source: icGREP/icgrep-devel/llvm-3.6.1.src/lib/Transforms/Scalar/LoopRotation.cpp @ 4664

Last change on this file since 4664 was 4664, checked in by cameron, 4 years ago

Upgrade LLVM to 3.6.1

File size: 23.2 KB
Line 
1//===- LoopRotation.cpp - Loop Rotation 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 file implements Loop Rotation Pass.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Transforms/Scalar.h"
15#include "llvm/ADT/Statistic.h"
16#include "llvm/Analysis/AssumptionCache.h"
17#include "llvm/Analysis/CodeMetrics.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/LoopPass.h"
20#include "llvm/Analysis/ScalarEvolution.h"
21#include "llvm/Analysis/TargetTransformInfo.h"
22#include "llvm/Analysis/ValueTracking.h"
23#include "llvm/IR/CFG.h"
24#include "llvm/IR/Dominators.h"
25#include "llvm/IR/Function.h"
26#include "llvm/IR/IntrinsicInst.h"
27#include "llvm/Support/CommandLine.h"
28#include "llvm/Support/Debug.h"
29#include "llvm/Transforms/Utils/BasicBlockUtils.h"
30#include "llvm/Transforms/Utils/Local.h"
31#include "llvm/Transforms/Utils/SSAUpdater.h"
32#include "llvm/Transforms/Utils/ValueMapper.h"
33using namespace llvm;
34
35#define DEBUG_TYPE "loop-rotate"
36
37static cl::opt<unsigned>
38DefaultRotationThreshold("rotation-max-header-size", cl::init(16), cl::Hidden,
39       cl::desc("The default maximum header size for automatic loop rotation"));
40
41STATISTIC(NumRotated, "Number of loops rotated");
42namespace {
43
44  class LoopRotate : public LoopPass {
45  public:
46    static char ID; // Pass ID, replacement for typeid
47    LoopRotate(int SpecifiedMaxHeaderSize = -1) : LoopPass(ID) {
48      initializeLoopRotatePass(*PassRegistry::getPassRegistry());
49      if (SpecifiedMaxHeaderSize == -1)
50        MaxHeaderSize = DefaultRotationThreshold;
51      else
52        MaxHeaderSize = unsigned(SpecifiedMaxHeaderSize);
53    }
54
55    // LCSSA form makes instruction renaming easier.
56    void getAnalysisUsage(AnalysisUsage &AU) const override {
57      AU.addRequired<AssumptionCacheTracker>();
58      AU.addPreserved<DominatorTreeWrapperPass>();
59      AU.addRequired<LoopInfo>();
60      AU.addPreserved<LoopInfo>();
61      AU.addRequiredID(LoopSimplifyID);
62      AU.addPreservedID(LoopSimplifyID);
63      AU.addRequiredID(LCSSAID);
64      AU.addPreservedID(LCSSAID);
65      AU.addPreserved<ScalarEvolution>();
66      AU.addRequired<TargetTransformInfo>();
67    }
68
69    bool runOnLoop(Loop *L, LPPassManager &LPM) override;
70    bool simplifyLoopLatch(Loop *L);
71    bool rotateLoop(Loop *L, bool SimplifiedLatch);
72
73  private:
74    unsigned MaxHeaderSize;
75    LoopInfo *LI;
76    const TargetTransformInfo *TTI;
77    AssumptionCache *AC;
78  };
79}
80
81char LoopRotate::ID = 0;
82INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
83INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
84INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
85INITIALIZE_PASS_DEPENDENCY(LoopInfo)
86INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
87INITIALIZE_PASS_DEPENDENCY(LCSSA)
88INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
89
90Pass *llvm::createLoopRotatePass(int MaxHeaderSize) {
91  return new LoopRotate(MaxHeaderSize);
92}
93
94/// Rotate Loop L as many times as possible. Return true if
95/// the loop is rotated at least once.
96bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) {
97  if (skipOptnoneFunction(L))
98    return false;
99
100  // Save the loop metadata.
