
3112 LLVM_DEBUG(dbgs() << "\t\tA = " << *A << ", B = " << *B << ", C = " << *C
3113 << "\n");
3114 LLVM_DEBUG(dbgs() << "\t\tSrc = " << *Src << "\n");
3115 LLVM_DEBUG(dbgs() << "\t\tDst = " << *Dst << "\n");
3116 if (A->isZero()) {
3117 const SCEVConstant *Bconst = dyn_cast<SCEVConstant>(B);
3118 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3119 if (!Bconst || !Cconst) return false;
3120 APInt Beta = Bconst->getAPInt();
3121 APInt Charlie = Cconst->getAPInt();
3122 APInt CdivB = Charlie.sdiv(Beta);
3123 assert(Charlie.srem(Beta) == 0 && "C should be evenly divisible by B");
3124 const SCEV *AP_K = findCoefficient(Dst, CurLoop);
3125 // Src = SE->getAddExpr(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB)));
3126 Src = SE->getMinusSCEV(Src, SE->getMulExpr(AP_K, SE->getConstant(CdivB)));
3127 Dst = zeroCoefficient(Dst, CurLoop);
3128 if (!findCoefficient(Src, CurLoop)->isZero())
3129 Consistent = false;
3130 }
3131 else if (B->isZero()) {
3132 const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A);
3133 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3134 if (!Aconst || !Cconst) return false;
3135 APInt Alpha = Aconst->getAPInt();
3136 APInt Charlie = Cconst->getAPInt();
3137 APInt CdivA = Charlie.sdiv(Alpha);
3138 assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A");
3139 const SCEV *A_K = findCoefficient(Src, CurLoop);
3140 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA)));
3141 Src = zeroCoefficient(Src, CurLoop);
3142 if (!findCoefficient(Dst, CurLoop)->isZero())
3143 Consistent = false;
3144 }
3145 else if (isKnownPredicate(CmpInst::ICMP_EQ, A, B)) {
3146 const SCEVConstant *Aconst = dyn_cast<SCEVConstant>(A);
3147 const SCEVConstant *Cconst = dyn_cast<SCEVConstant>(C);
3148 if (!Aconst || !Cconst) return false;
3149 APInt Alpha = Aconst->getAPInt();
3150 APInt Charlie = Cconst->getAPInt();
3151 APInt CdivA = Charlie.sdiv(Alpha);
3152 assert(Charlie.srem(Alpha) == 0 && "C should be evenly divisible by A");
3153 const SCEV *A_K = findCoefficient(Src, CurLoop);
3154 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, SE->getConstant(CdivA)));
3155 Src = zeroCoefficient(Src, CurLoop);
3156 Dst = addToCoefficient(Dst, CurLoop, A_K);
3157 if (!findCoefficient(Dst, CurLoop)->isZero())
3158 Consistent = false;
3159 }
3160 else {
3161 // paper is incorrect here, or perhaps just misleading
3162 const SCEV *A_K = findCoefficient(Src, CurLoop);
3163 Src = SE->getMulExpr(Src, A);
3164 Dst = SE->getMulExpr(Dst, A);
3165 Src = SE->getAddExpr(Src, SE->getMulExpr(A_K, C));
3166 Src = zeroCoefficient(Src, CurLoop);
3167 Dst = addToCoefficient(Dst, CurLoop, SE->getMulExpr(A_K, B));
3168 if (!findCoefficient(Dst, CurLoop)->isZero())
3169 Consistent = false;
3170 }
3171 LLVM_DEBUG(dbgs() << "\t\tnew Src = " << *Src << "\n");
3172 LLVM_DEBUG(dbgs() << "\t\tnew Dst = " << *Dst << "\n");
3173 return true;
3174 }
3175
3176
3177 // Attempt to propagate a point
3178 // constraint into a subscript pair (Src and Dst).
3179 // Return true if some simplification occurs.
propagatePoint(const SCEV * & Src,const SCEV * & Dst,Constraint & CurConstraint)3180 bool DependenceInfo::propagatePoint(const SCEV *&Src, const SCEV *&Dst,
3181 Constraint &CurConstraint) {
3182 const Loop *CurLoop = CurConstraint.getAssociatedLoop();
3183 const SCEV *A_K = findCoefficient(Src, CurLoop);
3184 const SCEV *AP_K = findCoefficient(Dst, CurLoop);
3185 const SCEV *XA_K = SE->getMulExpr(A_K, CurConstraint.getX());
3186 const SCEV *YAP_K = SE->getMulExpr(AP_K, CurConstraint.getY());
3187 LLVM_DEBUG(dbgs() << "\t\tSrc is " << *Src << "\n");
3188 Src = SE->getAddExpr(Src, SE->getMinusSCEV(XA_K, YAP_K));
3189 Src = zeroCoefficient(Src, CurLoop);
3190 LLVM_DEBUG(dbgs() << "\t\tnew Src is " << *Src << "\n");
3191 LLVM_DEBUG(dbgs() << "\t\tDst is " << *Dst << "\n");
3192 Dst = zeroCoefficient(Dst, CurLoop);
3193 LLVM_DEBUG(dbgs() << "\t\tnew Dst is " << *Dst << "\n");
3194 return true;
3195 }
3196
3197
3198 // Update direction vector entry based on the current constraint.
