LLVM  9.0.0svn
SimpleLoopUnswitch.cpp
Go to the documentation of this file.
1 ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===//
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 
10 #include "llvm/ADT/DenseMap.h"
11 #include "llvm/ADT/STLExtras.h"
12 #include "llvm/ADT/Sequence.h"
13 #include "llvm/ADT/SetVector.h"
14 #include "llvm/ADT/SmallPtrSet.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/ADT/Twine.h"
19 #include "llvm/Analysis/CFG.h"
24 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/LoopPass.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Constant.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/InstrTypes.h"
36 #include "llvm/IR/Instruction.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Use.h"
40 #include "llvm/IR/Value.h"
41 #include "llvm/Pass.h"
42 #include "llvm/Support/Casting.h"
43 #include "llvm/Support/Debug.h"
52 #include <algorithm>
53 #include <cassert>
54 #include <iterator>
55 #include <numeric>
56 #include <utility>
57 
58 #define DEBUG_TYPE "simple-loop-unswitch"
59 
60 using namespace llvm;
61 
62 STATISTIC(NumBranches, "Number of branches unswitched");
63 STATISTIC(NumSwitches, "Number of switches unswitched");
64 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
65 STATISTIC(NumTrivial, "Number of unswitches that are trivial");
66 STATISTIC(
67  NumCostMultiplierSkipped,
68  "Number of unswitch candidates that had their cost multiplier skipped");
69 
71  "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
72  cl::desc("Forcibly enables non-trivial loop unswitching rather than "
73  "following the configuration passed into the pass."));
74 
75 static cl::opt<int>
76  UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
77  cl::desc("The cost threshold for unswitching a loop."));
78 
80  "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden,
81  cl::desc("Enable unswitch cost multiplier that prohibits exponential "
82  "explosion in nontrivial unswitch."));
84  "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden,
85  cl::desc("Toplevel siblings divisor for cost multiplier."));
87  "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden,
88  cl::desc("Number of unswitch candidates that are ignored when calculating "
89  "cost multiplier."));
91  "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
92  cl::desc("If enabled, simple loop unswitching will also consider "
93  "llvm.experimental.guard intrinsics as unswitch candidates."));
94 
95 /// Collect all of the loop invariant input values transitively used by the
96 /// homogeneous instruction graph from a given root.
97 ///
98 /// This essentially walks from a root recursively through loop variant operands
99 /// which have the exact same opcode and finds all inputs which are loop
100 /// invariant. For some operations these can be re-associated and unswitched out
101 /// of the loop entirely.
104  LoopInfo &LI) {
105  assert(!L.isLoopInvariant(&Root) &&
106  "Only need to walk the graph if root itself is not invariant.");
107  TinyPtrVector<Value *> Invariants;
108 
109  // Build a worklist and recurse through operators collecting invariants.
112  Worklist.push_back(&Root);
113  Visited.insert(&Root);
114  do {
115  Instruction &I = *Worklist.pop_back_val();
116  for (Value *OpV : I.operand_values()) {
117  // Skip constants as unswitching isn't interesting for them.
118  if (isa<Constant>(OpV))
119  continue;
120 
121  // Add it to our result if loop invariant.
122  if (L.isLoopInvariant(OpV)) {
123  Invariants.push_back(OpV);
124  continue;
125  }
126 
127  // If not an instruction with the same opcode, nothing we can do.
128  Instruction *OpI = dyn_cast<Instruction>(OpV);
129  if (!OpI || OpI->getOpcode() != Root.getOpcode())
130  continue;
131 
132  // Visit this operand.
133  if (Visited.insert(OpI).second)
134  Worklist.push_back(OpI);
135  }
136  } while (!Worklist.empty());
137 
138  return Invariants;
139 }
140 
141 static void replaceLoopInvariantUses(Loop &L, Value *Invariant,
142  Constant &Replacement) {
143  assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");
144 
145  // Replace uses of LIC in the loop with the given constant.
146  for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); UI != UE;) {
147  // Grab the use and walk past it so we can clobber it in the use list.
148  Use *U = &*UI++;
149  Instruction *UserI = dyn_cast<Instruction>(U->getUser());
150 
151  // Replace this use within the loop body.
152  if (UserI && L.contains(UserI))
153  U->set(&Replacement);
154  }
155 }
156 
157 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial
158 /// incoming values along this edge.
159 static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB,
160  BasicBlock &ExitBB) {
161  for (Instruction &I : ExitBB) {
162  auto *PN = dyn_cast<PHINode>(&I);
163  if (!PN)
164  // No more PHIs to check.
165  return true;
166 
167  // If the incoming value for this edge isn't loop invariant the unswitch
168  // won't be trivial.
169  if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))
170  return false;
171  }
172  llvm_unreachable("Basic blocks should never be empty!");
173 }
174 
175 /// Insert code to test a set of loop invariant values, and conditionally branch
176 /// on them.
178  ArrayRef<Value *> Invariants,
179  bool Direction,
180  BasicBlock &UnswitchedSucc,
181  BasicBlock &NormalSucc) {
182  IRBuilder<> IRB(&BB);
183  Value *Cond = Invariants.front();
184  for (Value *Invariant :
185  make_range(std::next(Invariants.begin()), Invariants.end()))
186  if (Direction)
187  Cond = IRB.CreateOr(Cond, Invariant);
188  else
189  Cond = IRB.CreateAnd(Cond, Invariant);
190 
191  IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
192  Direction ? &NormalSucc : &UnswitchedSucc);
193 }
194 
195 /// Rewrite the PHI nodes in an unswitched loop exit basic block.
196 ///
197 /// Requires that the loop exit and unswitched basic block are the same, and
198 /// that the exiting block was a unique predecessor of that block. Rewrites the
199 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
200 /// PHI nodes from the old preheader that now contains the unswitched
201 /// terminator.
203  BasicBlock &OldExitingBB,
204  BasicBlock &OldPH) {
205  for (PHINode &PN : UnswitchedBB.phis()) {
206  // When the loop exit is directly unswitched we just need to update the
207  // incoming basic block. We loop to handle weird cases with repeated
208  // incoming blocks, but expect to typically only have one operand here.
209  for (auto i : seq<int>(0, PN.getNumOperands())) {
210  assert(PN.getIncomingBlock(i) == &OldExitingBB &&
211  "Found incoming block different from unique predecessor!");
212  PN.setIncomingBlock(i, &OldPH);
213  }
214  }
215 }
216 
217 /// Rewrite the PHI nodes in the loop exit basic block and the split off
218 /// unswitched block.
219 ///
220 /// Because the exit block remains an exit from the loop, this rewrites the
221 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
222 /// nodes into the unswitched basic block to select between the value in the
223 /// old preheader and the loop exit.
225  BasicBlock &UnswitchedBB,
226  BasicBlock &OldExitingBB,
227  BasicBlock &OldPH,
228  bool FullUnswitch) {
229  assert(&ExitBB != &UnswitchedBB &&
230  "Must have different loop exit and unswitched blocks!");
231  Instruction *InsertPt = &*UnswitchedBB.begin();
232  for (PHINode &PN : ExitBB.phis()) {
233  auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,
234  PN.getName() + ".split", InsertPt);
235 
236  // Walk backwards over the old PHI node's inputs to minimize the cost of
237  // removing each one. We have to do this weird loop manually so that we
238  // create the same number of new incoming edges in the new PHI as we expect
239  // each case-based edge to be included in the unswitched switch in some
240  // cases.
241  // FIXME: This is really, really gross. It would be much cleaner if LLVM
242  // allowed us to create a single entry for a predecessor block without
243  // having separate entries for each "edge" even though these edges are
244  // required to produce identical results.
245  for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
246  if (PN.getIncomingBlock(i) != &OldExitingBB)
247  continue;
248 
249  Value *Incoming = PN.getIncomingValue(i);
250  if (FullUnswitch)
251  // No more edge from the old exiting block to the exit block.
252  PN.removeIncomingValue(i);
253 
254  NewPN->addIncoming(Incoming, &OldPH);
255  }
256 
257  // Now replace the old PHI with the new one and wire the old one in as an
258  // input to the new one.
259  PN.replaceAllUsesWith(NewPN);
260  NewPN->addIncoming(&PN, &ExitBB);
261  }
262 }
263 
264 /// Hoist the current loop up to the innermost loop containing a remaining exit.
265 ///
266 /// Because we've removed an exit from the loop, we may have changed the set of
267 /// loops reachable and need to move the current loop up the loop nest or even
268 /// to an entirely separate nest.
269 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
270  DominatorTree &DT, LoopInfo &LI) {
271  // If the loop is already at the top level, we can't hoist it anywhere.
272  Loop *OldParentL = L.getParentLoop();
273  if (!OldParentL)
274  return;
275 
277  L.getExitBlocks(Exits);
278  Loop *NewParentL = nullptr;
279  for (auto *ExitBB : Exits)
280  if (Loop *ExitL = LI.getLoopFor(ExitBB))
281  if (!NewParentL || NewParentL->contains(ExitL))
282  NewParentL = ExitL;
283 
284  if (NewParentL == OldParentL)
285  return;
286 
287  // The new parent loop (if different) should always contain the old one.
288  if (NewParentL)
289  assert(NewParentL->contains(OldParentL) &&
290  "Can only hoist this loop up the nest!");
291 
292  // The preheader will need to move with the body of this loop. However,
293  // because it isn't in this loop we also need to update the primary loop map.
294  assert(OldParentL == LI.getLoopFor(&Preheader) &&
295  "Parent loop of this loop should contain this loop's preheader!");
296  LI.changeLoopFor(&Preheader, NewParentL);
297 
298  // Remove this loop from its old parent.
299  OldParentL->removeChildLoop(&L);
300 
301  // Add the loop either to the new parent or as a top-level loop.
302  if (NewParentL)
303  NewParentL->addChildLoop(&L);
304  else
305  LI.addTopLevelLoop(&L);
306 
307  // Remove this loops blocks from the old parent and every other loop up the
308  // nest until reaching the new parent. Also update all of these
309  // no-longer-containing loops to reflect the nesting change.
310  for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
311  OldContainingL = OldContainingL->getParentLoop()) {
312  llvm::erase_if(OldContainingL->getBlocksVector(),
313  [&](const BasicBlock *BB) {
314  return BB == &Preheader || L.contains(BB);
315  });
316 
317  OldContainingL->getBlocksSet().erase(&Preheader);
318  for (BasicBlock *BB : L.blocks())
319  OldContainingL->getBlocksSet().erase(BB);
320 
321  // Because we just hoisted a loop out of this one, we have essentially
322  // created new exit paths from it. That means we need to form LCSSA PHI
323  // nodes for values used in the no-longer-nested loop.
324  formLCSSA(*OldContainingL, DT, &LI, nullptr);
325 
326  // We shouldn't need to form dedicated exits because the exit introduced
327  // here is the (just split by unswitching) preheader. However, after trivial
328  // unswitching it is possible to get new non-dedicated exits out of parent
329  // loop so let's conservatively form dedicated exit blocks and figure out
330  // if we can optimize later.
331  formDedicatedExitBlocks(OldContainingL, &DT, &LI, /*PreserveLCSSA*/ true);
332  }
333 }
334 
335 /// Unswitch a trivial branch if the condition is loop invariant.
336 ///
337 /// This routine should only be called when loop code leading to the branch has
338 /// been validated as trivial (no side effects). This routine checks if the
339 /// condition is invariant and one of the successors is a loop exit. This
340 /// allows us to unswitch without duplicating the loop, making it trivial.
341 ///
342 /// If this routine fails to unswitch the branch it returns false.
343 ///
344 /// If the branch can be unswitched, this routine splits the preheader and
345 /// hoists the branch above that split. Preserves loop simplified form
346 /// (splitting the exit block as necessary). It simplifies the branch within
347 /// the loop to an unconditional branch but doesn't remove it entirely. Further
348 /// cleanup can be done with some simplify-cfg like pass.
