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.isValid()) {
314  LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains instructions"
315  " with invalid cost: ";
316  L->dump());
317  return Rotated;
318  }
319  if (*Metrics.NumInsts.getValue() > MaxHeaderSize) {
320  LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains "
321  << Metrics.NumInsts
322  << " instructions, which is more than the threshold ("
323  << MaxHeaderSize << " instructions): ";
324  L->dump());
325  ++NumNotRotatedDueToHeaderSize;
326  return Rotated;
327  }
328 
329  // When preparing for LTO, avoid rotating loops with calls that could be
330  // inlined during the LTO stage.
331  if (PrepareForLTO && Metrics.NumInlineCandidates > 0)
332  return Rotated;
333  }
334 
335  // Now, this loop is suitable for rotation.
336  BasicBlock *OrigPreheader = L->getLoopPreheader();
337 
338  // If the loop could not be converted to canonical form, it must have an
339  // indirectbr in it, just give up.
340  if (!OrigPreheader || !L->hasDedicatedExits())
341  return Rotated;
342 
343  // Anything ScalarEvolution may know about this loop or the PHI nodes
344  // in its header will soon be invalidated. We should also invalidate
345  // all outer loops because insertion and deletion of blocks that happens
346  // during the rotation may violate invariants related to backedge taken
347  // infos in them.
348  if (SE)
349  SE->forgetTopmostLoop(L);
350 
351  LLVM_DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
352  if (MSSAU && VerifyMemorySSA)
353  MSSAU->getMemorySSA()->verifyMemorySSA();
354 
355  // Find new Loop header. NewHeader is a Header's one and only successor
356  // that is inside loop. Header's other successor is outside the
357  // loop. Otherwise loop is not suitable for rotation.
358  BasicBlock *Exit = BI->getSuccessor(0);
359  BasicBlock *NewHeader = BI->getSuccessor(1);
360  if (L->contains(Exit))
361  std::swap(Exit, NewHeader);
362  assert(NewHeader && "Unable to determine new loop header");
363  assert(L->contains(NewHeader) && !L->contains(Exit) &&
364  "Unable to determine loop header and exit blocks");
365 
366  // This code assumes that the new header has exactly one predecessor.
367  // Remove any single-entry PHI nodes in it.
368  assert(NewHeader->getSinglePredecessor() &&
369  "New header doesn't have one pred!");
370  FoldSingleEntryPHINodes(NewHeader);
371 
372  // Begin by walking OrigHeader and populating ValueMap with an entry for
373  // each Instruction.
374  BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
375  ValueToValueMapTy ValueMap, ValueMapMSSA;
376 
377  // For PHI nodes, the value available in OldPreHeader is just the
378  // incoming value from OldPreHeader.
379  for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
381  PN->getIncomingValueForBlock(OrigPreheader));
382 
383  // For the rest of the instructions, either hoist to the OrigPreheader if
384  // possible or create a clone in the OldPreHeader if not.
385  Instruction *LoopEntryBranch = OrigPreheader->getTerminator();
386 
387  // Record all debug intrinsics preceding LoopEntryBranch to avoid
388  // duplication.
389  using DbgIntrinsicHash =
390  std::pair<std::pair<hash_code, DILocalVariable *>, DIExpression *>;
391  auto makeHash = [](DbgVariableIntrinsic *D) -> DbgIntrinsicHash {
392  auto VarLocOps = D->location_ops();
393  return {{hash_combine_range(VarLocOps.begin(), VarLocOps.end()),
394  D->getVariable()},
395  D->getExpression()};
396  };
398  for (Instruction &I : llvm::drop_begin(llvm::reverse(*OrigPreheader))) {
399  if (auto *DII = dyn_cast<DbgVariableIntrinsic>(&I))
400  DbgIntrinsics.insert(makeHash(DII));
401  else
402  break;
403  }
404 
405  // Remember the local noalias scope declarations in the header. After the
406  // rotation, they must be duplicated and the scope must be cloned. This
407  // avoids unwanted interaction across iterations.
408  SmallVector<NoAliasScopeDeclInst *, 6> NoAliasDeclInstructions;
409  for (Instruction &I : *OrigHeader)
410  if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
411  NoAliasDeclInstructions.push_back(Decl);
412 
413  while (I != E) {
414  Instruction *Inst = &*I++;
415 
416  // If the instruction's operands are invariant and it doesn't read or write
417  // memory, then it is safe to hoist. Doing this doesn't change the order of
418  // execution in the preheader, but does prevent the instruction from
419  // executing in each iteration of the loop. This means it is safe to hoist
420  // something that might trap, but isn't safe to hoist something that reads
421  // memory (without proving that the loop doesn't write).