updateDirection(Dependence::DVEntry & Level,const Constraint & CurConstraint) const3199 void DependenceInfo::updateDirection(Dependence::DVEntry &Level,
3200 const Constraint &CurConstraint) const {
3201 LLVM_DEBUG(dbgs() << "\tUpdate direction, constraint =");
3202 LLVM_DEBUG(CurConstraint.dump(dbgs()));
3203 if (CurConstraint.isAny())
3204 ; // use defaults
3205 else if (CurConstraint.isDistance()) {
3206 // this one is consistent, the others aren't
3207 Level.Scalar = false;
3208 Level.Distance = CurConstraint.getD();
3209 unsigned NewDirection = Dependence::DVEntry::NONE;
3210 if (!SE->isKnownNonZero(Level.Distance)) // if may be zero
3211 NewDirection = Dependence::DVEntry::EQ;
3212 if (!SE->isKnownNonPositive(Level.Distance)) // if may be positive
3213 NewDirection |= Dependence::DVEntry::LT;
3214 if (!SE->isKnownNonNegative(Level.Distance)) // if may be negative
3215 NewDirection |= Dependence::DVEntry::GT;
3216 Level.Direction &= NewDirection;
3217 }
3218 else if (CurConstraint.isLine()) {
3219 Level.Scalar = false;
3220 Level.Distance = nullptr;
3221 // direction should be accurate
3222 }
3223 else if (CurConstraint.isPoint()) {
3224 Level.Scalar = false;
3225 Level.Distance = nullptr;
3226 unsigned NewDirection = Dependence::DVEntry::NONE;
3227 if (!isKnownPredicate(CmpInst::ICMP_NE,
3228 CurConstraint.getY(),
3229 CurConstraint.getX()))
3230 // if X may be = Y
3231 NewDirection |= Dependence::DVEntry::EQ;
3232 if (!isKnownPredicate(CmpInst::ICMP_SLE,
3233 CurConstraint.getY(),
3234 CurConstraint.getX()))
3235 // if Y may be > X
3236 NewDirection |= Dependence::DVEntry::LT;
3237 if (!isKnownPredicate(CmpInst::ICMP_SGE,
3238 CurConstraint.getY(),
3239 CurConstraint.getX()))
3240 // if Y may be < X
3241 NewDirection |= Dependence::DVEntry::GT;
3242 Level.Direction &= NewDirection;
3243 }
3244 else
3245 llvm_unreachable("constraint has unexpected kind");
3246 }
3247
3248 /// Check if we can delinearize the subscripts. If the SCEVs representing the
3249 /// source and destination array references are recurrences on a nested loop,
3250 /// this function flattens the nested recurrences into separate recurrences
3251 /// for each loop level.
tryDelinearize(Instruction * Src,Instruction * Dst,SmallVectorImpl<Subscript> & Pair)3252 bool DependenceInfo::tryDelinearize(Instruction *Src, Instruction *Dst,
3253 SmallVectorImpl<Subscript> &Pair) {
3254 assert(isLoadOrStore(Src) && "instruction is not load or store");
3255 assert(isLoadOrStore(Dst) && "instruction is not load or store");
3256 Value *SrcPtr = getLoadStorePointerOperand(Src);
3257 Value *DstPtr = getLoadStorePointerOperand(Dst);
3258 Loop *SrcLoop = LI->getLoopFor(Src->getParent());
3259 Loop *DstLoop = LI->getLoopFor(Dst->getParent());
3260 const SCEV *SrcAccessFn = SE->getSCEVAtScope(SrcPtr, SrcLoop);
3261 const SCEV *DstAccessFn = SE->getSCEVAtScope(DstPtr, DstLoop);
3262 const SCEVUnknown *SrcBase =
3263 dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
3264 const SCEVUnknown *DstBase =
3265 dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
3266
3267 if (!SrcBase || !DstBase || SrcBase != DstBase)
3268 return false;
3269
3270 SmallVector<const SCEV *, 4> SrcSubscripts, DstSubscripts;
3271
3272 if (!tryDelinearizeFixedSize(Src, Dst, SrcAccessFn, DstAccessFn,
3273 SrcSubscripts, DstSubscripts) &&
3274 !tryDelinearizeParametricSize(Src, Dst, SrcAccessFn, DstAccessFn,
3275 SrcSubscripts, DstSubscripts))
3276 return false;
3277
3278 int Size = SrcSubscripts.size();
3279 LLVM_DEBUG({
3280 dbgs() << "\nSrcSubscripts: ";
3281 for (int I = 0; I < Size; I++)
3282 dbgs() << *SrcSubscripts[I];
3283 dbgs() << "\nDstSubscripts: ";
3284 for (int I = 0; I < Size; I++)
3285 dbgs() << *DstSubscripts[I];
3286 });
3287
3288 // The delinearization transforms a single-subscript MIV dependence test into
3289 // a multi-subscript SIV dependence test that is easier to compute. So we
3290 // resize Pair to contain as many pairs of subscripts as the delinearization
3291 // has found, and then initialize the pairs following the delinearization.
3292 Pair.resize(Size);
3293 for (int I = 0; I < Size; ++I) {
3294 Pair[I].Src = SrcSubscripts[I];
3295 Pair[I].Dst = DstSubscripts[I];
3296 unifySubscriptType(&Pair[I]);
3297 }
3298
3299 return true;
3300 }
3301
tryDelinearizeFixedSize(Instruction * Src,Instruction * Dst,const SCEV * SrcAccessFn,const SCEV * DstAccessFn,SmallVectorImpl<const SCEV * > & SrcSubscripts,SmallVectorImpl<const SCEV * > & DstSubscripts)3302 bool DependenceInfo::tryDelinearizeFixedSize(
3303 Instruction *Src, Instruction *Dst, const SCEV *SrcAccessFn,
3304 const SCEV *DstAccessFn, SmallVectorImpl<const SCEV *> &SrcSubscripts,
3305 SmallVectorImpl<const SCEV *> &DstSubscripts) {
3306
3307 Value *SrcPtr = getLoadStorePointerOperand(Src);
3308 Value *DstPtr = getLoadStorePointerOperand(Dst);
3309 const SCEVUnknown *SrcBase =
3310 dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
3311 const SCEVUnknown *DstBase =
3312 dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
3313 assert(SrcBase && DstBase && SrcBase == DstBase &&
3314 "expected src and dst scev unknowns to be equal");
3315
3316 // Check the simple case where the array dimensions are fixed size.
3317 auto *SrcGEP = dyn_cast<GetElementPtrInst>(SrcPtr);
3318 auto *DstGEP = dyn_cast<GetElementPtrInst>(DstPtr);
3319 if (!SrcGEP || !DstGEP)
3320 return false;
3321
3322 SmallVector<int, 4> SrcSizes, DstSizes;
3323 SE->getIndexExpressionsFromGEP(SrcGEP, SrcSubscripts, SrcSizes);
3324 SE->getIndexExpressionsFromGEP(DstGEP, DstSubscripts, DstSizes);
3325
3326 // Check that the two size arrays are non-empty and equal in length and
3327 // value.
