Bug Summary

File:lib/Transforms/Scalar/LoopInterchange.cpp
Warning:line 875, column 5
Value stored to 'InnerLoopPreHeader' is never read

Annotated Source Code

1//===- LoopInterchange.cpp - Loop interchange pass------------------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This Pass handles loop interchange transform.
11// This pass interchanges loops to provide a more cache-friendly memory access
12// patterns.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/ADT/SmallVector.h"
17#include "llvm/Analysis/AliasAnalysis.h"
18#include "llvm/Analysis/AssumptionCache.h"
19#include "llvm/Analysis/BlockFrequencyInfo.h"
20#include "llvm/Analysis/CodeMetrics.h"
21#include "llvm/Analysis/DependenceAnalysis.h"
22#include "llvm/Analysis/LoopInfo.h"
23#include "llvm/Analysis/LoopIterator.h"
24#include "llvm/Analysis/LoopPass.h"
25#include "llvm/Analysis/ScalarEvolution.h"
26#include "llvm/Analysis/ScalarEvolutionExpander.h"
27#include "llvm/Analysis/ScalarEvolutionExpressions.h"
28#include "llvm/Analysis/TargetTransformInfo.h"
29#include "llvm/Analysis/ValueTracking.h"
30#include "llvm/IR/Dominators.h"
31#include "llvm/IR/Function.h"
32#include "llvm/IR/IRBuilder.h"
33#include "llvm/IR/InstIterator.h"
34#include "llvm/IR/IntrinsicInst.h"
35#include "llvm/IR/Module.h"
36#include "llvm/Pass.h"
37#include "llvm/Support/Debug.h"
38#include "llvm/Support/raw_ostream.h"
39#include "llvm/Transforms/Scalar.h"
40#include "llvm/Transforms/Utils/BasicBlockUtils.h"
41#include "llvm/Transforms/Utils/LoopUtils.h"
42
43using namespace llvm;
44
45#define DEBUG_TYPE"loop-interchange" "loop-interchange"
46
47static cl::opt<int> LoopInterchangeCostThreshold(
48 "loop-interchange-threshold", cl::init(0), cl::Hidden,
49 cl::desc("Interchange if you gain more than this number"));
50
51namespace {
52
53typedef SmallVector<Loop *, 8> LoopVector;
54
55// TODO: Check if we can use a sparse matrix here.
56typedef std::vector<std::vector<char>> CharMatrix;
57
58// Maximum number of dependencies that can be handled in the dependency matrix.
59static const unsigned MaxMemInstrCount = 100;
60
61// Maximum loop depth supported.
62static const unsigned MaxLoopNestDepth = 10;
63
64struct LoopInterchange;
65
66#ifdef DUMP_DEP_MATRICIES
67void printDepMatrix(CharMatrix &DepMatrix) {
68 for (auto I = DepMatrix.begin(), E = DepMatrix.end(); I != E; ++I) {
69 std::vector<char> Vec = *I;
70 for (auto II = Vec.begin(), EE = Vec.end(); II != EE; ++II)
71 DEBUG(dbgs() << *II << " ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << *II << " "; } }
while (false)
;
72 DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "\n"; } } while (false
)
;
73 }
74}
75#endif
76
77static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level,
78 Loop *L, DependenceInfo *DI) {
79 typedef SmallVector<Value *, 16> ValueVector;
80 ValueVector MemInstr;
81
82 // For each block.
83 for (Loop::block_iterator BB = L->block_begin(), BE = L->block_end();
84 BB != BE; ++BB) {
85 // Scan the BB and collect legal loads and stores.
86 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E;
87 ++I) {
88 if (!isa<Instruction>(I))
89 return false;
90 if (LoadInst *Ld = dyn_cast<LoadInst>(I)) {
91 if (!Ld->isSimple())
92 return false;
93 MemInstr.push_back(&*I);
94 } else if (StoreInst *St = dyn_cast<StoreInst>(I)) {
95 if (!St->isSimple())
96 return false;
97 MemInstr.push_back(&*I);
98 }
99 }
100 }
101
102 DEBUG(dbgs() << "Found " << MemInstr.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Found " << MemInstr
.size() << " Loads and Stores to analyze\n"; } } while (
false)
103 << " Loads and Stores to analyze\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Found " << MemInstr
.size() << " Loads and Stores to analyze\n"; } } while (
false)
;
104
105 ValueVector::iterator I, IE, J, JE;
106
107 for (I = MemInstr.begin(), IE = MemInstr.end(); I != IE; ++I) {
108 for (J = I, JE = MemInstr.end(); J != JE; ++J) {
109 std::vector<char> Dep;
110 Instruction *Src = cast<Instruction>(*I);
111 Instruction *Dst = cast<Instruction>(*J);
112 if (Src == Dst)
113 continue;
114 // Ignore Input dependencies.
115 if (isa<LoadInst>(Src) && isa<LoadInst>(Dst))
116 continue;
117 // Track Output, Flow, and Anti dependencies.
118 if (auto D = DI->depends(Src, Dst, true)) {
119 assert(D->isOrdered() && "Expected an output, flow or anti dep.")((D->isOrdered() && "Expected an output, flow or anti dep."
) ? static_cast<void> (0) : __assert_fail ("D->isOrdered() && \"Expected an output, flow or anti dep.\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301389/lib/Transforms/Scalar/LoopInterchange.cpp"
, 119, __PRETTY_FUNCTION__))
;
120 DEBUG(StringRef DepType =do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow"
: D->isAnti() ? "anti" : "output"; dbgs() << "Found "
<< DepType << " dependency between Src and Dst\n"
<< " Src:" << *Src << "\n Dst:" << *
Dst << '\n'; } } while (false)
121 D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow"
: D->isAnti() ? "anti" : "output"; dbgs() << "Found "
<< DepType << " dependency between Src and Dst\n"
<< " Src:" << *Src << "\n Dst:" << *
Dst << '\n'; } } while (false)
122 dbgs() << "Found " << DepTypedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow"
: D->isAnti() ? "anti" : "output"; dbgs() << "Found "
<< DepType << " dependency between Src and Dst\n"
<< " Src:" << *Src << "\n Dst:" << *
Dst << '\n'; } } while (false)
123 << " dependency between Src and Dst\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow"
: D->isAnti() ? "anti" : "output"; dbgs() << "Found "
<< DepType << " dependency between Src and Dst\n"
<< " Src:" << *Src << "\n Dst:" << *
Dst << '\n'; } } while (false)
124 << " Src:" << *Src << "\n Dst:" << *Dst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { StringRef DepType = D->isFlow() ? "flow"
: D->isAnti() ? "anti" : "output"; dbgs() << "Found "
<< DepType << " dependency between Src and Dst\n"
<< " Src:" << *Src << "\n Dst:" << *
Dst << '\n'; } } while (false)
;
125 unsigned Levels = D->getLevels();
126 char Direction;
127 for (unsigned II = 1; II <= Levels; ++II) {
128 const SCEV *Distance = D->getDistance(II);
129 const SCEVConstant *SCEVConst =
130 dyn_cast_or_null<SCEVConstant>(Distance);
131 if (SCEVConst) {
132 const ConstantInt *CI = SCEVConst->getValue();
133 if (CI->isNegative())
134 Direction = '<';
135 else if (CI->isZero())
136 Direction = '=';
137 else
138 Direction = '>';
139 Dep.push_back(Direction);
140 } else if (D->isScalar(II)) {
141 Direction = 'S';
142 Dep.push_back(Direction);
143 } else {
144 unsigned Dir = D->getDirection(II);
145 if (Dir == Dependence::DVEntry::LT ||
146 Dir == Dependence::DVEntry::LE)
147 Direction = '<';
148 else if (Dir == Dependence::DVEntry::GT ||
149 Dir == Dependence::DVEntry::GE)
150 Direction = '>';
151 else if (Dir == Dependence::DVEntry::EQ)
152 Direction = '=';
153 else
154 Direction = '*';
155 Dep.push_back(Direction);
156 }
157 }
158 while (Dep.size() != Level) {
159 Dep.push_back('I');
160 }
161
162 DepMatrix.push_back(Dep);
163 if (DepMatrix.size() > MaxMemInstrCount) {
164 DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCountdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Cannot handle more than "
<< MaxMemInstrCount << " dependencies inside loop\n"
; } } while (false)
165 << " dependencies inside loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Cannot handle more than "
<< MaxMemInstrCount << " dependencies inside loop\n"
; } } while (false)
;
166 return false;
167 }
168 }
169 }
170 }
171
172 // We don't have a DepMatrix to check legality return false.
