LLVM  15.0.0git
LoopRotationUtils.cpp
Go to the documentation of this file.
1 //===----------------- LoopRotationUtils.cpp -----------------------------===//
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 file provides utilities to convert a loop into a loop with bottom test.
10 //
11 //===----------------------------------------------------------------------===//
12 
14 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/IR/CFG.h"
25 #include "llvm/IR/DebugInfo.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/Support/Debug.h"
36 using namespace llvm;
37 
38 #define DEBUG_TYPE "loop-rotate"
39 
40 STATISTIC(NumNotRotatedDueToHeaderSize,
41  "Number of loops not rotated due to the header size");
42 STATISTIC(NumInstrsHoisted,
43  "Number of instructions hoisted into loop preheader");
44 STATISTIC(NumInstrsDuplicated,
45  "Number of instructions cloned into loop preheader");
46 STATISTIC(NumRotated, "Number of loops rotated");
47 
48 static cl::opt<bool>
49  MultiRotate("loop-rotate-multi", cl::init(false), cl::Hidden,
50  cl::desc("Allow loop rotation multiple times in order to reach "
51  "a better latch exit"));
52 
53 namespace {
54 /// A simple loop rotation transformation.
55 class LoopRotate {
56  const unsigned MaxHeaderSize;
57  LoopInfo *LI;
58  const TargetTransformInfo *TTI;
59  AssumptionCache *AC;
60  DominatorTree *DT;
61  ScalarEvolution *SE;
62  MemorySSAUpdater *MSSAU;
63  const SimplifyQuery &SQ;
64  bool RotationOnly;
65  bool IsUtilMode;
66  bool PrepareForLTO;
67 
68 public:
69  LoopRotate(unsigned MaxHeaderSize, LoopInfo *LI,
72  const SimplifyQuery &SQ, bool RotationOnly, bool IsUtilMode,
73  bool PrepareForLTO)
74  : MaxHeaderSize(MaxHeaderSize), LI(LI), TTI(TTI), AC(AC), DT(DT), SE(SE),
75  MSSAU(MSSAU), SQ(SQ), RotationOnly(RotationOnly),
76  IsUtilMode(IsUtilMode), PrepareForLTO(PrepareForLTO) {}
77  bool processLoop(Loop *L);
78 
79 private:
80  bool rotateLoop(Loop *L, bool SimplifiedLatch);
81  bool simplifyLoopLatch(Loop *L);
82 };
83 } // end anonymous namespace
84 
85 /// Insert (K, V) pair into the ValueToValueMap, and verify the key did not
86 /// previously exist in the map, and the value was inserted.
88  bool Inserted = VM.insert({K, V}).second;
89  assert(Inserted);
90  (void)Inserted;
91 }
92 /// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
93 /// old header into the preheader. If there were uses of the values produced by
94 /// these instruction that were outside of the loop, we have to insert PHI nodes
95 /// to merge the two values. Do this now.
97  BasicBlock *OrigPreheader,
99  ScalarEvolution *SE,
100  SmallVectorImpl<PHINode*> *InsertedPHIs) {
101  // Remove PHI node entries that are no longer live.
102  BasicBlock::iterator I, E = OrigHeader->end();
103  for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
104  PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
105 
106  // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
107  // as necessary.
108  SSAUpdater SSA(InsertedPHIs);
109  for (I = OrigHeader->begin(); I != E; ++I) {
110  Value *OrigHeaderVal = &*I;
111 
112  // If there are no uses of the value (e.g. because it returns void), there
113  // is nothing to rewrite.
114  if (OrigHeaderVal->use_empty())
115  continue;
116 
117  Value *OrigPreHeaderVal = ValueMap.lookup(OrigHeaderVal);
118 
119  // The value now exits in two versions: the initial value in the preheader
120  // and the loop "next" value in the original header.
121  SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
122  // Force re-computation of OrigHeaderVal, as some users now need to use the
123  // new PHI node.
124  if (SE)
125  SE->forgetValue(OrigHeaderVal);
126  SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
127  SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal);
128 
129  // Visit each use of the OrigHeader instruction.
130  for (Use &U : llvm::make_early_inc_range(OrigHeaderVal->uses())) {
131  // SSAUpdater can't handle a non-PHI use in the same block as an
132  // earlier def. We can easily handle those cases manually.
133  Instruction *UserInst = cast<Instruction>(U.getUser());
134  if (!isa<PHINode>(UserInst)) {
135  BasicBlock *UserBB = UserInst->getParent();
136 
137  // The original users in the OrigHeader are already using the
138  // original definitions.
139  if (UserBB == OrigHeader)
140  continue;
141 
142  // Users in the OrigPreHeader need to use the value to which the
143  // original definitions are mapped.
144  if (UserBB == OrigPreheader) {
145  U = OrigPreHeaderVal;
146  continue;
147  }
148  }
149 
150  // Anything else can be handled by SSAUpdater.
151  SSA.RewriteUse(U);
152  }
153 
154  // Replace MetadataAsValue(ValueAsMetadata(OrigHeaderVal)) uses in debug
155  // intrinsics.
157  llvm::findDbgValues(DbgValues, OrigHeaderVal);
158  for (auto &DbgValue : DbgValues) {
159  // The original users in the OrigHeader are already using the original
160  // definitions.
161  BasicBlock *UserBB = DbgValue->getParent();
162  if (UserBB == OrigHeader)
163  continue;
164 
165  // Users in the OrigPreHeader need to use the value to which the
166  // original definitions are mapped and anything else can be handled by
167  // the SSAUpdater. To avoid adding PHINodes, check if the value is
168  // available in UserBB, if not substitute undef.
169  Value *NewVal;
170  if (UserBB == OrigPreheader)
171  NewVal = OrigPreHeaderVal;
172  else if (SSA.HasValueForBlock(UserBB))
173  NewVal = SSA.GetValueInMiddleOfBlock(UserBB);
174  else
175  NewVal = UndefValue::get(OrigHeaderVal->getType());
176  DbgValue->replaceVariableLocationOp(OrigHeaderVal, NewVal);
177  }
178  }
179 }
180 
181 // Assuming both header and latch are exiting, look for a phi which is only
182 // used outside the loop (via a LCSSA phi) in the exit from the header.
183 // This means that rotating the loop can remove the phi.
