1 //===- LoopInterchange.cpp - Loop interchange pass-------------------------===//
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 // This Pass handles loop interchange transform.
10 // This pass interchanges loops to provide a more cache-friendly memory access
11 // patterns.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/Transforms/Scalar/LoopInterchange.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/ADT/StringSet.h"
22 #include "llvm/Analysis/DependenceAnalysis.h"
23 #include "llvm/Analysis/LoopCacheAnalysis.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/LoopNestAnalysis.h"
26 #include "llvm/Analysis/LoopPass.h"
27 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
28 #include "llvm/Analysis/ScalarEvolution.h"
29 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/InstrTypes.h"
35 #include "llvm/IR/Instruction.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/User.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/Support/Casting.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Transforms/Scalar/LoopPassManager.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/LoopUtils.h"
47 #include <cassert>
48 #include <utility>
49 #include <vector>
50
51 using namespace llvm;
52
53 #define DEBUG_TYPE "loop-interchange"
54
55 STATISTIC(LoopsInterchanged, "Number of loops interchanged");
56
57 static cl::opt<int> LoopInterchangeCostThreshold(
58 "loop-interchange-threshold", cl::init(0), cl::Hidden,
59 cl::desc("Interchange if you gain more than this number"));
60
61 // Maximum number of load-stores that can be handled in the dependency matrix.
62 static cl::opt<unsigned int> MaxMemInstrCount(
63 "loop-interchange-max-meminstr-count", cl::init(64), cl::Hidden,
64 cl::desc(
65 "Maximum number of load-store instructions that should be handled "
66 "in the dependency matrix. Higher value may lead to more interchanges "
67 "at the cost of compile-time"));
68
69 namespace {
70
71 using LoopVector = SmallVector<Loop *, 8>;
72
73 // TODO: Check if we can use a sparse matrix here.
74 using CharMatrix = std::vector<std::vector<char>>;
75
76 /// Types of rules used in profitability check.
77 enum class RuleTy {
78 PerLoopCacheAnalysis,
79 PerInstrOrderCost,
80 ForVectorization,
81 };
82
83 } // end anonymous namespace
84
85 // Minimum loop depth supported.
86 static cl::opt<unsigned int> MinLoopNestDepth(
87 "loop-interchange-min-loop-nest-depth", cl::init(2), cl::Hidden,
88 cl::desc("Minimum depth of loop nest considered for the transform"));
89
90 // Maximum loop depth supported.
91 static cl::opt<unsigned int> MaxLoopNestDepth(
92 "loop-interchange-max-loop-nest-depth", cl::init(10), cl::Hidden,
93 cl::desc("Maximum depth of loop nest considered for the transform"));
94
95 // We prefer cache cost to vectorization by default.
96 static cl::list<RuleTy> Profitabilities(
97 "loop-interchange-profitabilities", cl::ZeroOrMore,
98 cl::MiscFlags::CommaSeparated, cl::Hidden,
99 cl::desc("List of profitability heuristics to be used. They are applied in "
100 "the given order"),
101 cl::list_init<RuleTy>({RuleTy::PerLoopCacheAnalysis,
102 RuleTy::PerInstrOrderCost,
103 RuleTy::ForVectorization}),
104 cl::values(clEnumValN(RuleTy::PerLoopCacheAnalysis, "cache",
105 "Prioritize loop cache cost"),
106 clEnumValN(RuleTy::PerInstrOrderCost, "instorder",
107 "Prioritize the IVs order of each instruction"),
108 clEnumValN(RuleTy::ForVectorization, "vectorize",
109 "Prioritize vectorization")));
110
111 #ifndef NDEBUG
noDuplicateRules(ArrayRef<RuleTy> Rules)112 static bool noDuplicateRules(ArrayRef<RuleTy> Rules) {
113 SmallSet<RuleTy, 4> Set;
114 for (RuleTy Rule : Rules)
115 if (!Set.insert(Rule).second)
116 return false;
117 return true;
118 }
119
printDepMatrix(CharMatrix & DepMatrix)120 static void printDepMatrix(CharMatrix &DepMatrix) {
121 for (auto &Row : DepMatrix) {
122 for (auto D : Row)
123 LLVM_DEBUG(dbgs() << D << " ");
124 LLVM_DEBUG(dbgs() << "\n");
125 }
126 }
127 #endif
128
populateDependencyMatrix(CharMatrix & DepMatrix,unsigned Level,Loop * L,DependenceInfo * DI,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)129 static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level,
130 Loop *L, DependenceInfo *DI,
131 ScalarEvolution *SE,
132 OptimizationRemarkEmitter *ORE) {
133 using ValueVector = SmallVector<Value *, 16>;
134
135 ValueVector MemInstr;
136
137 // For each block.
138 for (BasicBlock *BB : L->blocks()) {
139 // Scan the BB and collect legal loads and stores.
140 for (Instruction &I : *BB) {
141 if (!isa<Instruction>(I))
142 return false;
143 if (auto *Ld = dyn_cast<LoadInst>(&I)) {
144 if (!Ld->isSimple())
145 return false;
146 MemInstr.push_back(&I);
147 } else if (auto *St = dyn_cast<StoreInst>(&I)) {
148 if (!St->isSimple())
149 return false;
150 MemInstr.push_back(&I);
151 }
152 }
153 }
154
155 LLVM_DEBUG(dbgs() << "Found " << MemInstr.size()
156 << " Loads and Stores to analyze\n");
157 if (MemInstr.size() > MaxMemInstrCount) {
158 LLVM_DEBUG(dbgs() << "The transform doesn't support more than "
159 << MaxMemInstrCount << " load/stores in a loop\n");
160 ORE->emit([&]() {
161 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedLoop",
162 L->getStartLoc(), L->getHeader())
163 << "Number of loads/stores exceeded, the supported maximum "
164 "can be increased with option "
165 "-loop-interchange-maxmeminstr-count.";
166 });
167 return false;
168 }
169 ValueVector::iterator I, IE, J, JE;
170 StringSet<> Seen;
171
172 for (I = MemInstr.begin(), IE = MemInstr.end(); I != IE; ++I) {
173 for (J = I, JE = MemInstr.end(); J != JE; ++J) {
174 std::vector<char> Dep;
175 Instruction *Src = cast<Instruction>(*I);
176 Instruction *Dst = cast<Instruction>(*J);
177 // Ignore Input dependencies.
178 if (isa<LoadInst>(Src) && isa<LoadInst>(Dst))
179 continue;
180 // Track Output, Flow, and Anti dependencies.
181 if (auto D = DI->depends(Src, Dst)) {
182 assert(D->isOrdered() && "Expected an output, flow or anti dep.");
183 // If the direction vector is negative, normalize it to
184 // make it non-negative.
185 if (D->normalize(SE))
186 LLVM_DEBUG(dbgs() << "Negative dependence vector normalized.\n");
187 LLVM_DEBUG(StringRef DepType =
188 D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";
189 dbgs() << "Found " << DepType
190 << " dependency between Src and Dst\n"
191 << " Src:" << *Src << "\n Dst:" << *Dst << '\n');
192 unsigned Levels = D->getLevels();
193 char Direction;
194 for (unsigned II = 1; II <= Levels; ++II) {
195 // `DVEntry::LE` is converted to `*`. This is because `LE` means `<`
196 // or `=`, for which we don't have an equivalent representation, so
197 // that the conservative approximation is necessary. The same goes for
198 // `DVEntry::GE`.
199 // TODO: Use of fine-grained expressions allows for more accurate
200 // analysis.
201 unsigned Dir = D->getDirection(II);
202 if (Dir == Dependence::DVEntry::LT)
203 Direction = '<';
204 else if (Dir == Dependence::DVEntry::GT)
205 Direction = '>';
206 else if (Dir == Dependence::DVEntry::EQ)
207 Direction = '=';
208 else
209 Direction = '*';
210 Dep.push_back(Direction);
211 }
212
213 // If the Dependence object doesn't have any information, fill the
214 // dependency vector with '*'.
215 if (D->isConfused()) {
216 assert(Dep.empty() && "Expected empty dependency vector");
217 Dep.assign(Level, '*');
218 }
219
220 while (Dep.size() != Level) {
221 Dep.push_back('I');
222 }
223
224 // Make sure we only add unique entries to the dependency matrix.
225 if (Seen.insert(StringRef(Dep.data(), Dep.size())).second)
226 DepMatrix.push_back(Dep);
227 }
228 }
229 }
230
231 return true;
232 }
233
234 // A loop is moved from index 'from' to an index 'to'. Update the Dependence
235 // matrix by exchanging the two columns.
interChangeDependencies(CharMatrix & DepMatrix,unsigned FromIndx,unsigned ToIndx)236 static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx,
237 unsigned ToIndx) {
238 for (auto &Row : DepMatrix)
239 std::swap(Row[ToIndx], Row[FromIndx]);
240 }
241
242 // Check if a direction vector is lexicographically positive. Return true if it
243 // is positive, nullopt if it is "zero", otherwise false.
244 // [Theorem] A permutation of the loops in a perfect nest is legal if and only
245 // if the direction matrix, after the same permutation is applied to its
246 // columns, has no ">" direction as the leftmost non-"=" direction in any row.
247 static std::optional<bool>
isLexicographicallyPositive(ArrayRef<char> DV,unsigned Begin,unsigned End)248 isLexicographicallyPositive(ArrayRef<char> DV, unsigned Begin, unsigned End) {
249 for (unsigned char Direction : DV.slice(Begin, End - Begin)) {
250 if (Direction == '<')
251 return true;
252 if (Direction == '>' || Direction == '*')
253 return false;
254 }
255 return std::nullopt;
256 }
257
258 // Checks if it is legal to interchange 2 loops.
isLegalToInterChangeLoops(CharMatrix & DepMatrix,unsigned InnerLoopId,unsigned OuterLoopId)259 static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix,
260 unsigned InnerLoopId,
261 unsigned OuterLoopId) {
262 unsigned NumRows = DepMatrix.size();
263 std::vector<char> Cur;
264 // For each row check if it is valid to interchange.
265 for (unsigned Row = 0; Row < NumRows; ++Row) {
266 // Create temporary DepVector check its lexicographical order
267 // before and after swapping OuterLoop vs InnerLoop
268 Cur = DepMatrix[Row];
269
270 // If the surrounding loops already ensure that the direction vector is
271 // lexicographically positive, nothing within the loop will be able to break
272 // the dependence. In such a case we can skip the subsequent check.
273 if (isLexicographicallyPositive(Cur, 0, OuterLoopId) == true)
274 continue;
275
276 // Check if the direction vector is lexicographically positive (or zero)
277 // for both before/after exchanged.