349 ///
350 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
351 /// invalidated by this.
353  LoopInfo &LI, ScalarEvolution *SE,
354  MemorySSAUpdater *MSSAU) {
355  assert(BI.isConditional() && "Can only unswitch a conditional branch!");
356  LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n");
357 
358  // The loop invariant values that we want to unswitch.
359  TinyPtrVector<Value *> Invariants;
360 
361  // When true, we're fully unswitching the branch rather than just unswitching
362  // some input conditions to the branch.
363  bool FullUnswitch = false;
364 
365  if (L.isLoopInvariant(BI.getCondition())) {
366  Invariants.push_back(BI.getCondition());
367  FullUnswitch = true;
368  } else {
369  if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition()))
370  Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);
371  if (Invariants.empty())
372  // Couldn't find invariant inputs!
373  return false;
374  }
375 
376  // Check that one of the branch's successors exits, and which one.
377  bool ExitDirection = true;
378  int LoopExitSuccIdx = 0;
379  auto *LoopExitBB = BI.getSuccessor(0);
380  if (L.contains(LoopExitBB)) {
381  ExitDirection = false;
382  LoopExitSuccIdx = 1;
383  LoopExitBB = BI.getSuccessor(1);
384  if (L.contains(LoopExitBB))
385  return false;
386  }
387  auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
388  auto *ParentBB = BI.getParent();
389  if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB))
390  return false;
391 
392  // When unswitching only part of the branch's condition, we need the exit
393  // block to be reached directly from the partially unswitched input. This can
394  // be done when the exit block is along the true edge and the branch condition
395  // is a graph of `or` operations, or the exit block is along the false edge
396  // and the condition is a graph of `and` operations.
397  if (!FullUnswitch) {
398  if (ExitDirection) {
399  if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::Or)
400  return false;
401  } else {
402  if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::And)
403  return false;
404  }
405  }
406 
407  LLVM_DEBUG({
408  dbgs() << " unswitching trivial invariant conditions for: " << BI
409  << "\n";
410  for (Value *Invariant : Invariants) {
411  dbgs() << " " << *Invariant << " == true";
412  if (Invariant != Invariants.back())
413  dbgs() << " ||";
414  dbgs() << "\n";
415  }
416  });
417 
418  // If we have scalar evolutions, we need to invalidate them including this
419  // loop and the loop containing the exit block.
420  if (SE) {
421  if (Loop *ExitL = LI.getLoopFor(LoopExitBB))
422  SE->forgetLoop(ExitL);
423  else
424  // Forget the entire nest as this exits the entire nest.
425  SE->forgetTopmostLoop(&L);
426  }
427 
428  if (MSSAU && VerifyMemorySSA)
429  MSSAU->getMemorySSA()->verifyMemorySSA();
430 
431  // Split the preheader, so that we know that there is a safe place to insert
432  // the conditional branch. We will change the preheader to have a conditional
433  // branch on LoopCond.
434  BasicBlock *OldPH = L.getLoopPreheader();
435  BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
436 
437  // Now that we have a place to insert the conditional branch, create a place
438  // to branch to: this is the exit block out of the loop that we are
439  // unswitching. We need to split this if there are other loop predecessors.
440  // Because the loop is in simplified form, *any* other predecessor is enough.
441  BasicBlock *UnswitchedBB;
442  if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
443  assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
444  "A branch's parent isn't a predecessor!");
445  UnswitchedBB = LoopExitBB;
446  } else {
447  UnswitchedBB =
448  SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI, MSSAU);
449  }
450 
451  if (MSSAU && VerifyMemorySSA)
452  MSSAU->getMemorySSA()->verifyMemorySSA();
453 
454  // Actually move the invariant uses into the unswitched position. If possible,
455  // we do this by moving the instructions, but when doing partial unswitching
456  // we do it by building a new merge of the values in the unswitched position.
457  OldPH->getTerminator()->eraseFromParent();
458  if (FullUnswitch) {
459  // If fully unswitching, we can use the existing branch instruction.
460  // Splice it into the old PH to gate reaching the new preheader and re-point
461  // its successors.
462  OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(),
463  BI);
464  if (MSSAU) {
465  // Temporarily clone the terminator, to make MSSA update cheaper by
466  // separating "insert edge" updates from "remove edge" ones.
467  ParentBB->getInstList().push_back(BI.clone());
468  } else {
469  // Create a new unconditional branch that will continue the loop as a new
470  // terminator.
471  BranchInst::Create(ContinueBB, ParentBB);
472  }
473  BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
474  BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
475  } else {
476  // Only unswitching a subset of inputs to the condition, so we will need to
477  // build a new branch that merges the invariant inputs.
478  if (ExitDirection)
479  assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
480  Instruction::Or &&
481  "Must have an `or` of `i1`s for the condition!");
482  else
483  assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
484  Instruction::And &&
485  "Must have an `and` of `i1`s for the condition!");
486  buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection,
487  *UnswitchedBB, *NewPH);
488  }
489 
490  // Update the dominator tree with the added edge.
491  DT.insertEdge(OldPH, UnswitchedBB);
492 
493  // After the dominator tree was updated with the added edge, update MemorySSA
494  // if available.
495  if (MSSAU) {
497  Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
498  MSSAU->applyInsertUpdates(Updates, DT);
499  }
500 
501  // Finish updating dominator tree and memory ssa for full unswitch.
502  if (FullUnswitch) {
503  if (MSSAU) {
504  // Remove the cloned branch instruction.
505  ParentBB->getTerminator()->eraseFromParent();
506  // Create unconditional branch now.
507  BranchInst::Create(ContinueBB, ParentBB);
508  MSSAU->removeEdge(ParentBB, LoopExitBB);
509  }
510  DT.deleteEdge(ParentBB, LoopExitBB);
511  }
512 
513  if (MSSAU && VerifyMemorySSA)
514  MSSAU->getMemorySSA()->verifyMemorySSA();
515 
516  // Rewrite the relevant PHI nodes.
517  if (UnswitchedBB == LoopExitBB)
518  rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
519  else
520  rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
521  *ParentBB, *OldPH, FullUnswitch);
522 
523  // The constant we can replace all of our invariants with inside the loop
524  // body. If any of the invariants have a value other than this the loop won't
525  // be entered.
526  ConstantInt *Replacement = ExitDirection
529 
530  // Since this is an i1 condition we can also trivially replace uses of it
531  // within the loop with a constant.
532  for (Value *Invariant : Invariants)
533  replaceLoopInvariantUses(L, Invariant, *Replacement);
534 
535  // If this was full unswitching, we may have changed the nesting relationship
536  // for this loop so hoist it to its correct parent if needed.
537  if (FullUnswitch)
538  hoistLoopToNewParent(L, *NewPH, DT, LI);
539 
540  LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n");
541  ++NumTrivial;
542  ++NumBranches;
543  return true;
544 }
545 
546 /// Unswitch a trivial switch if the condition is loop invariant.
547 ///
548 /// This routine should only be called when loop code leading to the switch has
549 /// been validated as trivial (no side effects). This routine checks if the
550 /// condition is invariant and that at least one of the successors is a loop
551 /// exit. This allows us to unswitch without duplicating the loop, making it
552 /// trivial.
553 ///
554 /// If this routine fails to unswitch the switch it returns false.
555 ///
556 /// If the switch can be unswitched, this routine splits the preheader and
557 /// copies the switch above that split. If the default case is one of the
558 /// exiting cases, it copies the non-exiting cases and points them at the new
559 /// preheader. If the default case is not exiting, it copies the exiting cases
560 /// and points the default at the preheader. It preserves loop simplified form
561 /// (splitting the exit blocks as necessary). It simplifies the switch within
562 /// the loop by removing now-dead cases. If the default case is one of those
563 /// unswitched, it replaces its destination with a new basic block containing
564 /// only unreachable. Such basic blocks, while technically loop exits, are not
565 /// considered for unswitching so this is a stable transform and the same
566 /// switch will not be revisited. If after unswitching there is only a single
567 /// in-loop successor, the switch is further simplified to an unconditional
568 /// branch. Still more cleanup can be done with some simplify-cfg like pass.
569 ///
570 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
571 /// invalidated by this.
573  LoopInfo &LI, ScalarEvolution *SE,
574  MemorySSAUpdater *MSSAU) {
575  LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n");
576  Value *LoopCond = SI.getCondition();
577 
578  // If this isn't switching on an invariant condition, we can't unswitch it.
579  if (!L.isLoopInvariant(LoopCond))
580  return false;
581 
582  auto *ParentBB = SI.getParent();
583 
584  SmallVector<int, 4> ExitCaseIndices;
585  for (auto Case : SI.cases()) {
586  auto *SuccBB = Case.getCaseSuccessor();
587  if (!L.contains(SuccBB) &&
588  areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB))
589  ExitCaseIndices.push_back(Case.getCaseIndex());
590  }
591  BasicBlock *DefaultExitBB = nullptr;
592  if (!L.contains(SI.getDefaultDest()) &&
593  areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) &&
594  !isa<UnreachableInst>(SI.getDefaultDest()->getTerminator()))
595  DefaultExitBB = SI.getDefaultDest();
596  else if (ExitCaseIndices.empty())
597  return false;
598 
599  LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n");
600 
601  if (MSSAU && VerifyMemorySSA)
602  MSSAU->getMemorySSA()->verifyMemorySSA();
603 
604  // We may need to invalidate SCEVs for the outermost loop reached by any of
605  // the exits.
606  Loop *OuterL = &L;
607 
608  if (DefaultExitBB) {
609  // Clear out the default destination temporarily to allow accurate
610  // predecessor lists to be examined below.
611  SI.setDefaultDest(nullptr);
612  // Check the loop containing this exit.
613  Loop *ExitL = LI.getLoopFor(DefaultExitBB);
614  if (!ExitL || ExitL->contains(OuterL))
615  OuterL = ExitL;
616  }
617 
618  // Store the exit cases into a separate data structure and remove them from
619  // the switch.
621  ExitCases.reserve(ExitCaseIndices.size());
622  // We walk the case indices backwards so that we remove the last case first
623  // and don't disrupt the earlier indices.
624  for (unsigned Index : reverse(ExitCaseIndices)) {
625  auto CaseI = SI.case_begin() + Index;
626  // Compute the outer loop from this exit.
627  Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor());
628  if (!ExitL || ExitL->contains(OuterL))
629  OuterL = ExitL;
630  // Save the value of this case.
631  ExitCases.push_back({CaseI->getCaseValue(), CaseI->getCaseSuccessor()});
632  // Delete the unswitched cases.
633  SI.removeCase(CaseI);
634  }
635 
636  if (SE) {
637  if (OuterL)
638  SE->forgetLoop(OuterL);
639  else
640  SE->forgetTopmostLoop(&L);
641  }
642 
643  // Check if after this all of the remaining cases point at the same
644  // successor.
645  BasicBlock *CommonSuccBB = nullptr;
646  if (SI.getNumCases() > 0 &&
647  std::all_of(std::next(SI.case_begin()), SI.case_end(),
648  [&SI](const SwitchInst::CaseHandle &Case) {
649  return Case.getCaseSuccessor() ==
650  SI.case_begin()->getCaseSuccessor();
651  }))
652  CommonSuccBB = SI.case_begin()->getCaseSuccessor();
653  if (!DefaultExitBB) {
654  // If we're not unswitching the default, we need it to match any cases to
655  // have a common successor or if we have no cases it is the common
656  // successor.
657  if (SI.getNumCases() == 0)
658  CommonSuccBB = SI.getDefaultDest();
659  else if (SI.getDefaultDest() != CommonSuccBB)
660  CommonSuccBB = nullptr;
661  }
662 
663  // Split the preheader, so that we know that there is a safe place to insert
664  // the switch.
665  BasicBlock *OldPH = L.getLoopPreheader();
666  BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
667  OldPH->getTerminator()->eraseFromParent();
668 
669  // Now add the unswitched switch.