3328 if (SrcSizes.empty() || SrcSubscripts.size() <= 1 ||
3329 SrcSizes.size() != DstSizes.size() ||
3330 !std::equal(SrcSizes.begin(), SrcSizes.end(), DstSizes.begin())) {
3331 SrcSubscripts.clear();
3332 DstSubscripts.clear();
3333 return false;
3334 }
3335
3336 Value *SrcBasePtr = SrcGEP->getOperand(0);
3337 Value *DstBasePtr = DstGEP->getOperand(0);
3338 while (auto *PCast = dyn_cast<BitCastInst>(SrcBasePtr))
3339 SrcBasePtr = PCast->getOperand(0);
3340 while (auto *PCast = dyn_cast<BitCastInst>(DstBasePtr))
3341 DstBasePtr = PCast->getOperand(0);
3342
3343 // Check that for identical base pointers we do not miss index offsets
3344 // that have been added before this GEP is applied.
3345 if (SrcBasePtr != SrcBase->getValue() || DstBasePtr != DstBase->getValue()) {
3346 SrcSubscripts.clear();
3347 DstSubscripts.clear();
3348 return false;
3349 }
3350
3351 assert(SrcSubscripts.size() == DstSubscripts.size() &&
3352 SrcSubscripts.size() == SrcSizes.size() + 1 &&
3353 "Expected equal number of entries in the list of sizes and "
3354 "subscripts.");
3355
3356 // In general we cannot safely assume that the subscripts recovered from GEPs
3357 // are in the range of values defined for their corresponding array
3358 // dimensions. For example some C language usage/interpretation make it
3359 // impossible to verify this at compile-time. As such we can only delinearize
3360 // iff the subscripts are positive and are less than the range of the
3361 // dimension.
3362 if (!DisableDelinearizationChecks) {
3363 auto AllIndiciesInRange = [&](SmallVector<int, 4> &DimensionSizes,
3364 SmallVectorImpl<const SCEV *> &Subscripts,
3365 Value *Ptr) {
3366 size_t SSize = Subscripts.size();
3367 for (size_t I = 1; I < SSize; ++I) {
3368 const SCEV *S = Subscripts[I];
3369 if (!isKnownNonNegative(S, Ptr))
3370 return false;
3371 if (auto *SType = dyn_cast<IntegerType>(S->getType())) {
3372 const SCEV *Range = SE->getConstant(
3373 ConstantInt::get(SType, DimensionSizes[I - 1], false));
3374 if (!isKnownLessThan(S, Range))
3375 return false;
3376 }
3377 }
3378 return true;
3379 };
3380
3381 if (!AllIndiciesInRange(SrcSizes, SrcSubscripts, SrcPtr) ||
3382 !AllIndiciesInRange(DstSizes, DstSubscripts, DstPtr)) {
3383 SrcSubscripts.clear();
3384 DstSubscripts.clear();
3385 return false;
3386 }
3387 }
3388 LLVM_DEBUG({
3389 dbgs() << "Delinearized subscripts of fixed-size array\n"
3390 << "SrcGEP:" << *SrcGEP << "\n"
3391 << "DstGEP:" << *DstGEP << "\n";
3392 });
3393 return true;
3394 }
3395
tryDelinearizeParametricSize(Instruction * Src,Instruction * Dst,const SCEV * SrcAccessFn,const SCEV * DstAccessFn,SmallVectorImpl<const SCEV * > & SrcSubscripts,SmallVectorImpl<const SCEV * > & DstSubscripts)3396 bool DependenceInfo::tryDelinearizeParametricSize(
3397 Instruction *Src, Instruction *Dst, const SCEV *SrcAccessFn,
3398 const SCEV *DstAccessFn, SmallVectorImpl<const SCEV *> &SrcSubscripts,
3399 SmallVectorImpl<const SCEV *> &DstSubscripts) {
3400
3401 Value *SrcPtr = getLoadStorePointerOperand(Src);
3402 Value *DstPtr = getLoadStorePointerOperand(Dst);
3403 const SCEVUnknown *SrcBase =
3404 dyn_cast<SCEVUnknown>(SE->getPointerBase(SrcAccessFn));
3405 const SCEVUnknown *DstBase =
3406 dyn_cast<SCEVUnknown>(SE->getPointerBase(DstAccessFn));
3407 assert(SrcBase && DstBase && SrcBase == DstBase &&
3408 "expected src and dst scev unknowns to be equal");
3409
3410 const SCEV *ElementSize = SE->getElementSize(Src);
3411 if (ElementSize != SE->getElementSize(Dst))
3412 return false;
3413
3414 const SCEV *SrcSCEV = SE->getMinusSCEV(SrcAccessFn, SrcBase);
3415 const SCEV *DstSCEV = SE->getMinusSCEV(DstAccessFn, DstBase);
3416
3417 const SCEVAddRecExpr *SrcAR = dyn_cast<SCEVAddRecExpr>(SrcSCEV);
3418 const SCEVAddRecExpr *DstAR = dyn_cast<SCEVAddRecExpr>(DstSCEV);
3419 if (!SrcAR || !DstAR || !SrcAR->isAffine() || !DstAR->isAffine())
3420 return false;
3421
3422 // First step: collect parametric terms in both array references.
3423 SmallVector<const SCEV *, 4> Terms;
3424 SE->collectParametricTerms(SrcAR, Terms);
3425 SE->collectParametricTerms(DstAR, Terms);
3426
3427 // Second step: find subscript sizes.
3428 SmallVector<const SCEV *, 4> Sizes;
3429 SE->findArrayDimensions(Terms, Sizes, ElementSize);
3430
3431 // Third step: compute the access functions for each subscript.
3432 SE->computeAccessFunctions(SrcAR, SrcSubscripts, Sizes);
3433 SE->computeAccessFunctions(DstAR, DstSubscripts, Sizes);
3434
3435 // Fail when there is only a subscript: that's a linearized access function.
3436 if (SrcSubscripts.size() < 2 || DstSubscripts.size() < 2 ||
3437 SrcSubscripts.size() != DstSubscripts.size())
3438 return false;
3439
3440 size_t Size = SrcSubscripts.size();
3441
3442 // Statically check that the array bounds are in-range. The first subscript we
3443 // don't have a size for and it cannot overflow into another subscript, so is
3444 // always safe. The others need to be 0 <= subscript[i] < bound, for both src
3445 // and dst.