278 if (isLexicographicallyPositive(Cur, OuterLoopId, Cur.size()) == false)
279 return false;
280 std::swap(Cur[InnerLoopId], Cur[OuterLoopId]);
281 if (isLexicographicallyPositive(Cur, OuterLoopId, Cur.size()) == false)
282 return false;
283 }
284 return true;
285 }
286
populateWorklist(Loop & L,LoopVector & LoopList)287 static void populateWorklist(Loop &L, LoopVector &LoopList) {
288 LLVM_DEBUG(dbgs() << "Calling populateWorklist on Func: "
289 << L.getHeader()->getParent()->getName() << " Loop: %"
290 << L.getHeader()->getName() << '\n');
291 assert(LoopList.empty() && "LoopList should initially be empty!");
292 Loop *CurrentLoop = &L;
293 const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops();
294 while (!Vec->empty()) {
295 // The current loop has multiple subloops in it hence it is not tightly
296 // nested.
297 // Discard all loops above it added into Worklist.
298 if (Vec->size() != 1) {
299 LoopList = {};
300 return;
301 }
302
303 LoopList.push_back(CurrentLoop);
304 CurrentLoop = Vec->front();
305 Vec = &CurrentLoop->getSubLoops();
306 }
307 LoopList.push_back(CurrentLoop);
308 }
309
hasSupportedLoopDepth(ArrayRef<Loop * > LoopList,OptimizationRemarkEmitter & ORE)310 static bool hasSupportedLoopDepth(ArrayRef<Loop *> LoopList,
311 OptimizationRemarkEmitter &ORE) {
312 unsigned LoopNestDepth = LoopList.size();
313 if (LoopNestDepth < MinLoopNestDepth || LoopNestDepth > MaxLoopNestDepth) {
314 LLVM_DEBUG(dbgs() << "Unsupported depth of loop nest " << LoopNestDepth
315 << ", the supported range is [" << MinLoopNestDepth
316 << ", " << MaxLoopNestDepth << "].\n");
317 Loop *OuterLoop = LoopList.front();
318 ORE.emit([&]() {
319 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedLoopNestDepth",
320 OuterLoop->getStartLoc(),
321 OuterLoop->getHeader())
322 << "Unsupported depth of loop nest, the supported range is ["
323 << std::to_string(MinLoopNestDepth) << ", "
324 << std::to_string(MaxLoopNestDepth) << "].\n";
325 });
326 return false;
327 }
328 return true;
329 }
330
isComputableLoopNest(ScalarEvolution * SE,ArrayRef<Loop * > LoopList)331 static bool isComputableLoopNest(ScalarEvolution *SE,
332 ArrayRef<Loop *> LoopList) {
333 for (Loop *L : LoopList) {
334 const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
335 if (isa<SCEVCouldNotCompute>(ExitCountOuter)) {
336 LLVM_DEBUG(dbgs() << "Couldn't compute backedge count\n");
337 return false;
338 }
339 if (L->getNumBackEdges() != 1) {
340 LLVM_DEBUG(dbgs() << "NumBackEdges is not equal to 1\n");
341 return false;
342 }
343 if (!L->getExitingBlock()) {
344 LLVM_DEBUG(dbgs() << "Loop doesn't have unique exit block\n");
345 return false;
346 }
347 }
348 return true;
349 }
350
351 namespace {
352
353 /// LoopInterchangeLegality checks if it is legal to interchange the loop.
354 class LoopInterchangeLegality {
355 public:
LoopInterchangeLegality(Loop * Outer,Loop * Inner,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)356 LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
357 OptimizationRemarkEmitter *ORE)
358 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
359
360 /// Check if the loops can be interchanged.
361 bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
362 CharMatrix &DepMatrix);
363
364 /// Discover induction PHIs in the header of \p L. Induction
365 /// PHIs are added to \p Inductions.
366 bool findInductions(Loop *L, SmallVectorImpl<PHINode *> &Inductions);
367
368 /// Check if the loop structure is understood. We do not handle triangular
369 /// loops for now.
370 bool isLoopStructureUnderstood();
371
372 bool currentLimitations();
373
getOuterInnerReductions() const374 const SmallPtrSetImpl<PHINode *> &getOuterInnerReductions() const {
375 return OuterInnerReductions;
376 }
377
getInnerLoopInductions() const378 const ArrayRef<PHINode *> getInnerLoopInductions() const {
379 return InnerLoopInductions;
380 }
381
getHasNoWrapReductions() const382 ArrayRef<Instruction *> getHasNoWrapReductions() const {
383 return HasNoWrapReductions;
384 }
385
386 private:
387 bool tightlyNested(Loop *Outer, Loop *Inner);
388 bool containsUnsafeInstructions(BasicBlock *BB);
389
390 /// Discover induction and reduction PHIs in the header of \p L. Induction
391 /// PHIs are added to \p Inductions, reductions are added to
392 /// OuterInnerReductions. When the outer loop is passed, the inner loop needs
393 /// to be passed as \p InnerLoop.
394 bool findInductionAndReductions(Loop *L,
395 SmallVector<PHINode *, 8> &Inductions,
396 Loop *InnerLoop);
397
398 Loop *OuterLoop;
399 Loop *InnerLoop;
400
401 ScalarEvolution *SE;
402
403 /// Interface to emit optimization remarks.
404 OptimizationRemarkEmitter *ORE;
405
406 /// Set of reduction PHIs taking part of a reduction across the inner and
407 /// outer loop.
408 SmallPtrSet<PHINode *, 4> OuterInnerReductions;
409
410 /// Set of inner loop induction PHIs
411 SmallVector<PHINode *, 8> InnerLoopInductions;
412
413 /// Hold instructions that have nuw/nsw flags and involved in reductions,
414 /// like integer addition/multiplication. Those flags must be dropped when
415 /// interchanging the loops.
416 SmallVector<Instruction *, 4> HasNoWrapReductions;
417 };
418
419 /// Manages information utilized by the profitability check for cache. The main
420 /// purpose of this class is to delay the computation of CacheCost until it is
421 /// actually needed.
422 class CacheCostManager {
423 Loop *OutermostLoop;
424 LoopStandardAnalysisResults *AR;
425 DependenceInfo *DI;
426
427 /// CacheCost for \ref OutermostLoop. Once it is computed, it is cached. Note
428 /// that the result can be nullptr.
429 std::optional<std::unique_ptr<CacheCost>> CC;
430
431 /// Maps each loop to an index representing the optimal position within the
432 /// loop-nest, as determined by the cache cost analysis.
433 DenseMap<const Loop *, unsigned> CostMap;
434
435 void computeIfUnitinialized();
436
437 public:
CacheCostManager(Loop * OutermostLoop,LoopStandardAnalysisResults * AR,DependenceInfo * DI)438 CacheCostManager(Loop *OutermostLoop, LoopStandardAnalysisResults *AR,
439 DependenceInfo *DI)
440 : OutermostLoop(OutermostLoop), AR(AR), DI(DI) {}
441 CacheCost *getCacheCost();
442 const DenseMap<const Loop *, unsigned> &getCostMap();
443 };
444
445 /// LoopInterchangeProfitability checks if it is profitable to interchange the
446 /// loop.
447 class LoopInterchangeProfitability {
448 public:
LoopInterchangeProfitability(Loop * Outer,Loop * Inner,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)449 LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
450 OptimizationRemarkEmitter *ORE)
451 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
452
453 /// Check if the loop interchange is profitable.
454 bool isProfitable(const Loop *InnerLoop, const Loop *OuterLoop,
455 unsigned InnerLoopId, unsigned OuterLoopId,
456 CharMatrix &DepMatrix, CacheCostManager &CCM);
457
458 private:
459 int getInstrOrderCost();
460 std::optional<bool> isProfitablePerLoopCacheAnalysis(
461 const DenseMap<const Loop *, unsigned> &CostMap, CacheCost *CC);
462 std::optional<bool> isProfitablePerInstrOrderCost();
463 std::optional<bool> isProfitableForVectorization(unsigned InnerLoopId,
464 unsigned OuterLoopId,
465 CharMatrix &DepMatrix);
466 Loop *OuterLoop;
467 Loop *InnerLoop;
468
469 /// Scev analysis.
470 ScalarEvolution *SE;
471
472 /// Interface to emit optimization remarks.
473 OptimizationRemarkEmitter *ORE;
474 };
475
476 /// LoopInterchangeTransform interchanges the loop.
477 class LoopInterchangeTransform {
478 public:
LoopInterchangeTransform(Loop * Outer,Loop * Inner,ScalarEvolution * SE,LoopInfo * LI,DominatorTree * DT,const LoopInterchangeLegality & LIL)479 LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
480 LoopInfo *LI, DominatorTree *DT,
481 const LoopInterchangeLegality &LIL)
482 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT), LIL(LIL) {}
483
484 /// Interchange OuterLoop and InnerLoop.
485 bool transform(ArrayRef<Instruction *> DropNoWrapInsts);
486 void restructureLoops(Loop *NewInner, Loop *NewOuter,
487 BasicBlock *OrigInnerPreHeader,
488 BasicBlock *OrigOuterPreHeader);
489 void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
490
491 private:
492 bool adjustLoopLinks();
493 bool adjustLoopBranches();
494
495 Loop *OuterLoop;
496 Loop *InnerLoop;
497
498 /// Scev analysis.
499 ScalarEvolution *SE;
500
501 LoopInfo *LI;
502 DominatorTree *DT;
503
504 const LoopInterchangeLegality &LIL;
505 };
506
507 struct LoopInterchange {
508 ScalarEvolution *SE = nullptr;
509 LoopInfo *LI = nullptr;
510 DependenceInfo *DI = nullptr;
511 DominatorTree *DT = nullptr;
512 LoopStandardAnalysisResults *AR = nullptr;
513
514 /// Interface to emit optimization remarks.
515 OptimizationRemarkEmitter *ORE;
516
LoopInterchange__anon815ea8750411::LoopInterchange517 LoopInterchange(ScalarEvolution *SE, LoopInfo *LI, DependenceInfo *DI,
518 DominatorTree *DT, LoopStandardAnalysisResults *AR,
519 OptimizationRemarkEmitter *ORE)
520 : SE(SE), LI(LI), DI(DI), DT(DT), AR(AR), ORE(ORE) {}
521
run__anon815ea8750411::LoopInterchange522 bool run(Loop *L) {
523 if (L->getParentLoop())
524 return false;
525 SmallVector<Loop *, 8> LoopList;
526 populateWorklist(*L, LoopList);
527 return processLoopList(LoopList);
528 }
529
run__anon815ea8750411::LoopInterchange530 bool run(LoopNest &LN) {
531 SmallVector<Loop *, 8> LoopList(LN.getLoops());
532 for (unsigned I = 1; I < LoopList.size(); ++I)
533 if (LoopList[I]->getParentLoop() != LoopList[I - 1])
534 return false;
535 return processLoopList(LoopList);
536 }
537
selectLoopForInterchange__anon815ea8750411::LoopInterchange538 unsigned selectLoopForInterchange(ArrayRef<Loop *> LoopList) {
539 // TODO: Add a better heuristic to select the loop to be interchanged based
540 // on the dependence matrix. Currently we select the innermost loop.