101  MDNode *LoopMD = L->getLoopID();
102
103  LI = &getAnalysis<LoopInfo>();
104  TTI = &getAnalysis<TargetTransformInfo>();
105  AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
106      *L->getHeader()->getParent());
107
108  // Simplify the loop latch before attempting to rotate the header
109  // upward. Rotation may not be needed if the loop tail can be folded into the
110  // loop exit.
111  bool SimplifiedLatch = simplifyLoopLatch(L);
112
113  // One loop can be rotated multiple times.
114  bool MadeChange = false;
115  while (rotateLoop(L, SimplifiedLatch)) {
116    MadeChange = true;
117    SimplifiedLatch = false;
118  }
119
120  // Restore the loop metadata.
121  // NB! We presume LoopRotation DOESN'T ADD its own metadata.
122  if ((MadeChange || SimplifiedLatch) && LoopMD)
123    L->setLoopID(LoopMD);
124
125  return MadeChange;
126}
127
128/// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
129/// old header into the preheader.  If there were uses of the values produced by
130/// these instruction that were outside of the loop, we have to insert PHI nodes
131/// to merge the two values.  Do this now.
132static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader,
133                                            BasicBlock *OrigPreheader,
134                                            ValueToValueMapTy &ValueMap) {
135  // Remove PHI node entries that are no longer live.
136  BasicBlock::iterator I, E = OrigHeader->end();
137  for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
138    PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
139
140  // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
141  // as necessary.
142  SSAUpdater SSA;
143  for (I = OrigHeader->begin(); I != E; ++I) {
144    Value *OrigHeaderVal = I;
145
146    // If there are no uses of the value (e.g. because it returns void), there
147    // is nothing to rewrite.
148    if (OrigHeaderVal->use_empty())
149      continue;
150
151    Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal];
152
153    // The value now exits in two versions: the initial value in the preheader
154    // and the loop "next" value in the original header.
155    SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
156    SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
157    SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal);
158
159    // Visit each use of the OrigHeader instruction.
160    for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
161         UE = OrigHeaderVal->use_end(); UI != UE; ) {
162      // Grab the use before incrementing the iterator.
163      Use &U = *UI;
164
165      // Increment the iterator before removing the use from the list.
166      ++UI;
167
168      // SSAUpdater can't handle a non-PHI use in the same block as an
169      // earlier def. We can easily handle those cases manually.
170      Instruction *UserInst = cast<Instruction>(U.getUser());
171      if (!isa<PHINode>(UserInst)) {
172        BasicBlock *UserBB = UserInst->getParent();
173
174        // The original users in the OrigHeader are already using the
175        // original definitions.
176        if (UserBB == OrigHeader)
177          continue;
178
179        // Users in the OrigPreHeader need to use the value to which the
180        // original definitions are mapped.
181        if (UserBB == OrigPreheader) {
182          U = OrigPreHeaderVal;
183          continue;
184        }
185      }
186
187      // Anything else can be handled by SSAUpdater.
188      SSA.RewriteUse(U);
189    }
190  }
191}
192
193/// Determine whether the instructions in this range may be safely and cheaply
194/// speculated. This is not an important enough situation to develop complex
195/// heuristics. We handle a single arithmetic instruction along with any type
196/// conversions.
197static bool shouldSpeculateInstrs(BasicBlock::iterator Begin,
198                                  BasicBlock::iterator End, Loop *L) {
199  bool seenIncrement = false;
200  bool MultiExitLoop = false;
201
202  if (!L->getExitingBlock())
203    MultiExitLoop = true;
204
205  for (BasicBlock::iterator I = Begin; I != End; ++I) {
206
207    if (!isSafeToSpeculativelyExecute(I))
208      return false;
209
210    if (isa<DbgInfoIntrinsic>(I))
211      continue;
212
213    switch (I->getOpcode()) {
214    default:
215      return false;
216    case Instruction::GetElementPtr:
217      // GEPs are cheap if all indices are constant.