3446 // FIXME: It may be better to record these sizes and add them as constraints
3447 // to the dependency checks.
3448 if (!DisableDelinearizationChecks)
3449 for (size_t I = 1; I < Size; ++I) {
3450 if (!isKnownNonNegative(SrcSubscripts[I], SrcPtr))
3451 return false;
3452
3453 if (!isKnownLessThan(SrcSubscripts[I], Sizes[I - 1]))
3454 return false;
3455
3456 if (!isKnownNonNegative(DstSubscripts[I], DstPtr))
3457 return false;
3458
3459 if (!isKnownLessThan(DstSubscripts[I], Sizes[I - 1]))
3460 return false;
3461 }
3462
3463 return true;
3464 }
3465
3466 //===----------------------------------------------------------------------===//
3467
3468 #ifndef NDEBUG
3469 // For debugging purposes, dump a small bit vector to dbgs().
dumpSmallBitVector(SmallBitVector & BV)3470 static void dumpSmallBitVector(SmallBitVector &BV) {
3471 dbgs() << "{";
3472 for (unsigned VI : BV.set_bits()) {
3473 dbgs() << VI;
3474 if (BV.find_next(VI) >= 0)
3475 dbgs() << ' ';
3476 }
3477 dbgs() << "}\n";
3478 }
3479 #endif
3480
invalidate(Function & F,const PreservedAnalyses & PA,FunctionAnalysisManager::Invalidator & Inv)3481 bool DependenceInfo::invalidate(Function &F, const PreservedAnalyses &PA,
3482 FunctionAnalysisManager::Invalidator &Inv) {
3483 // Check if the analysis itself has been invalidated.
3484 auto PAC = PA.getChecker<DependenceAnalysis>();
3485 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
3486 return true;
3487
3488 // Check transitive dependencies.
3489 return Inv.invalidate<AAManager>(F, PA) ||
3490 Inv.invalidate<ScalarEvolutionAnalysis>(F, PA) ||
3491 Inv.invalidate<LoopAnalysis>(F, PA);
3492 }
3493
3494 // depends -
3495 // Returns NULL if there is no dependence.
3496 // Otherwise, return a Dependence with as many details as possible.
3497 // Corresponds to Section 3.1 in the paper
3498 //
3499 // Practical Dependence Testing
3500 // Goff, Kennedy, Tseng
3501 // PLDI 1991
3502 //
3503 // Care is required to keep the routine below, getSplitIteration(),
3504 // up to date with respect to this routine.
3505 std::unique_ptr<Dependence>
depends(Instruction * Src,Instruction * Dst,bool PossiblyLoopIndependent)3506 DependenceInfo::depends(Instruction *Src, Instruction *Dst,
3507 bool PossiblyLoopIndependent) {
3508 if (Src == Dst)
3509 PossiblyLoopIndependent = false;
3510
3511 if (!(Src->mayReadOrWriteMemory() && Dst->mayReadOrWriteMemory()))
3512 // if both instructions don't reference memory, there's no dependence
3513 return nullptr;
3514
3515 if (!isLoadOrStore(Src) || !isLoadOrStore(Dst)) {
3516 // can only analyze simple loads and stores, i.e., no calls, invokes, etc.
3517 LLVM_DEBUG(dbgs() << "can only handle simple loads and stores\n");
3518 return std::make_unique<Dependence>(Src, Dst);
3519 }
3520
3521 assert(isLoadOrStore(Src) && "instruction is not load or store");
3522 assert(isLoadOrStore(Dst) && "instruction is not load or store");
3523 Value *SrcPtr = getLoadStorePointerOperand(Src);
3524 Value *DstPtr = getLoadStorePointerOperand(Dst);
3525
3526 switch (underlyingObjectsAlias(AA, F->getParent()->getDataLayout(),
3527 MemoryLocation::get(Dst),
3528 MemoryLocation::get(Src))) {
3529 case AliasResult::MayAlias:
3530 case AliasResult::PartialAlias:
3531 // cannot analyse objects if we don't understand their aliasing.
3532 LLVM_DEBUG(dbgs() << "can't analyze may or partial alias\n");
3533 return std::make_unique<Dependence>(Src, Dst);
3534 case AliasResult::NoAlias:
3535 // If the objects noalias, they are distinct, accesses are independent.
3536 LLVM_DEBUG(dbgs() << "no alias\n");
3537 return nullptr;
3538 case AliasResult::MustAlias:
3539 break; // The underlying objects alias; test accesses for dependence.
3540 }
3541
3542 // establish loop nesting levels
3543 establishNestingLevels(Src, Dst);
3544 LLVM_DEBUG(dbgs() << " common nesting levels = " << CommonLevels << "\n");
3545 LLVM_DEBUG(dbgs() << " maximum nesting levels = " << MaxLevels << "\n");
3546
3547 FullDependence Result(Src, Dst, PossiblyLoopIndependent, CommonLevels);
3548 ++TotalArrayPairs;
3549
3550 unsigned Pairs = 1;
3551 SmallVector<Subscript, 2> Pair(Pairs);
3552 const SCEV *SrcSCEV = SE->getSCEV(SrcPtr);
3553 const SCEV *DstSCEV = SE->getSCEV(DstPtr);
3554 LLVM_DEBUG(dbgs() << " SrcSCEV = " << *SrcSCEV << "\n");
3555 LLVM_DEBUG(dbgs() << " DstSCEV = " << *DstSCEV << "\n");
3556 if (SE->getPointerBase(SrcSCEV) != SE->getPointerBase(DstSCEV)) {
3557 // If two pointers have different bases, trying to analyze indexes won't
3558 // work; we can't compare them to each other. This can happen, for example,
3559 // if one is produced by an LCSSA PHI node.
3560 //
3561 // We check this upfront so we don't crash in cases where getMinusSCEV()
3562 // returns a SCEVCouldNotCompute.