670  auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
671 
672  // Rewrite the IR for the unswitched basic blocks. This requires two steps.
673  // First, we split any exit blocks with remaining in-loop predecessors. Then
674  // we update the PHIs in one of two ways depending on if there was a split.
675  // We walk in reverse so that we split in the same order as the cases
676  // appeared. This is purely for convenience of reading the resulting IR, but
677  // it doesn't cost anything really.
678  SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
680  // Handle the default exit if necessary.
681  // FIXME: It'd be great if we could merge this with the loop below but LLVM's
682  // ranges aren't quite powerful enough yet.
683  if (DefaultExitBB) {
684  if (pred_empty(DefaultExitBB)) {
685  UnswitchedExitBBs.insert(DefaultExitBB);
686  rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
687  } else {
688  auto *SplitBB =
689  SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI, MSSAU);
690  rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,
691  *ParentBB, *OldPH,
692  /*FullUnswitch*/ true);
693  DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
694  }
695  }
696  // Note that we must use a reference in the for loop so that we update the
697  // container.
698  for (auto &CasePair : reverse(ExitCases)) {
699  // Grab a reference to the exit block in the pair so that we can update it.
700  BasicBlock *ExitBB = CasePair.second;
701 
702  // If this case is the last edge into the exit block, we can simply reuse it
703  // as it will no longer be a loop exit. No mapping necessary.
704  if (pred_empty(ExitBB)) {
705  // Only rewrite once.
706  if (UnswitchedExitBBs.insert(ExitBB).second)
707  rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
708  continue;
709  }
710 
711  // Otherwise we need to split the exit block so that we retain an exit
712  // block from the loop and a target for the unswitched condition.
713  BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
714  if (!SplitExitBB) {
715  // If this is the first time we see this, do the split and remember it.
716  SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
717  rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,
718  *ParentBB, *OldPH,
719  /*FullUnswitch*/ true);
720  }
721  // Update the case pair to point to the split block.
722  CasePair.second = SplitExitBB;
723  }
724 
725  // Now add the unswitched cases. We do this in reverse order as we built them
726  // in reverse order.
727  for (auto CasePair : reverse(ExitCases)) {
728  ConstantInt *CaseVal = CasePair.first;
729  BasicBlock *UnswitchedBB = CasePair.second;
730 
731  NewSI->addCase(CaseVal, UnswitchedBB);
732  }
733 
734  // If the default was unswitched, re-point it and add explicit cases for
735  // entering the loop.
736  if (DefaultExitBB) {
737  NewSI->setDefaultDest(DefaultExitBB);
738 
739  // We removed all the exit cases, so we just copy the cases to the
740  // unswitched switch.
741  for (auto Case : SI.cases())
742  NewSI->addCase(Case.getCaseValue(), NewPH);
743  }
744 
745  // If we ended up with a common successor for every path through the switch
746  // after unswitching, rewrite it to an unconditional branch to make it easy
747  // to recognize. Otherwise we potentially have to recognize the default case
748  // pointing at unreachable and other complexity.
749  if (CommonSuccBB) {
750  BasicBlock *BB = SI.getParent();
751  // We may have had multiple edges to this common successor block, so remove
752  // them as predecessors. We skip the first one, either the default or the
753  // actual first case.
754  bool SkippedFirst = DefaultExitBB == nullptr;
755  for (auto Case : SI.cases()) {
756  assert(Case.getCaseSuccessor() == CommonSuccBB &&
757  "Non-common successor!");
758  (void)Case;
759  if (!SkippedFirst) {
760  SkippedFirst = true;
761  continue;
762  }
763  CommonSuccBB->removePredecessor(BB,
764  /*DontDeleteUselessPHIs*/ true);
765  }
766  // Now nuke the switch and replace it with a direct branch.
767  SI.eraseFromParent();
768  BranchInst::Create(CommonSuccBB, BB);
769  } else if (DefaultExitBB) {
770  assert(SI.getNumCases() > 0 &&
771  "If we had no cases we'd have a common successor!");
772  // Move the last case to the default successor. This is valid as if the
773  // default got unswitched it cannot be reached. This has the advantage of
774  // being simple and keeping the number of edges from this switch to
775  // successors the same, and avoiding any PHI update complexity.
776  auto LastCaseI = std::prev(SI.case_end());
777  SI.setDefaultDest(LastCaseI->getCaseSuccessor());
778  SI.removeCase(LastCaseI);
779  }
780 
781  // Walk the unswitched exit blocks and the unswitched split blocks and update
782  // the dominator tree based on the CFG edits. While we are walking unordered
783  // containers here, the API for applyUpdates takes an unordered list of
784  // updates and requires them to not contain duplicates.
786  for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
787  DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});
788  DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});
789  }
790  for (auto SplitUnswitchedPair : SplitExitBBMap) {
791  auto *UnswitchedBB = SplitUnswitchedPair.second;
792  DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedBB});
793  DTUpdates.push_back({DT.Insert, OldPH, UnswitchedBB});
794  }
795  DT.applyUpdates(DTUpdates);
796 
797  if (MSSAU) {
798  MSSAU->applyUpdates(DTUpdates, DT);
799  if (VerifyMemorySSA)
800  MSSAU->getMemorySSA()->verifyMemorySSA();
801  }
802 
804 
805  // We may have changed the nesting relationship for this loop so hoist it to
806  // its correct parent if needed.
807  hoistLoopToNewParent(L, *NewPH, DT, LI);
808 
809  ++NumTrivial;
810  ++NumSwitches;
811  LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n");
812  return true;
813 }
814 
815 /// This routine scans the loop to find a branch or switch which occurs before
816 /// any side effects occur. These can potentially be unswitched without
817 /// duplicating the loop. If a branch or switch is successfully unswitched the
818 /// scanning continues to see if subsequent branches or switches have become
819 /// trivial. Once all trivial candidates have been unswitched, this routine
820 /// returns.
821 ///
822 /// The return value indicates whether anything was unswitched (and therefore
823 /// changed).
824 ///
825 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
826 /// invalidated by this.
828  LoopInfo &LI, ScalarEvolution *SE,
829  MemorySSAUpdater *MSSAU) {
830  bool Changed = false;
831 
832  // If loop header has only one reachable successor we should keep looking for
833  // trivial condition candidates in the successor as well. An alternative is
834  // to constant fold conditions and merge successors into loop header (then we
835  // only need to check header's terminator). The reason for not doing this in
836  // LoopUnswitch pass is that it could potentially break LoopPassManager's
837  // invariants. Folding dead branches could either eliminate the current loop
838  // or make other loops unreachable. LCSSA form might also not be preserved
839  // after deleting branches. The following code keeps traversing loop header's
840  // successors until it finds the trivial condition candidate (condition that
841  // is not a constant). Since unswitching generates branches with constant
842  // conditions, this scenario could be very common in practice.
843  BasicBlock *CurrentBB = L.getHeader();
845  Visited.insert(CurrentBB);
846  do {
847  // Check if there are any side-effecting instructions (e.g. stores, calls,
848  // volatile loads) in the part of the loop that the code *would* execute
849  // without unswitching.
850  if (llvm::any_of(*CurrentBB,
851  [](Instruction &I) { return I.mayHaveSideEffects(); }))
852  return Changed;
853 
854  Instruction *CurrentTerm = CurrentBB->getTerminator();
855 
856  if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
857  // Don't bother trying to unswitch past a switch with a constant
858  // condition. This should be removed prior to running this pass by
859  // simplify-cfg.
860  if (isa<Constant>(SI->getCondition()))
861  return Changed;
862 
863  if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU))
864  // Couldn't unswitch this one so we're done.
865  return Changed;
866 
867  // Mark that we managed to unswitch something.
868  Changed = true;
869 
870  // If unswitching turned the terminator into an unconditional branch then
871  // we can continue. The unswitching logic specifically works to fold any
872  // cases it can into an unconditional branch to make it easier to
873  // recognize here.
874  auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
875  if (!BI || BI->isConditional())
876  return Changed;
877 
878  CurrentBB = BI->getSuccessor(0);
879  continue;
880  }
881 
882  auto *BI = dyn_cast<BranchInst>(CurrentTerm);
883  if (!BI)
884  // We do not understand other terminator instructions.
885  return Changed;
886 
887  // Don't bother trying to unswitch past an unconditional branch or a branch
888  // with a constant value. These should be removed by simplify-cfg prior to
889  // running this pass.
890  if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
891  return Changed;
892 
893  // Found a trivial condition candidate: non-foldable conditional branch. If
894  // we fail to unswitch this, we can't do anything else that is trivial.
895  if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU))
896  return Changed;
897 
898  // Mark that we managed to unswitch something.
899  Changed = true;
900 
901  // If we only unswitched some of the conditions feeding the branch, we won't
902  // have collapsed it to a single successor.
903  BI = cast<BranchInst>(CurrentBB->getTerminator());
904  if (BI->isConditional())
905  return Changed;
906 
907  // Follow the newly unconditional branch into its successor.
908  CurrentBB = BI->getSuccessor(0);
909 
910  // When continuing, if we exit the loop or reach a previous visited block,
911  // then we can not reach any trivial condition candidates (unfoldable
912  // branch instructions or switch instructions) and no unswitch can happen.
913  } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
914 
915  return Changed;
916 }
917 
918 /// Build the cloned blocks for an unswitched copy of the given loop.
919 ///
920 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
921 /// after the split block (`SplitBB`) that will be used to select between the
922 /// cloned and original loop.
923 ///
924 /// This routine handles cloning all of the necessary loop blocks and exit
925 /// blocks including rewriting their instructions and the relevant PHI nodes.
926 /// Any loop blocks or exit blocks which are dominated by a different successor
927 /// than the one for this clone of the loop blocks can be trivially skipped. We
928 /// use the `DominatingSucc` map to determine whether a block satisfies that
929 /// property with a simple map lookup.
930 ///
931 /// It also correctly creates the unconditional branch in the cloned
932 /// unswitched parent block to only point at the unswitched successor.
933 ///
934 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit
935 /// block splitting is correctly reflected in `LoopInfo`, essentially all of
936 /// the cloned blocks (and their loops) are left without full `LoopInfo`
937 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
938 /// blocks to them but doesn't create the cloned `DominatorTree` structure and
939 /// instead the caller must recompute an accurate DT. It *does* correctly
940 /// update the `AssumptionCache` provided in `AC`.
942  Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
943  ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
944  BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
945  const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
946  ValueToValueMapTy &VMap,
948  DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
950  NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());
951 
952  // We will need to clone a bunch of blocks, wrap up the clone operation in
953  // a helper.
954  auto CloneBlock = [&](BasicBlock *OldBB) {
955  // Clone the basic block and insert it before the new preheader.
956  BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());
957  NewBB->moveBefore(LoopPH);
958 
959  // Record this block and the mapping.
960  NewBlocks.push_back(NewBB);
961  VMap[OldBB] = NewBB;
962 
963  return NewBB;
964  };
965 
966  // We skip cloning blocks when they have a dominating succ that is not the
967  // succ we are cloning for.
968  auto SkipBlock = [&](BasicBlock *BB) {
969  auto It = DominatingSucc.find(BB);
970  return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
971  };
972 
973  // First, clone the preheader.
974  auto *ClonedPH = CloneBlock(LoopPH);
975 
976  // Then clone all the loop blocks, skipping the ones that aren't necessary.
977  for (auto *LoopBB : L.blocks())
978  if (!SkipBlock(LoopBB))
979  CloneBlock(LoopBB);
980 
981  // Split all the loop exit edges so that when we clone the exit blocks, if
982  // any of the exit blocks are *also* a preheader for some other loop, we
983  // don't create multiple predecessors entering the loop header.
984  for (auto *ExitBB : ExitBlocks) {
985  if (SkipBlock(ExitBB))
986  continue;
987 
988  // When we are going to clone an exit, we don't need to clone all the
989  // instructions in the exit block and we want to ensure we have an easy
990  // place to merge the CFG, so split the exit first. This is always safe to
991  // do because there cannot be any non-loop predecessors of a loop exit in
992  // loop simplified form.