185  BasicBlock *Header = L->getHeader();
186  BranchInst *BI = dyn_cast<BranchInst>(Header->getTerminator());
187  assert(BI && BI->isConditional() && "need header with conditional exit");
188  BasicBlock *HeaderExit = BI->getSuccessor(0);
189  if (L->contains(HeaderExit))
190  HeaderExit = BI->getSuccessor(1);
191 
192  for (auto &Phi : Header->phis()) {
193  // Look for uses of this phi in the loop/via exits other than the header.
194  if (llvm::any_of(Phi.users(), [HeaderExit](const User *U) {
195  return cast<Instruction>(U)->getParent() != HeaderExit;
196  }))
197  continue;
198  return true;
199  }
200  return false;
201 }
202 
203 // Check that latch exit is deoptimizing (which means - very unlikely to happen)
204 // and there is another exit from the loop which is non-deoptimizing.
205 // If we rotate latch to that exit our loop has a better chance of being fully
206 // canonical.
207 //
208 // It can give false positives in some rare cases.
210  BasicBlock *Latch = L->getLoopLatch();
211  assert(Latch && "need latch");
212  BranchInst *BI = dyn_cast<BranchInst>(Latch->getTerminator());
213  // Need normal exiting latch.
214  if (!BI || !BI->isConditional())
215  return false;
216 
217  BasicBlock *Exit = BI->getSuccessor(1);
218  if (L->contains(Exit))
219  Exit = BI->getSuccessor(0);
220 
221  // Latch exit is non-deoptimizing, no need to rotate.
222  if (!Exit->getPostdominatingDeoptimizeCall())
223  return false;
224 
226  L->getUniqueExitBlocks(Exits);
227  if (!Exits.empty()) {
228  // There is at least one non-deoptimizing exit.
229  //
230  // Note, that BasicBlock::getPostdominatingDeoptimizeCall is not exact,
231  // as it can conservatively return false for deoptimizing exits with
232  // complex enough control flow down to deoptimize call.
233  //
234  // That means here we can report success for a case where
235  // all exits are deoptimizing but one of them has complex enough
236  // control flow (e.g. with loops).
237  //
238  // That should be a very rare case and false positives for this function
239  // have compile-time effect only.
240  return any_of(Exits, [](const BasicBlock *BB) {
241  return !BB->getPostdominatingDeoptimizeCall();
242  });
243  }
244  return false;
245 }
246 
247 /// Rotate loop LP. Return true if the loop is rotated.
248 ///
249 /// \param SimplifiedLatch is true if the latch was just folded into the final
250 /// loop exit. In this case we may want to rotate even though the new latch is
251 /// now an exiting branch. This rotation would have happened had the latch not
252 /// been simplified. However, if SimplifiedLatch is false, then we avoid
253 /// rotating loops in which the latch exits to avoid excessive or endless
254 /// rotation. LoopRotate should be repeatable and converge to a canonical
255 /// form. This property is satisfied because simplifying the loop latch can only
256 /// happen once across multiple invocations of the LoopRotate pass.
257 ///
258 /// If -loop-rotate-multi is enabled we can do multiple rotations in one go
259 /// so to reach a suitable (non-deoptimizing) exit.
260 bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
261  // If the loop has only one block then there is not much to rotate.
262  if (L->getBlocks().size() == 1)
263  return false;
264 
265  bool Rotated = false;
266  do {
267  BasicBlock *OrigHeader = L->getHeader();
268  BasicBlock *OrigLatch = L->getLoopLatch();
269 
270  BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
271  if (!BI || BI->isUnconditional())
272  return Rotated;
273 
274  // If the loop header is not one of the loop exiting blocks then
275  // either this loop is already rotated or it is not
276  // suitable for loop rotation transformations.
277  if (!L->isLoopExiting(OrigHeader))
278  return Rotated;
279 
280  // If the loop latch already contains a branch that leaves the loop then the
281  // loop is already rotated.
282  if (!OrigLatch)
283  return Rotated;
284 
285  // Rotate if either the loop latch does *not* exit the loop, or if the loop
286  // latch was just simplified. Or if we think it will be profitable.
287  if (L->isLoopExiting(OrigLatch) && !SimplifiedLatch && IsUtilMode == false &&
290  return Rotated;
291 
292  // Check size of original header and reject loop if it is very big or we can't
293  // duplicate blocks inside it.
294  {
296  CodeMetrics::collectEphemeralValues(L, AC, EphValues);
297 
299  Metrics.analyzeBasicBlock(OrigHeader, *TTI, EphValues, PrepareForLTO);
300  if (Metrics.notDuplicatable) {
301  LLVM_DEBUG(
302  dbgs() << "LoopRotation: NOT rotating - contains non-duplicatable"
303  << " instructions: ";
304  L->dump());
305  return Rotated;
306  }
307  if (Metrics.convergent) {
308  LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains convergent "
309  "instructions: ";
310  L->dump());
311  return Rotated;
312  }
313  if (Metrics.NumInsts > MaxHeaderSize) {
314  LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains "
315  << Metrics.NumInsts
316  << " instructions, which is more than the threshold ("
317  << MaxHeaderSize << " instructions): ";
318  L->dump());
319  ++NumNotRotatedDueToHeaderSize;
320  return Rotated;
321  }
322 
323  // When preparing for LTO, avoid rotating loops with calls that could be
324  // inlined during the LTO stage.
325  if (PrepareForLTO && Metrics.NumInlineCandidates > 0)
326  return Rotated;
327  }
328 
329  // Now, this loop is suitable for rotation.
330  BasicBlock *OrigPreheader = L->getLoopPreheader();
331 
332  // If the loop could not be converted to canonical form, it must have an
333  // indirectbr in it, just give up.
334  if (!OrigPreheader || !L->hasDedicatedExits())
335  return Rotated;
336 
337  // Anything ScalarEvolution may know about this loop or the PHI nodes
338  // in its header will soon be invalidated. We should also invalidate
339  // all outer loops because insertion and deletion of blocks that happens
340  // during the rotation may violate invariants related to backedge taken
341  // infos in them.
342  if (SE)
343  SE->forgetTopmostLoop(L);
344 
345  LLVM_DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
346  if (MSSAU && VerifyMemorySSA)
347  MSSAU->getMemorySSA()->verifyMemorySSA();
348 
349  // Find new Loop header. NewHeader is a Header's one and only successor
350  // that is inside loop. Header's other successor is outside the
351  // loop. Otherwise loop is not suitable for rotation.