218      if (!cast<GEPOperator>(I)->hasAllConstantIndices())
219        return false;
220      // fall-thru to increment case
221    case Instruction::Add:
222    case Instruction::Sub:
223    case Instruction::And:
224    case Instruction::Or:
225    case Instruction::Xor:
226    case Instruction::Shl:
227    case Instruction::LShr:
228    case Instruction::AShr: {
229      Value *IVOpnd = nullptr;
230      if (isa<ConstantInt>(I->getOperand(0)))
231        IVOpnd = I->getOperand(1);
232
233      if (isa<ConstantInt>(I->getOperand(1))) {
234        if (IVOpnd)
235          return false;
236
237        IVOpnd = I->getOperand(0);
238      }
239
240      // If increment operand is used outside of the loop, this speculation
241      // could cause extra live range interference.
242      if (MultiExitLoop && IVOpnd) {
243        for (User *UseI : IVOpnd->users()) {
244          auto *UserInst = cast<Instruction>(UseI);
245          if (!L->contains(UserInst))
246            return false;
247        }
248      }
249
250      if (seenIncrement)
251        return false;
252      seenIncrement = true;
253      break;
254    }
255    case Instruction::Trunc:
256    case Instruction::ZExt:
257    case Instruction::SExt:
258      // ignore type conversions
259      break;
260    }
261  }
262  return true;
263}
264
265/// Fold the loop tail into the loop exit by speculating the loop tail
266/// instructions. Typically, this is a single post-increment. In the case of a
267/// simple 2-block loop, hoisting the increment can be much better than
268/// duplicating the entire loop header. In the case of loops with early exits,
269/// rotation will not work anyway, but simplifyLoopLatch will put the loop in
270/// canonical form so downstream passes can handle it.
271///
272/// I don't believe this invalidates SCEV.
273bool LoopRotate::simplifyLoopLatch(Loop *L) {
274  BasicBlock *Latch = L->getLoopLatch();
275  if (!Latch || Latch->hasAddressTaken())
276    return false;
277
278  BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
279  if (!Jmp || !Jmp->isUnconditional())
280    return false;
281
282  BasicBlock *LastExit = Latch->getSinglePredecessor();
283  if (!LastExit || !L->isLoopExiting(LastExit))
284    return false;
285
286  BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
287  if (!BI)
288    return false;
289
290  if (!shouldSpeculateInstrs(Latch->begin(), Jmp, L))
291    return false;
292
293  DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
294        << LastExit->getName() << "\n");
295
296  // Hoist the instructions from Latch into LastExit.
297  LastExit->getInstList().splice(BI, Latch->getInstList(), Latch->begin(), Jmp);
298
299  unsigned FallThruPath = BI->getSuccessor(0) == Latch ? 0 : 1;
300  BasicBlock *Header = Jmp->getSuccessor(0);
301  assert(Header == L->getHeader() && "expected a backward branch");
302
303  // Remove Latch from the CFG so that LastExit becomes the new Latch.
304  BI->setSuccessor(FallThruPath, Header);
305  Latch->replaceSuccessorsPhiUsesWith(LastExit);
306  Jmp->eraseFromParent();
307
308  // Nuke the Latch block.
309  assert(Latch->empty() && "unable to evacuate Latch");
310  LI->removeBlock(Latch);
311  if (DominatorTreeWrapperPass *DTWP =
312          getAnalysisIfAvailable<DominatorTreeWrapperPass>())
313    DTWP->getDomTree().eraseNode(Latch);
314  Latch->eraseFromParent();
315  return true;
316}
317
318/// Rotate loop LP. Return true if the loop is rotated.
319///
320/// \param SimplifiedLatch is true if the latch was just folded into the final
321/// loop exit. In this case we may want to rotate even though the new latch is
322/// now an exiting branch. This rotation would have happened had the latch not
323/// been simplified. However, if SimplifiedLatch is false, then we avoid
324/// rotating loops in which the latch exits to avoid excessive or endless
325/// rotation. LoopRotate should be repeatable and converge to a canonical
326/// form. This property is satisfied because simplifying the loop latch can only
327/// happen once across multiple invocations of the LoopRotate pass.
328bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
329  // If the loop has only one block then there is not much to rotate.
330  if (L->getBlocks().size() == 1)
331    return false;
332
333  BasicBlock *OrigHeader = L->getHeader();
334  BasicBlock *OrigLatch = L->getLoopLatch();
335
336  BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
337  if (!BI || BI->isUnconditional())
338    return false;
339
340  // If the loop header is not one of the loop exiting blocks then
341  // either this loop is already rotated or it is not
342  // suitable for loop rotation transformations.
343  if (!L->isLoopExiting(OrigHeader))
344    return false;
345
346  // If the loop latch already contains a branch that leaves the loop then the
347  // loop is already rotated.
348  if (!OrigLatch)
349    return false;
350
351  // Rotate if either the loop latch does *not* exit the loop, or if the loop
352  // latch was just simplified.
353  if (L->isLoopExiting(OrigLatch) && !SimplifiedLatch)
354    return false;
355
356  // Check size of original header and reject loop if it is very big or we can't
357  // duplicate blocks inside it.
358  {
359    SmallPtrSet<const Value *, 32> EphValues;
360    CodeMetrics::collectEphemeralValues(L, AC, EphValues);
361
362    CodeMetrics Metrics;
363    Metrics.analyzeBasicBlock(OrigHeader, *TTI, EphValues);
364    if (Metrics.notDuplicatable) {
365      DEBUG(dbgs() << "LoopRotation: NOT rotating - contains non-duplicatable"
366            << " instructions: "; L->dump());
367      return false;
368    }
369    if (Metrics.NumInsts > MaxHeaderSize)
370      return false;
371  }
372
373  // Now, this loop is suitable for rotation.
374  BasicBlock *OrigPreheader = L->getLoopPreheader();
375
376  // If the loop could not be converted to canonical form, it must have an
377  // indirectbr in it, just give up.
378  if (!OrigPreheader)
379    return false;
380
381  // Anything ScalarEvolution may know about this loop or the PHI nodes
382  // in its header will soon be invalidated.
383  if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
384    SE->forgetLoop(L);
385
386  DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
387
388  // Find new Loop header. NewHeader is a Header's one and only successor
389  // that is inside loop.  Header's other successor is outside the
390  // loop.  Otherwise loop is not suitable for rotation.
391  BasicBlock *Exit = BI->getSuccessor(0);
392  BasicBlock *NewHeader = BI->getSuccessor(1);
393  if (L->contains(Exit))
394    std::swap(Exit, NewHeader);
395  assert(NewHeader && "Unable to determine new loop header");
396  assert(L->contains(NewHeader) && !L->contains(Exit) &&
397         "Unable to determine loop header and exit blocks");
398
399  // This code assumes that the new header has exactly one predecessor.
400  // Remove any single-entry PHI nodes in it.
401  assert(NewHeader->getSinglePredecessor() &&
402         "New header doesn't have one pred!");
403  FoldSingleEntryPHINodes(NewHeader);
404
405  // Begin by walking OrigHeader and populating ValueMap with an entry for
406  // each Instruction.
407  BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
408  ValueToValueMapTy ValueMap;
409
410  // For PHI nodes, the value available in OldPreHeader is just the
411  // incoming value from OldPreHeader.
412  for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
413    ValueMap[PN] = PN->getIncomingValueForBlock(OrigPreheader);
414
415  // For the rest of the instructions, either hoist to the OrigPreheader if
416  // possible or create a clone in the OldPreHeader if not.
417  TerminatorInst *LoopEntryBranch = OrigPreheader->getTerminator();
418  while (I != E) {
419    Instruction *Inst = I++;
420
421    // If the instruction's operands are invariant and it doesn't read or write
422    // memory, then it is safe to hoist.  Doing this doesn't change the order of
423    // execution in the preheader, but does prevent the instruction from
424    // executing in each iteration of the loop.  This means it is safe to hoist
425    // something that might trap, but isn't safe to hoist something that reads
426    // memory (without proving that the loop doesn't write).