173 if (DepMatrix.size() == 0)
174 return false;
175 return true;
176}
177
178// A loop is moved from index 'from' to an index 'to'. Update the Dependence
179// matrix by exchanging the two columns.
180static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx,
181 unsigned ToIndx) {
182 unsigned numRows = DepMatrix.size();
183 for (unsigned i = 0; i < numRows; ++i) {
184 char TmpVal = DepMatrix[i][ToIndx];
185 DepMatrix[i][ToIndx] = DepMatrix[i][FromIndx];
186 DepMatrix[i][FromIndx] = TmpVal;
187 }
188}
189
190// Checks if outermost non '=','S'or'I' dependence in the dependence matrix is
191// '>'
192static bool isOuterMostDepPositive(CharMatrix &DepMatrix, unsigned Row,
193 unsigned Column) {
194 for (unsigned i = 0; i <= Column; ++i) {
195 if (DepMatrix[Row][i] == '<')
196 return false;
197 if (DepMatrix[Row][i] == '>')
198 return true;
199 }
200 // All dependencies were '=','S' or 'I'
201 return false;
202}
203
204// Checks if no dependence exist in the dependency matrix in Row before Column.
205static bool containsNoDependence(CharMatrix &DepMatrix, unsigned Row,
206 unsigned Column) {
207 for (unsigned i = 0; i < Column; ++i) {
208 if (DepMatrix[Row][i] != '=' && DepMatrix[Row][i] != 'S' &&
209 DepMatrix[Row][i] != 'I')
210 return false;
211 }
212 return true;
213}
214
215static bool validDepInterchange(CharMatrix &DepMatrix, unsigned Row,
216 unsigned OuterLoopId, char InnerDep,
217 char OuterDep) {
218
219 if (isOuterMostDepPositive(DepMatrix, Row, OuterLoopId))
220 return false;
221
222 if (InnerDep == OuterDep)
223 return true;
224
225 // It is legal to interchange if and only if after interchange no row has a
226 // '>' direction as the leftmost non-'='.
227
228 if (InnerDep == '=' || InnerDep == 'S' || InnerDep == 'I')
229 return true;
230
231 if (InnerDep == '<')
232 return true;
233
234 if (InnerDep == '>') {
235 // If OuterLoopId represents outermost loop then interchanging will make the
236 // 1st dependency as '>'
237 if (OuterLoopId == 0)
238 return false;
239
240 // If all dependencies before OuterloopId are '=','S'or 'I'. Then
241 // interchanging will result in this row having an outermost non '='
242 // dependency of '>'
243 if (!containsNoDependence(DepMatrix, Row, OuterLoopId))
244 return true;
245 }
246
247 return false;
248}
249
250// Checks if it is legal to interchange 2 loops.
251// [Theorem] A permutation of the loops in a perfect nest is legal if and only
252// if the direction matrix, after the same permutation is applied to its
253// columns, has no ">" direction as the leftmost non-"=" direction in any row.
254static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix,
255 unsigned InnerLoopId,
256 unsigned OuterLoopId) {
257
258 unsigned NumRows = DepMatrix.size();
259 // For each row check if it is valid to interchange.
260 for (unsigned Row = 0; Row < NumRows; ++Row) {
261 char InnerDep = DepMatrix[Row][InnerLoopId];
262 char OuterDep = DepMatrix[Row][OuterLoopId];
263 if (InnerDep == '*' || OuterDep == '*')
264 return false;
265 if (!validDepInterchange(DepMatrix, Row, OuterLoopId, InnerDep, OuterDep))
266 return false;
267 }
268 return true;
269}
270
271static void populateWorklist(Loop &L, SmallVector<LoopVector, 8> &V) {
272
273 DEBUG(dbgs() << "Calling populateWorklist on Func: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Calling populateWorklist on Func: "
<< L.getHeader()->getParent()->getName() <<
" Loop: %" << L.getHeader()->getName() << '\n'
; } } while (false)
274 << L.getHeader()->getParent()->getName() << " Loop: %"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Calling populateWorklist on Func: "
<< L.getHeader()->getParent()->getName() <<
" Loop: %" << L.getHeader()->getName() << '\n'
; } } while (false)
275 << L.getHeader()->getName() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Calling populateWorklist on Func: "
<< L.getHeader()->getParent()->getName() <<
" Loop: %" << L.getHeader()->getName() << '\n'
; } } while (false)
;
276 LoopVector LoopList;
277 Loop *CurrentLoop = &L;
278 const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops();
279 while (!Vec->empty()) {
280 // The current loop has multiple subloops in it hence it is not tightly
281 // nested.
282 // Discard all loops above it added into Worklist.
283 if (Vec->size() != 1) {
284 LoopList.clear();
285 return;
286 }
287 LoopList.push_back(CurrentLoop);
288 CurrentLoop = Vec->front();
289 Vec = &CurrentLoop->getSubLoops();
290 }
291 LoopList.push_back(CurrentLoop);
292 V.push_back(std::move(LoopList));
293}
294
295static PHINode *getInductionVariable(Loop *L, ScalarEvolution *SE) {
296 PHINode *InnerIndexVar = L->getCanonicalInductionVariable();
297 if (InnerIndexVar)
298 return InnerIndexVar;
299 if (L->getLoopLatch() == nullptr || L->getLoopPredecessor() == nullptr)
300 return nullptr;
301 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
302 PHINode *PhiVar = cast<PHINode>(I);
303 Type *PhiTy = PhiVar->getType();
304 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
305 !PhiTy->isPointerTy())
306 return nullptr;
307 const SCEVAddRecExpr *AddRec =
308 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(PhiVar));
309 if (!AddRec || !AddRec->isAffine())
310 continue;
311 const SCEV *Step = AddRec->getStepRecurrence(*SE);
312 if (!isa<SCEVConstant>(Step))
313 continue;
314 // Found the induction variable.