352  BasicBlock *Exit = BI->getSuccessor(0);
353  BasicBlock *NewHeader = BI->getSuccessor(1);
354  if (L->contains(Exit))
355  std::swap(Exit, NewHeader);
356  assert(NewHeader && "Unable to determine new loop header");
357  assert(L->contains(NewHeader) && !L->contains(Exit) &&
358  "Unable to determine loop header and exit blocks");
359 
360  // This code assumes that the new header has exactly one predecessor.
361  // Remove any single-entry PHI nodes in it.
362  assert(NewHeader->getSinglePredecessor() &&
363  "New header doesn't have one pred!");
364  FoldSingleEntryPHINodes(NewHeader);
365 
366  // Begin by walking OrigHeader and populating ValueMap with an entry for
367  // each Instruction.
368  BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
369  ValueToValueMapTy ValueMap, ValueMapMSSA;
370 
371  // For PHI nodes, the value available in OldPreHeader is just the
372  // incoming value from OldPreHeader.
373  for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
375  PN->getIncomingValueForBlock(OrigPreheader));
376 
377  // For the rest of the instructions, either hoist to the OrigPreheader if
378  // possible or create a clone in the OldPreHeader if not.
379  Instruction *LoopEntryBranch = OrigPreheader->getTerminator();
380 
381  // Record all debug intrinsics preceding LoopEntryBranch to avoid
382  // duplication.
383  using DbgIntrinsicHash =
384  std::pair<std::pair<hash_code, DILocalVariable *>, DIExpression *>;
385  auto makeHash = [](DbgVariableIntrinsic *D) -> DbgIntrinsicHash {
386  auto VarLocOps = D->location_ops();
387  return {{hash_combine_range(VarLocOps.begin(), VarLocOps.end()),
388  D->getVariable()},
389  D->getExpression()};
390  };
392  for (Instruction &I : llvm::drop_begin(llvm::reverse(*OrigPreheader))) {
393  if (auto *DII = dyn_cast<DbgVariableIntrinsic>(&I))
394  DbgIntrinsics.insert(makeHash(DII));
395  else
396  break;
397  }
398 
399  // Remember the local noalias scope declarations in the header. After the
400  // rotation, they must be duplicated and the scope must be cloned. This
401  // avoids unwanted interaction across iterations.
402  SmallVector<NoAliasScopeDeclInst *, 6> NoAliasDeclInstructions;
403  for (Instruction &I : *OrigHeader)
404  if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
405  NoAliasDeclInstructions.push_back(Decl);
406 
407  while (I != E) {
408  Instruction *Inst = &*I++;
409 
410  // If the instruction's operands are invariant and it doesn't read or write
411  // memory, then it is safe to hoist. Doing this doesn't change the order of
412  // execution in the preheader, but does prevent the instruction from
413  // executing in each iteration of the loop. This means it is safe to hoist
414  // something that might trap, but isn't safe to hoist something that reads
415  // memory (without proving that the loop doesn't write).
416  if (L->hasLoopInvariantOperands(Inst) && !Inst->mayReadFromMemory() &&
417  !Inst->mayWriteToMemory() && !Inst->isTerminator() &&
418  !isa<DbgInfoIntrinsic>(Inst) && !isa<AllocaInst>(Inst)) {
419  Inst->moveBefore(LoopEntryBranch);
420  ++NumInstrsHoisted;
421  continue;
422  }
423 
424  // Otherwise, create a duplicate of the instruction.
425  Instruction *C = Inst->clone();
426  ++NumInstrsDuplicated;
427 
428  // Eagerly remap the operands of the instruction.
431 
432  // Avoid inserting the same intrinsic twice.
433  if (auto *DII = dyn_cast<DbgVariableIntrinsic>(C))
434  if (DbgIntrinsics.count(makeHash(DII))) {
435  C->deleteValue();
436  continue;
437  }
438 
439  // With the operands remapped, see if the instruction constant folds or is
440  // otherwise simplifyable. This commonly occurs because the entry from PHI
441  // nodes allows icmps and other instructions to fold.
442  Value *V = SimplifyInstruction(C, SQ);
443  if (V && LI->replacementPreservesLCSSAForm(C, V)) {
444  // If so, then delete the temporary instruction and stick the folded value
445  // in the map.
447  if (!C->mayHaveSideEffects()) {
448  C->deleteValue();
449  C = nullptr;
450  }
451  } else {
453  }
454  if (C) {
455  // Otherwise, stick the new instruction into the new block!
456  C->setName(Inst->getName());
457  C->insertBefore(LoopEntryBranch);
458 
459  if (auto *II = dyn_cast<AssumeInst>(C))
460  AC->registerAssumption(II);
461  // MemorySSA cares whether the cloned instruction was inserted or not, and
462  // not whether it can be remapped to a simplified value.
463  if (MSSAU)
464  InsertNewValueIntoMap(ValueMapMSSA, Inst, C);
465  }
466  }
467 
468  if (!NoAliasDeclInstructions.empty()) {
469  // There are noalias scope declarations:
470  // (general):
471  // Original: OrigPre { OrigHeader NewHeader ... Latch }
472  // after: (OrigPre+OrigHeader') { NewHeader ... Latch OrigHeader }
473  //
474  // with D: llvm.experimental.noalias.scope.decl,
475  // U: !noalias or !alias.scope depending on D
476  // ... { D U1 U2 } can transform into:
477  // (0) : ... { D U1 U2 } // no relevant rotation for this part
478  // (1) : ... D' { U1 U2 D } // D is part of OrigHeader
479  // (2) : ... D' U1' { U2 D U1 } // D, U1 are part of OrigHeader
480  //
481  // We now want to transform:
482  // (1) -> : ... D' { D U1 U2 D'' }
483  // (2) -> : ... D' U1' { D U2 D'' U1'' }
484  // D: original llvm.experimental.noalias.scope.decl
485  // D', U1': duplicate with replaced scopes
486  // D'', U1'': different duplicate with replaced scopes
487  // This ensures a safe fallback to 'may_alias' introduced by the rotate,
488  // as U1'' and U1' scopes will not be compatible wrt to the local restrict
489 
490  // Clone the llvm.experimental.noalias.decl again for the NewHeader.