541 return LoopList.size() - 1;
542 }
543
processLoopList__anon815ea8750411::LoopInterchange544 bool processLoopList(SmallVectorImpl<Loop *> &LoopList) {
545 bool Changed = false;
546
547 // Ensure proper loop nest depth.
548 assert(hasSupportedLoopDepth(LoopList, *ORE) &&
549 "Unsupported depth of loop nest.");
550
551 unsigned LoopNestDepth = LoopList.size();
552
553 LLVM_DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth
554 << "\n");
555
556 CharMatrix DependencyMatrix;
557 Loop *OuterMostLoop = *(LoopList.begin());
558 if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
559 OuterMostLoop, DI, SE, ORE)) {
560 LLVM_DEBUG(dbgs() << "Populating dependency matrix failed\n");
561 return false;
562 }
563
564 LLVM_DEBUG(dbgs() << "Dependency matrix before interchange:\n";
565 printDepMatrix(DependencyMatrix));
566
567 // Get the Outermost loop exit.
568 BasicBlock *LoopNestExit = OuterMostLoop->getExitBlock();
569 if (!LoopNestExit) {
570 LLVM_DEBUG(dbgs() << "OuterMostLoop needs an unique exit block");
571 return false;
572 }
573
574 unsigned SelecLoopId = selectLoopForInterchange(LoopList);
575 CacheCostManager CCM(LoopList[0], AR, DI);
576 // We try to achieve the globally optimal memory access for the loopnest,
577 // and do interchange based on a bubble-sort fasion. We start from
578 // the innermost loop, move it outwards to the best possible position
579 // and repeat this process.
580 for (unsigned j = SelecLoopId; j > 0; j--) {
581 bool ChangedPerIter = false;
582 for (unsigned i = SelecLoopId; i > SelecLoopId - j; i--) {
583 bool Interchanged =
584 processLoop(LoopList, i, i - 1, DependencyMatrix, CCM);
585 ChangedPerIter |= Interchanged;
586 Changed |= Interchanged;
587 }
588 // Early abort if there was no interchange during an entire round of
589 // moving loops outwards.
590 if (!ChangedPerIter)
591 break;
592 }
593 return Changed;
594 }
595
processLoop__anon815ea8750411::LoopInterchange596 bool processLoop(SmallVectorImpl<Loop *> &LoopList, unsigned InnerLoopId,
597 unsigned OuterLoopId,
598 std::vector<std::vector<char>> &DependencyMatrix,
599 CacheCostManager &CCM) {
600 Loop *OuterLoop = LoopList[OuterLoopId];
601 Loop *InnerLoop = LoopList[InnerLoopId];
602 LLVM_DEBUG(dbgs() << "Processing InnerLoopId = " << InnerLoopId
603 << " and OuterLoopId = " << OuterLoopId << "\n");
604 LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, ORE);
605 if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
606 LLVM_DEBUG(dbgs() << "Not interchanging loops. Cannot prove legality.\n");
607 return false;
608 }
609 LLVM_DEBUG(dbgs() << "Loops are legal to interchange\n");
610 LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE);
611 if (!LIP.isProfitable(InnerLoop, OuterLoop, InnerLoopId, OuterLoopId,
612 DependencyMatrix, CCM)) {
613 LLVM_DEBUG(dbgs() << "Interchanging loops not profitable.\n");
614 return false;
615 }
616
617 ORE->emit([&]() {
618 return OptimizationRemark(DEBUG_TYPE, "Interchanged",
619 InnerLoop->getStartLoc(),
620 InnerLoop->getHeader())
621 << "Loop interchanged with enclosing loop.";
622 });
623
624 LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, LIL);
625 LIT.transform(LIL.getHasNoWrapReductions());
626 LLVM_DEBUG(dbgs() << "Loops interchanged.\n");
627 LoopsInterchanged++;
628
629 llvm::formLCSSARecursively(*OuterLoop, *DT, LI, SE);
630
631 // Loops interchanged, update LoopList accordingly.
632 std::swap(LoopList[OuterLoopId], LoopList[InnerLoopId]);
633 // Update the DependencyMatrix
634 interChangeDependencies(DependencyMatrix, InnerLoopId, OuterLoopId);
635
636 LLVM_DEBUG(dbgs() << "Dependency matrix after interchange:\n";
637 printDepMatrix(DependencyMatrix));
638
639 return true;
640 }
641 };
642
643 } // end anonymous namespace
644
containsUnsafeInstructions(BasicBlock * BB)645 bool LoopInterchangeLegality::containsUnsafeInstructions(BasicBlock *BB) {
646 return any_of(*BB, [](const Instruction &I) {
647 return I.mayHaveSideEffects() || I.mayReadFromMemory();
648 });
649 }
650
tightlyNested(Loop * OuterLoop,Loop * InnerLoop)651 bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
652 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
653 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
654 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
655
656 LLVM_DEBUG(dbgs() << "Checking if loops are tightly nested\n");
657
658 // A perfectly nested loop will not have any branch in between the outer and
659 // inner block i.e. outer header will branch to either inner preheader and
660 // outerloop latch.
661 BranchInst *OuterLoopHeaderBI =
662 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
663 if (!OuterLoopHeaderBI)
664 return false;
665
666 for (BasicBlock *Succ : successors(OuterLoopHeaderBI))
667 if (Succ != InnerLoopPreHeader && Succ != InnerLoop->getHeader() &&
668 Succ != OuterLoopLatch)
669 return false;
670
671 LLVM_DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n");
672 // We do not have any basic block in between now make sure the outer header
673 // and outer loop latch doesn't contain any unsafe instructions.
674 if (containsUnsafeInstructions(OuterLoopHeader) ||
675 containsUnsafeInstructions(OuterLoopLatch))
676 return false;
677
678 // Also make sure the inner loop preheader does not contain any unsafe
679 // instructions. Note that all instructions in the preheader will be moved to
680 // the outer loop header when interchanging.
681 if (InnerLoopPreHeader != OuterLoopHeader &&
682 containsUnsafeInstructions(InnerLoopPreHeader))
683 return false;
684
685 BasicBlock *InnerLoopExit = InnerLoop->getExitBlock();
686 // Ensure the inner loop exit block flows to the outer loop latch possibly
687 // through empty blocks.
688 const BasicBlock &SuccInner =
689 LoopNest::skipEmptyBlockUntil(InnerLoopExit, OuterLoopLatch);
690 if (&SuccInner != OuterLoopLatch) {
691 LLVM_DEBUG(dbgs() << "Inner loop exit block " << *InnerLoopExit
692 << " does not lead to the outer loop latch.\n";);
693 return false;
694 }
695 // The inner loop exit block does flow to the outer loop latch and not some
696 // other BBs, now make sure it contains safe instructions, since it will be
697 // moved into the (new) inner loop after interchange.
698 if (containsUnsafeInstructions(InnerLoopExit))
699 return false;
700
701 LLVM_DEBUG(dbgs() << "Loops are perfectly nested\n");
702 // We have a perfect loop nest.
703 return true;
704 }
705
isLoopStructureUnderstood()706 bool LoopInterchangeLegality::isLoopStructureUnderstood() {
707 BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
708 for (PHINode *InnerInduction : InnerLoopInductions) {
709 unsigned Num = InnerInduction->getNumOperands();
710 for (unsigned i = 0; i < Num; ++i) {
711 Value *Val = InnerInduction->getOperand(i);
712 if (isa<Constant>(Val))
713 continue;
714 Instruction *I = dyn_cast<Instruction>(Val);
715 if (!I)
716 return false;
717 // TODO: Handle triangular loops.
718 // e.g. for(int i=0;i<N;i++)
719 // for(int j=i;j<N;j++)
720 unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
721 if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
722 InnerLoopPreheader &&
723 !OuterLoop->isLoopInvariant(I)) {
724 return false;
725 }
726 }
727 }
728
729 // TODO: Handle triangular loops of another form.
730 // e.g. for(int i=0;i<N;i++)
731 // for(int j=0;j<i;j++)
732 // or,
733 // for(int i=0;i<N;i++)
734 // for(int j=0;j*i<N;j++)
735 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
736 BranchInst *InnerLoopLatchBI =
737 dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
738 if (!InnerLoopLatchBI->isConditional())
739 return false;
740 if (CmpInst *InnerLoopCmp =
741 dyn_cast<CmpInst>(InnerLoopLatchBI->getCondition())) {
742 Value *Op0 = InnerLoopCmp->getOperand(0);
743 Value *Op1 = InnerLoopCmp->getOperand(1);
744
745 // LHS and RHS of the inner loop exit condition, e.g.,
746 // in "for(int j=0;j<i;j++)", LHS is j and RHS is i.
747 Value *Left = nullptr;
748 Value *Right = nullptr;
749
750 // Check if V only involves inner loop induction variable.
751 // Return true if V is InnerInduction, or a cast from
752 // InnerInduction, or a binary operator that involves
753 // InnerInduction and a constant.
754 std::function<bool(Value *)> IsPathToInnerIndVar;
755 IsPathToInnerIndVar = [this, &IsPathToInnerIndVar](const Value *V) -> bool {
756 if (llvm::is_contained(InnerLoopInductions, V))
757 return true;
758 if (isa<Constant>(V))
759 return true;
760 const Instruction *I = dyn_cast<Instruction>(V);
761 if (!I)
762 return false;
763 if (isa<CastInst>(I))
764 return IsPathToInnerIndVar(I->getOperand(0));
765 if (isa<BinaryOperator>(I))
766 return IsPathToInnerIndVar(I->getOperand(0)) &&
767 IsPathToInnerIndVar(I->getOperand(1));
768 return false;
769 };
770
771 // In case of multiple inner loop indvars, it is okay if LHS and RHS
772 // are both inner indvar related variables.
773 if (IsPathToInnerIndVar(Op0) && IsPathToInnerIndVar(Op1))
774 return true;
775
776 // Otherwise we check if the cmp instruction compares an inner indvar
777 // related variable (Left) with a outer loop invariant (Right).