993  auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
994 
995  // Rearrange the names to make it easier to write test cases by having the
996  // exit block carry the suffix rather than the merge block carrying the
997  // suffix.
998  MergeBB->takeName(ExitBB);
999  ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1000 
1001  // Now clone the original exit block.
1002  auto *ClonedExitBB = CloneBlock(ExitBB);
1003  assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1004  "Exit block should have been split to have one successor!");
1005  assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1006  "Cloned exit block has the wrong successor!");
1007 
1008  // Remap any cloned instructions and create a merge phi node for them.
1009  for (auto ZippedInsts : llvm::zip_first(
1010  llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),
1011  llvm::make_range(ClonedExitBB->begin(),
1012  std::prev(ClonedExitBB->end())))) {
1013  Instruction &I = std::get<0>(ZippedInsts);
1014  Instruction &ClonedI = std::get<1>(ZippedInsts);
1015 
1016  // The only instructions in the exit block should be PHI nodes and
1017  // potentially a landing pad.
1018  assert(
1019  (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1020  "Bad instruction in exit block!");
1021  // We should have a value map between the instruction and its clone.
1022  assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1023 
1024  auto *MergePN =
1025  PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi",
1026  &*MergeBB->getFirstInsertionPt());
1027  I.replaceAllUsesWith(MergePN);
1028  MergePN->addIncoming(&I, ExitBB);
1029  MergePN->addIncoming(&ClonedI, ClonedExitBB);
1030  }
1031  }
1032 
1033  // Rewrite the instructions in the cloned blocks to refer to the instructions
1034  // in the cloned blocks. We have to do this as a second pass so that we have
1035  // everything available. Also, we have inserted new instructions which may
1036  // include assume intrinsics, so we update the assumption cache while
1037  // processing this.
1038  for (auto *ClonedBB : NewBlocks)
1039  for (Instruction &I : *ClonedBB) {
1040  RemapInstruction(&I, VMap,
1042  if (auto *II = dyn_cast<IntrinsicInst>(&I))
1043  if (II->getIntrinsicID() == Intrinsic::assume)
1044  AC.registerAssumption(II);
1045  }
1046 
1047  // Update any PHI nodes in the cloned successors of the skipped blocks to not
1048  // have spurious incoming values.
1049  for (auto *LoopBB : L.blocks())
1050  if (SkipBlock(LoopBB))
1051  for (auto *SuccBB : successors(LoopBB))
1052  if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))
1053  for (PHINode &PN : ClonedSuccBB->phis())
1054  PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);
1055 
1056  // Remove the cloned parent as a predecessor of any successor we ended up
1057  // cloning other than the unswitched one.
1058  auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));
1059  for (auto *SuccBB : successors(ParentBB)) {
1060  if (SuccBB == UnswitchedSuccBB)
1061  continue;
1062 
1063  auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));
1064  if (!ClonedSuccBB)
1065  continue;
1066 
1067  ClonedSuccBB->removePredecessor(ClonedParentBB,
1068  /*DontDeleteUselessPHIs*/ true);
1069  }
1070 
1071  // Replace the cloned branch with an unconditional branch to the cloned
1072  // unswitched successor.
1073  auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));
1074  ClonedParentBB->getTerminator()->eraseFromParent();
1075  BranchInst::Create(ClonedSuccBB, ClonedParentBB);
1076 
1077  // If there are duplicate entries in the PHI nodes because of multiple edges
1078  // to the unswitched successor, we need to nuke all but one as we replaced it
1079  // with a direct branch.
1080  for (PHINode &PN : ClonedSuccBB->phis()) {
1081  bool Found = false;
1082  // Loop over the incoming operands backwards so we can easily delete as we
1083  // go without invalidating the index.
1084  for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1085  if (PN.getIncomingBlock(i) != ClonedParentBB)
1086  continue;
1087  if (!Found) {
1088  Found = true;
1089  continue;
1090  }
1091  PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);
1092  }
1093  }
1094 
1095  // Record the domtree updates for the new blocks.
1097  for (auto *ClonedBB : NewBlocks) {
1098  for (auto *SuccBB : successors(ClonedBB))
1099  if (SuccSet.insert(SuccBB).second)
1100  DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});
1101  SuccSet.clear();
1102  }
1103 
1104  return ClonedPH;
1105 }
1106 
1107 /// Recursively clone the specified loop and all of its children.
1108 ///
1109 /// The target parent loop for the clone should be provided, or can be null if
1110 /// the clone is a top-level loop. While cloning, all the blocks are mapped
1111 /// with the provided value map. The entire original loop must be present in
1112 /// the value map. The cloned loop is returned.
1113 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1114  const ValueToValueMapTy &VMap, LoopInfo &LI) {
1115  auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1116  assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1117  ClonedL.reserveBlocks(OrigL.getNumBlocks());
1118  for (auto *BB : OrigL.blocks()) {
1119  auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));
1120  ClonedL.addBlockEntry(ClonedBB);
1121  if (LI.getLoopFor(BB) == &OrigL)
1122  LI.changeLoopFor(ClonedBB, &ClonedL);
1123  }
1124  };
1125 
1126  // We specially handle the first loop because it may get cloned into
1127  // a different parent and because we most commonly are cloning leaf loops.
1128  Loop *ClonedRootL = LI.AllocateLoop();
1129  if (RootParentL)
1130  RootParentL->addChildLoop(ClonedRootL);
1131  else
1132  LI.addTopLevelLoop(ClonedRootL);
1133  AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1134 
1135  if (OrigRootL.empty())
1136  return ClonedRootL;
1137 
1138  // If we have a nest, we can quickly clone the entire loop nest using an
1139  // iterative approach because it is a tree. We keep the cloned parent in the
1140  // data structure to avoid repeatedly querying through a map to find it.
1141  SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1142  // Build up the loops to clone in reverse order as we'll clone them from the
1143  // back.
1144  for (Loop *ChildL : llvm::reverse(OrigRootL))
1145  LoopsToClone.push_back({ClonedRootL, ChildL});
1146  do {
1147  Loop *ClonedParentL, *L;
1148  std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();
1149  Loop *ClonedL = LI.AllocateLoop();
1150  ClonedParentL->addChildLoop(ClonedL);
1151  AddClonedBlocksToLoop(*L, *ClonedL);
1152  for (Loop *ChildL : llvm::reverse(*L))
1153  LoopsToClone.push_back({ClonedL, ChildL});
1154  } while (!LoopsToClone.empty());
1155 
1156  return ClonedRootL;
1157 }
1158 
1159 /// Build the cloned loops of an original loop from unswitching.
1160 ///
1161 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1162 /// operation. We need to re-verify that there even is a loop (as the backedge
1163 /// may not have been cloned), and even if there are remaining backedges the
1164 /// backedge set may be different. However, we know that each child loop is
1165 /// undisturbed, we only need to find where to place each child loop within
1166 /// either any parent loop or within a cloned version of the original loop.
1167 ///
1168 /// Because child loops may end up cloned outside of any cloned version of the
1169 /// original loop, multiple cloned sibling loops may be created. All of them
1170 /// are returned so that the newly introduced loop nest roots can be
1171 /// identified.
1172 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1173  const ValueToValueMapTy &VMap, LoopInfo &LI,
1174  SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1175  Loop *ClonedL = nullptr;
1176 
1177  auto *OrigPH = OrigL.getLoopPreheader();
1178  auto *OrigHeader = OrigL.getHeader();
1179 
1180  auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));
1181  auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));
1182 
1183  // We need to know the loops of the cloned exit blocks to even compute the
1184  // accurate parent loop. If we only clone exits to some parent of the
1185  // original parent, we want to clone into that outer loop. We also keep track
1186  // of the loops that our cloned exit blocks participate in.
1187  Loop *ParentL = nullptr;
1188  SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1190  ClonedExitsInLoops.reserve(ExitBlocks.size());
1191  for (auto *ExitBB : ExitBlocks)
1192  if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))
1193  if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1194  ExitLoopMap[ClonedExitBB] = ExitL;
1195  ClonedExitsInLoops.push_back(ClonedExitBB);
1196  if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1197  ParentL = ExitL;
1198  }
1199  assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1200  ParentL->contains(OrigL.getParentLoop())) &&
1201  "The computed parent loop should always contain (or be) the parent of "
1202  "the original loop.");
1203 
1204  // We build the set of blocks dominated by the cloned header from the set of
1205  // cloned blocks out of the original loop. While not all of these will
1206  // necessarily be in the cloned loop, it is enough to establish that they
1207  // aren't in unreachable cycles, etc.
1208  SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1209  for (auto *BB : OrigL.blocks())
1210  if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))
1211  ClonedLoopBlocks.insert(ClonedBB);
1212 
1213  // Rebuild the set of blocks that will end up in the cloned loop. We may have
1214  // skipped cloning some region of this loop which can in turn skip some of
1215  // the backedges so we have to rebuild the blocks in the loop based on the
1216  // backedges that remain after cloning.
1218  SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1219  for (auto *Pred : predecessors(ClonedHeader)) {
1220  // The only possible non-loop header predecessor is the preheader because
1221  // we know we cloned the loop in simplified form.
1222  if (Pred == ClonedPH)
1223  continue;
1224 
1225  // Because the loop was in simplified form, the only non-loop predecessor
1226  // should be the preheader.
1227  assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1228  "header other than the preheader "
1229  "that is not part of the loop!");
1230 
1231  // Insert this block into the loop set and on the first visit (and if it
1232  // isn't the header we're currently walking) put it into the worklist to
1233  // recurse through.
1234  if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)
1235  Worklist.push_back(Pred);
1236  }
1237 
1238  // If we had any backedges then there *is* a cloned loop. Put the header into
1239  // the loop set and then walk the worklist backwards to find all the blocks
1240  // that remain within the loop after cloning.
1241  if (!BlocksInClonedLoop.empty()) {
1242  BlocksInClonedLoop.insert(ClonedHeader);
1243 
1244  while (!Worklist.empty()) {
1245  BasicBlock *BB = Worklist.pop_back_val();
1246  assert(BlocksInClonedLoop.count(BB) &&
1247  "Didn't put block into the loop set!");
1248 
1249  // Insert any predecessors that are in the possible set into the cloned
1250  // set, and if the insert is successful, add them to the worklist. Note
1251  // that we filter on the blocks that are definitely reachable via the
1252  // backedge to the loop header so we may prune out dead code within the
1253  // cloned loop.
1254  for (auto *Pred : predecessors(BB))
1255  if (ClonedLoopBlocks.count(Pred) &&
1256  BlocksInClonedLoop.insert(Pred).second)
1257  Worklist.push_back(Pred);
1258  }
1259 
1260  ClonedL = LI.AllocateLoop();
1261  if (ParentL) {
1262  ParentL->addBasicBlockToLoop(ClonedPH, LI);
1263  ParentL->addChildLoop(ClonedL);
1264  } else {
1265  LI.addTopLevelLoop(ClonedL);
1266  }
1267  NonChildClonedLoops.push_back(ClonedL);
1268 
1269  ClonedL->reserveBlocks(BlocksInClonedLoop.size());
1270  // We don't want to just add the cloned loop blocks based on how we
1271  // discovered them. The original order of blocks was carefully built in
1272  // a way that doesn't rely on predecessor ordering. Rather than re-invent
1273  // that logic, we just re-walk the original blocks (and those of the child
1274  // loops) and filter them as we add them into the cloned loop.
1275  for (auto *BB : OrigL.blocks()) {
1276  auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));
1277  if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))
1278  continue;
1279 
1280  // Directly add the blocks that are only in this loop.
1281  if (LI.getLoopFor(BB) == &OrigL) {
1282  ClonedL->addBasicBlockToLoop(ClonedBB, LI);
1283  continue;
1284  }
1285 
1286  // We want to manually add it to this loop and parents.
1287  // Registering it with LoopInfo will happen when we clone the top
1288  // loop for this block.