491  Instruction *NewHeaderInsertionPoint = &(*NewHeader->getFirstNonPHI());
492  for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions) {
493  LLVM_DEBUG(dbgs() << " Cloning llvm.experimental.noalias.scope.decl:"
494  << *NAD << "\n");
495  Instruction *NewNAD = NAD->clone();
496  NewNAD->insertBefore(NewHeaderInsertionPoint);
497  }
498 
499  // Scopes must now be duplicated, once for OrigHeader and once for
500  // OrigPreHeader'.
501  {
502  auto &Context = NewHeader->getContext();
503 
504  SmallVector<MDNode *, 8> NoAliasDeclScopes;
505  for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions)
506  NoAliasDeclScopes.push_back(NAD->getScopeList());
507 
508  LLVM_DEBUG(dbgs() << " Updating OrigHeader scopes\n");
509  cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, {OrigHeader}, Context,
510  "h.rot");
511  LLVM_DEBUG(OrigHeader->dump());
512 
513  // Keep the compile time impact low by only adapting the inserted block
514  // of instructions in the OrigPreHeader. This might result in slightly
515  // more aliasing between these instructions and those that were already
516  // present, but it will be much faster when the original PreHeader is
517  // large.
518  LLVM_DEBUG(dbgs() << " Updating part of OrigPreheader scopes\n");
519  auto *FirstDecl =
520  cast<Instruction>(ValueMap[*NoAliasDeclInstructions.begin()]);
521  auto *LastInst = &OrigPreheader->back();
522  cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, FirstDecl, LastInst,
523  Context, "pre.rot");
524  LLVM_DEBUG(OrigPreheader->dump());
525 
526  LLVM_DEBUG(dbgs() << " Updated NewHeader:\n");
527  LLVM_DEBUG(NewHeader->dump());
528  }
529  }
530 
531  // Along with all the other instructions, we just cloned OrigHeader's
532  // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
533  // successors by duplicating their incoming values for OrigHeader.
534  for (BasicBlock *SuccBB : successors(OrigHeader))
535  for (BasicBlock::iterator BI = SuccBB->begin();
536  PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
537  PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
538 
539  // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
540  // OrigPreHeader's old terminator (the original branch into the loop), and
541  // remove the corresponding incoming values from the PHI nodes in OrigHeader.
542  LoopEntryBranch->eraseFromParent();
543 
544  // Update MemorySSA before the rewrite call below changes the 1:1
545  // instruction:cloned_instruction_or_value mapping.
546  if (MSSAU) {
547  InsertNewValueIntoMap(ValueMapMSSA, OrigHeader, OrigPreheader);
548  MSSAU->updateForClonedBlockIntoPred(OrigHeader, OrigPreheader,
549  ValueMapMSSA);
550  }
551 
552  SmallVector<PHINode*, 2> InsertedPHIs;
553  // If there were any uses of instructions in the duplicated block outside the
554  // loop, update them, inserting PHI nodes as required
555  RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap, SE,
556  &InsertedPHIs);
557 
558  // Attach dbg.value intrinsics to the new phis if that phi uses a value that
559  // previously had debug metadata attached. This keeps the debug info
560  // up-to-date in the loop body.
561  if (!InsertedPHIs.empty())
562  insertDebugValuesForPHIs(OrigHeader, InsertedPHIs);
563 
564  // NewHeader is now the header of the loop.
565  L->moveToHeader(NewHeader);
566  assert(L->getHeader() == NewHeader && "Latch block is our new header");
567 
568  // Inform DT about changes to the CFG.
569  if (DT) {
570  // The OrigPreheader branches to the NewHeader and Exit now. Then, inform
571  // the DT about the removed edge to the OrigHeader (that got removed).
573  Updates.push_back({DominatorTree::Insert, OrigPreheader, Exit});
574  Updates.push_back({DominatorTree::Insert, OrigPreheader, NewHeader});
575  Updates.push_back({DominatorTree::Delete, OrigPreheader, OrigHeader});
576 
577  if (MSSAU) {
578  MSSAU->applyUpdates(Updates, *DT, /*UpdateDT=*/true);
579  if (VerifyMemorySSA)
580  MSSAU->getMemorySSA()->verifyMemorySSA();
581  } else {
582  DT->applyUpdates(Updates);
583  }
584  }
585 
586  // At this point, we've finished our major CFG changes. As part of cloning
587  // the loop into the preheader we've simplified instructions and the
588  // duplicated conditional branch may now be branching on a constant. If it is
589  // branching on a constant and if that constant means that we enter the loop,
590  // then we fold away the cond branch to an uncond branch. This simplifies the
591  // loop in cases important for nested loops, and it also means we don't have
592  // to split as many edges.
593  BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
594  assert(PHBI->isConditional() && "Should be clone of BI condbr!");
595  if (!isa<ConstantInt>(PHBI->getCondition()) ||
596  PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero()) !=
597  NewHeader) {
598  // The conditional branch can't be folded, handle the general case.
599  // Split edges as necessary to preserve LoopSimplify form.
600 
601  // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
602  // thus is not a preheader anymore.
603  // Split the edge to form a real preheader.
604  BasicBlock *NewPH = SplitCriticalEdge(
605  OrigPreheader, NewHeader,
606  CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
607  NewPH->setName(NewHeader->getName() + ".lr.ph");
608 
609  // Preserve canonical loop form, which means that 'Exit' should have only
610  // one predecessor. Note that Exit could be an exit block for multiple
611  // nested loops, causing both of the edges to now be critical and need to
612  // be split.
614  bool SplitLatchEdge = false;
615  for (BasicBlock *ExitPred : ExitPreds) {
616  // We only need to split loop exit edges.
617  Loop *PredLoop = LI->getLoopFor(ExitPred);
618  if (!PredLoop || PredLoop->contains(Exit) ||
619  ExitPred->getTerminator()->isIndirectTerminator())
620  continue;
621  SplitLatchEdge |= L->getLoopLatch() == ExitPred;
622  BasicBlock *ExitSplit = SplitCriticalEdge(
623  ExitPred, Exit,
624  CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
625  ExitSplit->moveBefore(Exit);
626  }
627  assert(SplitLatchEdge &&
628  "Despite splitting all preds, failed to split latch exit?");
629  (void)SplitLatchEdge;
630  } else {
631  // We can fold the conditional branch in the preheader, this makes things
632  // simpler. The first step is to remove the extra edge to the Exit block.
633  Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
634  BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI);
635  NewBI->setDebugLoc(PHBI->getDebugLoc());
636  PHBI->eraseFromParent();
637 
638  // With our CFG finalized, update DomTree if it is available.