778 if (IsPathToInnerIndVar(Op0) && !isa<Constant>(Op0)) {
779 Left = Op0;
780 Right = Op1;
781 } else if (IsPathToInnerIndVar(Op1) && !isa<Constant>(Op1)) {
782 Left = Op1;
783 Right = Op0;
784 }
785
786 if (Left == nullptr)
787 return false;
788
789 const SCEV *S = SE->getSCEV(Right);
790 if (!SE->isLoopInvariant(S, OuterLoop))
791 return false;
792 }
793
794 return true;
795 }
796
797 // If SV is a LCSSA PHI node with a single incoming value, return the incoming
798 // value.
followLCSSA(Value * SV)799 static Value *followLCSSA(Value *SV) {
800 PHINode *PHI = dyn_cast<PHINode>(SV);
801 if (!PHI)
802 return SV;
803
804 if (PHI->getNumIncomingValues() != 1)
805 return SV;
806 return followLCSSA(PHI->getIncomingValue(0));
807 }
808
809 // Check V's users to see if it is involved in a reduction in L.
810 static PHINode *
findInnerReductionPhi(Loop * L,Value * V,SmallVectorImpl<Instruction * > & HasNoWrapInsts)811 findInnerReductionPhi(Loop *L, Value *V,
812 SmallVectorImpl<Instruction *> &HasNoWrapInsts) {
813 // Reduction variables cannot be constants.
814 if (isa<Constant>(V))
815 return nullptr;
816
817 for (Value *User : V->users()) {
818 if (PHINode *PHI = dyn_cast<PHINode>(User)) {
819 if (PHI->getNumIncomingValues() == 1)
820 continue;
821 RecurrenceDescriptor RD;
822 if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD)) {
823 // Detect floating point reduction only when it can be reordered.
824 if (RD.getExactFPMathInst() != nullptr)
825 return nullptr;
826
827 RecurKind RK = RD.getRecurrenceKind();
828 switch (RK) {
829 case RecurKind::Or:
830 case RecurKind::And:
831 case RecurKind::Xor:
832 case RecurKind::SMin:
833 case RecurKind::SMax:
834 case RecurKind::UMin:
835 case RecurKind::UMax:
836 case RecurKind::FAdd:
837 case RecurKind::FMul:
838 case RecurKind::FMin:
839 case RecurKind::FMax:
840 case RecurKind::FMinimum:
841 case RecurKind::FMaximum:
842 case RecurKind::FMinimumNum:
843 case RecurKind::FMaximumNum:
844 case RecurKind::FMulAdd:
845 case RecurKind::AnyOf:
846 return PHI;
847
848 // Change the order of integer addition/multiplication may change the
849 // semantics. Consider the following case:
850 //
851 // int A[2][2] = {{ INT_MAX, INT_MAX }, { INT_MIN, INT_MIN }};
852 // int sum = 0;
853 // for (int i = 0; i < 2; i++)
854 // for (int j = 0; j < 2; j++)
855 // sum += A[j][i];
856 //
857 // If the above loops are exchanged, the addition will cause an
858 // overflow. To prevent this, we must drop the nuw/nsw flags from the
859 // addition/multiplication instructions when we actually exchanges the
860 // loops.
861 case RecurKind::Add:
862 case RecurKind::Mul: {
863 unsigned OpCode = RecurrenceDescriptor::getOpcode(RK);
864 SmallVector<Instruction *, 4> Ops = RD.getReductionOpChain(PHI, L);
865
866 // Bail out when we fail to collect reduction instructions chain.
867 if (Ops.empty())
868 return nullptr;
869
870 for (Instruction *I : Ops) {
871 assert(I->getOpcode() == OpCode &&
872 "Expected the instruction to be the reduction operation");
873
874 // If the instruction has nuw/nsw flags, we must drop them when the
875 // transformation is actually performed.
876 if (I->hasNoSignedWrap() || I->hasNoUnsignedWrap())
877 HasNoWrapInsts.push_back(I);
878 }
879 return PHI;
880 }
881
882 default:
883 return nullptr;
884 }
885 }
886 return nullptr;
887 }
888 }
889
890 return nullptr;
891 }
892
findInductionAndReductions(Loop * L,SmallVector<PHINode *,8> & Inductions,Loop * InnerLoop)893 bool LoopInterchangeLegality::findInductionAndReductions(
894 Loop *L, SmallVector<PHINode *, 8> &Inductions, Loop *InnerLoop) {
895 if (!L->getLoopLatch() || !L->getLoopPredecessor())
896 return false;
897 for (PHINode &PHI : L->getHeader()->phis()) {
898 InductionDescriptor ID;
899 if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
900 Inductions.push_back(&PHI);
901 else {
902 // PHIs in inner loops need to be part of a reduction in the outer loop,
903 // discovered when checking the PHIs of the outer loop earlier.
904 if (!InnerLoop) {
905 if (!OuterInnerReductions.count(&PHI)) {
906 LLVM_DEBUG(dbgs() << "Inner loop PHI is not part of reductions "
907 "across the outer loop.\n");
908 return false;
909 }
910 } else {
911 assert(PHI.getNumIncomingValues() == 2 &&
912 "Phis in loop header should have exactly 2 incoming values");
913 // Check if we have a PHI node in the outer loop that has a reduction
914 // result from the inner loop as an incoming value.
915 Value *V = followLCSSA(PHI.getIncomingValueForBlock(L->getLoopLatch()));
916 PHINode *InnerRedPhi =
917 findInnerReductionPhi(InnerLoop, V, HasNoWrapReductions);
918 if (!InnerRedPhi ||
919 !llvm::is_contained(InnerRedPhi->incoming_values(), &PHI)) {
920 LLVM_DEBUG(
921 dbgs()
922 << "Failed to recognize PHI as an induction or reduction.\n");
923 return false;
924 }
925 OuterInnerReductions.insert(&PHI);
926 OuterInnerReductions.insert(InnerRedPhi);
927 }
928 }
929 }
930 return true;
931 }
932
933 // This function indicates the current limitations in the transform as a result
934 // of which we do not proceed.
currentLimitations()935 bool LoopInterchangeLegality::currentLimitations() {
936 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
937
938 // transform currently expects the loop latches to also be the exiting
939 // blocks.
940 if (InnerLoop->getExitingBlock() != InnerLoopLatch ||
941 OuterLoop->getExitingBlock() != OuterLoop->getLoopLatch() ||
942 !isa<BranchInst>(InnerLoopLatch->getTerminator()) ||
943 !isa<BranchInst>(OuterLoop->getLoopLatch()->getTerminator())) {
944 LLVM_DEBUG(
945 dbgs() << "Loops where the latch is not the exiting block are not"
946 << " supported currently.\n");
947 ORE->emit([&]() {
948 return OptimizationRemarkMissed(DEBUG_TYPE, "ExitingNotLatch",
949 OuterLoop->getStartLoc(),
950 OuterLoop->getHeader())
951 << "Loops where the latch is not the exiting block cannot be"
952 " interchange currently.";
953 });
954 return true;
955 }
956
957 SmallVector<PHINode *, 8> Inductions;
958 if (!findInductionAndReductions(OuterLoop, Inductions, InnerLoop)) {
959 LLVM_DEBUG(
960 dbgs() << "Only outer loops with induction or reduction PHI nodes "
961 << "are supported currently.\n");
962 ORE->emit([&]() {
963 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIOuter",
964 OuterLoop->getStartLoc(),
965 OuterLoop->getHeader())
966 << "Only outer loops with induction or reduction PHI nodes can be"
967 " interchanged currently.";
968 });
969 return true;
970 }
971
972 Inductions.clear();
973 // For multi-level loop nests, make sure that all phi nodes for inner loops
974 // at all levels can be recognized as a induction or reduction phi. Bail out
975 // if a phi node at a certain nesting level cannot be properly recognized.
976 Loop *CurLevelLoop = OuterLoop;
977 while (!CurLevelLoop->getSubLoops().empty()) {
978 // We already made sure that the loop nest is tightly nested.
979 CurLevelLoop = CurLevelLoop->getSubLoops().front();
980 if (!findInductionAndReductions(CurLevelLoop, Inductions, nullptr)) {
981 LLVM_DEBUG(
982 dbgs() << "Only inner loops with induction or reduction PHI nodes "
983 << "are supported currently.\n");
984 ORE->emit([&]() {
985 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIInner",
986 CurLevelLoop->getStartLoc(),
987 CurLevelLoop->getHeader())
988 << "Only inner loops with induction or reduction PHI nodes can be"
989 " interchange currently.";
990 });
991 return true;
992 }
993 }
994
995 // TODO: Triangular loops are not handled for now.
996 if (!isLoopStructureUnderstood()) {
997 LLVM_DEBUG(dbgs() << "Loop structure not understood by pass\n");
998 ORE->emit([&]() {
999 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedStructureInner",
1000 InnerLoop->getStartLoc(),
1001 InnerLoop->getHeader())
1002 << "Inner loop structure not understood currently.";
1003 });
1004 return true;
1005 }
1006
1007 return false;
1008 }
1009
findInductions(Loop * L,SmallVectorImpl<PHINode * > & Inductions)1010 bool LoopInterchangeLegality::findInductions(
1011 Loop *L, SmallVectorImpl<PHINode *> &Inductions) {
1012 for (PHINode &PHI : L->getHeader()->phis()) {
1013 InductionDescriptor ID;
1014 if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
1015 Inductions.push_back(&PHI);
1016 }
1017 return !Inductions.empty();
1018 }
1019
1020 // We currently only support LCSSA PHI nodes in the inner loop exit, if their
1021 // users are either reduction PHIs or PHIs outside the outer loop (which means
1022 // the we are only interested in the final value after the loop).
1023 static bool
areInnerLoopExitPHIsSupported(Loop * InnerL,Loop * OuterL,SmallPtrSetImpl<PHINode * > & Reductions)1024 areInnerLoopExitPHIsSupported(Loop *InnerL, Loop *OuterL,
1025 SmallPtrSetImpl<PHINode *> &Reductions) {
1026 BasicBlock *InnerExit = OuterL->getUniqueExitBlock();
1027 for (PHINode &PHI : InnerExit->phis()) {
1028 // Reduction lcssa phi will have only 1 incoming block that from loop latch.