422  if (L->hasLoopInvariantOperands(Inst) && !Inst->mayReadFromMemory() &&
423  !Inst->mayWriteToMemory() && !Inst->isTerminator() &&
424  !isa<DbgInfoIntrinsic>(Inst) && !isa<AllocaInst>(Inst)) {
425  Inst->moveBefore(LoopEntryBranch);
426  ++NumInstrsHoisted;
427  continue;
428  }
429 
430  // Otherwise, create a duplicate of the instruction.
431  Instruction *C = Inst->clone();
432  ++NumInstrsDuplicated;
433 
434  // Eagerly remap the operands of the instruction.
437 
438  // Avoid inserting the same intrinsic twice.
439  if (auto *DII = dyn_cast<DbgVariableIntrinsic>(C))
440  if (DbgIntrinsics.count(makeHash(DII))) {
441  C->deleteValue();
442  continue;
443  }
444 
445  // With the operands remapped, see if the instruction constant folds or is
446  // otherwise simplifyable. This commonly occurs because the entry from PHI
447  // nodes allows icmps and other instructions to fold.
448  Value *V = simplifyInstruction(C, SQ);
449  if (V && LI->replacementPreservesLCSSAForm(C, V)) {
450  // If so, then delete the temporary instruction and stick the folded value
451  // in the map.
453  if (!C->mayHaveSideEffects()) {
454  C->deleteValue();
455  C = nullptr;
456  }
457  } else {
459  }
460  if (C) {
461  // Otherwise, stick the new instruction into the new block!
462  C->setName(Inst->getName());
463  C->insertBefore(LoopEntryBranch);
464 
465  if (auto *II = dyn_cast<AssumeInst>(C))
466  AC->registerAssumption(II);
467  // MemorySSA cares whether the cloned instruction was inserted or not, and
468  // not whether it can be remapped to a simplified value.
469  if (MSSAU)
470  InsertNewValueIntoMap(ValueMapMSSA, Inst, C);
471  }
472  }
473 
474  if (!NoAliasDeclInstructions.empty()) {
475  // There are noalias scope declarations:
476  // (general):
477  // Original: OrigPre { OrigHeader NewHeader ... Latch }
478  // after: (OrigPre+OrigHeader') { NewHeader ... Latch OrigHeader }
479  //
480  // with D: llvm.experimental.noalias.scope.decl,
481  // U: !noalias or !alias.scope depending on D
482  // ... { D U1 U2 } can transform into:
483  // (0) : ... { D U1 U2 } // no relevant rotation for this part
484  // (1) : ... D' { U1 U2 D } // D is part of OrigHeader
485  // (2) : ... D' U1' { U2 D U1 } // D, U1 are part of OrigHeader
486  //
487  // We now want to transform:
488  // (1) -> : ... D' { D U1 U2 D'' }
489  // (2) -> : ... D' U1' { D U2 D'' U1'' }
490  // D: original llvm.experimental.noalias.scope.decl
491  // D', U1': duplicate with replaced scopes
492  // D'', U1'': different duplicate with replaced scopes
493  // This ensures a safe fallback to 'may_alias' introduced by the rotate,
494  // as U1'' and U1' scopes will not be compatible wrt to the local restrict
495 
496  // Clone the llvm.experimental.noalias.decl again for the NewHeader.
497  Instruction *NewHeaderInsertionPoint = &(*NewHeader->getFirstNonPHI());
498  for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions) {
499  LLVM_DEBUG(dbgs() << " Cloning llvm.experimental.noalias.scope.decl:"
500  << *NAD << "\n");
501  Instruction *NewNAD = NAD->clone();
502  NewNAD->insertBefore(NewHeaderInsertionPoint);
503  }
504 
505  // Scopes must now be duplicated, once for OrigHeader and once for
506  // OrigPreHeader'.
507  {
508  auto &Context = NewHeader->getContext();
509 
510  SmallVector<MDNode *, 8> NoAliasDeclScopes;
511  for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions)
512  NoAliasDeclScopes.push_back(NAD->getScopeList());
513 
514  LLVM_DEBUG(dbgs() << " Updating OrigHeader scopes\n");
515  cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, {OrigHeader}, Context,
516  "h.rot");
517  LLVM_DEBUG(OrigHeader->dump());
518 
519  // Keep the compile time impact low by only adapting the inserted block
520  // of instructions in the OrigPreHeader. This might result in slightly
521  // more aliasing between these instructions and those that were already
522  // present, but it will be much faster when the original PreHeader is
523  // large.