639  if (DT) DT->deleteEdge(OrigPreheader, Exit);
640 
641  // Update MSSA too, if available.
642  if (MSSAU)
643  MSSAU->removeEdge(OrigPreheader, Exit);
644  }
645 
646  assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
647  assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
648 
649  if (MSSAU && VerifyMemorySSA)
650  MSSAU->getMemorySSA()->verifyMemorySSA();
651 
652  // Now that the CFG and DomTree are in a consistent state again, try to merge
653  // the OrigHeader block into OrigLatch. This will succeed if they are
654  // connected by an unconditional branch. This is just a cleanup so the
655  // emitted code isn't too gross in this common case.
657  BasicBlock *PredBB = OrigHeader->getUniquePredecessor();
658  bool DidMerge = MergeBlockIntoPredecessor(OrigHeader, &DTU, LI, MSSAU);
659  if (DidMerge)
660  RemoveRedundantDbgInstrs(PredBB);
661 
662  if (MSSAU && VerifyMemorySSA)
663  MSSAU->getMemorySSA()->verifyMemorySSA();
664 
665  LLVM_DEBUG(dbgs() << "LoopRotation: into "; L->dump());
666 
667  ++NumRotated;
668 
669  Rotated = true;
670  SimplifiedLatch = false;
671 
672  // Check that new latch is a deoptimizing exit and then repeat rotation if possible.
673  // Deoptimizing latch exit is not a generally typical case, so we just loop over.
674  // TODO: if it becomes a performance bottleneck extend rotation algorithm
675  // to handle multiple rotations in one go.
677 
678 
679  return true;
680 }
681 
682 /// Determine whether the instructions in this range may be safely and cheaply
683 /// speculated. This is not an important enough situation to develop complex
684 /// heuristics. We handle a single arithmetic instruction along with any type
685 /// conversions.
687  BasicBlock::iterator End, Loop *L) {
688  bool seenIncrement = false;
689  bool MultiExitLoop = false;
690 
691  if (!L->getExitingBlock())
692  MultiExitLoop = true;
693 
694  for (BasicBlock::iterator I = Begin; I != End; ++I) {
695 
697  return false;
698 
699  if (isa<DbgInfoIntrinsic>(I))
700  continue;
701 
702  switch (I->getOpcode()) {
703  default:
704  return false;
705  case Instruction::GetElementPtr:
706  // GEPs are cheap if all indices are constant.
707  if (!cast<GEPOperator>(I)->hasAllConstantIndices())
708  return false;
709  // fall-thru to increment case
711  case Instruction::Add:
712  case Instruction::Sub:
713  case Instruction::And:
714  case Instruction::Or:
715  case Instruction::Xor:
716  case Instruction::Shl:
717  case Instruction::LShr:
718  case Instruction::AShr: {
719  Value *IVOpnd =
720  !isa<Constant>(I->getOperand(0))
721  ? I->getOperand(0)
722  : !isa<Constant>(I->getOperand(1)) ? I->getOperand(1) : nullptr;
723  if (!IVOpnd)
724  return false;
725 
726  // If increment operand is used outside of the loop, this speculation
727  // could cause extra live range interference.
728  if (MultiExitLoop) {
729  for (User *UseI : IVOpnd->users()) {
730  auto *UserInst = cast<Instruction>(UseI);
731  if (!L->contains(UserInst))
732  return false;
733  }
734  }
735 
736  if (seenIncrement)
737  return false;
738  seenIncrement = true;
739  break;
740  }
741  case Instruction::Trunc:
742  case Instruction::ZExt:
743  case Instruction::SExt:
744  // ignore type conversions
745  break;
746  }
747  }
748  return true;
749 }
750 
751 /// Fold the loop tail into the loop exit by speculating the loop tail
752 /// instructions. Typically, this is a single post-increment. In the case of a
753 /// simple 2-block loop, hoisting the increment can be much better than
754 /// duplicating the entire loop header. In the case of loops with early exits,
755 /// rotation will not work anyway, but simplifyLoopLatch will put the loop in
756 /// canonical form so downstream passes can handle it.
757 ///
758 /// I don't believe this invalidates SCEV.
759 bool LoopRotate::simplifyLoopLatch(Loop *L) {
760  BasicBlock *Latch = L->getLoopLatch();
761  if (!Latch || Latch->hasAddressTaken())
762  return false;
763 
764  BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
765  if (!Jmp || !Jmp->isUnconditional())
766  return false;
767 
768  BasicBlock *LastExit = Latch->getSinglePredecessor();
769  if (!LastExit || !L->isLoopExiting(LastExit))
770  return false;
771 
772  BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
773  if (!BI)
774  return false;
775 
776  if (!shouldSpeculateInstrs(Latch->begin(), Jmp->getIterator(), L))
777  return false;
778 
779  LLVM_DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
780  << LastExit->getName() << "\n");
781 
783  MergeBlockIntoPredecessor(Latch, &DTU, LI, MSSAU, nullptr,
784  /*PredecessorWithTwoSuccessors=*/true);
785 
786  if (MSSAU && VerifyMemorySSA)
787  MSSAU->getMemorySSA()->verifyMemorySSA();
788 
789  return true;
790 }
791 
792 /// Rotate \c L, and return true if any modification was made.
793 bool LoopRotate::processLoop(Loop *L) {
794  // Save the loop metadata.
795  MDNode *LoopMD = L->getLoopID();
796 
797  bool SimplifiedLatch = false;
798 
799  // Simplify the loop latch before attempting to rotate the header
800  // upward. Rotation may not be needed if the loop tail can be folded into the
801  // loop exit.
802  if (!RotationOnly)
803  SimplifiedLatch = simplifyLoopLatch(L);
804 
805  bool MadeChange = rotateLoop(L, SimplifiedLatch);
806  assert((!MadeChange || L->isLoopExiting(L->getLoopLatch())) &&
807  "Loop latch should be exiting after loop-rotate.");
808 
809  // Restore the loop metadata.
810  // NB! We presume LoopRotation DOESN'T ADD its own metadata.
811  if ((MadeChange || SimplifiedLatch) && LoopMD)
812  L->setLoopID(LoopMD);
813 
814  return MadeChange || SimplifiedLatch;
815 }
816 
817 
818 /// The utility to convert a loop into a loop with bottom test.