1029 if (PHI.getNumIncomingValues() > 1)
1030 return false;
1031 if (any_of(PHI.users(), [&Reductions, OuterL](User *U) {
1032 PHINode *PN = dyn_cast<PHINode>(U);
1033 return !PN ||
1034 (!Reductions.count(PN) && OuterL->contains(PN->getParent()));
1035 })) {
1036 return false;
1037 }
1038 }
1039 return true;
1040 }
1041
1042 // We currently support LCSSA PHI nodes in the outer loop exit, if their
1043 // incoming values do not come from the outer loop latch or if the
1044 // outer loop latch has a single predecessor. In that case, the value will
1045 // be available if both the inner and outer loop conditions are true, which
1046 // will still be true after interchanging. If we have multiple predecessor,
1047 // that may not be the case, e.g. because the outer loop latch may be executed
1048 // if the inner loop is not executed.
areOuterLoopExitPHIsSupported(Loop * OuterLoop,Loop * InnerLoop)1049 static bool areOuterLoopExitPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
1050 BasicBlock *LoopNestExit = OuterLoop->getUniqueExitBlock();
1051 for (PHINode &PHI : LoopNestExit->phis()) {
1052 for (Value *Incoming : PHI.incoming_values()) {
1053 Instruction *IncomingI = dyn_cast<Instruction>(Incoming);
1054 if (!IncomingI || IncomingI->getParent() != OuterLoop->getLoopLatch())
1055 continue;
1056
1057 // The incoming value is defined in the outer loop latch. Currently we
1058 // only support that in case the outer loop latch has a single predecessor.
1059 // This guarantees that the outer loop latch is executed if and only if
1060 // the inner loop is executed (because tightlyNested() guarantees that the
1061 // outer loop header only branches to the inner loop or the outer loop
1062 // latch).
1063 // FIXME: We could weaken this logic and allow multiple predecessors,
1064 // if the values are produced outside the loop latch. We would need
1065 // additional logic to update the PHI nodes in the exit block as
1066 // well.
1067 if (OuterLoop->getLoopLatch()->getUniquePredecessor() == nullptr)
1068 return false;
1069 }
1070 }
1071 return true;
1072 }
1073
1074 // In case of multi-level nested loops, it may occur that lcssa phis exist in
1075 // the latch of InnerLoop, i.e., when defs of the incoming values are further
1076 // inside the loopnest. Sometimes those incoming values are not available
1077 // after interchange, since the original inner latch will become the new outer
1078 // latch which may have predecessor paths that do not include those incoming
1079 // values.
1080 // TODO: Handle transformation of lcssa phis in the InnerLoop latch in case of
1081 // multi-level loop nests.
areInnerLoopLatchPHIsSupported(Loop * OuterLoop,Loop * InnerLoop)1082 static bool areInnerLoopLatchPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
1083 if (InnerLoop->getSubLoops().empty())
1084 return true;
1085 // If the original outer latch has only one predecessor, then values defined
1086 // further inside the looploop, e.g., in the innermost loop, will be available
1087 // at the new outer latch after interchange.
1088 if (OuterLoop->getLoopLatch()->getUniquePredecessor() != nullptr)
1089 return true;
1090
1091 // The outer latch has more than one predecessors, i.e., the inner
1092 // exit and the inner header.
1093 // PHI nodes in the inner latch are lcssa phis where the incoming values
1094 // are defined further inside the loopnest. Check if those phis are used
1095 // in the original inner latch. If that is the case then bail out since
1096 // those incoming values may not be available at the new outer latch.
1097 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1098 for (PHINode &PHI : InnerLoopLatch->phis()) {
1099 for (auto *U : PHI.users()) {
1100 Instruction *UI = cast<Instruction>(U);
1101 if (InnerLoopLatch == UI->getParent())
1102 return false;
1103 }
1104 }
1105 return true;
1106 }
1107
canInterchangeLoops(unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix)1108 bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
1109 unsigned OuterLoopId,
1110 CharMatrix &DepMatrix) {
1111 if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
1112 LLVM_DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId
1113 << " and OuterLoopId = " << OuterLoopId
1114 << " due to dependence\n");
1115 ORE->emit([&]() {
1116 return OptimizationRemarkMissed(DEBUG_TYPE, "Dependence",
1117 InnerLoop->getStartLoc(),
1118 InnerLoop->getHeader())
1119 << "Cannot interchange loops due to dependences.";
1120 });
1121 return false;
1122 }
1123 // Check if outer and inner loop contain legal instructions only.
1124 for (auto *BB : OuterLoop->blocks())
1125 for (Instruction &I : BB->instructionsWithoutDebug())
1126 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1127 // readnone functions do not prevent interchanging.
1128 if (CI->onlyWritesMemory())
1129 continue;
1130 LLVM_DEBUG(
1131 dbgs() << "Loops with call instructions cannot be interchanged "
1132 << "safely.");
1133 ORE->emit([&]() {
1134 return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst",
1135 CI->getDebugLoc(),
1136 CI->getParent())
1137 << "Cannot interchange loops due to call instruction.";
1138 });
1139
1140 return false;
1141 }
1142
1143 if (!findInductions(InnerLoop, InnerLoopInductions)) {
1144 LLVM_DEBUG(dbgs() << "Could not find inner loop induction variables.\n");
1145 return false;
1146 }
1147
1148 if (!areInnerLoopLatchPHIsSupported(OuterLoop, InnerLoop)) {
1149 LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop latch.\n");
1150 ORE->emit([&]() {
1151 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedInnerLatchPHI",
1152 InnerLoop->getStartLoc(),
1153 InnerLoop->getHeader())
1154 << "Cannot interchange loops because unsupported PHI nodes found "
1155 "in inner loop latch.";
1156 });
1157 return false;
1158 }
1159
1160 // TODO: The loops could not be interchanged due to current limitations in the
1161 // transform module.
1162 if (currentLimitations()) {
1163 LLVM_DEBUG(dbgs() << "Not legal because of current transform limitation\n");
1164 return false;
1165 }
1166
1167 // Check if the loops are tightly nested.
1168 if (!tightlyNested(OuterLoop, InnerLoop)) {
1169 LLVM_DEBUG(dbgs() << "Loops not tightly nested\n");
1170 ORE->emit([&]() {
1171 return OptimizationRemarkMissed(DEBUG_TYPE, "NotTightlyNested",
1172 InnerLoop->getStartLoc(),
1173 InnerLoop->getHeader())
1174 << "Cannot interchange loops because they are not tightly "
1175 "nested.";
1176 });
1177 return false;
1178 }
1179
1180 if (!areInnerLoopExitPHIsSupported(OuterLoop, InnerLoop,
1181 OuterInnerReductions)) {
1182 LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop exit.\n");
1183 ORE->emit([&]() {
1184 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1185 InnerLoop->getStartLoc(),
1186 InnerLoop->getHeader())
1187 << "Found unsupported PHI node in loop exit.";
1188 });
1189 return false;
1190 }
1191
1192 if (!areOuterLoopExitPHIsSupported(OuterLoop, InnerLoop)) {
1193 LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in outer loop exit.\n");
1194 ORE->emit([&]() {
1195 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1196 OuterLoop->getStartLoc(),
1197 OuterLoop->getHeader())
1198 << "Found unsupported PHI node in loop exit.";
1199 });
1200 return false;
1201 }
1202
1203 return true;
1204 }
1205
computeIfUnitinialized()1206 void CacheCostManager::computeIfUnitinialized() {
1207 if (CC.has_value())
1208 return;
1209
1210 LLVM_DEBUG(dbgs() << "Compute CacheCost.\n");
1211 CC = CacheCost::getCacheCost(*OutermostLoop, *AR, *DI);
1212 // Obtain the loop vector returned from loop cache analysis beforehand,
1213 // and put each <Loop, index> pair into a map for constant time query
1214 // later. Indices in loop vector reprsent the optimal order of the
1215 // corresponding loop, e.g., given a loopnest with depth N, index 0
1216 // indicates the loop should be placed as the outermost loop and index N
1217 // indicates the loop should be placed as the innermost loop.
1218 //
1219 // For the old pass manager CacheCost would be null.
1220 if (*CC != nullptr)
1221 for (const auto &[Idx, Cost] : enumerate((*CC)->getLoopCosts()))
1222 CostMap[Cost.first] = Idx;
1223 }
1224
getCacheCost()1225 CacheCost *CacheCostManager::getCacheCost() {
1226 computeIfUnitinialized();
1227 return CC->get();
1228 }
1229
getCostMap()1230 const DenseMap<const Loop *, unsigned> &CacheCostManager::getCostMap() {
1231 computeIfUnitinialized();
1232 return CostMap;
1233 }
1234
getInstrOrderCost()1235 int LoopInterchangeProfitability::getInstrOrderCost() {
1236 unsigned GoodOrder, BadOrder;
1237 BadOrder = GoodOrder = 0;
1238 for (BasicBlock *BB : InnerLoop->blocks()) {
1239 for (Instruction &Ins : *BB) {
1240 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
1241 bool FoundInnerInduction = false;
1242 bool FoundOuterInduction = false;
1243 for (Value *Op : GEP->operands()) {
1244 // Skip operands that are not SCEV-able.
1245 if (!SE->isSCEVable(Op->getType()))
1246 continue;
1247
1248 const SCEV *OperandVal = SE->getSCEV(Op);
1249 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
1250 if (!AR)
1251 continue;
1252
1253 // If we find the inner induction after an outer induction e.g.
1254 // for(int i=0;i<N;i++)
1255 // for(int j=0;j<N;j++)
1256 // A[i][j] = A[i-1][j-1]+k;
1257 // then it is a good order.
1258 if (AR->getLoop() == InnerLoop) {
1259 // We found an InnerLoop induction after OuterLoop induction. It is
1260 // a good order.
1261 FoundInnerInduction = true;
1262 if (FoundOuterInduction) {
1263 GoodOrder++;
1264 break;
1265 }
1266 }
1267 // If we find the outer induction after an inner induction e.g.
1268 // for(int i=0;i<N;i++)
1269 // for(int j=0;j<N;j++)
1270 // A[j][i] = A[j-1][i-1]+k;
1271 // then it is a bad order.
1272 if (AR->getLoop() == OuterLoop) {
1273 // We found an OuterLoop induction after InnerLoop induction. It is
1274 // a bad order.
1275 FoundOuterInduction = true;
1276 if (FoundInnerInduction) {
1277 BadOrder++;
1278 break;
1279 }
1280 }
1281 }
1282 }
1283 }
1284 }
1285 return GoodOrder - BadOrder;
1286 }
1287
1288 std::optional<bool>
isProfitablePerLoopCacheAnalysis(const DenseMap<const Loop *,unsigned> & CostMap,CacheCost * CC)1289 LoopInterchangeProfitability::isProfitablePerLoopCacheAnalysis(
1290 const DenseMap<const Loop *, unsigned> &CostMap, CacheCost *CC) {
1291 // This is the new cost model returned from loop cache analysis.
1292 // A smaller index means the loop should be placed an outer loop, and vice
1293 // versa.