524  LLVM_DEBUG(dbgs() << " Updating part of OrigPreheader scopes\n");
525  auto *FirstDecl =
526  cast<Instruction>(ValueMap[*NoAliasDeclInstructions.begin()]);
527  auto *LastInst = &OrigPreheader->back();
528  cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, FirstDecl, LastInst,
529  Context, "pre.rot");
530  LLVM_DEBUG(OrigPreheader->dump());
531 
532  LLVM_DEBUG(dbgs() << " Updated NewHeader:\n");
533  LLVM_DEBUG(NewHeader->dump());
534  }
535  }
536 
537  // Along with all the other instructions, we just cloned OrigHeader's
538  // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
539  // successors by duplicating their incoming values for OrigHeader.
540  for (BasicBlock *SuccBB : successors(OrigHeader))
541  for (BasicBlock::iterator BI = SuccBB->begin();
542  PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
543  PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
544 
545  // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
546  // OrigPreHeader's old terminator (the original branch into the loop), and
547  // remove the corresponding incoming values from the PHI nodes in OrigHeader.
548  LoopEntryBranch->eraseFromParent();
549 
550  // Update MemorySSA before the rewrite call below changes the 1:1
551  // instruction:cloned_instruction_or_value mapping.
552  if (MSSAU) {
553  InsertNewValueIntoMap(ValueMapMSSA, OrigHeader, OrigPreheader);
554  MSSAU->updateForClonedBlockIntoPred(OrigHeader, OrigPreheader,
555  ValueMapMSSA);
556  }
557 
558  SmallVector<PHINode*, 2> InsertedPHIs;
559  // If there were any uses of instructions in the duplicated block outside the
560  // loop, update them, inserting PHI nodes as required
561  RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap, SE,
562  &InsertedPHIs);
563 
564  // Attach dbg.value intrinsics to the new phis if that phi uses a value that
565  // previously had debug metadata attached. This keeps the debug info
566  // up-to-date in the loop body.
567  if (!InsertedPHIs.empty())
568  insertDebugValuesForPHIs(OrigHeader, InsertedPHIs);
569 
570  // NewHeader is now the header of the loop.
571  L->moveToHeader(NewHeader);
572  assert(L->getHeader() == NewHeader && "Latch block is our new header");
573 
574  // Inform DT about changes to the CFG.
575  if (DT) {
576  // The OrigPreheader branches to the NewHeader and Exit now. Then, inform
577  // the DT about the removed edge to the OrigHeader (that got removed).
579  Updates.push_back({DominatorTree::Insert, OrigPreheader, Exit});
580  Updates.push_back({DominatorTree::Insert, OrigPreheader, NewHeader});
581  Updates.push_back({DominatorTree::Delete, OrigPreheader, OrigHeader});
582 
583  if (MSSAU) {
584  MSSAU->applyUpdates(Updates, *DT, /*UpdateDT=*/true);
585  if (VerifyMemorySSA)
586  MSSAU->getMemorySSA()->verifyMemorySSA();
587  } else {
588  DT->applyUpdates(Updates);
589  }
590  }
591 
592  // At this point, we've finished our major CFG changes. As part of cloning
593  // the loop into the preheader we've simplified instructions and the
594  // duplicated conditional branch may now be branching on a constant. If it is
595  // branching on a constant and if that constant means that we enter the loop,
596  // then we fold away the cond branch to an uncond branch. This simplifies the
597  // loop in cases important for nested loops, and it also means we don't have
598  // to split as many edges.
599  BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
600  assert(PHBI->isConditional() && "Should be clone of BI condbr!");
601  if (!isa<ConstantInt>(PHBI->getCondition()) ||
602  PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero()) !=
603  NewHeader) {
604  // The conditional branch can't be folded, handle the general case.
605  // Split edges as necessary to preserve LoopSimplify form.
606 
607  // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
608  // thus is not a preheader anymore.
609  // Split the edge to form a real preheader.
610  BasicBlock *NewPH = SplitCriticalEdge(
611  OrigPreheader, NewHeader,
612  CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
613  NewPH->setName(NewHeader->getName() + ".lr.ph");
614 
615  // Preserve canonical loop form, which means that 'Exit' should have only
616  // one predecessor. Note that Exit could be an exit block for multiple
617  // nested loops, causing both of the edges to now be critical and need to
618  // be split.
620  bool SplitLatchEdge = false;
621  for (BasicBlock *ExitPred : ExitPreds) {
622  // We only need to split loop exit edges.