821  ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
822  const SimplifyQuery &SQ, bool RotationOnly = true,
823  unsigned Threshold = unsigned(-1),
824  bool IsUtilMode = true, bool PrepareForLTO) {
825  LoopRotate LR(Threshold, LI, TTI, AC, DT, SE, MSSAU, SQ, RotationOnly,
826  IsUtilMode, PrepareForLTO);
827  return LR.processLoop(L);
828 }
llvm::NoAliasScopeDeclInst
Definition: IntrinsicInst.h:1318
llvm::Instruction::isTerminator
bool isTerminator() const
Definition: Instruction.h:160
AssumptionCache.h
llvm::BasicBlock::end
iterator end()
Definition: BasicBlock.h:299
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:17
llvm::drop_begin
auto drop_begin(T &&RangeOrContainer, size_t N=1)
Return a range covering RangeOrContainer with the first N elements excluded.
Definition: STLExtras.h:280
ValueMapper.h
llvm::BasicBlock::iterator
InstListType::iterator iterator
Instruction iterators...
Definition: BasicBlock.h:87
llvm::isSafeToSpeculativelyExecute
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, const TargetLibraryInfo *TLI=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
Definition: ValueTracking.cpp:4576
IntrinsicInst.h
llvm::SimplifyQuery
Definition: InstructionSimplify.h:93
MemorySSAUpdater.h
shouldSpeculateInstrs
static bool shouldSpeculateInstrs(BasicBlock::iterator Begin, BasicBlock::iterator End, Loop *L)
Determine whether the instructions in this range may be safely and cheaply speculated.
Definition: LoopRotationUtils.cpp:686
llvm::Loop
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:530
llvm::Value::dump
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:4823
llvm::LoopBase::contains
bool contains(const LoopT *L) const
Return true if the specified loop is contained within in this loop.
Definition: LoopInfo.h:122
llvm::SmallVector
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1185
Statistic.h
llvm::CodeMetrics
Utility to calculate the size and a few similar metrics for a set of basic blocks.
Definition: CodeMetrics.h:30
llvm::TargetTransformInfo
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
Definition: TargetTransformInfo.h:167
llvm::SmallDenseSet
Implements a dense probed hash-table based set with some number of buckets stored inline.
Definition: DenseSet.h:286
DomTreeUpdater.h
llvm::ScalarEvolution
The main scalar evolution driver.
Definition: ScalarEvolution.h:449
ValueTracking.h
Local.h
llvm::Instruction::insertBefore
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction.
Definition: Instruction.cpp:83
llvm::DominatorTree
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Definition: Dominators.h:166
profitableToRotateLoopExitingLatch
static bool profitableToRotateLoopExitingLatch(Loop *L)
Definition: LoopRotationUtils.cpp:184
llvm::cl::Hidden
@ Hidden
Definition: CommandLine.h:139
ScalarEvolution.h
llvm::SplitCriticalEdge
BasicBlock * SplitCriticalEdge(Instruction *TI, unsigned SuccNum, const CriticalEdgeSplittingOptions &Options=CriticalEdgeSplittingOptions(), const Twine &BBName="")
If this edge is a critical edge, insert a new node to split the critical edge.
Definition: BreakCriticalEdges.cpp:101
llvm::reverse
auto reverse(ContainerTy &&C, std::enable_if_t< has_rbegin< ContainerTy >::value > *=nullptr)
Definition: STLExtras.h:380
llvm::DominatorTreeBase< BasicBlock, false >::Insert
static constexpr UpdateKind Insert
Definition: GenericDomTree.h:242
llvm::SmallPtrSet
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements.
Definition: SmallPtrSet.h:450
llvm::successors
auto successors(MachineBasicBlock *BB)
Definition: MachineSSAContext.h:29
llvm::BasicBlock::getSinglePredecessor
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:261
llvm::DIExpression
DWARF expression.
Definition: DebugInfoMetadata.h:2548
llvm::detail::DenseSetImpl< ValueT, SmallDenseMap< ValueT, detail::DenseSetEmpty, 4, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::insert
std::pair< iterator, bool > insert(const ValueT &V)
Definition: DenseSet.h:206
llvm::detail::DenseSetImpl< ValueT, SmallDenseMap< ValueT, detail::DenseSetEmpty, 4, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::count
size_type count(const_arg_type_t< ValueT > V) const
Return 1 if the specified key is in the set, 0 otherwise.
Definition: DenseSet.h:97
LLVM_DEBUG
#define LLVM_DEBUG(X)
Definition: Debug.h:101
MultiRotate
static cl::opt< bool > MultiRotate("loop-rotate-multi", cl::init(false), cl::Hidden, cl::desc("Allow loop rotation multiple times in order to reach " "a better latch exit"))
llvm::BasicBlock
LLVM Basic Block Representation.
Definition: BasicBlock.h:55
Context
LLVMContext & Context
Definition: NVVMIntrRange.cpp:66
llvm::dbgs
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
CommandLine.h
CodeMetrics.h
llvm::RemapInstruction
void RemapInstruction(Instruction *I, ValueToValueMapTy &VM, RemapFlags Flags=RF_None, ValueMapTypeRemapper *TypeMapper=nullptr, ValueMaterializer *Materializer=nullptr)
Convert the instruction operands from referencing the current values into those specified by VM.
Definition: ValueMapper.h:254
E
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
llvm::User
Definition: User.h:44
C
(vector float) vec_cmpeq(*A, *B) C
Definition: README_ALTIVEC.txt:86
llvm::BasicBlock::begin
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:297
llvm::CriticalEdgeSplittingOptions
Option class for critical edge splitting.
Definition: BasicBlockUtils.h:142
llvm::Value::uses
iterator_range< use_iterator > uses()
Definition: Value.h:376
llvm::PHINode::getIncomingValueForBlock
Value * getIncomingValueForBlock(const BasicBlock *BB) const
Definition: Instructions.h:2849
llvm::LoopBase::getBlocks
ArrayRef< BlockT * > getBlocks() const
Get a list of the basic blocks which make up this loop.
Definition: LoopInfo.h:171
llvm::Instruction
Definition: Instruction.h:42
llvm::SimplifyInstruction
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
Definition: InstructionSimplify.cpp:6460
llvm::BasicBlock::phis
iterator_range< const_phi_iterator > phis() const
Returns a range that iterates over the phis in the basic block.