1289  for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1290  PL->addBlockEntry(ClonedBB);
1291  }
1292 
1293  // Now add each child loop whose header remains within the cloned loop. All
1294  // of the blocks within the loop must satisfy the same constraints as the
1295  // header so once we pass the header checks we can just clone the entire
1296  // child loop nest.
1297  for (Loop *ChildL : OrigL) {
1298  auto *ClonedChildHeader =
1299  cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1300  if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))
1301  continue;
1302 
1303 #ifndef NDEBUG
1304  // We should never have a cloned child loop header but fail to have
1305  // all of the blocks for that child loop.
1306  for (auto *ChildLoopBB : ChildL->blocks())
1307  assert(BlocksInClonedLoop.count(
1308  cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1309  "Child cloned loop has a header within the cloned outer "
1310  "loop but not all of its blocks!");
1311 #endif
1312 
1313  cloneLoopNest(*ChildL, ClonedL, VMap, LI);
1314  }
1315  }
1316 
1317  // Now that we've handled all the components of the original loop that were
1318  // cloned into a new loop, we still need to handle anything from the original
1319  // loop that wasn't in a cloned loop.
1320 
1321  // Figure out what blocks are left to place within any loop nest containing
1322  // the unswitched loop. If we never formed a loop, the cloned PH is one of
1323  // them.
1324  SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1325  if (BlocksInClonedLoop.empty())
1326  UnloopedBlockSet.insert(ClonedPH);
1327  for (auto *ClonedBB : ClonedLoopBlocks)
1328  if (!BlocksInClonedLoop.count(ClonedBB))
1329  UnloopedBlockSet.insert(ClonedBB);
1330 
1331  // Copy the cloned exits and sort them in ascending loop depth, we'll work
1332  // backwards across these to process them inside out. The order shouldn't
1333  // matter as we're just trying to build up the map from inside-out; we use
1334  // the map in a more stably ordered way below.
1335  auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1336  llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1337  return ExitLoopMap.lookup(LHS)->getLoopDepth() <
1338  ExitLoopMap.lookup(RHS)->getLoopDepth();
1339  });
1340 
1341  // Populate the existing ExitLoopMap with everything reachable from each
1342  // exit, starting from the inner most exit.
1343  while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1344  assert(Worklist.empty() && "Didn't clear worklist!");
1345 
1346  BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1347  Loop *ExitL = ExitLoopMap.lookup(ExitBB);
1348 
1349  // Walk the CFG back until we hit the cloned PH adding everything reachable
1350  // and in the unlooped set to this exit block's loop.
1351  Worklist.push_back(ExitBB);
1352  do {
1353  BasicBlock *BB = Worklist.pop_back_val();
1354  // We can stop recursing at the cloned preheader (if we get there).
1355  if (BB == ClonedPH)
1356  continue;
1357 
1358  for (BasicBlock *PredBB : predecessors(BB)) {
1359  // If this pred has already been moved to our set or is part of some
1360  // (inner) loop, no update needed.
1361  if (!UnloopedBlockSet.erase(PredBB)) {
1362  assert(
1363  (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1364  "Predecessor not mapped to a loop!");
1365  continue;
1366  }
1367 
1368  // We just insert into the loop set here. We'll add these blocks to the
1369  // exit loop after we build up the set in an order that doesn't rely on
1370  // predecessor order (which in turn relies on use list order).
1371  bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;
1372  (void)Inserted;
1373  assert(Inserted && "Should only visit an unlooped block once!");
1374 
1375  // And recurse through to its predecessors.
1376  Worklist.push_back(PredBB);
1377  }
1378  } while (!Worklist.empty());
1379  }
1380 
1381  // Now that the ExitLoopMap gives as mapping for all the non-looping cloned
1382  // blocks to their outer loops, walk the cloned blocks and the cloned exits
1383  // in their original order adding them to the correct loop.
1384 
1385  // We need a stable insertion order. We use the order of the original loop
1386  // order and map into the correct parent loop.
1387  for (auto *BB : llvm::concat<BasicBlock *const>(
1388  makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))
1389  if (Loop *OuterL = ExitLoopMap.lookup(BB))
1390  OuterL->addBasicBlockToLoop(BB, LI);
1391 
1392 #ifndef NDEBUG
1393  for (auto &BBAndL : ExitLoopMap) {
1394  auto *BB = BBAndL.first;
1395  auto *OuterL = BBAndL.second;
1396  assert(LI.getLoopFor(BB) == OuterL &&
1397  "Failed to put all blocks into outer loops!");
1398  }
1399 #endif
1400 
1401  // Now that all the blocks are placed into the correct containing loop in the
1402  // absence of child loops, find all the potentially cloned child loops and
1403  // clone them into whatever outer loop we placed their header into.
1404  for (Loop *ChildL : OrigL) {
1405  auto *ClonedChildHeader =
1406  cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1407  if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))
1408  continue;
1409 
1410 #ifndef NDEBUG
1411  for (auto *ChildLoopBB : ChildL->blocks())
1412  assert(VMap.count(ChildLoopBB) &&
1413  "Cloned a child loop header but not all of that loops blocks!");
1414 #endif
1415 
1416  NonChildClonedLoops.push_back(cloneLoopNest(
1417  *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));
1418  }
1419 }
1420 
1421 static void
1423  ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1424  DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1425  // Find all the dead clones, and remove them from their successors.
1426  SmallVector<BasicBlock *, 16> DeadBlocks;
1427  for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))
1428  for (auto &VMap : VMaps)
1429  if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))
1430  if (!DT.isReachableFromEntry(ClonedBB)) {
1431  for (BasicBlock *SuccBB : successors(ClonedBB))
1432  SuccBB->removePredecessor(ClonedBB);
1433  DeadBlocks.push_back(ClonedBB);
1434  }
1435 
1436  // Remove all MemorySSA in the dead blocks
1437  if (MSSAU) {
1438  SmallPtrSet<BasicBlock *, 16> DeadBlockSet(DeadBlocks.begin(),
1439  DeadBlocks.end());
1440  MSSAU->removeBlocks(DeadBlockSet);
1441  }
1442 
1443  // Drop any remaining references to break cycles.
1444  for (BasicBlock *BB : DeadBlocks)
1445  BB->dropAllReferences();
1446  // Erase them from the IR.
1447  for (BasicBlock *BB : DeadBlocks)
1448  BB->eraseFromParent();
1449 }
1450 
1452  SmallVectorImpl<BasicBlock *> &ExitBlocks,
1453  DominatorTree &DT, LoopInfo &LI,
1454  MemorySSAUpdater *MSSAU) {
1455  // Find all the dead blocks tied to this loop, and remove them from their
1456  // successors.
1457  SmallPtrSet<BasicBlock *, 16> DeadBlockSet;
1458 
1459  // Start with loop/exit blocks and get a transitive closure of reachable dead
1460  // blocks.
1461  SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1462  ExitBlocks.end());
1463  DeathCandidates.append(L.blocks().begin(), L.blocks().end());
1464  while (!DeathCandidates.empty()) {
1465  auto *BB = DeathCandidates.pop_back_val();
1466  if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {
1467  for (BasicBlock *SuccBB : successors(BB)) {
1468  SuccBB->removePredecessor(BB);
1469  DeathCandidates.push_back(SuccBB);
1470  }
1471  DeadBlockSet.insert(BB);
1472  }
1473  }
1474 
1475  // Remove all MemorySSA in the dead blocks
1476  if (MSSAU)
1477  MSSAU->removeBlocks(DeadBlockSet);
1478 
1479  // Filter out the dead blocks from the exit blocks list so that it can be
1480  // used in the caller.
1481  llvm::erase_if(ExitBlocks,
1482  [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1483 
1484  // Walk from this loop up through its parents removing all of the dead blocks.
1485  for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1486  for (auto *BB : DeadBlockSet)
1487  ParentL->getBlocksSet().erase(BB);
1488  llvm::erase_if(ParentL->getBlocksVector(),
1489  [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1490  }
1491 
1492  // Now delete the dead child loops. This raw delete will clear them
1493  // recursively.
1494  llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {
1495  if (!DeadBlockSet.count(ChildL->getHeader()))
1496  return false;
1497 
1498  assert(llvm::all_of(ChildL->blocks(),
1499  [&](BasicBlock *ChildBB) {
1500  return DeadBlockSet.count(ChildBB);
1501  }) &&
1502  "If the child loop header is dead all blocks in the child loop must "
1503  "be dead as well!");
1504  LI.destroy(ChildL);
1505  return true;
1506  });
1507 
1508  // Remove the loop mappings for the dead blocks and drop all the references
1509  // from these blocks to others to handle cyclic references as we start
1510  // deleting the blocks themselves.
1511  for (auto *BB : DeadBlockSet) {
1512  // Check that the dominator tree has already been updated.
1513  assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1514  LI.changeLoopFor(BB, nullptr);
1515  BB->dropAllReferences();
1516  }
1517 
1518  // Actually delete the blocks now that they've been fully unhooked from the
1519  // IR.
1520  for (auto *BB : DeadBlockSet)
1521  BB->eraseFromParent();
1522 }
1523 
1524 /// Recompute the set of blocks in a loop after unswitching.
1525 ///
1526 /// This walks from the original headers predecessors to rebuild the loop. We
1527 /// take advantage of the fact that new blocks can't have been added, and so we
1528 /// filter by the original loop's blocks. This also handles potentially
1529 /// unreachable code that we don't want to explore but might be found examining
1530 /// the predecessors of the header.
1531 ///
1532 /// If the original loop is no longer a loop, this will return an empty set. If
1533 /// it remains a loop, all the blocks within it will be added to the set
1534 /// (including those blocks in inner loops).
1536  LoopInfo &LI) {
1538 
1539  auto *PH = L.getLoopPreheader();
1540  auto *Header = L.getHeader();
1541 
1542  // A worklist to use while walking backwards from the header.
1544 
1545  // First walk the predecessors of the header to find the backedges. This will
1546  // form the basis of our walk.
1547  for (auto *Pred : predecessors(Header)) {
1548  // Skip the preheader.
1549  if (Pred == PH)
1550  continue;
1551 
1552  // Because the loop was in simplified form, the only non-loop predecessor
1553  // is the preheader.
1554  assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1555  "than the preheader that is not part of the "
1556  "loop!");
1557 
1558  // Insert this block into the loop set and on the first visit and, if it
1559  // isn't the header we're currently walking, put it into the worklist to
1560  // recurse through.
1561  if (LoopBlockSet.insert(Pred).second && Pred != Header)
1562  Worklist.push_back(Pred);
1563  }
1564 
1565  // If no backedges were found, we're done.
1566  if (LoopBlockSet.empty())
1567  return LoopBlockSet;
1568 
1569  // We found backedges, recurse through them to identify the loop blocks.
1570  while (!Worklist.empty()) {
1571  BasicBlock *BB = Worklist.pop_back_val();
1572  assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1573 
1574  // No need to walk past the header.
1575  if (BB == Header)
1576  continue;
1577 
1578  // Because we know the inner loop structure remains valid we can use the
1579  // loop structure to jump immediately across the entire nested loop.
1580  // Further, because it is in loop simplified form, we can directly jump
1581  // to its preheader afterward.
1582  if (Loop *InnerL = LI.getLoopFor(BB))
1583  if (InnerL != &L) {
1584  assert(L.contains(InnerL) &&
1585  "Should not reach a loop *outside* this loop!");
1586  // The preheader is the only possible predecessor of the loop so
1587  // insert it into the set and check whether it was already handled.
1588  auto *InnerPH = InnerL->getLoopPreheader();
1589  assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1590  "but not contain the inner loop "
1591  "preheader!");
1592  if (!LoopBlockSet.insert(InnerPH).second)
1593  // The only way to reach the preheader is through the loop body
1594  // itself so if it has been visited the loop is already handled.