623  Loop *PredLoop = LI->getLoopFor(ExitPred);
624  if (!PredLoop || PredLoop->contains(Exit) ||
625  ExitPred->getTerminator()->isIndirectTerminator())
626  continue;
627  SplitLatchEdge |= L->getLoopLatch() == ExitPred;
628  BasicBlock *ExitSplit = SplitCriticalEdge(
629  ExitPred, Exit,
630  CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
631  ExitSplit->moveBefore(Exit);
632  }
633  assert(SplitLatchEdge &&
634  "Despite splitting all preds, failed to split latch exit?");
635  (void)SplitLatchEdge;
636  } else {
637  // We can fold the conditional branch in the preheader, this makes things
638  // simpler. The first step is to remove the extra edge to the Exit block.
639  Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
640  BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI);
641  NewBI->setDebugLoc(PHBI->getDebugLoc());
642  PHBI->eraseFromParent();
643 
644  // With our CFG finalized, update DomTree if it is available.
645  if (DT) DT->deleteEdge(OrigPreheader, Exit);
646 
647  // Update MSSA too, if available.
648  if (MSSAU)
649  MSSAU->removeEdge(OrigPreheader, Exit);
650  }
651 
652  assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
653  assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
654 
655  if (MSSAU && VerifyMemorySSA)
656  MSSAU->getMemorySSA()->verifyMemorySSA();
657 
658  // Now that the CFG and DomTree are in a consistent state again, try to merge
659  // the OrigHeader block into OrigLatch. This will succeed if they are
660  // connected by an unconditional branch. This is just a cleanup so the
661  // emitted code isn't too gross in this common case.
663  BasicBlock *PredBB = OrigHeader->getUniquePredecessor();
664  bool DidMerge = MergeBlockIntoPredecessor(OrigHeader, &DTU, LI, MSSAU);
665  if (DidMerge)
666  RemoveRedundantDbgInstrs(PredBB);
667 
668  if (MSSAU && VerifyMemorySSA)
669  MSSAU->getMemorySSA()->verifyMemorySSA();
670 
671  LLVM_DEBUG(dbgs() << "LoopRotation: into "; L->dump());
672 
673  ++NumRotated;
674 
675  Rotated = true;
676  SimplifiedLatch = false;
677 
678  // Check that new latch is a deoptimizing exit and then repeat rotation if possible.
679  // Deoptimizing latch exit is not a generally typical case, so we just loop over.
680  // TODO: if it becomes a performance bottleneck extend rotation algorithm
681  // to handle multiple rotations in one go.
683 
684 
685  return true;
686 }
687 
688 /// Determine whether the instructions in this range may be safely and cheaply
689 /// speculated. This is not an important enough situation to develop complex
690 /// heuristics. We handle a single arithmetic instruction along with any type
691 /// conversions.
693  BasicBlock::iterator End, Loop *L) {
694  bool seenIncrement = false;
695  bool MultiExitLoop = false;
696 
697  if (!L->getExitingBlock())
698  MultiExitLoop = true;
699 
700  for (BasicBlock::iterator I = Begin; I != End; ++I) {
701 
703  return false;
704 
705  if (isa<DbgInfoIntrinsic>(I))
706  continue;
707 
708  switch (I->getOpcode()) {
709  default:
710  return false;
711  case Instruction::GetElementPtr:
712  // GEPs are cheap if all indices are constant.
713  if (!cast<GEPOperator>(I)->hasAllConstantIndices())
714  return false;
715  // fall-thru to increment case
717  case Instruction::Add:
718  case Instruction::Sub:
719  case Instruction::And:
720  case Instruction::Or:
721  case Instruction::Xor:
722  case Instruction::Shl:
723  case Instruction::LShr:
724  case Instruction::AShr: {
725  Value *IVOpnd =
726  !isa<Constant>(I->getOperand(0))
727  ? I->getOperand(0)
728  : !isa<Constant>(I->getOperand(1)) ? I->getOperand(1) : nullptr;
729  if (!IVOpnd)
730  return false;
731 
732  // If increment operand is used outside of the loop, this speculation
733  // could cause extra live range interference.
734  if (MultiExitLoop) {
735  for (User *UseI : IVOpnd->users()) {
736  auto *UserInst = cast<Instruction>(UseI);
737  if (!L->contains(UserInst))
738  return false;
739  }
740  }
741 
742  if (seenIncrement)
743  return false;
744  seenIncrement = true;
745  break;
746  }
747  case Instruction::Trunc:
748  case Instruction::ZExt:
749  case Instruction::SExt:
750  // ignore type conversions
751  break;
752  }
753  }
754  return true;
755 }
756 
757 /// Fold the loop tail into the loop exit by speculating the loop tail
758 /// instructions. Typically, this is a single post-increment. In the case of a
759 /// simple 2-block loop, hoisting the increment can be much better than
760 /// duplicating the entire loop header. In the case of loops with early exits,
761 /// rotation will not work anyway, but simplifyLoopLatch will put the loop in
762 /// canonical form so downstream passes can handle it.