Definition: BasicBlock.h:355
llvm::STATISTIC
STATISTIC(NumFunctions, "Total number of functions")
llvm::predecessors
auto predecessors(MachineBasicBlock *BB)
Definition: MachineSSAContext.h:30
llvm::Value::setName
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:372
llvm::UndefValue::get
static UndefValue * get(Type *T)
Static factory methods - Return an 'undef' object of the specified type.
Definition: Constants.cpp:1769
llvm::DomTreeUpdater
Definition: DomTreeUpdater.h:28
llvm::CodeMetrics::collectEphemeralValues
static void collectEphemeralValues(const Loop *L, AssumptionCache *AC, SmallPtrSetImpl< const Value * > &EphValues)
Collect a loop's ephemeral values (those used only by an assume or similar intrinsics in the loop).
Definition: CodeMetrics.cpp:70
llvm::LoopBase::getExitingBlock
BlockT * getExitingBlock() const
If getExitingBlocks would return exactly one block, return that block.
Definition: LoopInfoImpl.h:48
llvm::BasicBlock::getFirstNonPHI
const Instruction * getFirstNonPHI() const
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
Definition: BasicBlock.cpp:209
llvm::RemoveRedundantDbgInstrs
bool RemoveRedundantDbgInstrs(BasicBlock *BB)
Try to remove redundant dbg.value instructions from given basic block.
Definition: BasicBlockUtils.cpp:441
llvm::Instruction::mayWriteToMemory
bool mayWriteToMemory() const
Return true if this instruction may modify memory.
Definition: Instruction.cpp:596
llvm::RF_NoModuleLevelChanges
@ RF_NoModuleLevelChanges
If this flag is set, the remapper knows that only local values within a function (such as an instruct...
Definition: ValueMapper.h:71
llvm::Value::use_empty
bool use_empty() const
Definition: Value.h:344
llvm::BasicBlock::getPostdominatingDeoptimizeCall
const CallInst * getPostdominatingDeoptimizeCall() const
Returns the call instruction calling @llvm.experimental.deoptimize that is present either in current ...
Definition: BasicBlock.cpp:197
CFG.h
LoopInfo.h
llvm::BranchInst::getCondition
Value * getCondition() const
Definition: Instructions.h:3180
llvm::cloneAndAdaptNoAliasScopes
void cloneAndAdaptNoAliasScopes(ArrayRef< MDNode * > NoAliasDeclScopes, ArrayRef< BasicBlock * > NewBlocks, LLVMContext &Context, StringRef Ext)
Clone the specified noalias decl scopes.
Definition: CloneFunction.cpp:1086
llvm::BasicBlock::hasAddressTaken
bool hasAddressTaken() const
Returns true if there are any uses of this basic block other than direct branches,...
Definition: BasicBlock.h:449
llvm::cl::opt< bool >
llvm::DbgVariableIntrinsic
This is the common base class for debug info intrinsics for variables.
Definition: IntrinsicInst.h:149
SSA
Memory SSA
Definition: MemorySSA.cpp:73
LiveDebugValues::DbgValue
Class recording the (high level) value of a variable.
Definition: InstrRefBasedImpl.h:229
llvm::Instruction::eraseFromParent
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks 'this' from the containing basic block and deletes it.
Definition: Instruction.cpp:77
llvm::LoopBase::moveToHeader
void moveToHeader(BlockT *BB)
This method is used to move BB (which must be part of this loop) to be the loop header of the loop (t...
Definition: LoopInfo.h:443
D
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
llvm::MemorySSAUpdater
Definition: MemorySSAUpdater.h:54
llvm::LoopBase::getUniqueExitBlocks
void getUniqueExitBlocks(SmallVectorImpl< BlockT * > &ExitBlocks) const
Return all unique successor blocks of this loop.
Definition: LoopInfoImpl.h:121
llvm::PHINode::addIncoming
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
Definition: Instructions.h:2814
llvm::BranchInst::Create
static BranchInst * Create(BasicBlock *IfTrue, Instruction *InsertBefore=nullptr)
Definition: Instructions.h:3155
DebugInfo.h
I
#define I(x, y, z)
Definition: MD5.cpp:58
Cloning.h
llvm::cl::init
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:432
llvm::make_early_inc_range
iterator_range< early_inc_iterator_impl< detail::IterOfRange< RangeT > > > make_early_inc_range(RangeT &&Range)
Make a range that does early increment to allow mutation of the underlying range without disrupting i...
Definition: STLExtras.h:608
Metrics
Machine Trace Metrics
Definition: MachineTraceMetrics.cpp:53
llvm::LoopBase::getLoopPreheader
BlockT * getLoopPreheader() const
If there is a preheader for this loop, return it.
Definition: LoopInfoImpl.h:166
RewriteUsesOfClonedInstructions
static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader, BasicBlock *OrigPreheader, ValueToValueMapTy &ValueMap, ScalarEvolution *SE, SmallVectorImpl< PHINode * > *InsertedPHIs)
RewriteUsesOfClonedInstructions - We just cloned the instructions from the old header into the prehea...
Definition: LoopRotationUtils.cpp:96
llvm::insertDebugValuesForPHIs
void insertDebugValuesForPHIs(BasicBlock *BB, SmallVectorImpl< PHINode * > &InsertedPHIs)
Propagate dbg.value intrinsics through the newly inserted PHIs.
Definition: Local.cpp:1625
llvm::findDbgValues
void findDbgValues(SmallVectorImpl< DbgValueInst * > &DbgValues, Value *V)
Finds the llvm.dbg.value intrinsics describing a value.
Definition: DebugInfo.cpp:75
llvm::Instruction::setDebugLoc
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:364
llvm::ScalarEvolution::forgetValue
void forgetValue(Value *V)
This method should be called by the client when it has changed a value in a way that may effect its v...
Definition: ScalarEvolution.cpp:8099
llvm::LoopBase::getLoopLatch
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
Definition: LoopInfoImpl.h:215
llvm::LoopBase::hasDedicatedExits
bool hasDedicatedExits() const
Return true if no exit block for the loop has a predecessor that is outside the loop.
Definition: LoopInfoImpl.h:91
assert
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
std::swap
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:853
llvm::ValueMap::insert
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: ValueMap.h:173
llvm::TTI
TargetTransformInfo TTI
Definition: TargetTransformInfo.h:162
llvm::MDNode
Metadata node.