427    if (L->hasLoopInvariantOperands(Inst) &&
428        !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() &&
429        !isa<TerminatorInst>(Inst) && !isa<DbgInfoIntrinsic>(Inst) &&
430        !isa<AllocaInst>(Inst)) {
431      Inst->moveBefore(LoopEntryBranch);
432      continue;
433    }
434
435    // Otherwise, create a duplicate of the instruction.
436    Instruction *C = Inst->clone();
437
438    // Eagerly remap the operands of the instruction.
439    RemapInstruction(C, ValueMap,
440                     RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
441
442    // With the operands remapped, see if the instruction constant folds or is
443    // otherwise simplifyable.  This commonly occurs because the entry from PHI
444    // nodes allows icmps and other instructions to fold.
445    // FIXME: Provide DL, TLI, DT, AC to SimplifyInstruction.
446    Value *V = SimplifyInstruction(C);
447    if (V && LI->replacementPreservesLCSSAForm(C, V)) {
448      // If so, then delete the temporary instruction and stick the folded value
449      // in the map.
450      delete C;
451      ValueMap[Inst] = V;
452    } else {
453      // Otherwise, stick the new instruction into the new block!
454      C->setName(Inst->getName());
455      C->insertBefore(LoopEntryBranch);
456      ValueMap[Inst] = C;
457    }
458  }
459
460  // Along with all the other instructions, we just cloned OrigHeader's
461  // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
462  // successors by duplicating their incoming values for OrigHeader.
463  TerminatorInst *TI = OrigHeader->getTerminator();
464  for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
465    for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin();
466         PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
467      PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
468
469  // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
470  // OrigPreHeader's old terminator (the original branch into the loop), and
471  // remove the corresponding incoming values from the PHI nodes in OrigHeader.
472  LoopEntryBranch->eraseFromParent();
473
474  // If there were any uses of instructions in the duplicated block outside the
475  // loop, update them, inserting PHI nodes as required
476  RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap);
477
478  // NewHeader is now the header of the loop.
479  L->moveToHeader(NewHeader);
480  assert(L->getHeader() == NewHeader && "Latch block is our new header");
481
482
483  // At this point, we've finished our major CFG changes.  As part of cloning
484  // the loop into the preheader we've simplified instructions and the
485  // duplicated conditional branch may now be branching on a constant.  If it is
486  // branching on a constant and if that constant means that we enter the loop,
487  // then we fold away the cond branch to an uncond branch.  This simplifies the
488  // loop in cases important for nested loops, and it also means we don't have
489  // to split as many edges.
490  BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
491  assert(PHBI->isConditional() && "Should be clone of BI condbr!");
492  if (!isa<ConstantInt>(PHBI->getCondition()) ||
493      PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero())
494          != NewHeader) {
495    // The conditional branch can't be folded, handle the general case.
496    // Update DominatorTree to reflect the CFG change we just made.  Then split
497    // edges as necessary to preserve LoopSimplify form.
498    if (DominatorTreeWrapperPass *DTWP =
499            getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
500      DominatorTree &DT = DTWP->getDomTree();
501      // Everything that was dominated by the old loop header is now dominated
502      // by the original loop preheader. Conceptually the header was merged
503      // into the preheader, even though we reuse the actual block as a new
504      // loop latch.
505      DomTreeNode *OrigHeaderNode = DT.getNode(OrigHeader);
506      SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
507                                                   OrigHeaderNode->end());
508      DomTreeNode *OrigPreheaderNode = DT.getNode(OrigPreheader);
509      for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I)
510        DT.changeImmediateDominator(HeaderChildren[I], OrigPreheaderNode);
511
512      assert(DT.getNode(Exit)->getIDom() == OrigPreheaderNode);
513      assert(DT.getNode(NewHeader)->getIDom() == OrigPreheaderNode);
514
515      // Update OrigHeader to be dominated by the new header block.