1595  continue;
1596 
1597  // Insert all of the blocks (other than those already present) into
1598  // the loop set. We expect at least the block that led us to find the
1599  // inner loop to be in the block set, but we may also have other loop
1600  // blocks if they were already enqueued as predecessors of some other
1601  // outer loop block.
1602  for (auto *InnerBB : InnerL->blocks()) {
1603  if (InnerBB == BB) {
1604  assert(LoopBlockSet.count(InnerBB) &&
1605  "Block should already be in the set!");
1606  continue;
1607  }
1608 
1609  LoopBlockSet.insert(InnerBB);
1610  }
1611 
1612  // Add the preheader to the worklist so we will continue past the
1613  // loop body.
1614  Worklist.push_back(InnerPH);
1615  continue;
1616  }
1617 
1618  // Insert any predecessors that were in the original loop into the new
1619  // set, and if the insert is successful, add them to the worklist.
1620  for (auto *Pred : predecessors(BB))
1621  if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)
1622  Worklist.push_back(Pred);
1623  }
1624 
1625  assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1626 
1627  // We've found all the blocks participating in the loop, return our completed
1628  // set.
1629  return LoopBlockSet;
1630 }
1631 
1632 /// Rebuild a loop after unswitching removes some subset of blocks and edges.
1633 ///
1634 /// The removal may have removed some child loops entirely but cannot have
1635 /// disturbed any remaining child loops. However, they may need to be hoisted
1636 /// to the parent loop (or to be top-level loops). The original loop may be
1637 /// completely removed.
1638 ///
1639 /// The sibling loops resulting from this update are returned. If the original
1640 /// loop remains a valid loop, it will be the first entry in this list with all
1641 /// of the newly sibling loops following it.
1642 ///
1643 /// Returns true if the loop remains a loop after unswitching, and false if it
1644 /// is no longer a loop after unswitching (and should not continue to be
1645 /// referenced).
1647  LoopInfo &LI,
1648  SmallVectorImpl<Loop *> &HoistedLoops) {
1649  auto *PH = L.getLoopPreheader();
1650 
1651  // Compute the actual parent loop from the exit blocks. Because we may have
1652  // pruned some exits the loop may be different from the original parent.
1653  Loop *ParentL = nullptr;
1654  SmallVector<Loop *, 4> ExitLoops;
1655  SmallVector<BasicBlock *, 4> ExitsInLoops;
1656  ExitsInLoops.reserve(ExitBlocks.size());
1657  for (auto *ExitBB : ExitBlocks)
1658  if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1659  ExitLoops.push_back(ExitL);
1660  ExitsInLoops.push_back(ExitBB);
1661  if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1662  ParentL = ExitL;
1663  }
1664 
1665  // Recompute the blocks participating in this loop. This may be empty if it
1666  // is no longer a loop.
1667  auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1668 
1669  // If we still have a loop, we need to re-set the loop's parent as the exit
1670  // block set changing may have moved it within the loop nest. Note that this
1671  // can only happen when this loop has a parent as it can only hoist the loop
1672  // *up* the nest.
1673  if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1674  // Remove this loop's (original) blocks from all of the intervening loops.
1675  for (Loop *IL = L.getParentLoop(); IL != ParentL;
1676  IL = IL->getParentLoop()) {
1677  IL->getBlocksSet().erase(PH);
1678  for (auto *BB : L.blocks())
1679  IL->getBlocksSet().erase(BB);
1680  llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {
1681  return BB == PH || L.contains(BB);
1682  });
1683  }
1684 
1685  LI.changeLoopFor(PH, ParentL);
1686  L.getParentLoop()->removeChildLoop(&L);
1687  if (ParentL)
1688  ParentL->addChildLoop(&L);
1689  else
1690  LI.addTopLevelLoop(&L);
1691  }
1692 
1693  // Now we update all the blocks which are no longer within the loop.
1694  auto &Blocks = L.getBlocksVector();
1695  auto BlocksSplitI =
1696  LoopBlockSet.empty()
1697  ? Blocks.begin()
1698  : std::stable_partition(
1699  Blocks.begin(), Blocks.end(),
1700  [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });
1701 
1702  // Before we erase the list of unlooped blocks, build a set of them.
1703  SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1704  if (LoopBlockSet.empty())
1705  UnloopedBlocks.insert(PH);
1706 
1707  // Now erase these blocks from the loop.
1708  for (auto *BB : make_range(BlocksSplitI, Blocks.end()))
1709  L.getBlocksSet().erase(BB);
1710  Blocks.erase(BlocksSplitI, Blocks.end());
1711 
1712  // Sort the exits in ascending loop depth, we'll work backwards across these
1713  // to process them inside out.
1714  std::stable_sort(ExitsInLoops.begin(), ExitsInLoops.end(),
1715  [&](BasicBlock *LHS, BasicBlock *RHS) {
1716  return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);
1717  });
1718 
1719  // We'll build up a set for each exit loop.
1720  SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1721  Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1722 
1723  auto RemoveUnloopedBlocksFromLoop =
1724  [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1725  for (auto *BB : UnloopedBlocks)
1726  L.getBlocksSet().erase(BB);
1727  llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {
1728  return UnloopedBlocks.count(BB);
1729  });
1730  };
1731 
1733  while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1734  assert(Worklist.empty() && "Didn't clear worklist!");
1735  assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
1736 
1737  // Grab the next exit block, in decreasing loop depth order.
1738  BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1739  Loop &ExitL = *LI.getLoopFor(ExitBB);
1740  assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
1741 
1742  // Erase all of the unlooped blocks from the loops between the previous
1743  // exit loop and this exit loop. This works because the ExitInLoops list is
1744  // sorted in increasing order of loop depth and thus we visit loops in
1745  // decreasing order of loop depth.
1746  for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
1747  RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1748 
1749  // Walk the CFG back until we hit the cloned PH adding everything reachable
1750  // and in the unlooped set to this exit block's loop.
1751  Worklist.push_back(ExitBB);
1752  do {
1753  BasicBlock *BB = Worklist.pop_back_val();
1754  // We can stop recursing at the cloned preheader (if we get there).
1755  if (BB == PH)
1756  continue;
1757 
1758  for (BasicBlock *PredBB : predecessors(BB)) {
1759  // If this pred has already been moved to our set or is part of some
1760  // (inner) loop, no update needed.
1761  if (!UnloopedBlocks.erase(PredBB)) {
1762  assert((NewExitLoopBlocks.count(PredBB) ||
1763  ExitL.contains(LI.getLoopFor(PredBB))) &&
1764  "Predecessor not in a nested loop (or already visited)!");
1765  continue;
1766  }
1767 
1768  // We just insert into the loop set here. We'll add these blocks to the
1769  // exit loop after we build up the set in a deterministic order rather
1770  // than the predecessor-influenced visit order.
1771  bool Inserted = NewExitLoopBlocks.insert(PredBB).second;
1772  (void)Inserted;
1773  assert(Inserted && "Should only visit an unlooped block once!");
1774 
1775  // And recurse through to its predecessors.
1776  Worklist.push_back(PredBB);
1777  }
1778  } while (!Worklist.empty());
1779 
1780  // If blocks in this exit loop were directly part of the original loop (as
1781  // opposed to a child loop) update the map to point to this exit loop. This
1782  // just updates a map and so the fact that the order is unstable is fine.
1783  for (auto *BB : NewExitLoopBlocks)
1784  if (Loop *BBL = LI.getLoopFor(BB))
1785  if (BBL == &L || !L.contains(BBL))
1786  LI.changeLoopFor(BB, &ExitL);
1787 
1788  // We will remove the remaining unlooped blocks from this loop in the next
1789  // iteration or below.
1790  NewExitLoopBlocks.clear();
1791  }
1792 
1793  // Any remaining unlooped blocks are no longer part of any loop unless they
1794  // are part of some child loop.
1795  for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
1796  RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1797  for (auto *BB : UnloopedBlocks)
1798  if (Loop *BBL = LI.getLoopFor(BB))
1799  if (BBL == &L || !L.contains(BBL))
1800  LI.changeLoopFor(BB, nullptr);
1801 
1802  // Sink all the child loops whose headers are no longer in the loop set to
1803  // the parent (or to be top level loops). We reach into the loop and directly
1804  // update its subloop vector to make this batch update efficient.
1805  auto &SubLoops = L.getSubLoopsVector();
1806  auto SubLoopsSplitI =
1807  LoopBlockSet.empty()
1808  ? SubLoops.begin()
1809  : std::stable_partition(
1810  SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {
1811  return LoopBlockSet.count(SubL->getHeader());
1812  });
1813  for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {
1814  HoistedLoops.push_back(HoistedL);
1815  HoistedL->setParentLoop(nullptr);
1816 
1817  // To compute the new parent of this hoisted loop we look at where we
1818  // placed the preheader above. We can't lookup the header itself because we
1819  // retained the mapping from the header to the hoisted loop. But the
1820  // preheader and header should have the exact same new parent computed
1821  // based on the set of exit blocks from the original loop as the preheader
1822  // is a predecessor of the header and so reached in the reverse walk. And
1823  // because the loops were all in simplified form the preheader of the
1824  // hoisted loop can't be part of some *other* loop.
1825  if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))
1826  NewParentL->addChildLoop(HoistedL);
1827  else
1828  LI.addTopLevelLoop(HoistedL);
1829  }
1830  SubLoops.erase(SubLoopsSplitI, SubLoops.end());
1831 
1832  // Actually delete the loop if nothing remained within it.
1833  if (Blocks.empty()) {
1834  assert(SubLoops.empty() &&
1835  "Failed to remove all subloops from the original loop!");
1836  if (Loop *ParentL = L.getParentLoop())
1837  ParentL->removeChildLoop(llvm::find(*ParentL, &L));
1838  else
1839  LI.removeLoop(llvm::find(LI, &L));
1840  LI.destroy(&L);
1841  return false;
1842  }
1843 
1844  return true;
1845 }
1846 
1847 /// Helper to visit a dominator subtree, invoking a callable on each node.
1848 ///
1849 /// Returning false at any point will stop walking past that node of the tree.
1850 template <typename CallableT>
1851 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
1852  SmallVector<DomTreeNode *, 4> DomWorklist;
1853  DomWorklist.push_back(DT[BB]);
1854 #ifndef NDEBUG
1856  Visited.insert(DT[BB]);
1857 #endif
1858  do {
1859  DomTreeNode *N = DomWorklist.pop_back_val();
1860 
1861  // Visit this node.
1862  if (!Callable(N->getBlock()))
1863  continue;
1864 
1865  // Accumulate the child nodes.
1866  for (DomTreeNode *ChildN : *N) {
1867  assert(Visited.insert(ChildN).second &&
1868  "Cannot visit a node twice when walking a tree!");
1869  DomWorklist.push_back(ChildN);
1870  }
1871  } while (!DomWorklist.empty());
1872 }
1873 
1875  Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
1877  AssumptionCache &AC, function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
1878  ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
1879  auto *ParentBB = TI.getParent();
1880  BranchInst *BI = dyn_cast<BranchInst>(&TI);
1881  SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
1882 
1883  // We can only unswitch switches, conditional branches with an invariant
1884  // condition, or combining invariant conditions with an instruction.
1885  assert((SI || BI->isConditional()) &&
1886  "Can only unswitch switches and conditional branch!");
1887  bool FullUnswitch = SI || BI->getCondition() == Invariants[0];
1888  if (FullUnswitch)
1889  assert(Invariants.size() == 1 &&
1890  "Cannot have other invariants with full unswitching!");
1891  else
1892  assert(isa<Instruction>(BI->getCondition()) &&
1893  "Partial unswitching requires an instruction as the condition!");
1894 
1895  if (MSSAU && VerifyMemorySSA)
1896  MSSAU->getMemorySSA()->verifyMemorySSA();
1897 
1898  // Constant and BBs tracking the cloned and continuing successor. When we are
1899  // unswitching the entire condition, this can just be trivially chosen to
1900  // unswitch towards `true`. However, when we are unswitching a set of
1901  // invariants combined with `and` or `or`, the combining operation determines
1902  // the best direction to unswitch: we want to unswitch the direction that will
1903  // collapse the branch.