315 // FIXME: Handle loops with more than one induction variable. Note that,
316 // currently, legality makes sure we have only one induction variable.
317 return PhiVar;
318 }
319 return nullptr;
320}
321
322/// LoopInterchangeLegality checks if it is legal to interchange the loop.
323class LoopInterchangeLegality {
324public:
325 LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
326 LoopInfo *LI, DominatorTree *DT, bool PreserveLCSSA)
327 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT),
328 PreserveLCSSA(PreserveLCSSA), InnerLoopHasReduction(false) {}
329
330 /// Check if the loops can be interchanged.
331 bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
332 CharMatrix &DepMatrix);
333 /// Check if the loop structure is understood. We do not handle triangular
334 /// loops for now.
335 bool isLoopStructureUnderstood(PHINode *InnerInductionVar);
336
337 bool currentLimitations();
338
339 bool hasInnerLoopReduction() { return InnerLoopHasReduction; }
340
341private:
342 bool tightlyNested(Loop *Outer, Loop *Inner);
343 bool containsUnsafeInstructionsInHeader(BasicBlock *BB);
344 bool areAllUsesReductions(Instruction *Ins, Loop *L);
345 bool containsUnsafeInstructionsInLatch(BasicBlock *BB);
346 bool findInductionAndReductions(Loop *L,
347 SmallVector<PHINode *, 8> &Inductions,
348 SmallVector<PHINode *, 8> &Reductions);
349 Loop *OuterLoop;
350 Loop *InnerLoop;
351
352 ScalarEvolution *SE;
353 LoopInfo *LI;
354 DominatorTree *DT;
355 bool PreserveLCSSA;
356
357 bool InnerLoopHasReduction;
358};
359
360/// LoopInterchangeProfitability checks if it is profitable to interchange the
361/// loop.
362class LoopInterchangeProfitability {
363public:
364 LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE)
365 : OuterLoop(Outer), InnerLoop(Inner), SE(SE) {}
366
367 /// Check if the loop interchange is profitable.
368 bool isProfitable(unsigned InnerLoopId, unsigned OuterLoopId,
369 CharMatrix &DepMatrix);
370
371private:
372 int getInstrOrderCost();
373
374 Loop *OuterLoop;
375 Loop *InnerLoop;
376
377 /// Scev analysis.
378 ScalarEvolution *SE;
379};
380
381/// LoopInterchangeTransform interchanges the loop.
382class LoopInterchangeTransform {
383public:
384 LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
385 LoopInfo *LI, DominatorTree *DT,
386 BasicBlock *LoopNestExit,
387 bool InnerLoopContainsReductions)
388 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT),
389 LoopExit(LoopNestExit),
390 InnerLoopHasReduction(InnerLoopContainsReductions) {}
391
392 /// Interchange OuterLoop and InnerLoop.
393 bool transform();
394 void restructureLoops(Loop *InnerLoop, Loop *OuterLoop);
395 void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
396
397private:
398 void splitInnerLoopLatch(Instruction *);
399 void splitInnerLoopHeader();
400 bool adjustLoopLinks();
401 void adjustLoopPreheaders();
402 bool adjustLoopBranches();
403 void updateIncomingBlock(BasicBlock *CurrBlock, BasicBlock *OldPred,
404 BasicBlock *NewPred);
405
406 Loop *OuterLoop;
407 Loop *InnerLoop;
408
409 /// Scev analysis.
410 ScalarEvolution *SE;
411 LoopInfo *LI;
412 DominatorTree *DT;
413 BasicBlock *LoopExit;
414 bool InnerLoopHasReduction;
415};
416
417// Main LoopInterchange Pass.
418struct LoopInterchange : public FunctionPass {
419 static char ID;
420 ScalarEvolution *SE;
421 LoopInfo *LI;
422 DependenceInfo *DI;
423 DominatorTree *DT;
424 bool PreserveLCSSA;
425 LoopInterchange()
426 : FunctionPass(ID), SE(nullptr), LI(nullptr), DI(nullptr), DT(nullptr) {
427 initializeLoopInterchangePass(*PassRegistry::getPassRegistry());
428 }
429
430 void getAnalysisUsage(AnalysisUsage &AU) const override {
431 AU.addRequired<ScalarEvolutionWrapperPass>();
432 AU.addRequired<AAResultsWrapperPass>();
433 AU.addRequired<DominatorTreeWrapperPass>();
434 AU.addRequired<LoopInfoWrapperPass>();
435 AU.addRequired<DependenceAnalysisWrapperPass>();
436 AU.addRequiredID(LoopSimplifyID);
437 AU.addRequiredID(LCSSAID);
438 }
439
440 bool runOnFunction(Function &F) override {
441 if (skipFunction(F))
442 return false;
443
444 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
445 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
446 DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI();
447 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
448 DT = DTWP ? &DTWP->getDomTree() : nullptr;
449 PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
450
451 // Build up a worklist of loop pairs to analyze.
452 SmallVector<LoopVector, 8> Worklist;
453
454 for (Loop *L : *LI)
455 populateWorklist(*L, Worklist);
456
457 DEBUG(dbgs() << "Worklist size = " << Worklist.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Worklist size = " <<
Worklist.size() << "\n"; } } while (false)
;
458 bool Changed = true;
459 while (!Worklist.empty()) {
460 LoopVector LoopList = Worklist.pop_back_val();
461 Changed = processLoopList(LoopList, F);
462 }
463 return Changed;
464 }
465
466 bool isComputableLoopNest(LoopVector LoopList) {
467 for (Loop *L : LoopList) {
468 const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
469 if (ExitCountOuter == SE->getCouldNotCompute()) {
470 DEBUG(dbgs() << "Couldn't compute backedge count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Couldn't compute backedge count\n"
; } } while (false)
;
471 return false;
472 }
473 if (L->getNumBackEdges() != 1) {
474 DEBUG(dbgs() << "NumBackEdges is not equal to 1\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "NumBackEdges is not equal to 1\n"
; } } while (false)
;
475 return false;
476 }
477 if (!L->getExitingBlock()) {
478 DEBUG(dbgs() << "Loop doesn't have unique exit block\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Loop doesn't have unique exit block\n"
; } } while (false)
;
479 return false;
480 }
481 }
482 return true;
483 }
484
485 unsigned selectLoopForInterchange(const LoopVector &LoopList) {
486 // TODO: Add a better heuristic to select the loop to be interchanged based
487 // on the dependence matrix. Currently we select the innermost loop.