3563 LLVM_DEBUG(dbgs() << "can't analyze SCEV with different pointer base\n");
3564 return std::make_unique<Dependence>(Src, Dst);
3565 }
3566 Pair[0].Src = SrcSCEV;
3567 Pair[0].Dst = DstSCEV;
3568
3569 if (Delinearize) {
3570 if (tryDelinearize(Src, Dst, Pair)) {
3571 LLVM_DEBUG(dbgs() << " delinearized\n");
3572 Pairs = Pair.size();
3573 }
3574 }
3575
3576 for (unsigned P = 0; P < Pairs; ++P) {
3577 Pair[P].Loops.resize(MaxLevels + 1);
3578 Pair[P].GroupLoops.resize(MaxLevels + 1);
3579 Pair[P].Group.resize(Pairs);
3580 removeMatchingExtensions(&Pair[P]);
3581 Pair[P].Classification =
3582 classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()),
3583 Pair[P].Dst, LI->getLoopFor(Dst->getParent()),
3584 Pair[P].Loops);
3585 Pair[P].GroupLoops = Pair[P].Loops;
3586 Pair[P].Group.set(P);
3587 LLVM_DEBUG(dbgs() << " subscript " << P << "\n");
3588 LLVM_DEBUG(dbgs() << "\tsrc = " << *Pair[P].Src << "\n");
3589 LLVM_DEBUG(dbgs() << "\tdst = " << *Pair[P].Dst << "\n");
3590 LLVM_DEBUG(dbgs() << "\tclass = " << Pair[P].Classification << "\n");
3591 LLVM_DEBUG(dbgs() << "\tloops = ");
3592 LLVM_DEBUG(dumpSmallBitVector(Pair[P].Loops));
3593 }
3594
3595 SmallBitVector Separable(Pairs);
3596 SmallBitVector Coupled(Pairs);
3597
3598 // Partition subscripts into separable and minimally-coupled groups
3599 // Algorithm in paper is algorithmically better;
3600 // this may be faster in practice. Check someday.
3601 //
3602 // Here's an example of how it works. Consider this code:
3603 //
3604 // for (i = ...) {
3605 // for (j = ...) {
3606 // for (k = ...) {
3607 // for (l = ...) {
3608 // for (m = ...) {
3609 // A[i][j][k][m] = ...;
3610 // ... = A[0][j][l][i + j];
3611 // }
3612 // }
3613 // }
3614 // }
3615 // }
3616 //
3617 // There are 4 subscripts here:
3618 // 0 [i] and [0]
3619 // 1 [j] and [j]
3620 // 2 [k] and [l]
3621 // 3 [m] and [i + j]
3622 //
3623 // We've already classified each subscript pair as ZIV, SIV, etc.,
3624 // and collected all the loops mentioned by pair P in Pair[P].Loops.
3625 // In addition, we've initialized Pair[P].GroupLoops to Pair[P].Loops
3626 // and set Pair[P].Group = {P}.
3627 //
3628 // Src Dst Classification Loops GroupLoops Group
3629 // 0 [i] [0] SIV {1} {1} {0}
3630 // 1 [j] [j] SIV {2} {2} {1}
3631 // 2 [k] [l] RDIV {3,4} {3,4} {2}
3632 // 3 [m] [i + j] MIV {1,2,5} {1,2,5} {3}
3633 //
3634 // For each subscript SI 0 .. 3, we consider each remaining subscript, SJ.
3635 // So, 0 is compared against 1, 2, and 3; 1 is compared against 2 and 3, etc.
3636 //
3637 // We begin by comparing 0 and 1. The intersection of the GroupLoops is empty.
3638 // Next, 0 and 2. Again, the intersection of their GroupLoops is empty.
3639 // Next 0 and 3. The intersection of their GroupLoop = {1}, not empty,
3640 // so Pair[3].Group = {0,3} and Done = false (that is, 0 will not be added
3641 // to either Separable or Coupled).
3642 //
3643 // Next, we consider 1 and 2. The intersection of the GroupLoops is empty.
3644 // Next, 1 and 3. The intersection of their GroupLoops = {2}, not empty,
3645 // so Pair[3].Group = {0, 1, 3} and Done = false.
3646 //
3647 // Next, we compare 2 against 3. The intersection of the GroupLoops is empty.
3648 // Since Done remains true, we add 2 to the set of Separable pairs.
3649 //
3650 // Finally, we consider 3. There's nothing to compare it with,
3651 // so Done remains true and we add it to the Coupled set.
3652 // Pair[3].Group = {0, 1, 3} and GroupLoops = {1, 2, 5}.
3653 //
3654 // In the end, we've got 1 separable subscript and 1 coupled group.