1904  bool Direction = true;
1905  int ClonedSucc = 0;
1906  if (!FullUnswitch) {
1907  if (cast<Instruction>(BI->getCondition())->getOpcode() != Instruction::Or) {
1908  assert(cast<Instruction>(BI->getCondition())->getOpcode() ==
1909  Instruction::And &&
1910  "Only `or` and `and` instructions can combine invariants being "
1911  "unswitched.");
1912  Direction = false;
1913  ClonedSucc = 1;
1914  }
1915  }
1916 
1917  BasicBlock *RetainedSuccBB =
1918  BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();
1919  SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
1920  if (BI)
1921  UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));
1922  else
1923  for (auto Case : SI->cases())
1924  if (Case.getCaseSuccessor() != RetainedSuccBB)
1925  UnswitchedSuccBBs.insert(Case.getCaseSuccessor());
1926 
1927  assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
1928  "Should not unswitch the same successor we are retaining!");
1929 
1930  // The branch should be in this exact loop. Any inner loop's invariant branch
1931  // should be handled by unswitching that inner loop. The caller of this
1932  // routine should filter out any candidates that remain (but were skipped for
1933  // whatever reason).
1934  assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
1935 
1936  // Compute the parent loop now before we start hacking on things.
1937  Loop *ParentL = L.getParentLoop();
1938  // Get blocks in RPO order for MSSA update, before changing the CFG.
1939  LoopBlocksRPO LBRPO(&L);
1940  if (MSSAU)
1941  LBRPO.perform(&LI);
1942 
1943  // Compute the outer-most loop containing one of our exit blocks. This is the
1944  // furthest up our loopnest which can be mutated, which we will use below to
1945  // update things.
1946  Loop *OuterExitL = &L;
1947  for (auto *ExitBB : ExitBlocks) {
1948  Loop *NewOuterExitL = LI.getLoopFor(ExitBB);
1949  if (!NewOuterExitL) {
1950  // We exited the entire nest with this block, so we're done.
1951  OuterExitL = nullptr;
1952  break;
1953  }
1954  if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))
1955  OuterExitL = NewOuterExitL;
1956  }
1957 
1958  // At this point, we're definitely going to unswitch something so invalidate
1959  // any cached information in ScalarEvolution for the outer most loop
1960  // containing an exit block and all nested loops.
1961  if (SE) {
1962  if (OuterExitL)
1963  SE->forgetLoop(OuterExitL);
1964  else
1965  SE->forgetTopmostLoop(&L);
1966  }
1967 
1968  // If the edge from this terminator to a successor dominates that successor,
1969  // store a map from each block in its dominator subtree to it. This lets us
1970  // tell when cloning for a particular successor if a block is dominated by
1971  // some *other* successor with a single data structure. We use this to
1972  // significantly reduce cloning.
1974  for (auto *SuccBB : llvm::concat<BasicBlock *const>(
1975  makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs))
1976  if (SuccBB->getUniquePredecessor() ||
1977  llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
1978  return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);
1979  }))
1980  visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {
1981  DominatingSucc[BB] = SuccBB;
1982  return true;
1983  });
1984 
1985  // Split the preheader, so that we know that there is a safe place to insert
1986  // the conditional branch. We will change the preheader to have a conditional
1987  // branch on LoopCond. The original preheader will become the split point
1988  // between the unswitched versions, and we will have a new preheader for the
1989  // original loop.
1990  BasicBlock *SplitBB = L.getLoopPreheader();
1991  BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU);
1992 
1993  // Keep track of the dominator tree updates needed.
1995 
1996  // Clone the loop for each unswitched successor.
1998  VMaps.reserve(UnswitchedSuccBBs.size());
2000  for (auto *SuccBB : UnswitchedSuccBBs) {
2001  VMaps.emplace_back(new ValueToValueMapTy());
2002  ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2003  L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,
2004  DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU);
2005  }
2006 
2007  // The stitching of the branched code back together depends on whether we're
2008  // doing full unswitching or not with the exception that we always want to
2009  // nuke the initial terminator placed in the split block.
2010  SplitBB->getTerminator()->eraseFromParent();
2011  if (FullUnswitch) {
2012  // Splice the terminator from the original loop and rewrite its
2013  // successors.
2014  SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI);
2015 
2016  // Keep a clone of the terminator for MSSA updates.
2017  Instruction *NewTI = TI.clone();
2018  ParentBB->getInstList().push_back(NewTI);
2019 
2020  // First wire up the moved terminator to the preheaders.
2021  if (BI) {
2022  BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2023  BI->setSuccessor(ClonedSucc, ClonedPH);
2024  BI->setSuccessor(1 - ClonedSucc, LoopPH);
2025  DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2026  } else {
2027  assert(SI && "Must either be a branch or switch!");
2028 
2029  // Walk the cases and directly update their successors.
2030  assert(SI->getDefaultDest() == RetainedSuccBB &&
2031  "Not retaining default successor!");
2032  SI->setDefaultDest(LoopPH);
2033  for (auto &Case : SI->cases())
2034  if (Case.getCaseSuccessor() == RetainedSuccBB)
2035  Case.setSuccessor(LoopPH);
2036  else
2037  Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);
2038 
2039  // We need to use the set to populate domtree updates as even when there
2040  // are multiple cases pointing at the same successor we only want to
2041  // remove and insert one edge in the domtree.
2042  for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2043  DTUpdates.push_back(
2044  {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});
2045  }
2046 
2047  if (MSSAU) {
2048  DT.applyUpdates(DTUpdates);
2049  DTUpdates.clear();
2050 
2051  // Remove all but one edge to the retained block and all unswitched
2052  // blocks. This is to avoid having duplicate entries in the cloned Phis,
2053  // when we know we only keep a single edge for each case.
2054  MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB);
2055  for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2056  MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB);
2057 
2058  for (auto &VMap : VMaps)
2059  MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2060  /*IgnoreIncomingWithNoClones=*/true);
2061  MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2062 
2063  // Remove all edges to unswitched blocks.
2064  for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2065  MSSAU->removeEdge(ParentBB, SuccBB);
2066  }
2067 
2068  // Now unhook the successor relationship as we'll be replacing
2069  // the terminator with a direct branch. This is much simpler for branches
2070  // than switches so we handle those first.
2071  if (BI) {
2072  // Remove the parent as a predecessor of the unswitched successor.
2073  assert(UnswitchedSuccBBs.size() == 1 &&
2074  "Only one possible unswitched block for a branch!");
2075  BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2076  UnswitchedSuccBB->removePredecessor(ParentBB,
2077  /*DontDeleteUselessPHIs*/ true);
2078  DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2079  } else {
2080  // Note that we actually want to remove the parent block as a predecessor
2081  // of *every* case successor. The case successor is either unswitched,
2082  // completely eliminating an edge from the parent to that successor, or it
2083  // is a duplicate edge to the retained successor as the retained successor
2084  // is always the default successor and as we'll replace this with a direct
2085  // branch we no longer need the duplicate entries in the PHI nodes.
2086  SwitchInst *NewSI = cast<SwitchInst>(NewTI);
2087  assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2088  "Not retaining default successor!");
2089  for (auto &Case : NewSI->cases())
2090  Case.getCaseSuccessor()->removePredecessor(
2091  ParentBB,
2092  /*DontDeleteUselessPHIs*/ true);
2093 
2094  // We need to use the set to populate domtree updates as even when there
2095  // are multiple cases pointing at the same successor we only want to
2096  // remove and insert one edge in the domtree.
2097  for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2098  DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});
2099  }
2100 
2101  // After MSSAU update, remove the cloned terminator instruction NewTI.
2102  ParentBB->getTerminator()->eraseFromParent();
2103 
2104  // Create a new unconditional branch to the continuing block (as opposed to
2105  // the one cloned).
2106  BranchInst::Create(RetainedSuccBB, ParentBB);
2107  } else {
2108  assert(BI && "Only branches have partial unswitching.");
2109  assert(UnswitchedSuccBBs.size() == 1 &&
2110  "Only one possible unswitched block for a branch!");
2111  BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2112  // When doing a partial unswitch, we have to do a bit more work to build up
2113  // the branch in the split block.
2114  buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction,
2115  *ClonedPH, *LoopPH);
2116  DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2117  }
2118 
2119  // Apply the updates accumulated above to get an up-to-date dominator tree.
2120  DT.applyUpdates(DTUpdates);
2121  if (!FullUnswitch && MSSAU) {
2122  // Update MSSA for partial unswitch, after DT update.
2123  SmallVector<CFGUpdate, 1> Updates;
2124  Updates.push_back(
2125  {cfg::UpdateKind::Insert, SplitBB, ClonedPHs.begin()->second});
2126  MSSAU->applyInsertUpdates(Updates, DT);
2127  }
2128 
2129  // Now that we have an accurate dominator tree, first delete the dead cloned
2130  // blocks so that we can accurately build any cloned loops. It is important to
2131  // not delete the blocks from the original loop yet because we still want to
2132  // reference the original loop to understand the cloned loop's structure.
2133  deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2134 
2135  // Build the cloned loop structure itself. This may be substantially
2136  // different from the original structure due to the simplified CFG. This also
2137  // handles inserting all the cloned blocks into the correct loops.
2138  SmallVector<Loop *, 4> NonChildClonedLoops;
2139  for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2140  buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);
2141 
2142  // Now that our cloned loops have been built, we can update the original loop.
2143  // First we delete the dead blocks from it and then we rebuild the loop
2144  // structure taking these deletions into account.
2145  deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU);
2146 
2147  if (MSSAU && VerifyMemorySSA)
2148  MSSAU->getMemorySSA()->verifyMemorySSA();
2149 
2150  SmallVector<Loop *, 4> HoistedLoops;
2151  bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops);
2152 
2153  if (MSSAU && VerifyMemorySSA)
2154  MSSAU->getMemorySSA()->verifyMemorySSA();
2155 
2156  // This transformation has a high risk of corrupting the dominator tree, and
2157  // the below steps to rebuild loop structures will result in hard to debug
2158  // errors in that case so verify that the dominator tree is sane first.
2159  // FIXME: Remove this when the bugs stop showing up and rely on existing
2160  // verification steps.
2162 
2163  if (BI) {
2164  // If we unswitched a branch which collapses the condition to a known
2165  // constant we want to replace all the uses of the invariants within both
2166  // the original and cloned blocks. We do this here so that we can use the
2167  // now updated dominator tree to identify which side the users are on.
2168  assert(UnswitchedSuccBBs.size() == 1 &&
2169  "Only one possible unswitched block for a branch!");
2170  BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2171 
2172  // When considering multiple partially-unswitched invariants
2173  // we cant just go replace them with constants in both branches.
2174  //
2175  // For 'AND' we infer that true branch ("continue") means true
2176  // for each invariant operand.
2177  // For 'OR' we can infer that false branch ("continue") means false
2178  // for each invariant operand.
2179  // So it happens that for multiple-partial case we dont replace
2180  // in the unswitched branch.
2181  bool ReplaceUnswitched = FullUnswitch || (Invariants.size() == 1);
2182 
2183  ConstantInt *UnswitchedReplacement =
2184  Direction ? ConstantInt::getTrue(BI->getContext())
2186  ConstantInt *ContinueReplacement =
2187  Direction ? ConstantInt::getFalse(BI->getContext())
2189  for (Value *Invariant : Invariants)
2190  for (auto UI = Invariant->use_begin(), UE = Invariant->use_end();
2191  UI != UE;) {
2192  // Grab the use and walk past it so we can clobber it in the use list.
2193  Use *U = &*UI++;
2194  Instruction *UserI = dyn_cast<Instruction>(U->getUser());
2195  if (!UserI)
2196  continue;
2197 
2198  // Replace it with the 'continue' side if in the main loop body, and the
2199  // unswitched if in the cloned blocks.