488 return LoopList.size() - 1;
489 }
490
491 bool processLoopList(LoopVector LoopList, Function &F) {
492
493 bool Changed = false;
494 unsigned LoopNestDepth = LoopList.size();
495 if (LoopNestDepth < 2) {
496 DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Loop doesn't contain minimum nesting level.\n"
; } } while (false)
;
497 return false;
498 }
499 if (LoopNestDepth > MaxLoopNestDepth) {
500 DEBUG(dbgs() << "Cannot handle loops of depth greater than "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Cannot handle loops of depth greater than "
<< MaxLoopNestDepth << "\n"; } } while (false)
501 << MaxLoopNestDepth << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Cannot handle loops of depth greater than "
<< MaxLoopNestDepth << "\n"; } } while (false)
;
502 return false;
503 }
504 if (!isComputableLoopNest(LoopList)) {
505 DEBUG(dbgs() << "Not valid loop candidate for interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Not valid loop candidate for interchange\n"
; } } while (false)
;
506 return false;
507 }
508
509 DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Processing LoopList of size = "
<< LoopNestDepth << "\n"; } } while (false)
;
510
511 CharMatrix DependencyMatrix;
512 Loop *OuterMostLoop = *(LoopList.begin());
513 if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
514 OuterMostLoop, DI)) {
515 DEBUG(dbgs() << "Populating dependency matrix failed\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Populating dependency matrix failed\n"
; } } while (false)
;
516 return false;
517 }
518#ifdef DUMP_DEP_MATRICIES
519 DEBUG(dbgs() << "Dependence before interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Dependence before interchange\n"
; } } while (false)
;
520 printDepMatrix(DependencyMatrix);
521#endif
522
523 BasicBlock *OuterMostLoopLatch = OuterMostLoop->getLoopLatch();
524 BranchInst *OuterMostLoopLatchBI =
525 dyn_cast<BranchInst>(OuterMostLoopLatch->getTerminator());
526 if (!OuterMostLoopLatchBI)
527 return false;
528
529 // Since we currently do not handle LCSSA PHI's any failure in loop
530 // condition will now branch to LoopNestExit.
531 // TODO: This should be removed once we handle LCSSA PHI nodes.
532
533 // Get the Outermost loop exit.
534 BasicBlock *LoopNestExit;
535 if (OuterMostLoopLatchBI->getSuccessor(0) == OuterMostLoop->getHeader())
536 LoopNestExit = OuterMostLoopLatchBI->getSuccessor(1);
537 else
538 LoopNestExit = OuterMostLoopLatchBI->getSuccessor(0);
539
540 if (isa<PHINode>(LoopNestExit->begin())) {
541 DEBUG(dbgs() << "PHI Nodes in loop nest exit is not handled for now "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "PHI Nodes in loop nest exit is not handled for now "
"since on failure all loops branch to loop nest exit.\n"; } }
while (false)
542 "since on failure all loops branch to loop nest exit.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "PHI Nodes in loop nest exit is not handled for now "
"since on failure all loops branch to loop nest exit.\n"; } }
while (false)
;
543 return false;
544 }
545
546 unsigned SelecLoopId = selectLoopForInterchange(LoopList);
547 // Move the selected loop outwards to the best possible position.
548 for (unsigned i = SelecLoopId; i > 0; i--) {
549 bool Interchanged =
550 processLoop(LoopList, i, i - 1, LoopNestExit, DependencyMatrix);
551 if (!Interchanged)
552 return Changed;
553 // Loops interchanged reflect the same in LoopList
554 std::swap(LoopList[i - 1], LoopList[i]);
555
556 // Update the DependencyMatrix
557 interChangeDependencies(DependencyMatrix, i, i - 1);
558 DT->recalculate(F);
559#ifdef DUMP_DEP_MATRICIES
560 DEBUG(dbgs() << "Dependence after interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Dependence after interchange\n"
; } } while (false)
;
561 printDepMatrix(DependencyMatrix);
562#endif
563 Changed |= Interchanged;
564 }
565 return Changed;
566 }
567
568 bool processLoop(LoopVector LoopList, unsigned InnerLoopId,
569 unsigned OuterLoopId, BasicBlock *LoopNestExit,
570 std::vector<std::vector<char>> &DependencyMatrix) {
571
572 DEBUG(dbgs() << "Processing Inner Loop Id = " << InnerLoopIddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Processing Inner Loop Id = "
<< InnerLoopId << " and OuterLoopId = " <<
OuterLoopId << "\n"; } } while (false)
573 << " and OuterLoopId = " << OuterLoopId << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Processing Inner Loop Id = "
<< InnerLoopId << " and OuterLoopId = " <<
OuterLoopId << "\n"; } } while (false)
;
574 Loop *InnerLoop = LoopList[InnerLoopId];
575 Loop *OuterLoop = LoopList[OuterLoopId];
576
577 LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, LI, DT,
578 PreserveLCSSA);
579 if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
580 DEBUG(dbgs() << "Not interchanging Loops. Cannot prove legality\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Not interchanging Loops. Cannot prove legality\n"
; } } while (false)
;
581 return false;
582 }
583 DEBUG(dbgs() << "Loops are legal to interchange\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Loops are legal to interchange\n"
; } } while (false)
;
584 LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE);
585 if (!LIP.isProfitable(InnerLoopId, OuterLoopId, DependencyMatrix)) {
586 DEBUG(dbgs() << "Interchanging loops not profitable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Interchanging loops not profitable\n"
; } } while (false)
;
587 return false;
588 }
589
590 LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT,
591 LoopNestExit, LIL.hasInnerLoopReduction());
592 LIT.transform();
593 DEBUG(dbgs() << "Loops interchanged\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Loops interchanged\n"
; } } while (false)
;
594 return true;
595 }
596};
597
598} // end of namespace
599bool LoopInterchangeLegality::areAllUsesReductions(Instruction *Ins, Loop *L) {
600 return none_of(Ins->users(), [=](User *U) -> bool {
601 auto *UserIns = dyn_cast<PHINode>(U);
602 RecurrenceDescriptor RD;
603 return !UserIns || !RecurrenceDescriptor::isReductionPHI(UserIns, L, RD);
604 });
605}
606
607bool LoopInterchangeLegality::containsUnsafeInstructionsInHeader(
608 BasicBlock *BB) {
609 for (auto I = BB->begin(), E = BB->end(); I != E; ++I) {
610 // Load corresponding to reduction PHI's are safe while concluding if
611 // tightly nested.
612 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
613 if (!areAllUsesReductions(L, InnerLoop))
614 return true;
615 } else if (I->mayHaveSideEffects() || I->mayReadFromMemory())
616 return true;
617 }
618 return false;
619}
620
621bool LoopInterchangeLegality::containsUnsafeInstructionsInLatch(
622 BasicBlock *BB) {
623 for (auto I = BB->begin(), E = BB->end(); I != E; ++I) {
624 // Stores corresponding to reductions are safe while concluding if tightly
625 // nested.
626 if (StoreInst *L = dyn_cast<StoreInst>(I)) {
627 if (!isa<PHINode>(L->getOperand(0)))
628 return true;
629 } else if (I->mayHaveSideEffects() || I->mayReadFromMemory())
630 return true;
631 }
632 return false;
633}
634
635bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
636 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
637 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
638 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
639
640 DEBUG(dbgs() << "Checking if loops are tightly nested\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Checking if loops are tightly nested\n"
; } } while (false)
;
641
642 // A perfectly nested loop will not have any branch in between the outer and
643 // inner block i.e. outer header will branch to either inner preheader and
644 // outerloop latch.