3655 for (unsigned SI = 0; SI < Pairs; ++SI) {
3656 if (Pair[SI].Classification == Subscript::NonLinear) {
3657 // ignore these, but collect loops for later
3658 ++NonlinearSubscriptPairs;
3659 collectCommonLoops(Pair[SI].Src,
3660 LI->getLoopFor(Src->getParent()),
3661 Pair[SI].Loops);
3662 collectCommonLoops(Pair[SI].Dst,
3663 LI->getLoopFor(Dst->getParent()),
3664 Pair[SI].Loops);
3665 Result.Consistent = false;
3666 } else if (Pair[SI].Classification == Subscript::ZIV) {
3667 // always separable
3668 Separable.set(SI);
3669 }
3670 else {
3671 // SIV, RDIV, or MIV, so check for coupled group
3672 bool Done = true;
3673 for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) {
3674 SmallBitVector Intersection = Pair[SI].GroupLoops;
3675 Intersection &= Pair[SJ].GroupLoops;
3676 if (Intersection.any()) {
3677 // accumulate set of all the loops in group
3678 Pair[SJ].GroupLoops |= Pair[SI].GroupLoops;
3679 // accumulate set of all subscripts in group
3680 Pair[SJ].Group |= Pair[SI].Group;
3681 Done = false;
3682 }
3683 }
3684 if (Done) {
3685 if (Pair[SI].Group.count() == 1) {
3686 Separable.set(SI);
3687 ++SeparableSubscriptPairs;
3688 }
3689 else {
3690 Coupled.set(SI);
3691 ++CoupledSubscriptPairs;
3692 }
3693 }
3694 }
3695 }
3696
3697 LLVM_DEBUG(dbgs() << " Separable = ");
3698 LLVM_DEBUG(dumpSmallBitVector(Separable));
3699 LLVM_DEBUG(dbgs() << " Coupled = ");
3700 LLVM_DEBUG(dumpSmallBitVector(Coupled));
3701
3702 Constraint NewConstraint;
3703 NewConstraint.setAny(SE);
3704
3705 // test separable subscripts
3706 for (unsigned SI : Separable.set_bits()) {
3707 LLVM_DEBUG(dbgs() << "testing subscript " << SI);
3708 switch (Pair[SI].Classification) {
3709 case Subscript::ZIV:
3710 LLVM_DEBUG(dbgs() << ", ZIV\n");
3711 if (testZIV(Pair[SI].Src, Pair[SI].Dst, Result))
3712 return nullptr;
3713 break;
3714 case Subscript::SIV: {
3715 LLVM_DEBUG(dbgs() << ", SIV\n");
3716 unsigned Level;
3717 const SCEV *SplitIter = nullptr;
3718 if (testSIV(Pair[SI].Src, Pair[SI].Dst, Level, Result, NewConstraint,
3719 SplitIter))
3720 return nullptr;
3721 break;
3722 }
3723 case Subscript::RDIV:
3724 LLVM_DEBUG(dbgs() << ", RDIV\n");
3725 if (testRDIV(Pair[SI].Src, Pair[SI].Dst, Result))
3726 return nullptr;
3727 break;
3728 case Subscript::MIV:
3729 LLVM_DEBUG(dbgs() << ", MIV\n");
3730 if (testMIV(Pair[SI].Src, Pair[SI].Dst, Pair[SI].Loops, Result))
3731 return nullptr;
3732 break;
3733 default:
3734 llvm_unreachable("subscript has unexpected classification");
3735 }
3736 }
3737
3738 if (Coupled.count()) {
3739 // test coupled subscript groups
3740 LLVM_DEBUG(dbgs() << "starting on coupled subscripts\n");
3741 LLVM_DEBUG(dbgs() << "MaxLevels + 1 = " << MaxLevels + 1 << "\n");
3742 SmallVector<Constraint, 4> Constraints(MaxLevels + 1);
3743 for (unsigned II = 0; II <= MaxLevels; ++II)
3744 Constraints[II].setAny(SE);
3745 for (unsigned SI : Coupled.set_bits()) {
3746 LLVM_DEBUG(dbgs() << "testing subscript group " << SI << " { ");
3747 SmallBitVector Group(Pair[SI].Group);
3748 SmallBitVector Sivs(Pairs);
3749 SmallBitVector Mivs(Pairs);
3750 SmallBitVector ConstrainedLevels(MaxLevels + 1);
3751 SmallVector<Subscript *, 4> PairsInGroup;
3752 for (unsigned SJ : Group.set_bits()) {
3753 LLVM_DEBUG(dbgs() << SJ << " ");
3754 if (Pair[SJ].Classification == Subscript::SIV)
3755 Sivs.set(SJ);
3756 else
3757 Mivs.set(SJ);
3758 PairsInGroup.push_back(&Pair[SJ]);
3759 }
3760 unifySubscriptType(PairsInGroup);
3761 LLVM_DEBUG(dbgs() << "}\n");
3762 while (Sivs.any()) {
3763 bool Changed = false;
3764 for (unsigned SJ : Sivs.set_bits()) {
3765 LLVM_DEBUG(dbgs() << "testing subscript " << SJ << ", SIV\n");
3766 // SJ is an SIV subscript that's part of the current coupled group
3767 unsigned Level;
3768 const SCEV *SplitIter = nullptr;
3769 LLVM_DEBUG(dbgs() << "SIV\n");
3770 if (testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level, Result, NewConstraint,
3771 SplitIter))
3772 return nullptr;
3773 ConstrainedLevels.set(Level);
3774 if (intersectConstraints(&Constraints[Level], &NewConstraint)) {
3775 if (Constraints[Level].isEmpty()) {
3776 ++DeltaIndependence;
3777 return nullptr;
3778 }
3779 Changed = true;
3780 }
3781 Sivs.reset(SJ);
3782 }
3783 if (Changed) {
3784 // propagate, possibly creating new SIVs and ZIVs
3785 LLVM_DEBUG(dbgs() << " propagating\n");
3786 LLVM_DEBUG(dbgs() << "\tMivs = ");
3787 LLVM_DEBUG(dumpSmallBitVector(Mivs));
3788 for (unsigned SJ : Mivs.set_bits()) {
3789 // SJ is an MIV subscript that's part of the current coupled group
3790 LLVM_DEBUG(dbgs() << "\tSJ = " << SJ << "\n");
3791 if (propagate(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops,
3792 Constraints, Result.Consistent)) {
3793 LLVM_DEBUG(dbgs() << "\t Changed\n");
3794 ++DeltaPropagations;
3795 Pair[SJ].Classification =
3796 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()),
3797 Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()),
3798 Pair[SJ].Loops);
3799 switch (Pair[SJ].Classification) {
3800 case Subscript::ZIV:
3801 LLVM_DEBUG(dbgs() << "ZIV\n");
3802 if (testZIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
3803 return nullptr;
3804 Mivs.reset(SJ);
3805 break;
3806 case Subscript::SIV:
3807 Sivs.set(SJ);
3808 Mivs.reset(SJ);
3809 break;
3810 case Subscript::RDIV:
3811 case Subscript::MIV:
3812 break;
3813 default:
3814 llvm_unreachable("bad subscript classification");
3815 }
3816 }
3817 }
3818 }
3819 }
3820
3821 // test & propagate remaining RDIVs
3822 for (unsigned SJ : Mivs.set_bits()) {
3823 if (Pair[SJ].Classification == Subscript::RDIV) {
3824 LLVM_DEBUG(dbgs() << "RDIV test\n");
3825 if (testRDIV(Pair[SJ].Src, Pair[SJ].Dst, Result))
3826 return nullptr;
3827 // I don't yet understand how to propagate RDIV results
3828 Mivs.reset(SJ);
3829 }
3830 }
3831
3832 // test remaining MIVs
3833 // This code is temporary.
3834 // Better to somehow test all remaining subscripts simultaneously.