1294 auto InnerLoopIt = CostMap.find(InnerLoop);
1295 if (InnerLoopIt == CostMap.end())
1296 return std::nullopt;
1297 auto OuterLoopIt = CostMap.find(OuterLoop);
1298 if (OuterLoopIt == CostMap.end())
1299 return std::nullopt;
1300
1301 if (CC->getLoopCost(*OuterLoop) == CC->getLoopCost(*InnerLoop))
1302 return std::nullopt;
1303 unsigned InnerIndex = InnerLoopIt->second;
1304 unsigned OuterIndex = OuterLoopIt->second;
1305 LLVM_DEBUG(dbgs() << "InnerIndex = " << InnerIndex
1306 << ", OuterIndex = " << OuterIndex << "\n");
1307 assert(InnerIndex != OuterIndex && "CostMap should assign unique "
1308 "numbers to each loop");
1309 return std::optional<bool>(InnerIndex < OuterIndex);
1310 }
1311
1312 std::optional<bool>
isProfitablePerInstrOrderCost()1313 LoopInterchangeProfitability::isProfitablePerInstrOrderCost() {
1314 // Legacy cost model: this is rough cost estimation algorithm. It counts the
1315 // good and bad order of induction variables in the instruction and allows
1316 // reordering if number of bad orders is more than good.
1317 int Cost = getInstrOrderCost();
1318 LLVM_DEBUG(dbgs() << "Cost = " << Cost << "\n");
1319 if (Cost < 0 && Cost < LoopInterchangeCostThreshold)
1320 return std::optional<bool>(true);
1321
1322 return std::nullopt;
1323 }
1324
1325 /// Return true if we can vectorize the loop specified by \p LoopId.
canVectorize(const CharMatrix & DepMatrix,unsigned LoopId)1326 static bool canVectorize(const CharMatrix &DepMatrix, unsigned LoopId) {
1327 for (const auto &Dep : DepMatrix) {
1328 char Dir = Dep[LoopId];
1329 if (Dir != 'I' && Dir != '=')
1330 return false;
1331 }
1332 return true;
1333 }
1334
isProfitableForVectorization(unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix)1335 std::optional<bool> LoopInterchangeProfitability::isProfitableForVectorization(
1336 unsigned InnerLoopId, unsigned OuterLoopId, CharMatrix &DepMatrix) {
1337 // If the outer loop is not loop independent it is not profitable to move
1338 // this to inner position, since doing so would not enable inner loop
1339 // parallelism.
1340 if (!canVectorize(DepMatrix, OuterLoopId))
1341 return false;
1342
1343 // If inner loop has dependence and outer loop is loop independent then it is
1344 // profitable to interchange to enable inner loop parallelism.
1345 if (!canVectorize(DepMatrix, InnerLoopId))
1346 return true;
1347
1348 // If both the inner and the outer loop can be vectorized, it is necessary to
1349 // check the cost of each vectorized loop for profitability decision. At this
1350 // time we do not have a cost model to estimate them, so return nullopt.
1351 // TODO: Estimate the cost of vectorized loop when both the outer and the
1352 // inner loop can be vectorized.
1353 return std::nullopt;
1354 }
1355
isProfitable(const Loop * InnerLoop,const Loop * OuterLoop,unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix,CacheCostManager & CCM)1356 bool LoopInterchangeProfitability::isProfitable(
1357 const Loop *InnerLoop, const Loop *OuterLoop, unsigned InnerLoopId,
1358 unsigned OuterLoopId, CharMatrix &DepMatrix, CacheCostManager &CCM) {
1359 // isProfitable() is structured to avoid endless loop interchange. If the
1360 // highest priority rule (isProfitablePerLoopCacheAnalysis by default) could
1361 // decide the profitability then, profitability check will stop and return the
1362 // analysis result. If it failed to determine it (e.g., cache analysis failed
1363 // to analyze the loopnest due to delinearization issues) then go ahead the
1364 // second highest priority rule (isProfitablePerInstrOrderCost by default).
1365 // Likewise, if it failed to analysis the profitability then only, the last
1366 // rule (isProfitableForVectorization by default) will decide.
1367 assert(noDuplicateRules(Profitabilities) && "Detect duplicate rules");
1368 std::optional<bool> shouldInterchange;
1369 for (RuleTy RT : Profitabilities) {
1370 switch (RT) {
1371 case RuleTy::PerLoopCacheAnalysis: {
1372 CacheCost *CC = CCM.getCacheCost();
1373 const DenseMap<const Loop *, unsigned> &CostMap = CCM.getCostMap();
1374 shouldInterchange = isProfitablePerLoopCacheAnalysis(CostMap, CC);
1375 break;
1376 }
1377 case RuleTy::PerInstrOrderCost:
1378 shouldInterchange = isProfitablePerInstrOrderCost();
1379 break;
1380 case RuleTy::ForVectorization:
1381 shouldInterchange =
1382 isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix);
1383 break;
1384 }
1385
1386 // If this rule could determine the profitability, don't call subsequent
1387 // rules.
1388 if (shouldInterchange.has_value())
1389 break;
1390 }
1391
1392 if (!shouldInterchange.has_value()) {
1393 ORE->emit([&]() {
1394 return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1395 InnerLoop->getStartLoc(),
1396 InnerLoop->getHeader())
1397 << "Insufficient information to calculate the cost of loop for "
1398 "interchange.";
1399 });
1400 return false;
1401 } else if (!shouldInterchange.value()) {
1402 ORE->emit([&]() {
1403 return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1404 InnerLoop->getStartLoc(),
1405 InnerLoop->getHeader())
1406 << "Interchanging loops is not considered to improve cache "
1407 "locality nor vectorization.";
1408 });
1409 return false;
1410 }
1411 return true;
1412 }
1413
removeChildLoop(Loop * OuterLoop,Loop * InnerLoop)1414 void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
1415 Loop *InnerLoop) {
1416 for (Loop *L : *OuterLoop)
1417 if (L == InnerLoop) {
1418 OuterLoop->removeChildLoop(L);
1419 return;
1420 }
1421 llvm_unreachable("Couldn't find loop");
1422 }
1423
1424 /// Update LoopInfo, after interchanging. NewInner and NewOuter refer to the
1425 /// new inner and outer loop after interchanging: NewInner is the original
1426 /// outer loop and NewOuter is the original inner loop.
1427 ///
1428 /// Before interchanging, we have the following structure
1429 /// Outer preheader
1430 // Outer header
1431 // Inner preheader
1432 // Inner header
1433 // Inner body
1434 // Inner latch
1435 // outer bbs
1436 // Outer latch
1437 //
1438 // After interchanging:
1439 // Inner preheader
1440 // Inner header
1441 // Outer preheader
1442 // Outer header
1443 // Inner body
1444 // outer bbs
1445 // Outer latch
1446 // Inner latch
restructureLoops(Loop * NewInner,Loop * NewOuter,BasicBlock * OrigInnerPreHeader,BasicBlock * OrigOuterPreHeader)1447 void LoopInterchangeTransform::restructureLoops(
1448 Loop *NewInner, Loop *NewOuter, BasicBlock *OrigInnerPreHeader,
1449 BasicBlock *OrigOuterPreHeader) {
1450 Loop *OuterLoopParent = OuterLoop->getParentLoop();
1451 // The original inner loop preheader moves from the new inner loop to
1452 // the parent loop, if there is one.
1453 NewInner->removeBlockFromLoop(OrigInnerPreHeader);
1454 LI->changeLoopFor(OrigInnerPreHeader, OuterLoopParent);
1455
1456 // Switch the loop levels.
1457 if (OuterLoopParent) {
1458 // Remove the loop from its parent loop.
1459 removeChildLoop(OuterLoopParent, NewInner);
1460 removeChildLoop(NewInner, NewOuter);
1461 OuterLoopParent->addChildLoop(NewOuter);
1462 } else {
1463 removeChildLoop(NewInner, NewOuter);
1464 LI->changeTopLevelLoop(NewInner, NewOuter);
1465 }
1466 while (!NewOuter->isInnermost())
1467 NewInner->addChildLoop(NewOuter->removeChildLoop(NewOuter->begin()));
1468 NewOuter->addChildLoop(NewInner);
1469
1470 // BBs from the original inner loop.
1471 SmallVector<BasicBlock *, 8> OrigInnerBBs(NewOuter->blocks());
1472
1473 // Add BBs from the original outer loop to the original inner loop (excluding
1474 // BBs already in inner loop)
1475 for (BasicBlock *BB : NewInner->blocks())
1476 if (LI->getLoopFor(BB) == NewInner)
1477 NewOuter->addBlockEntry(BB);
1478
1479 // Now remove inner loop header and latch from the new inner loop and move
1480 // other BBs (the loop body) to the new inner loop.
1481 BasicBlock *OuterHeader = NewOuter->getHeader();
1482 BasicBlock *OuterLatch = NewOuter->getLoopLatch();
1483 for (BasicBlock *BB : OrigInnerBBs) {
1484 // Nothing will change for BBs in child loops.
1485 if (LI->getLoopFor(BB) != NewOuter)
1486 continue;
1487 // Remove the new outer loop header and latch from the new inner loop.
1488 if (BB == OuterHeader || BB == OuterLatch)
1489 NewInner->removeBlockFromLoop(BB);
1490 else
1491 LI->changeLoopFor(BB, NewInner);
1492 }
1493
1494 // The preheader of the original outer loop becomes part of the new
1495 // outer loop.
1496 NewOuter->addBlockEntry(OrigOuterPreHeader);
1497 LI->changeLoopFor(OrigOuterPreHeader, NewOuter);
1498
1499 // Tell SE that we move the loops around.
1500 SE->forgetLoop(NewOuter);
1501 }
1502
transform(ArrayRef<Instruction * > DropNoWrapInsts)1503 bool LoopInterchangeTransform::transform(
1504 ArrayRef<Instruction *> DropNoWrapInsts) {
1505 bool Transformed = false;
1506
1507 if (InnerLoop->getSubLoops().empty()) {
1508 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1509 LLVM_DEBUG(dbgs() << "Splitting the inner loop latch\n");
1510 auto &InductionPHIs = LIL.getInnerLoopInductions();
1511 if (InductionPHIs.empty()) {
1512 LLVM_DEBUG(dbgs() << "Failed to find the point to split loop latch \n");
1513 return false;
1514 }
1515
1516 SmallVector<Instruction *, 8> InnerIndexVarList;
1517 for (PHINode *CurInductionPHI : InductionPHIs) {
1518 if (CurInductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1519 InnerIndexVarList.push_back(
1520 dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(1)));
1521 else
1522 InnerIndexVarList.push_back(
1523 dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(0)));
1524 }
1525
1526 // Create a new latch block for the inner loop. We split at the
1527 // current latch's terminator and then move the condition and all
1528 // operands that are not either loop-invariant or the induction PHI into the
1529 // new latch block.