645 BranchInst *OuterLoopHeaderBI =
646 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
647 if (!OuterLoopHeaderBI)
648 return false;
649
650 for (unsigned i = 0, e = OuterLoopHeaderBI->getNumSuccessors(); i < e; ++i) {
651 if (OuterLoopHeaderBI->getSuccessor(i) != InnerLoopPreHeader &&
652 OuterLoopHeaderBI->getSuccessor(i) != OuterLoopLatch)
653 return false;
654 }
655
656 DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Checking instructions in Loop header and Loop latch\n"
; } } while (false)
;
657 // We do not have any basic block in between now make sure the outer header
658 // and outer loop latch doesn't contain any unsafe instructions.
659 if (containsUnsafeInstructionsInHeader(OuterLoopHeader) ||
660 containsUnsafeInstructionsInLatch(OuterLoopLatch))
661 return false;
662
663 DEBUG(dbgs() << "Loops are perfectly nested\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Loops are perfectly nested\n"
; } } while (false)
;
664 // We have a perfect loop nest.
665 return true;
666}
667
668
669bool LoopInterchangeLegality::isLoopStructureUnderstood(
670 PHINode *InnerInduction) {
671
672 unsigned Num = InnerInduction->getNumOperands();
673 BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
674 for (unsigned i = 0; i < Num; ++i) {
675 Value *Val = InnerInduction->getOperand(i);
676 if (isa<Constant>(Val))
677 continue;
678 Instruction *I = dyn_cast<Instruction>(Val);
679 if (!I)
680 return false;
681 // TODO: Handle triangular loops.
682 // e.g. for(int i=0;i<N;i++)
683 // for(int j=i;j<N;j++)
684 unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
685 if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
686 InnerLoopPreheader &&
687 !OuterLoop->isLoopInvariant(I)) {
688 return false;
689 }
690 }
691 return true;
692}
693
694bool LoopInterchangeLegality::findInductionAndReductions(
695 Loop *L, SmallVector<PHINode *, 8> &Inductions,
696 SmallVector<PHINode *, 8> &Reductions) {
697 if (!L->getLoopLatch() || !L->getLoopPredecessor())
698 return false;
699 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
700 RecurrenceDescriptor RD;
701 InductionDescriptor ID;
702 PHINode *PHI = cast<PHINode>(I);
703 if (InductionDescriptor::isInductionPHI(PHI, L, SE, ID))
704 Inductions.push_back(PHI);
705 else if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD))
706 Reductions.push_back(PHI);
707 else {
708 DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Failed to recognize PHI as an induction or reduction.\n"
; } } while (false)
709 dbgs() << "Failed to recognize PHI as an induction or reduction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Failed to recognize PHI as an induction or reduction.\n"
; } } while (false)
;
710 return false;
711 }
712 }
713 return true;
714}
715
716static bool containsSafePHI(BasicBlock *Block, bool isOuterLoopExitBlock) {
717 for (auto I = Block->begin(); isa<PHINode>(I); ++I) {
718 PHINode *PHI = cast<PHINode>(I);
719 // Reduction lcssa phi will have only 1 incoming block that from loop latch.
720 if (PHI->getNumIncomingValues() > 1)
721 return false;
722 Instruction *Ins = dyn_cast<Instruction>(PHI->getIncomingValue(0));
723 if (!Ins)
724 return false;
725 // Incoming value for lcssa phi's in outer loop exit can only be inner loop
726 // exits lcssa phi else it would not be tightly nested.
727 if (!isa<PHINode>(Ins) && isOuterLoopExitBlock)
728 return false;
729 }
730 return true;
731}
732
733static BasicBlock *getLoopLatchExitBlock(BasicBlock *LatchBlock,
734 BasicBlock *LoopHeader) {
735 if (BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator())) {
736 unsigned Num = BI->getNumSuccessors();
737 assert(Num == 2)((Num == 2) ? static_cast<void> (0) : __assert_fail ("Num == 2"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301389/lib/Transforms/Scalar/LoopInterchange.cpp"
, 737, __PRETTY_FUNCTION__))
;
738 for (unsigned i = 0; i < Num; ++i) {
739 if (BI->getSuccessor(i) == LoopHeader)
740 continue;
741 return BI->getSuccessor(i);
742 }
743 }
744 return nullptr;
745}
746
747// This function indicates the current limitations in the transform as a result
748// of which we do not proceed.
749bool LoopInterchangeLegality::currentLimitations() {
750
751 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
752 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
753 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
754 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
755 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
756
757 PHINode *InnerInductionVar;
758 SmallVector<PHINode *, 8> Inductions;
759 SmallVector<PHINode *, 8> Reductions;
760 if (!findInductionAndReductions(InnerLoop, Inductions, Reductions))
761 return true;
762
763 // TODO: Currently we handle only loops with 1 induction variable.
764 if (Inductions.size() != 1) {
765 DEBUG(dbgs() << "We currently only support loops with 1 induction variable."do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "We currently only support loops with 1 induction variable."
<< "Failed to interchange due to current limitation\n"
; } } while (false)
766 << "Failed to interchange due to current limitation\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "We currently only support loops with 1 induction variable."
<< "Failed to interchange due to current limitation\n"
; } } while (false)
;
767 return true;
768 }
769 if (Reductions.size() > 0)
770 InnerLoopHasReduction = true;
771
772 InnerInductionVar = Inductions.pop_back_val();
773 Reductions.clear();
774 if (!findInductionAndReductions(OuterLoop, Inductions, Reductions))
775 return true;
776
777 // Outer loop cannot have reduction because then loops will not be tightly
778 // nested.
779 if (!Reductions.empty())
780 return true;
781 // TODO: Currently we handle only loops with 1 induction variable.
782 if (Inductions.size() != 1)
783 return true;
784
785 // TODO: Triangular loops are not handled for now.
786 if (!isLoopStructureUnderstood(InnerInductionVar)) {
787 DEBUG(dbgs() << "Loop structure not understood by pass\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Loop structure not understood by pass\n"
; } } while (false)
;
788 return true;
789 }
790
791 // TODO: We only handle LCSSA PHI's corresponding to reduction for now.
792 BasicBlock *LoopExitBlock =
793 getLoopLatchExitBlock(OuterLoopLatch, OuterLoopHeader);
794 if (!LoopExitBlock || !containsSafePHI(LoopExitBlock, true))
795 return true;
796
797 LoopExitBlock = getLoopLatchExitBlock(InnerLoopLatch, InnerLoopHeader);
798 if (!LoopExitBlock || !containsSafePHI(LoopExitBlock, false))
799 return true;
800
801 // TODO: Current limitation: Since we split the inner loop latch at the point
802 // were induction variable is incremented (induction.next); We cannot have
803 // more than 1 user of induction.next since it would result in broken code
804 // after split.