3835 for (unsigned SJ : Mivs.set_bits()) {
3836 if (Pair[SJ].Classification == Subscript::MIV) {
3837 LLVM_DEBUG(dbgs() << "MIV test\n");
3838 if (testMIV(Pair[SJ].Src, Pair[SJ].Dst, Pair[SJ].Loops, Result))
3839 return nullptr;
3840 }
3841 else
3842 llvm_unreachable("expected only MIV subscripts at this point");
3843 }
3844
3845 // update Result.DV from constraint vector
3846 LLVM_DEBUG(dbgs() << " updating\n");
3847 for (unsigned SJ : ConstrainedLevels.set_bits()) {
3848 if (SJ > CommonLevels)
3849 break;
3850 updateDirection(Result.DV[SJ - 1], Constraints[SJ]);
3851 if (Result.DV[SJ - 1].Direction == Dependence::DVEntry::NONE)
3852 return nullptr;
3853 }
3854 }
3855 }
3856
3857 // Make sure the Scalar flags are set correctly.
3858 SmallBitVector CompleteLoops(MaxLevels + 1);
3859 for (unsigned SI = 0; SI < Pairs; ++SI)
3860 CompleteLoops |= Pair[SI].Loops;
3861 for (unsigned II = 1; II <= CommonLevels; ++II)
3862 if (CompleteLoops[II])
3863 Result.DV[II - 1].Scalar = false;
3864
3865 if (PossiblyLoopIndependent) {
3866 // Make sure the LoopIndependent flag is set correctly.
3867 // All directions must include equal, otherwise no
3868 // loop-independent dependence is possible.
3869 for (unsigned II = 1; II <= CommonLevels; ++II) {
3870 if (!(Result.getDirection(II) & Dependence::DVEntry::EQ)) {
3871 Result.LoopIndependent = false;
3872 break;
3873 }
3874 }
3875 }
3876 else {
3877 // On the other hand, if all directions are equal and there's no
3878 // loop-independent dependence possible, then no dependence exists.
3879 bool AllEqual = true;
3880 for (unsigned II = 1; II <= CommonLevels; ++II) {
3881 if (Result.getDirection(II) != Dependence::DVEntry::EQ) {
3882 AllEqual = false;
3883 break;
3884 }
3885 }
3886 if (AllEqual)
3887 return nullptr;
3888 }
3889
3890 return std::make_unique<FullDependence>(std::move(Result));
3891 }
3892
3893 //===----------------------------------------------------------------------===//
3894 // getSplitIteration -
3895 // Rather than spend rarely-used space recording the splitting iteration
3896 // during the Weak-Crossing SIV test, we re-compute it on demand.
3897 // The re-computation is basically a repeat of the entire dependence test,
3898 // though simplified since we know that the dependence exists.
3899 // It's tedious, since we must go through all propagations, etc.
3900 //
3901 // Care is required to keep this code up to date with respect to the routine
3902 // above, depends().
3903 //
3904 // Generally, the dependence analyzer will be used to build
3905 // a dependence graph for a function (basically a map from instructions
3906 // to dependences). Looking for cycles in the graph shows us loops
3907 // that cannot be trivially vectorized/parallelized.
3908 //
3909 // We can try to improve the situation by examining all the dependences
3910 // that make up the cycle, looking for ones we can break.
3911 // Sometimes, peeling the first or last iteration of a loop will break
3912 // dependences, and we've got flags for those possibilities.
3913 // Sometimes, splitting a loop at some other iteration will do the trick,
3914 // and we've got a flag for that case. Rather than waste the space to
3915 // record the exact iteration (since we rarely know), we provide
3916 // a method that calculates the iteration. It's a drag that it must work
3917 // from scratch, but wonderful in that it's possible.
3918 //
3919 // Here's an example:
3920 //
3921 // for (i = 0; i < 10; i++)
3922 // A[i] = ...
3923 // ... = A[11 - i]
3924 //
3925 // There's a loop-carried flow dependence from the store to the load,
3926 // found by the weak-crossing SIV test. The dependence will have a flag,
3927 // indicating that the dependence can be broken by splitting the loop.
3928 // Calling getSplitIteration will return 5.
3929 // Splitting the loop breaks the dependence, like so:
3930 //
3931 // for (i = 0; i <= 5; i++)
3932 // A[i] = ...
3933 // ... = A[11 - i]
3934 // for (i = 6; i < 10; i++)
3935 // A[i] = ...
3936 // ... = A[11 - i]
3937 //
3938 // breaks the dependence and allows us to vectorize/parallelize
3939 // both loops.