1530 BasicBlock *NewLatch =
1531 SplitBlock(InnerLoop->getLoopLatch(),
1532 InnerLoop->getLoopLatch()->getTerminator(), DT, LI);
1533
1534 SmallSetVector<Instruction *, 4> WorkList;
1535 unsigned i = 0;
1536 auto MoveInstructions = [&i, &WorkList, this, &InductionPHIs, NewLatch]() {
1537 for (; i < WorkList.size(); i++) {
1538 // Duplicate instruction and move it the new latch. Update uses that
1539 // have been moved.
1540 Instruction *NewI = WorkList[i]->clone();
1541 NewI->insertBefore(NewLatch->getFirstNonPHIIt());
1542 assert(!NewI->mayHaveSideEffects() &&
1543 "Moving instructions with side-effects may change behavior of "
1544 "the loop nest!");
1545 for (Use &U : llvm::make_early_inc_range(WorkList[i]->uses())) {
1546 Instruction *UserI = cast<Instruction>(U.getUser());
1547 if (!InnerLoop->contains(UserI->getParent()) ||
1548 UserI->getParent() == NewLatch ||
1549 llvm::is_contained(InductionPHIs, UserI))
1550 U.set(NewI);
1551 }
1552 // Add operands of moved instruction to the worklist, except if they are
1553 // outside the inner loop or are the induction PHI.
1554 for (Value *Op : WorkList[i]->operands()) {
1555 Instruction *OpI = dyn_cast<Instruction>(Op);
1556 if (!OpI ||
1557 this->LI->getLoopFor(OpI->getParent()) != this->InnerLoop ||
1558 llvm::is_contained(InductionPHIs, OpI))
1559 continue;
1560 WorkList.insert(OpI);
1561 }
1562 }
1563 };
1564
1565 // FIXME: Should we interchange when we have a constant condition?
1566 Instruction *CondI = dyn_cast<Instruction>(
1567 cast<BranchInst>(InnerLoop->getLoopLatch()->getTerminator())
1568 ->getCondition());
1569 if (CondI)
1570 WorkList.insert(CondI);
1571 MoveInstructions();
1572 for (Instruction *InnerIndexVar : InnerIndexVarList)
1573 WorkList.insert(cast<Instruction>(InnerIndexVar));
1574 MoveInstructions();
1575 }
1576
1577 // Ensure the inner loop phi nodes have a separate basic block.
1578 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1579 if (&*InnerLoopHeader->getFirstNonPHIIt() !=
1580 InnerLoopHeader->getTerminator()) {
1581 SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHIIt(), DT, LI);
1582 LLVM_DEBUG(dbgs() << "splitting InnerLoopHeader done\n");
1583 }
1584
1585 // Instructions in the original inner loop preheader may depend on values
1586 // defined in the outer loop header. Move them there, because the original
1587 // inner loop preheader will become the entry into the interchanged loop nest.
1588 // Currently we move all instructions and rely on LICM to move invariant
1589 // instructions outside the loop nest.
1590 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1591 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1592 if (InnerLoopPreHeader != OuterLoopHeader) {
1593 for (Instruction &I :
1594 make_early_inc_range(make_range(InnerLoopPreHeader->begin(),
1595 std::prev(InnerLoopPreHeader->end()))))
1596 I.moveBeforePreserving(OuterLoopHeader->getTerminator()->getIterator());
1597 }
1598
1599 Transformed |= adjustLoopLinks();
1600 if (!Transformed) {
1601 LLVM_DEBUG(dbgs() << "adjustLoopLinks failed\n");
1602 return false;
1603 }
1604
1605 // Finally, drop the nsw/nuw flags from the instructions for reduction
1606 // calculations.
1607 for (Instruction *Reduction : DropNoWrapInsts) {
1608 Reduction->setHasNoSignedWrap(false);
1609 Reduction->setHasNoUnsignedWrap(false);
1610 }
1611
1612 return true;
1613 }
1614
1615 /// \brief Move all instructions except the terminator from FromBB right before
1616 /// InsertBefore
moveBBContents(BasicBlock * FromBB,Instruction * InsertBefore)1617 static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1618 BasicBlock *ToBB = InsertBefore->getParent();
1619
1620 ToBB->splice(InsertBefore->getIterator(), FromBB, FromBB->begin(),
1621 FromBB->getTerminator()->getIterator());
1622 }
1623
1624 /// Swap instructions between \p BB1 and \p BB2 but keep terminators intact.
swapBBContents(BasicBlock * BB1,BasicBlock * BB2)1625 static void swapBBContents(BasicBlock *BB1, BasicBlock *BB2) {
1626 // Save all non-terminator instructions of BB1 into TempInstrs and unlink them
1627 // from BB1 afterwards.
1628 auto Iter = map_range(*BB1, [](Instruction &I) { return &I; });
1629 SmallVector<Instruction *, 4> TempInstrs(Iter.begin(), std::prev(Iter.end()));
1630 for (Instruction *I : TempInstrs)
1631 I->removeFromParent();
1632
1633 // Move instructions from BB2 to BB1.
1634 moveBBContents(BB2, BB1->getTerminator());
1635
1636 // Move instructions from TempInstrs to BB2.
1637 for (Instruction *I : TempInstrs)
1638 I->insertBefore(BB2->getTerminator()->getIterator());
1639 }
1640
1641 // Update BI to jump to NewBB instead of OldBB. Records updates to the
1642 // dominator tree in DTUpdates. If \p MustUpdateOnce is true, assert that
1643 // \p OldBB is exactly once in BI's successor list.
updateSuccessor(BranchInst * BI,BasicBlock * OldBB,BasicBlock * NewBB,std::vector<DominatorTree::UpdateType> & DTUpdates,bool MustUpdateOnce=true)1644 static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB,
1645 BasicBlock *NewBB,
1646 std::vector<DominatorTree::UpdateType> &DTUpdates,
1647 bool MustUpdateOnce = true) {
1648 assert((!MustUpdateOnce || llvm::count(successors(BI), OldBB) == 1) &&
1649 "BI must jump to OldBB exactly once.");
1650 bool Changed = false;
1651 for (Use &Op : BI->operands())
1652 if (Op == OldBB) {
1653 Op.set(NewBB);
1654 Changed = true;
1655 }
1656
1657 if (Changed) {
1658 DTUpdates.push_back(
1659 {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB});
1660 DTUpdates.push_back(
1661 {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB});
1662 }
1663 assert(Changed && "Expected a successor to be updated");
1664 }
1665
1666 // Move Lcssa PHIs to the right place.
moveLCSSAPhis(BasicBlock * InnerExit,BasicBlock * InnerHeader,BasicBlock * InnerLatch,BasicBlock * OuterHeader,BasicBlock * OuterLatch,BasicBlock * OuterExit,Loop * InnerLoop,LoopInfo * LI)1667 static void moveLCSSAPhis(BasicBlock *InnerExit, BasicBlock *InnerHeader,
1668 BasicBlock *InnerLatch, BasicBlock *OuterHeader,
1669 BasicBlock *OuterLatch, BasicBlock *OuterExit,
1670 Loop *InnerLoop, LoopInfo *LI) {
1671
1672 // Deal with LCSSA PHI nodes in the exit block of the inner loop, that are
1673 // defined either in the header or latch. Those blocks will become header and
1674 // latch of the new outer loop, and the only possible users can PHI nodes
1675 // in the exit block of the loop nest or the outer loop header (reduction
1676 // PHIs, in that case, the incoming value must be defined in the inner loop
1677 // header). We can just substitute the user with the incoming value and remove
1678 // the PHI.
1679 for (PHINode &P : make_early_inc_range(InnerExit->phis())) {
1680 assert(P.getNumIncomingValues() == 1 &&
1681 "Only loops with a single exit are supported!");
1682
1683 // Incoming values are guaranteed be instructions currently.
1684 auto IncI = cast<Instruction>(P.getIncomingValueForBlock(InnerLatch));
1685 // In case of multi-level nested loops, follow LCSSA to find the incoming
1686 // value defined from the innermost loop.
1687 auto IncIInnerMost = cast<Instruction>(followLCSSA(IncI));
1688 // Skip phis with incoming values from the inner loop body, excluding the
1689 // header and latch.
1690 if (IncIInnerMost->getParent() != InnerLatch &&
1691 IncIInnerMost->getParent() != InnerHeader)
1692 continue;
1693
1694 assert(all_of(P.users(),
1695 [OuterHeader, OuterExit, IncI, InnerHeader](User *U) {
1696 return (cast<PHINode>(U)->getParent() == OuterHeader &&
1697 IncI->getParent() == InnerHeader) ||
1698 cast<PHINode>(U)->getParent() == OuterExit;
1699 }) &&
1700 "Can only replace phis iff the uses are in the loop nest exit or "
1701 "the incoming value is defined in the inner header (it will "
1702 "dominate all loop blocks after interchanging)");
1703 P.replaceAllUsesWith(IncI);
1704 P.eraseFromParent();
1705 }
1706
1707 SmallVector<PHINode *, 8> LcssaInnerExit(
1708 llvm::make_pointer_range(InnerExit->phis()));
1709
1710 SmallVector<PHINode *, 8> LcssaInnerLatch(
1711 llvm::make_pointer_range(InnerLatch->phis()));
1712
1713 // Lcssa PHIs for values used outside the inner loop are in InnerExit.
1714 // If a PHI node has users outside of InnerExit, it has a use outside the
1715 // interchanged loop and we have to preserve it. We move these to
1716 // InnerLatch, which will become the new exit block for the innermost
1717 // loop after interchanging.
1718 for (PHINode *P : LcssaInnerExit)
1719 P->moveBefore(InnerLatch->getFirstNonPHIIt());
1720
1721 // If the inner loop latch contains LCSSA PHIs, those come from a child loop
1722 // and we have to move them to the new inner latch.
1723 for (PHINode *P : LcssaInnerLatch)
1724 P->moveBefore(InnerExit->getFirstNonPHIIt());
1725
1726 // Deal with LCSSA PHI nodes in the loop nest exit block. For PHIs that have
1727 // incoming values defined in the outer loop, we have to add a new PHI
1728 // in the inner loop latch, which became the exit block of the outer loop,
1729 // after interchanging.
1730 if (OuterExit) {
1731 for (PHINode &P : OuterExit->phis()) {
1732 if (P.getNumIncomingValues() != 1)
1733 continue;
1734 // Skip Phis with incoming values defined in the inner loop. Those should
1735 // already have been updated.