805 // e.g.
806 // for(i=0;i<N;i++) {
807 // for(j = 0;j<M;j++) {
808 // A[j+1][i+2] = A[j][i]+k;
809 // }
810 // }
811 Instruction *InnerIndexVarInc = nullptr;
812 if (InnerInductionVar->getIncomingBlock(0) == InnerLoopPreHeader)
813 InnerIndexVarInc =
814 dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(1));
815 else
816 InnerIndexVarInc =
817 dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(0));
818
819 if (!InnerIndexVarInc)
820 return true;
821
822 // Since we split the inner loop latch on this induction variable. Make sure
823 // we do not have any instruction between the induction variable and branch
824 // instruction.
825
826 bool FoundInduction = false;
827 for (const Instruction &I : reverse(*InnerLoopLatch)) {
828 if (isa<BranchInst>(I) || isa<CmpInst>(I) || isa<TruncInst>(I))
829 continue;
830 // We found an instruction. If this is not induction variable then it is not
831 // safe to split this loop latch.
832 if (!I.isIdenticalTo(InnerIndexVarInc))
833 return true;
834
835 FoundInduction = true;
836 break;
837 }
838 // The loop latch ended and we didn't find the induction variable return as
839 // current limitation.
840 if (!FoundInduction)
841 return true;
842
843 return false;
844}
845
846bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
847 unsigned OuterLoopId,
848 CharMatrix &DepMatrix) {
849
850 if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
851 DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopIddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Failed interchange InnerLoopId = "
<< InnerLoopId << " and OuterLoopId = " <<
OuterLoopId << " due to dependence\n"; } } while (false
)
852 << " and OuterLoopId = " << OuterLoopIddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Failed interchange InnerLoopId = "
<< InnerLoopId << " and OuterLoopId = " <<
OuterLoopId << " due to dependence\n"; } } while (false
)
853 << " due to dependence\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Failed interchange InnerLoopId = "
<< InnerLoopId << " and OuterLoopId = " <<
OuterLoopId << " due to dependence\n"; } } while (false
)
;
854 return false;
855 }
856
857 // Create unique Preheaders if we already do not have one.
858 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
859 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
860
861 // Create a unique outer preheader -
862 // 1) If OuterLoop preheader is not present.
863 // 2) If OuterLoop Preheader is same as OuterLoop Header
864 // 3) If OuterLoop Preheader is same as Header of the previous loop.
865 // 4) If OuterLoop Preheader is Entry node.
866 if (!OuterLoopPreHeader || OuterLoopPreHeader == OuterLoop->getHeader() ||
867 isa<PHINode>(OuterLoopPreHeader->begin()) ||
868 !OuterLoopPreHeader->getUniquePredecessor()) {
869 OuterLoopPreHeader =
870 InsertPreheaderForLoop(OuterLoop, DT, LI, PreserveLCSSA);
871 }
872
873 if (!InnerLoopPreHeader || InnerLoopPreHeader == InnerLoop->getHeader() ||
874 InnerLoopPreHeader == OuterLoop->getHeader()) {
875 InnerLoopPreHeader =
Value stored to 'InnerLoopPreHeader' is never read
876 InsertPreheaderForLoop(InnerLoop, DT, LI, PreserveLCSSA);
877 }
878
879 // TODO: The loops could not be interchanged due to current limitations in the
880 // transform module.
881 if (currentLimitations()) {
882 DEBUG(dbgs() << "Not legal because of current transform limitation\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Not legal because of current transform limitation\n"
; } } while (false)
;
883 return false;
884 }
885
886 // Check if the loops are tightly nested.
887 if (!tightlyNested(OuterLoop, InnerLoop)) {
888 DEBUG(dbgs() << "Loops not tightly nested\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Loops not tightly nested\n"
; } } while (false)
;
889 return false;
890 }
891
892 return true;
893}
894
895int LoopInterchangeProfitability::getInstrOrderCost() {
896 unsigned GoodOrder, BadOrder;
897 BadOrder = GoodOrder = 0;
898 for (auto BI = InnerLoop->block_begin(), BE = InnerLoop->block_end();
899 BI != BE; ++BI) {
900 for (Instruction &Ins : **BI) {
901 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
902 unsigned NumOp = GEP->getNumOperands();
903 bool FoundInnerInduction = false;
904 bool FoundOuterInduction = false;
905 for (unsigned i = 0; i < NumOp; ++i) {
906 const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i));
907 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
908 if (!AR)
909 continue;
910
911 // If we find the inner induction after an outer induction e.g.
912 // for(int i=0;i<N;i++)
913 // for(int j=0;j<N;j++)
914 // A[i][j] = A[i-1][j-1]+k;
915 // then it is a good order.
916 if (AR->getLoop() == InnerLoop) {
917 // We found an InnerLoop induction after OuterLoop induction. It is
918 // a good order.
919 FoundInnerInduction = true;
920 if (FoundOuterInduction) {
921 GoodOrder++;
922 break;
923 }
924 }
925 // If we find the outer induction after an inner induction e.g.
926 // for(int i=0;i<N;i++)
927 // for(int j=0;j<N;j++)
928 // A[j][i] = A[j-1][i-1]+k;
929 // then it is a bad order.
930 if (AR->getLoop() == OuterLoop) {
931 // We found an OuterLoop induction after InnerLoop induction. It is
932 // a bad order.
933 FoundOuterInduction = true;
934 if (FoundInnerInduction) {
935 BadOrder++;
936 break;
937 }
938 }
939 }
940 }
941 }
942 }
943 return GoodOrder - BadOrder;
944}
945
946static bool isProfitableForVectorization(unsigned InnerLoopId,
947 unsigned OuterLoopId,
948 CharMatrix &DepMatrix) {
949 // TODO: Improve this heuristic to catch more cases.
950 // If the inner loop is loop independent or doesn't carry any dependency it is
951 // profitable to move this to outer position.
952 unsigned Row = DepMatrix.size();
953 for (unsigned i = 0; i < Row; ++i) {
954 if (DepMatrix[i][InnerLoopId] != 'S' && DepMatrix[i][InnerLoopId] != 'I')
955 return false;
956 // TODO: We need to improve this heuristic.
957 if (DepMatrix[i][OuterLoopId] != '=')
958 return false;
959 }
960 // If outer loop has dependence and inner loop is loop independent then it is
961 // profitable to interchange to enable parallelism.
962 return true;
963}
964
965bool LoopInterchangeProfitability::isProfitable(unsigned InnerLoopId,
966 unsigned OuterLoopId,
967 CharMatrix &DepMatrix) {
968
969 // TODO: Add better profitability checks.
970 // e.g
971 // 1) Construct dependency matrix and move the one with no loop carried dep
972 // inside to enable vectorization.
973
974 // This is rough cost estimation algorithm. It counts the good and bad order
975 // of induction variables in the instruction and allows reordering if number
976 // of bad orders is more than good.
977 int Cost = getInstrOrderCost();
978 DEBUG(dbgs() << "Cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Cost = " << Cost
<< "\n"; } } while (false)
;
979 if (Cost < -LoopInterchangeCostThreshold)
980 return true;
981
982 // It is not profitable as per current cache profitability model. But check if
983 // we can move this loop outside to improve parallelism.