763 ///
764 /// I don't believe this invalidates SCEV.
765 bool LoopRotate::simplifyLoopLatch(Loop *L) {
766  BasicBlock *Latch = L->getLoopLatch();
767  if (!Latch || Latch->hasAddressTaken())
768  return false;
769 
770  BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
771  if (!Jmp || !Jmp->isUnconditional())
772  return false;
773 
774  BasicBlock *LastExit = Latch->getSinglePredecessor();
775  if (!LastExit || !L->isLoopExiting(LastExit))
776  return false;
777 
778  BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
779  if (!BI)
780  return false;
781 
782  if (!shouldSpeculateInstrs(Latch->begin(), Jmp->getIterator(), L))
783  return false;
784 
785  LLVM_DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
786  << LastExit->getName() << "\n");
787 
789  MergeBlockIntoPredecessor(Latch, &DTU, LI, MSSAU, nullptr,
790  /*PredecessorWithTwoSuccessors=*/true);
791 
792  if (MSSAU && VerifyMemorySSA)
793  MSSAU->getMemorySSA()->verifyMemorySSA();
794 
795  return true;
796 }
797 
798 /// Rotate \c L, and return true if any modification was made.
799 bool LoopRotate::processLoop(Loop *L) {
800  // Save the loop metadata.
801  MDNode *LoopMD = L->getLoopID();
802 
803  bool SimplifiedLatch = false;
804 
805  // Simplify the loop latch before attempting to rotate the header
806  // upward. Rotation may not be needed if the loop tail can be folded into the
807  // loop exit.
808  if (!RotationOnly)
809  SimplifiedLatch = simplifyLoopLatch(L);
810 
811  bool MadeChange = rotateLoop(L, SimplifiedLatch);
812  assert((!MadeChange || L->isLoopExiting(L->getLoopLatch())) &&
813  "Loop latch should be exiting after loop-rotate.");
814 
815  // Restore the loop metadata.
816  // NB! We presume LoopRotation DOESN'T ADD its own metadata.
817  if ((MadeChange || SimplifiedLatch) && LoopMD)
818  L->setLoopID(LoopMD);
819 
820  return MadeChange || SimplifiedLatch;
821 }
822 
823 
824 /// The utility to convert a loop into a loop with bottom test.
827  ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
828  const SimplifyQuery &SQ, bool RotationOnly = true,
829  unsigned Threshold = unsigned(-1),
830  bool IsUtilMode = true, bool PrepareForLTO) {
831  LoopRotate LR(Threshold, LI, TTI, AC, DT, SE, MSSAU, SQ, RotationOnly,
832  IsUtilMode, PrepareForLTO);
833  return LR.processLoop(L);
834 }
llvm::NoAliasScopeDeclInst
Definition: IntrinsicInst.h:1337
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:4688
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:692
llvm::Loop
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:546
llvm::Value::dump
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:4833
llvm::LoopBase::contains
bool contains(const LoopT *L) const
Return true if the specified loop is contained within in this loop.
Definition: LoopInfo.h:138
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:31
llvm::TargetTransformInfo
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
Definition: TargetTransformInfo.h:168
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:2558
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:2836
llvm::LoopBase::getBlocks
ArrayRef< BlockT * > getBlocks() const
Get a list of the basic blocks which make up this loop.
Definition: LoopInfo.h:187
llvm::Instruction
Definition: Instruction.h:42
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:1777
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:3167
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:459
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:2801
llvm::BranchInst::Create
static BranchInst * Create(BasicBlock *IfTrue, Instruction *InsertBefore=nullptr)
Definition: Instructions.h:3142
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:618
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::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:6486
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:8091
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:163
llvm::MDNode
Metadata node.
Definition: Metadata.h:937
canRotateDeoptimizingLatchExit
static bool canRotateDeoptimizingLatchExit(Loop *L)
Definition: LoopRotationUtils.cpp:209
llvm::BranchInst::isUnconditional
bool isUnconditional() const
Definition: Instructions.h:3164
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:1102
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:1624
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:825
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:2651
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:241
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:3086
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:3165
llvm::BranchInst::getSuccessor
BasicBlock * getSuccessor(unsigned i) const
Definition: Instructions.h:3179
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