1736 auto I = dyn_cast<Instruction>(P.getIncomingValue(0));
1737 if (!I || LI->getLoopFor(I->getParent()) == InnerLoop)
1738 continue;
1739
1740 PHINode *NewPhi = dyn_cast<PHINode>(P.clone());
1741 NewPhi->setIncomingValue(0, P.getIncomingValue(0));
1742 NewPhi->setIncomingBlock(0, OuterLatch);
1743 // We might have incoming edges from other BBs, i.e., the original outer
1744 // header.
1745 for (auto *Pred : predecessors(InnerLatch)) {
1746 if (Pred == OuterLatch)
1747 continue;
1748 NewPhi->addIncoming(P.getIncomingValue(0), Pred);
1749 }
1750 NewPhi->insertBefore(InnerLatch->getFirstNonPHIIt());
1751 P.setIncomingValue(0, NewPhi);
1752 }
1753 }
1754
1755 // Now adjust the incoming blocks for the LCSSA PHIs.
1756 // For PHIs moved from Inner's exit block, we need to replace Inner's latch
1757 // with the new latch.
1758 InnerLatch->replacePhiUsesWith(InnerLatch, OuterLatch);
1759 }
1760
adjustLoopBranches()1761 bool LoopInterchangeTransform::adjustLoopBranches() {
1762 LLVM_DEBUG(dbgs() << "adjustLoopBranches called\n");
1763 std::vector<DominatorTree::UpdateType> DTUpdates;
1764
1765 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1766 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1767
1768 assert(OuterLoopPreHeader != OuterLoop->getHeader() &&
1769 InnerLoopPreHeader != InnerLoop->getHeader() && OuterLoopPreHeader &&
1770 InnerLoopPreHeader && "Guaranteed by loop-simplify form");
1771 // Ensure that both preheaders do not contain PHI nodes and have single
1772 // predecessors. This allows us to move them easily. We use
1773 // InsertPreHeaderForLoop to create an 'extra' preheader, if the existing
1774 // preheaders do not satisfy those conditions.
1775 if (isa<PHINode>(OuterLoopPreHeader->begin()) ||
1776 !OuterLoopPreHeader->getUniquePredecessor())
1777 OuterLoopPreHeader =
1778 InsertPreheaderForLoop(OuterLoop, DT, LI, nullptr, true);
1779 if (InnerLoopPreHeader == OuterLoop->getHeader())
1780 InnerLoopPreHeader =
1781 InsertPreheaderForLoop(InnerLoop, DT, LI, nullptr, true);
1782
1783 // Adjust the loop preheader
1784 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1785 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1786 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1787 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1788 BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1789 BasicBlock *InnerLoopLatchPredecessor =
1790 InnerLoopLatch->getUniquePredecessor();
1791 BasicBlock *InnerLoopLatchSuccessor;
1792 BasicBlock *OuterLoopLatchSuccessor;
1793
1794 BranchInst *OuterLoopLatchBI =
1795 dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1796 BranchInst *InnerLoopLatchBI =
1797 dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1798 BranchInst *OuterLoopHeaderBI =
1799 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1800 BranchInst *InnerLoopHeaderBI =
1801 dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1802
1803 if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1804 !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1805 !InnerLoopHeaderBI)
1806 return false;
1807
1808 BranchInst *InnerLoopLatchPredecessorBI =
1809 dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1810 BranchInst *OuterLoopPredecessorBI =
1811 dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1812
1813 if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1814 return false;
1815 BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1816 if (!InnerLoopHeaderSuccessor)
1817 return false;
1818
1819 // Adjust Loop Preheader and headers.
1820 // The branches in the outer loop predecessor and the outer loop header can
1821 // be unconditional branches or conditional branches with duplicates. Consider
1822 // this when updating the successors.
1823 updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader,
1824 InnerLoopPreHeader, DTUpdates, /*MustUpdateOnce=*/false);
1825 // The outer loop header might or might not branch to the outer latch.
1826 // We are guaranteed to branch to the inner loop preheader.
1827 if (llvm::is_contained(OuterLoopHeaderBI->successors(), OuterLoopLatch)) {
1828 // In this case the outerLoopHeader should branch to the InnerLoopLatch.
1829 updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, InnerLoopLatch,
1830 DTUpdates,
1831 /*MustUpdateOnce=*/false);
1832 }
1833 updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader,
1834 InnerLoopHeaderSuccessor, DTUpdates,
1835 /*MustUpdateOnce=*/false);
1836
1837 // Adjust reduction PHI's now that the incoming block has changed.
1838 InnerLoopHeaderSuccessor->replacePhiUsesWith(InnerLoopHeader,
1839 OuterLoopHeader);
1840
1841 updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor,
1842 OuterLoopPreHeader, DTUpdates);
1843
1844 // -------------Adjust loop latches-----------
1845 if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1846 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1847 else
1848 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1849
1850 updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch,
1851 InnerLoopLatchSuccessor, DTUpdates);
1852
1853 if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1854 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1855 else
1856 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1857
1858 updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor,
1859 OuterLoopLatchSuccessor, DTUpdates);
1860 updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch,
1861 DTUpdates);
1862
1863 DT->applyUpdates(DTUpdates);
1864 restructureLoops(OuterLoop, InnerLoop, InnerLoopPreHeader,
1865 OuterLoopPreHeader);
1866
1867 moveLCSSAPhis(InnerLoopLatchSuccessor, InnerLoopHeader, InnerLoopLatch,
1868 OuterLoopHeader, OuterLoopLatch, InnerLoop->getExitBlock(),
1869 InnerLoop, LI);
1870 // For PHIs in the exit block of the outer loop, outer's latch has been
1871 // replaced by Inners'.
1872 OuterLoopLatchSuccessor->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1873
1874 auto &OuterInnerReductions = LIL.getOuterInnerReductions();
1875 // Now update the reduction PHIs in the inner and outer loop headers.
1876 SmallVector<PHINode *, 4> InnerLoopPHIs, OuterLoopPHIs;
1877 for (PHINode &PHI : InnerLoopHeader->phis())
1878 if (OuterInnerReductions.contains(&PHI))
1879 InnerLoopPHIs.push_back(&PHI);
1880
1881 for (PHINode &PHI : OuterLoopHeader->phis())
1882 if (OuterInnerReductions.contains(&PHI))
1883 OuterLoopPHIs.push_back(&PHI);
1884
1885 // Now move the remaining reduction PHIs from outer to inner loop header and
1886 // vice versa. The PHI nodes must be part of a reduction across the inner and
1887 // outer loop and all the remains to do is and updating the incoming blocks.
1888 for (PHINode *PHI : OuterLoopPHIs) {
1889 LLVM_DEBUG(dbgs() << "Outer loop reduction PHIs:\n"; PHI->dump(););
1890 PHI->moveBefore(InnerLoopHeader->getFirstNonPHIIt());
1891 assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1892 }
1893 for (PHINode *PHI : InnerLoopPHIs) {
1894 LLVM_DEBUG(dbgs() << "Inner loop reduction PHIs:\n"; PHI->dump(););
1895 PHI->moveBefore(OuterLoopHeader->getFirstNonPHIIt());
1896 assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1897 }
1898
1899 // Update the incoming blocks for moved PHI nodes.
1900 OuterLoopHeader->replacePhiUsesWith(InnerLoopPreHeader, OuterLoopPreHeader);
1901 OuterLoopHeader->replacePhiUsesWith(InnerLoopLatch, OuterLoopLatch);
1902 InnerLoopHeader->replacePhiUsesWith(OuterLoopPreHeader, InnerLoopPreHeader);
1903 InnerLoopHeader->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1904
1905 // Values defined in the outer loop header could be used in the inner loop
1906 // latch. In that case, we need to create LCSSA phis for them, because after
1907 // interchanging they will be defined in the new inner loop and used in the
1908 // new outer loop.
1909 SmallVector<Instruction *, 4> MayNeedLCSSAPhis;
1910 for (Instruction &I :
1911 make_range(OuterLoopHeader->begin(), std::prev(OuterLoopHeader->end())))
1912 MayNeedLCSSAPhis.push_back(&I);
1913 formLCSSAForInstructions(MayNeedLCSSAPhis, *DT, *LI, SE);
1914
1915 return true;
1916 }
1917
adjustLoopLinks()1918 bool LoopInterchangeTransform::adjustLoopLinks() {
1919 // Adjust all branches in the inner and outer loop.
1920 bool Changed = adjustLoopBranches();
1921 if (Changed) {
1922 // We have interchanged the preheaders so we need to interchange the data in
1923 // the preheaders as well. This is because the content of the inner
1924 // preheader was previously executed inside the outer loop.
1925 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1926 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1927 swapBBContents(OuterLoopPreHeader, InnerLoopPreHeader);
1928 }
1929 return Changed;
1930 }
1931
run(LoopNest & LN,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)1932 PreservedAnalyses LoopInterchangePass::run(LoopNest &LN,
1933 LoopAnalysisManager &AM,
1934 LoopStandardAnalysisResults &AR,
1935 LPMUpdater &U) {
1936 Function &F = *LN.getParent();
1937 SmallVector<Loop *, 8> LoopList(LN.getLoops());
1938
1939 if (MaxMemInstrCount < 1) {
1940 LLVM_DEBUG(dbgs() << "MaxMemInstrCount should be at least 1");
1941 return PreservedAnalyses::all();
1942 }
1943 OptimizationRemarkEmitter ORE(&F);
1944
1945 // Ensure minimum depth of the loop nest to do the interchange.
1946 if (!hasSupportedLoopDepth(LoopList, ORE))
1947 return PreservedAnalyses::all();
1948 // Ensure computable loop nest.
1949 if (!isComputableLoopNest(&AR.SE, LoopList)) {
1950 LLVM_DEBUG(dbgs() << "Not valid loop candidate for interchange\n");
1951 return PreservedAnalyses::all();
1952 }
1953
1954 ORE.emit([&]() {
1955 return OptimizationRemarkAnalysis(DEBUG_TYPE, "Dependence",
1956 LN.getOutermostLoop().getStartLoc(),
1957 LN.getOutermostLoop().getHeader())
1958 << "Computed dependence info, invoking the transform.";
1959 });
1960
1961 DependenceInfo DI(&F, &AR.AA, &AR.SE, &AR.LI);
1962 if (!LoopInterchange(&AR.SE, &AR.LI, &DI, &AR.DT, &AR, &ORE).run(LN))
1963 return PreservedAnalyses::all();
1964 U.markLoopNestChanged(true);
1965 return getLoopPassPreservedAnalyses();
1966 }
1967