984 bool ImprovesPar =
985 isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix);
986 return ImprovesPar;
987}
988
989void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
990 Loop *InnerLoop) {
991 for (Loop::iterator I = OuterLoop->begin(), E = OuterLoop->end(); I != E;
992 ++I) {
993 if (*I == InnerLoop) {
994 OuterLoop->removeChildLoop(I);
995 return;
996 }
997 }
998 llvm_unreachable("Couldn't find loop")::llvm::llvm_unreachable_internal("Couldn't find loop", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn301389/lib/Transforms/Scalar/LoopInterchange.cpp"
, 998)
;
999}
1000
1001void LoopInterchangeTransform::restructureLoops(Loop *InnerLoop,
1002 Loop *OuterLoop) {
1003 Loop *OuterLoopParent = OuterLoop->getParentLoop();
1004 if (OuterLoopParent) {
1005 // Remove the loop from its parent loop.
1006 removeChildLoop(OuterLoopParent, OuterLoop);
1007 removeChildLoop(OuterLoop, InnerLoop);
1008 OuterLoopParent->addChildLoop(InnerLoop);
1009 } else {
1010 removeChildLoop(OuterLoop, InnerLoop);
1011 LI->changeTopLevelLoop(OuterLoop, InnerLoop);
1012 }
1013
1014 while (!InnerLoop->empty())
1015 OuterLoop->addChildLoop(InnerLoop->removeChildLoop(InnerLoop->begin()));
1016
1017 InnerLoop->addChildLoop(OuterLoop);
1018}
1019
1020bool LoopInterchangeTransform::transform() {
1021 bool Transformed = false;
1022 Instruction *InnerIndexVar;
1023
1024 if (InnerLoop->getSubLoops().size() == 0) {
1025 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1026 DEBUG(dbgs() << "Calling Split Inner Loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Calling Split Inner Loop\n"
; } } while (false)
;
1027 PHINode *InductionPHI = getInductionVariable(InnerLoop, SE);
1028 if (!InductionPHI) {
1029 DEBUG(dbgs() << "Failed to find the point to split loop latch \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Failed to find the point to split loop latch \n"
; } } while (false)
;
1030 return false;
1031 }
1032
1033 if (InductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1034 InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(1));
1035 else
1036 InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(0));
1037
1038 //
1039 // Split at the place were the induction variable is
1040 // incremented/decremented.
1041 // TODO: This splitting logic may not work always. Fix this.
1042 splitInnerLoopLatch(InnerIndexVar);
1043 DEBUG(dbgs() << "splitInnerLoopLatch done\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "splitInnerLoopLatch done\n"
; } } while (false)
;
1044
1045 // Splits the inner loops phi nodes out into a separate basic block.
1046 splitInnerLoopHeader();
1047 DEBUG(dbgs() << "splitInnerLoopHeader done\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "splitInnerLoopHeader done\n"
; } } while (false)
;
1048 }
1049
1050 Transformed |= adjustLoopLinks();
1051 if (!Transformed) {
1052 DEBUG(dbgs() << "adjustLoopLinks failed\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "adjustLoopLinks failed\n"
; } } while (false)
;
1053 return false;
1054 }
1055
1056 restructureLoops(InnerLoop, OuterLoop);
1057 return true;
1058}
1059
1060void LoopInterchangeTransform::splitInnerLoopLatch(Instruction *Inc) {
1061 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1062 BasicBlock *InnerLoopLatchPred = InnerLoopLatch;
1063 InnerLoopLatch = SplitBlock(InnerLoopLatchPred, Inc, DT, LI);
1064}
1065
1066void LoopInterchangeTransform::splitInnerLoopHeader() {
1067
1068 // Split the inner loop header out. Here make sure that the reduction PHI's
1069 // stay in the innerloop body.
1070 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1071 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1072 if (InnerLoopHasReduction) {
1073 // FIXME: Check if the induction PHI will always be the first PHI.
1074 BasicBlock *New = InnerLoopHeader->splitBasicBlock(
1075 ++(InnerLoopHeader->begin()), InnerLoopHeader->getName() + ".split");
1076 if (LI)
1077 if (Loop *L = LI->getLoopFor(InnerLoopHeader))
1078 L->addBasicBlockToLoop(New, *LI);
1079
1080 // Adjust Reduction PHI's in the block.
1081 SmallVector<PHINode *, 8> PHIVec;
1082 for (auto I = New->begin(); isa<PHINode>(I); ++I) {
1083 PHINode *PHI = dyn_cast<PHINode>(I);
1084 Value *V = PHI->getIncomingValueForBlock(InnerLoopPreHeader);
1085 PHI->replaceAllUsesWith(V);
1086 PHIVec.push_back((PHI));
1087 }
1088 for (PHINode *P : PHIVec) {
1089 P->eraseFromParent();
1090 }
1091 } else {
1092 SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI);
1093 }
1094
1095 DEBUG(dbgs() << "Output of splitInnerLoopHeader InnerLoopHeaderSucc & "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Output of splitInnerLoopHeader InnerLoopHeaderSucc & "
"InnerLoopHeader\n"; } } while (false)
1096 "InnerLoopHeader\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "Output of splitInnerLoopHeader InnerLoopHeaderSucc & "
"InnerLoopHeader\n"; } } while (false)
;
1097}
1098
1099/// \brief Move all instructions except the terminator from FromBB right before
1100/// InsertBefore
1101static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1102 auto &ToList = InsertBefore->getParent()->getInstList();
1103 auto &FromList = FromBB->getInstList();
1104
1105 ToList.splice(InsertBefore->getIterator(), FromList, FromList.begin(),
1106 FromBB->getTerminator()->getIterator());
1107}
1108
1109void LoopInterchangeTransform::updateIncomingBlock(BasicBlock *CurrBlock,
1110 BasicBlock *OldPred,
1111 BasicBlock *NewPred) {
1112 for (auto I = CurrBlock->begin(); isa<PHINode>(I); ++I) {
1113 PHINode *PHI = cast<PHINode>(I);
1114 unsigned Num = PHI->getNumIncomingValues();
1115 for (unsigned i = 0; i < Num; ++i) {
1116 if (PHI->getIncomingBlock(i) == OldPred)
1117 PHI->setIncomingBlock(i, NewPred);
1118 }
1119 }
1120}
1121
1122bool LoopInterchangeTransform::adjustLoopBranches() {
1123
1124 DEBUG(dbgs() << "adjustLoopBranches called\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-interchange")) { dbgs() << "adjustLoopBranches called\n"
; } } while (false)
;
1125 // Adjust the loop preheader
1126 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1127 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1128 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1129 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1130 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1131 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1132 BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1133 BasicBlock *InnerLoopLatchPredecessor =
1134 InnerLoopLatch->getUniquePredecessor();
1135 BasicBlock *InnerLoopLatchSuccessor;
1136 BasicBlock *OuterLoopLatchSuccessor;
1137
1138 BranchInst *OuterLoopLatchBI =
1139 dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1140 BranchInst *InnerLoopLatchBI =
1141 dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1142 BranchInst *OuterLoopHeaderBI =
1143 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1144 BranchInst *InnerLoopHeaderBI =
1145 dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1146
1147 if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1148 !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1149 !InnerLoopHeaderBI)
1150 return false;
1151
1152 BranchInst *InnerLoopLatchPredecessorBI =
1153 dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1154 BranchInst *OuterLoopPredecessorBI =
1155 dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1156
1157 if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1158 return false;
1159 BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1160 if (!InnerLoopHeaderSuccessor)
1161 return false;
1162
1163 // Adjust Loop Preheader and headers
1164
1165 unsigned NumSucc = OuterLoopPredecessorBI->getNumSuccessors();
1166 for (unsigned i = 0; i < NumSucc; ++i) {
1167 if (OuterLoopPredecessorBI->getSuccessor(i) == OuterLoopPreHeader)
1168 OuterLoopPredecessorBI->setSuccessor(i, InnerLoopPreHeader);
1169 }
1170
1171 NumSucc = OuterLoopHeaderBI->getNumSuccessors();
1172 for (unsigned i = 0; i < NumSucc; ++i) {
1173 if (OuterLoopHeaderBI->getSuccessor(i) == OuterLoopLatch)
1174 OuterLoopHeaderBI->setSuccessor(i, LoopExit);
1175 else if (OuterLoopHeaderBI->getSuccessor(i) == InnerLoopPreHeader)
1176 OuterLoopHeaderBI->setSuccessor(i, InnerLoopHeaderSuccessor);
1177 }
1178
1179 // Adjust reduction PHI's now that the incoming block has changed.