Definition: Metadata.h:926
canRotateDeoptimizingLatchExit
static bool canRotateDeoptimizingLatchExit(Loop *L)
Definition: LoopRotationUtils.cpp:209
llvm::BranchInst::isUnconditional
bool isUnconditional() const
Definition: Instructions.h:3177
llvm::Instruction::clone
Instruction * clone() const
Create a copy of 'this' instruction that is identical in all ways except the following:
Definition: Instruction.cpp:862
llvm::LoopInfo
Definition: LoopInfo.h:1086
llvm::MergeBlockIntoPredecessor
bool MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, MemoryDependenceResults *MemDep=nullptr, bool PredecessorWithTwoSuccessors=false)
Attempts to merge a block into its predecessor, if possible.
Definition: BasicBlockUtils.cpp:178
llvm::any_of
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1614
llvm::AssumptionCache
A cache of @llvm.assume calls within a function.
Definition: AssumptionCache.h:42
llvm::Value::getType
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
LoopRotationUtils.h
llvm::BasicBlock::getUniquePredecessor
const BasicBlock * getUniquePredecessor() const
Return the predecessor of this block if it has a unique predecessor block.
Definition: BasicBlock.cpp:269
llvm::ilist_node_impl::getIterator
self_iterator getIterator()
Definition: ilist_node.h:82
llvm::DomTreeUpdater::UpdateStrategy::Eager
@ Eager
SSAUpdater.h
llvm::Instruction::mayReadFromMemory
bool mayReadFromMemory() const
Return true if this instruction may read memory.
Definition: Instruction.cpp:576
llvm::Loop::hasLoopInvariantOperands
bool hasLoopInvariantOperands(const Instruction *I) const
Return true if all the operands of the specified instruction are loop invariant.
Definition: LoopInfo.cpp:66
LLVM_FALLTHROUGH
#define LLVM_FALLTHROUGH
LLVM_FALLTHROUGH - Mark fallthrough cases in switch statements.
Definition: Compiler.h:280
llvm::Value::getName
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:305
llvm::ValueMap< const Value *, WeakTrackingVH >
llvm::RF_IgnoreMissingLocals
@ RF_IgnoreMissingLocals
If this flag is set, the remapper ignores missing function-local entries (Argument,...
Definition: ValueMapper.h:89
llvm::BasicBlock::getContext
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:35
llvm::Loop::setLoopID
void setLoopID(MDNode *LoopID) const
Set the llvm.loop loop id metadata for this loop.
Definition: LoopInfo.cpp:524
llvm::Loop::getLoopID
MDNode * getLoopID() const
Return the llvm.loop loop id metadata node for this loop if it is present.
Definition: LoopInfo.cpp:500
llvm::LoopRotation
bool LoopRotation(Loop *L, LoopInfo *LI, const TargetTransformInfo *TTI, AssumptionCache *AC, DominatorTree *DT, ScalarEvolution *SE, MemorySSAUpdater *MSSAU, const SimplifyQuery &SQ, bool RotationOnly, unsigned Threshold, bool IsUtilMode, bool PrepareForLTO=false)
Convert a loop into a loop with bottom test.
Definition: LoopRotationUtils.cpp:819
llvm::LoopBase::getHeader
BlockT * getHeader() const
Definition: LoopInfo.h:104
llvm::Loop::dump
void dump() const
Definition: LoopInfo.cpp:666
llvm::MCID::Add
@ Add
Definition: MCInstrDesc.h:185
llvm::FoldSingleEntryPHINodes
bool FoldSingleEntryPHINodes(BasicBlock *BB, MemoryDependenceResults *MemDep=nullptr)
We know that BB has one predecessor.
Definition: BasicBlockUtils.cpp:143
llvm::BasicBlock::back
const Instruction & back() const
Definition: BasicBlock.h:311
MemorySSA.h
llvm::BasicBlock::moveBefore
void moveBefore(BasicBlock *MovePos)
Unlink this basic block from its current function and insert it into the function that MovePos lives ...
Definition: BasicBlock.cpp:136
llvm::Instruction::getDebugLoc
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:367
Dominators.h
llvm::hash_combine_range
hash_code hash_combine_range(InputIteratorT first, InputIteratorT last)
Compute a hash_code for a sequence of values.
Definition: Hashing.h:483
llvm::Instruction::getParent
const BasicBlock * getParent() const
Definition: Instruction.h:91
InstructionSimplify.h
llvm::PHINode
Definition: Instructions.h:2664
llvm::BasicBlock::getTerminator
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.h:119
llvm::BasicBlock::removePredecessor
void removePredecessor(BasicBlock *Pred, bool KeepOneInputPHIs=false)
Update PHI nodes in this BasicBlock before removal of predecessor Pred.
Definition: BasicBlock.cpp:318
llvm::SmallVectorImpl
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
InsertNewValueIntoMap
static void InsertNewValueIntoMap(ValueToValueMapTy &VM, Value *K, Value *V)
Insert (K, V) pair into the ValueToValueMap, and verify the key did not previously exist in the map,...
Definition: LoopRotationUtils.cpp:87
llvm::LoopBase::isLoopExiting
bool isLoopExiting(const BlockT *BB) const
True if terminator in the block can branch to another block that is outside of the current loop.
Definition: LoopInfo.h:225
BB
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM BB
Definition: README.txt:39
llvm::ValueMap::lookup
ValueT lookup(const KeyT &Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: ValueMap.h:165
llvm::VerifyMemorySSA
bool VerifyMemorySSA
Enables verification of MemorySSA.
Definition: MemorySSA.cpp:91
llvm::cl::desc
Definition: CommandLine.h:405
llvm::BranchInst
Conditional or Unconditional Branch instruction.
Definition: Instructions.h:3099
raw_ostream.h
llvm::SSAUpdater
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:38
BasicBlockUtils.h
llvm::Value
LLVM Value Representation.
Definition: Value.h:74
Debug.h
llvm::Value::users
iterator_range< user_iterator > users()
Definition: Value.h:421
llvm::BranchInst::isConditional
bool isConditional() const
Definition: Instructions.h:3178
llvm::BranchInst::getSuccessor
BasicBlock * getSuccessor(unsigned i) const
Definition: Instructions.h:3192
llvm::DominatorTreeBase< BasicBlock, false >::Delete
static constexpr UpdateKind Delete
Definition: GenericDomTree.h:243
llvm::Instruction::moveBefore
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:96
llvm::Use
A Use represents the edge between a Value definition and its users.
Definition: Use.h:43