getSplitIteration(const Dependence & Dep,unsigned SplitLevel)3940 const SCEV *DependenceInfo::getSplitIteration(const Dependence &Dep,
3941 unsigned SplitLevel) {
3942 assert(Dep.isSplitable(SplitLevel) &&
3943 "Dep should be splitable at SplitLevel");
3944 Instruction *Src = Dep.getSrc();
3945 Instruction *Dst = Dep.getDst();
3946 assert(Src->mayReadFromMemory() || Src->mayWriteToMemory());
3947 assert(Dst->mayReadFromMemory() || Dst->mayWriteToMemory());
3948 assert(isLoadOrStore(Src));
3949 assert(isLoadOrStore(Dst));
3950 Value *SrcPtr = getLoadStorePointerOperand(Src);
3951 Value *DstPtr = getLoadStorePointerOperand(Dst);
3952 assert(underlyingObjectsAlias(
3953 AA, F->getParent()->getDataLayout(), MemoryLocation::get(Dst),
3954 MemoryLocation::get(Src)) == AliasResult::MustAlias);
3955
3956 // establish loop nesting levels
3957 establishNestingLevels(Src, Dst);
3958
3959 FullDependence Result(Src, Dst, false, CommonLevels);
3960
3961 unsigned Pairs = 1;
3962 SmallVector<Subscript, 2> Pair(Pairs);
3963 const SCEV *SrcSCEV = SE->getSCEV(SrcPtr);
3964 const SCEV *DstSCEV = SE->getSCEV(DstPtr);
3965 Pair[0].Src = SrcSCEV;
3966 Pair[0].Dst = DstSCEV;
3967
3968 if (Delinearize) {
3969 if (tryDelinearize(Src, Dst, Pair)) {
3970 LLVM_DEBUG(dbgs() << " delinearized\n");
3971 Pairs = Pair.size();
3972 }
3973 }
3974
3975 for (unsigned P = 0; P < Pairs; ++P) {
3976 Pair[P].Loops.resize(MaxLevels + 1);
3977 Pair[P].GroupLoops.resize(MaxLevels + 1);
3978 Pair[P].Group.resize(Pairs);
3979 removeMatchingExtensions(&Pair[P]);
3980 Pair[P].Classification =
3981 classifyPair(Pair[P].Src, LI->getLoopFor(Src->getParent()),
3982 Pair[P].Dst, LI->getLoopFor(Dst->getParent()),
3983 Pair[P].Loops);
3984 Pair[P].GroupLoops = Pair[P].Loops;
3985 Pair[P].Group.set(P);
3986 }
3987
3988 SmallBitVector Separable(Pairs);
3989 SmallBitVector Coupled(Pairs);
3990
3991 // partition subscripts into separable and minimally-coupled groups
3992 for (unsigned SI = 0; SI < Pairs; ++SI) {
3993 if (Pair[SI].Classification == Subscript::NonLinear) {
3994 // ignore these, but collect loops for later
3995 collectCommonLoops(Pair[SI].Src,
3996 LI->getLoopFor(Src->getParent()),
3997 Pair[SI].Loops);
3998 collectCommonLoops(Pair[SI].Dst,
3999 LI->getLoopFor(Dst->getParent()),
4000 Pair[SI].Loops);
4001 Result.Consistent = false;
4002 }
4003 else if (Pair[SI].Classification == Subscript::ZIV)
4004 Separable.set(SI);
4005 else {
4006 // SIV, RDIV, or MIV, so check for coupled group
4007 bool Done = true;
4008 for (unsigned SJ = SI + 1; SJ < Pairs; ++SJ) {
4009 SmallBitVector Intersection = Pair[SI].GroupLoops;
4010 Intersection &= Pair[SJ].GroupLoops;
4011 if (Intersection.any()) {
4012 // accumulate set of all the loops in group
4013 Pair[SJ].GroupLoops |= Pair[SI].GroupLoops;
4014 // accumulate set of all subscripts in group
4015 Pair[SJ].Group |= Pair[SI].Group;
4016 Done = false;
4017 }
4018 }
4019 if (Done) {
4020 if (Pair[SI].Group.count() == 1)
4021 Separable.set(SI);
4022 else
4023 Coupled.set(SI);
4024 }
4025 }
4026 }
4027
4028 Constraint NewConstraint;
4029 NewConstraint.setAny(SE);
4030
4031 // test separable subscripts
4032 for (unsigned SI : Separable.set_bits()) {
4033 switch (Pair[SI].Classification) {
4034 case Subscript::SIV: {
4035 unsigned Level;
4036 const SCEV *SplitIter = nullptr;
4037 (void) testSIV(Pair[SI].Src, Pair[SI].Dst, Level,
4038 Result, NewConstraint, SplitIter);
4039 if (Level == SplitLevel) {
4040 assert(SplitIter != nullptr);
4041 return SplitIter;
4042 }
4043 break;
4044 }
4045 case Subscript::ZIV:
4046 case Subscript::RDIV:
4047 case Subscript::MIV:
4048 break;
4049 default:
4050 llvm_unreachable("subscript has unexpected classification");
4051 }
4052 }
4053
4054 if (Coupled.count()) {
4055 // test coupled subscript groups
4056 SmallVector<Constraint, 4> Constraints(MaxLevels + 1);
4057 for (unsigned II = 0; II <= MaxLevels; ++II)
4058 Constraints[II].setAny(SE);
4059 for (unsigned SI : Coupled.set_bits()) {
4060 SmallBitVector Group(Pair[SI].Group);
4061 SmallBitVector Sivs(Pairs);
4062 SmallBitVector Mivs(Pairs);
4063 SmallBitVector ConstrainedLevels(MaxLevels + 1);
4064 for (unsigned SJ : Group.set_bits()) {
4065 if (Pair[SJ].Classification == Subscript::SIV)
4066 Sivs.set(SJ);
4067 else
4068 Mivs.set(SJ);
4069 }
4070 while (Sivs.any()) {
4071 bool Changed = false;
4072 for (unsigned SJ : Sivs.set_bits()) {
4073 // SJ is an SIV subscript that's part of the current coupled group
4074 unsigned Level;
4075 const SCEV *SplitIter = nullptr;
4076 (void) testSIV(Pair[SJ].Src, Pair[SJ].Dst, Level,
4077 Result, NewConstraint, SplitIter);
4078 if (Level == SplitLevel && SplitIter)
4079 return SplitIter;
4080 ConstrainedLevels.set(Level);
4081 if (intersectConstraints(&Constraints[Level], &NewConstraint))
4082 Changed = true;
4083 Sivs.reset(SJ);
4084 }
4085 if (Changed) {
4086 // propagate, possibly creating new SIVs and ZIVs
4087 for (unsigned SJ : Mivs.set_bits()) {
4088 // SJ is an MIV subscript that's part of the current coupled group
4089 if (propagate(Pair[SJ].Src, Pair[SJ].Dst,
4090 Pair[SJ].Loops, Constraints, Result.Consistent)) {
4091 Pair[SJ].Classification =
4092 classifyPair(Pair[SJ].Src, LI->getLoopFor(Src->getParent()),
4093 Pair[SJ].Dst, LI->getLoopFor(Dst->getParent()),
4094 Pair[SJ].Loops);
4095 switch (Pair[SJ].Classification) {
4096 case Subscript::ZIV:
4097 Mivs.reset(SJ);
4098 break;
4099 case Subscript::SIV:
4100 Sivs.set(SJ);
4101 Mivs.reset(SJ);
4102 break;
4103 case Subscript::RDIV:
4104 case Subscript::MIV:
4105 break;
4106 default:
4107 llvm_unreachable("bad subscript classification");
4108 }
4109 }
4110 }
4111 }
4112 }
4113 }
4114 }
4115 llvm_unreachable("somehow reached end of routine");
4116 return nullptr;
4117 }
4118