1180 updateIncomingBlock(InnerLoopHeaderSuccessor, InnerLoopHeader,
1181 OuterLoopHeader);
1182
1183 BranchInst::Create(OuterLoopPreHeader, InnerLoopHeaderBI);
1184 InnerLoopHeaderBI->eraseFromParent();
1185
1186 // -------------Adjust loop latches-----------
1187 if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1188 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1189 else
1190 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1191
1192 NumSucc = InnerLoopLatchPredecessorBI->getNumSuccessors();
1193 for (unsigned i = 0; i < NumSucc; ++i) {
1194 if (InnerLoopLatchPredecessorBI->getSuccessor(i) == InnerLoopLatch)
1195 InnerLoopLatchPredecessorBI->setSuccessor(i, InnerLoopLatchSuccessor);
1196 }
1197
1198 // Adjust PHI nodes in InnerLoopLatchSuccessor. Update all uses of PHI with
1199 // the value and remove this PHI node from inner loop.
1200 SmallVector<PHINode *, 8> LcssaVec;
1201 for (auto I = InnerLoopLatchSuccessor->begin(); isa<PHINode>(I); ++I) {
1202 PHINode *LcssaPhi = cast<PHINode>(I);
1203 LcssaVec.push_back(LcssaPhi);
1204 }
1205 for (PHINode *P : LcssaVec) {
1206 Value *Incoming = P->getIncomingValueForBlock(InnerLoopLatch);
1207 P->replaceAllUsesWith(Incoming);
1208 P->eraseFromParent();
1209 }
1210
1211 if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1212 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1213 else
1214 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1215
1216 if (InnerLoopLatchBI->getSuccessor(1) == InnerLoopLatchSuccessor)
1217 InnerLoopLatchBI->setSuccessor(1, OuterLoopLatchSuccessor);
1218 else
1219 InnerLoopLatchBI->setSuccessor(0, OuterLoopLatchSuccessor);
1220
1221 updateIncomingBlock(OuterLoopLatchSuccessor, OuterLoopLatch, InnerLoopLatch);
1222
1223 if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopLatchSuccessor) {
1224 OuterLoopLatchBI->setSuccessor(0, InnerLoopLatch);
1225 } else {
1226 OuterLoopLatchBI->setSuccessor(1, InnerLoopLatch);
1227 }
1228
1229 return true;
1230}
1231void LoopInterchangeTransform::adjustLoopPreheaders() {
1232
1233 // We have interchanged the preheaders so we need to interchange the data in
1234 // the preheader as well.
1235 // This is because the content of inner preheader was previously executed
1236 // inside the outer loop.
1237 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1238 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1239 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1240 BranchInst *InnerTermBI =
1241 cast<BranchInst>(InnerLoopPreHeader->getTerminator());
1242
1243 // These instructions should now be executed inside the loop.
1244 // Move instruction into a new block after outer header.
1245 moveBBContents(InnerLoopPreHeader, OuterLoopHeader->getTerminator());
1246 // These instructions were not executed previously in the loop so move them to
1247 // the older inner loop preheader.
1248 moveBBContents(OuterLoopPreHeader, InnerTermBI);
1249}
1250
1251bool LoopInterchangeTransform::adjustLoopLinks() {
1252
1253 // Adjust all branches in the inner and outer loop.
1254 bool Changed = adjustLoopBranches();
1255 if (Changed)
1256 adjustLoopPreheaders();
1257 return Changed;
1258}
1259
1260char LoopInterchange::ID = 0;
1261INITIALIZE_PASS_BEGIN(LoopInterchange, "loop-interchange",static void *initializeLoopInterchangePassOnce(PassRegistry &
Registry) {
1262 "Interchanges loops for cache reuse", false, false)static void *initializeLoopInterchangePassOnce(PassRegistry &
Registry) {
1263INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
1264INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)initializeDependenceAnalysisWrapperPassPass(Registry);
1265INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
1266INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
1267INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
1268INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)initializeLCSSAWrapperPassPass(Registry);
1269INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
1270
1271INITIALIZE_PASS_END(LoopInterchange, "loop-interchange",PassInfo *PI = new PassInfo( "Interchanges loops for cache reuse"
, "loop-interchange", &LoopInterchange::ID, PassInfo::NormalCtor_t
(callDefaultCtor<LoopInterchange>), false, false); Registry
.registerPass(*PI, true); return PI; } static llvm::once_flag
InitializeLoopInterchangePassFlag; void llvm::initializeLoopInterchangePass
(PassRegistry &Registry) { llvm::call_once(InitializeLoopInterchangePassFlag
, initializeLoopInterchangePassOnce, std::ref(Registry)); }
1272 "Interchanges loops for cache reuse", false, false)PassInfo *PI = new PassInfo( "Interchanges loops for cache reuse"
, "loop-interchange", &LoopInterchange::ID, PassInfo::NormalCtor_t
(callDefaultCtor<LoopInterchange>), false, false); Registry
.registerPass(*PI, true); return PI; } static llvm::once_flag
InitializeLoopInterchangePassFlag; void llvm::initializeLoopInterchangePass
(PassRegistry &Registry) { llvm::call_once(InitializeLoopInterchangePassFlag
, initializeLoopInterchangePassOnce, std::ref(Registry)); }
1273
1274Pass *llvm::createLoopInterchangePass() { return new LoopInterchange(); }