516      DT.changeImmediateDominator(OrigHeader, OrigLatch);
517    }
518
519    // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
520    // thus is not a preheader anymore.
521    // Split the edge to form a real preheader.
522    BasicBlock *NewPH = SplitCriticalEdge(OrigPreheader, NewHeader, this);
523    NewPH->setName(NewHeader->getName() + ".lr.ph");
524
525    // Preserve canonical loop form, which means that 'Exit' should have only
526    // one predecessor. Note that Exit could be an exit block for multiple
527    // nested loops, causing both of the edges to now be critical and need to
528    // be split.
529    SmallVector<BasicBlock *, 4> ExitPreds(pred_begin(Exit), pred_end(Exit));
530    bool SplitLatchEdge = false;
531    for (SmallVectorImpl<BasicBlock *>::iterator PI = ExitPreds.begin(),
532                                                 PE = ExitPreds.end();
533         PI != PE; ++PI) {
534      // We only need to split loop exit edges.
535      Loop *PredLoop = LI->getLoopFor(*PI);
536      if (!PredLoop || PredLoop->contains(Exit))
537        continue;
538      if (isa<IndirectBrInst>((*PI)->getTerminator()))
539        continue;
540      SplitLatchEdge |= L->getLoopLatch() == *PI;
541      BasicBlock *ExitSplit = SplitCriticalEdge(*PI, Exit, this);
542      ExitSplit->moveBefore(Exit);
543    }
544    assert(SplitLatchEdge &&
545           "Despite splitting all preds, failed to split latch exit?");
546  } else {
547    // We can fold the conditional branch in the preheader, this makes things
548    // simpler. The first step is to remove the extra edge to the Exit block.
549    Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
550    BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI);
551    NewBI->setDebugLoc(PHBI->getDebugLoc());
552    PHBI->eraseFromParent();
553
554    // With our CFG finalized, update DomTree if it is available.
555    if (DominatorTreeWrapperPass *DTWP =
556            getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
557      DominatorTree &DT = DTWP->getDomTree();
558      // Update OrigHeader to be dominated by the new header block.
559      DT.changeImmediateDominator(NewHeader, OrigPreheader);
560      DT.changeImmediateDominator(OrigHeader, OrigLatch);
561
562      // Brute force incremental dominator tree update. Call
563      // findNearestCommonDominator on all CFG predecessors of each child of the
564      // original header.
565      DomTreeNode *OrigHeaderNode = DT.getNode(OrigHeader);
566      SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
567                                                   OrigHeaderNode->end());
568      bool Changed;
569      do {
570        Changed = false;
571        for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) {
572          DomTreeNode *Node = HeaderChildren[I];
573          BasicBlock *BB = Node->getBlock();
574
575          pred_iterator PI = pred_begin(BB);
576          BasicBlock *NearestDom = *PI;
577          for (pred_iterator PE = pred_end(BB); PI != PE; ++PI)
578            NearestDom = DT.findNearestCommonDominator(NearestDom, *PI);
579
580          // Remember if this changes the DomTree.
581          if (Node->getIDom()->getBlock() != NearestDom) {
582            DT.changeImmediateDominator(BB, NearestDom);
583            Changed = true;
584          }
585        }
586
587      // If the dominator changed, this may have an effect on other
588      // predecessors, continue until we reach a fixpoint.
589      } while (Changed);
590    }
591  }
592
593  assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
594  assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
595
596  // Now that the CFG and DomTree are in a consistent state again, try to merge
597  // the OrigHeader block into OrigLatch.  This will succeed if they are
598  // connected by an unconditional branch.  This is just a cleanup so the
599  // emitted code isn't too gross in this common case.
600  MergeBlockIntoPredecessor(OrigHeader, this);
601
602  DEBUG(dbgs() << "LoopRotation: into "; L->dump());
603
604  ++NumRotated;
605  return true;
606}
Note: See TracBrowser for help on using the repository browser.