2200  if (DT.dominates(LoopPH, UserI->getParent()))
2201  U->set(ContinueReplacement);
2202  else if (ReplaceUnswitched &&
2203  DT.dominates(ClonedPH, UserI->getParent()))
2204  U->set(UnswitchedReplacement);
2205  }
2206  }
2207 
2208  // We can change which blocks are exit blocks of all the cloned sibling
2209  // loops, the current loop, and any parent loops which shared exit blocks
2210  // with the current loop. As a consequence, we need to re-form LCSSA for
2211  // them. But we shouldn't need to re-form LCSSA for any child loops.
2212  // FIXME: This could be made more efficient by tracking which exit blocks are
2213  // new, and focusing on them, but that isn't likely to be necessary.
2214  //
2215  // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2216  // loop nest and update every loop that could have had its exits changed. We
2217  // also need to cover any intervening loops. We add all of these loops to
2218  // a list and sort them by loop depth to achieve this without updating
2219  // unnecessary loops.
2220  auto UpdateLoop = [&](Loop &UpdateL) {
2221 #ifndef NDEBUG
2222  UpdateL.verifyLoop();
2223  for (Loop *ChildL : UpdateL) {
2224  ChildL->verifyLoop();
2225  assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2226  "Perturbed a child loop's LCSSA form!");
2227  }
2228 #endif
2229  // First build LCSSA for this loop so that we can preserve it when
2230  // forming dedicated exits. We don't want to perturb some other loop's
2231  // LCSSA while doing that CFG edit.
2232  formLCSSA(UpdateL, DT, &LI, nullptr);
2233 
2234  // For loops reached by this loop's original exit blocks we may
2235  // introduced new, non-dedicated exits. At least try to re-form dedicated
2236  // exits for these loops. This may fail if they couldn't have dedicated
2237  // exits to start with.
2238  formDedicatedExitBlocks(&UpdateL, &DT, &LI, /*PreserveLCSSA*/ true);
2239  };
2240 
2241  // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2242  // and we can do it in any order as they don't nest relative to each other.
2243  //
2244  // Also check if any of the loops we have updated have become top-level loops
2245  // as that will necessitate widening the outer loop scope.
2246  for (Loop *UpdatedL :
2247  llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {
2248  UpdateLoop(*UpdatedL);
2249  if (!UpdatedL->getParentLoop())
2250  OuterExitL = nullptr;
2251  }
2252  if (IsStillLoop) {
2253  UpdateLoop(L);
2254  if (!L.getParentLoop())
2255  OuterExitL = nullptr;
2256  }
2257 
2258  // If the original loop had exit blocks, walk up through the outer most loop
2259  // of those exit blocks to update LCSSA and form updated dedicated exits.
2260  if (OuterExitL != &L)
2261  for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2262  OuterL = OuterL->getParentLoop())
2263  UpdateLoop(*OuterL);
2264 
2265 #ifndef NDEBUG
2266  // Verify the entire loop structure to catch any incorrect updates before we
2267  // progress in the pass pipeline.
2268  LI.verify(DT);
2269 #endif
2270 
2271  // Now that we've unswitched something, make callbacks to report the changes.
2272  // For that we need to merge together the updated loops and the cloned loops
2273  // and check whether the original loop survived.
2274  SmallVector<Loop *, 4> SibLoops;
2275  for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
2276  if (UpdatedL->getParentLoop() == ParentL)
2277  SibLoops.push_back(UpdatedL);
2278  UnswitchCB(IsStillLoop, SibLoops);
2279 
2280  if (MSSAU && VerifyMemorySSA)
2281  MSSAU->getMemorySSA()->verifyMemorySSA();
2282 
2283  if (BI)
2284  ++NumBranches;
2285  else
2286  ++NumSwitches;
2287 }
2288 
2289 /// Recursively compute the cost of a dominator subtree based on the per-block
2290 /// cost map provided.
2291 ///
2292 /// The recursive computation is memozied into the provided DT-indexed cost map
2293 /// to allow querying it for most nodes in the domtree without it becoming
2294 /// quadratic.
2295 static int
2297  const SmallDenseMap<BasicBlock *, int, 4> &BBCostMap,
2299  // Don't accumulate cost (or recurse through) blocks not in our block cost
2300  // map and thus not part of the duplication cost being considered.
2301  auto BBCostIt = BBCostMap.find(N.getBlock());
2302  if (BBCostIt == BBCostMap.end())
2303  return 0;
2304 
2305  // Lookup this node to see if we already computed its cost.
2306  auto DTCostIt = DTCostMap.find(&N);
2307  if (DTCostIt != DTCostMap.end())
2308  return DTCostIt->second;
2309 
2310  // If not, we have to compute it. We can't use insert above and update
2311  // because computing the cost may insert more things into the map.
2312  int Cost = std::accumulate(
2313  N.begin(), N.end(), BBCostIt->second, [&](int Sum, DomTreeNode *ChildN) {
2314  return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
2315  });
2316  bool Inserted = DTCostMap.insert({&N, Cost}).second;
2317  (void)Inserted;
2318  assert(Inserted && "Should not insert a node while visiting children!");
2319  return Cost;
2320 }
2321 
2322 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2323 /// making the following replacement:
2324 ///
2325 /// --code before guard--
2326 /// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2327 /// --code after guard--
2328 ///
2329 /// into
2330 ///
2331 /// --code before guard--
2332 /// br i1 %cond, label %guarded, label %deopt
2333 ///
2334 /// guarded:
2335 /// --code after guard--
2336 ///
2337 /// deopt:
2338 /// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2339 /// unreachable
2340 ///
2341 /// It also makes all relevant DT and LI updates, so that all structures are in
2342 /// valid state after this transform.
2343 static BranchInst *
2345  SmallVectorImpl<BasicBlock *> &ExitBlocks,
2346  DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
2348  LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2349  BasicBlock *CheckBB = GI->getParent();
2350 
2351  if (MSSAU && VerifyMemorySSA)
2352  MSSAU->getMemorySSA()->verifyMemorySSA();
2353 
2354  // Remove all CheckBB's successors from DomTree. A block can be seen among
2355  // successors more than once, but for DomTree it should be added only once.
2356  SmallPtrSet<BasicBlock *, 4> Successors;
2357  for (auto *Succ : successors(CheckBB))
2358  if (Successors.insert(Succ).second)
2359  DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ});
2360 
2361  Instruction *DeoptBlockTerm =
2362  SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true);
2363  BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
2364  // SplitBlockAndInsertIfThen inserts control flow that branches to
2365  // DeoptBlockTerm if the condition is true. We want the opposite.
2366  CheckBI->swapSuccessors();
2367 
2368  BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
2369  GuardedBlock->setName("guarded");
2370  CheckBI->getSuccessor(1)->setName("deopt");
2371  BasicBlock *DeoptBlock = CheckBI->getSuccessor(1);
2372 
2373  // We now have a new exit block.
2374  ExitBlocks.push_back(CheckBI->getSuccessor(1));
2375 
2376  if (MSSAU)
2377  MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI);
2378 
2379  GI->moveBefore(DeoptBlockTerm);
2381 
2382  // Add new successors of CheckBB into DomTree.
2383  for (auto *Succ : successors(CheckBB))
2384  DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ});
2385 
2386  // Now the blocks that used to be CheckBB's successors are GuardedBlock's
2387  // successors.
2388  for (auto *Succ : Successors)
2389  DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ});
2390 
2391  // Make proper changes to DT.
2392  DT.applyUpdates(DTUpdates);
2393  // Inform LI of a new loop block.
2394  L.addBasicBlockToLoop(GuardedBlock, LI);
2395 
2396  if (MSSAU) {
2397  MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI));
2398  MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::End);
2399  if (VerifyMemorySSA)
2400  MSSAU->getMemorySSA()->verifyMemorySSA();
2401  }
2402 
2403  ++NumGuards;
2404  return CheckBI;
2405 }
2406 
2407 /// Cost multiplier is a way to limit potentially exponential behavior
2408 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2409 /// candidates available. Also accounting for the number of "sibling" loops with
2410 /// the idea to account for previous unswitches that already happened on this
2411 /// cluster of loops. There was an attempt to keep this formula simple,
2412 /// just enough to limit the worst case behavior. Even if it is not that simple
2413 /// now it is still not an attempt to provide a detailed heuristic size
2414 /// prediction.
2415 ///
2416 /// TODO: Make a proper accounting of "explosion" effect for all kinds of
2417 /// unswitch candidates, making adequate predictions instead of wild guesses.
2418 /// That requires knowing not just the number of "remaining" candidates but
2419 /// also costs of unswitching for each of these candidates.
2421  Instruction &TI, Loop &L, LoopInfo &LI, DominatorTree &DT,
2423  UnswitchCandidates) {
2424 
2425  // Guards and other exiting conditions do not contribute to exponential
2426  // explosion as soon as they dominate the latch (otherwise there might be
2427  // another path to the latch remaining that does not allow to eliminate the
2428  // loop copy on unswitch).
2429  BasicBlock *Latch = L.getLoopLatch();
2430  BasicBlock *CondBlock = TI.getParent();
2431  if (DT.dominates(CondBlock, Latch) &&
2432  (isGuard(&TI) ||
2433  llvm::count_if(successors(&TI), [&L](BasicBlock *SuccBB) {
2434  return L.contains(SuccBB);
2435  }) <= 1)) {
2436  NumCostMultiplierSkipped++;
2437  return 1;
2438  }
2439 
2440  auto *ParentL = L.getParentLoop();
2441  int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2442  : std::distance(LI.begin(), LI.end()));
2443  // Count amount of clones that all the candidates might cause during
2444  // unswitching. Branch/guard counts as 1, switch counts as log2 of its cases.
2445  int UnswitchedClones = 0;
2446  for (auto Candidate : UnswitchCandidates) {
2447  Instruction *CI = Candidate.first;
2448  BasicBlock *CondBlock = CI->getParent();
2449  bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch);
2450  if (isGuard(CI)) {
2451  if (!SkipExitingSuccessors)
2452  UnswitchedClones++;
2453  continue;
2454  }
2455  int NonExitingSuccessors = llvm::count_if(
2456  successors(CondBlock), [SkipExitingSuccessors, &L](BasicBlock *SuccBB) {
2457  return !SkipExitingSuccessors || L.contains(SuccBB);
2458  });
2459  UnswitchedClones += Log2_32(NonExitingSuccessors);
2460  }
2461 
2462  // Ignore up to the "unscaled candidates" number of unswitch candidates
2463  // when calculating the power-of-two scaling of the cost. The main idea
2464  // with this control is to allow a small number of unswitches to happen
2465  // and rely more on siblings multiplier (see below) when the number
2466  // of candidates is small.
2467  unsigned ClonesPower =
2468  std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0);
2469 
2470  // Allowing top-level loops to spread a bit more than nested ones.
2471  int SiblingsMultiplier =
2472  std::max((ParentL ? SiblingsCount
2473  : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2474  1);
2475  // Compute the cost multiplier in a way that won't overflow by saturating
2476  // at an upper bound.
2477  int CostMultiplier;
2478  if (ClonesPower > Log2_32(UnswitchThreshold) ||
2479  SiblingsMultiplier > UnswitchThreshold)
2480  CostMultiplier = UnswitchThreshold;
2481  else
2482  CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower),
2483  (int)UnswitchThreshold);
2484 
2485  LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier
2486  << " (siblings " << SiblingsMultiplier << " * clones "
2487  << (1 << ClonesPower) << ")"
2488  << " for unswitch candidate: " << TI << "\n");
2489  return CostMultiplier;
2490 }
2491 
2492 static bool
2495  function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2496  ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
2497  // Collect all invariant conditions within this loop (as opposed to an inner
2498  // loop which would be handled when visiting that inner loop).
2500  UnswitchCandidates;
2501 
2502  // Whether or not we should also collect guards in the loop.
2503  bool CollectGuards = false;
2504  if (UnswitchGuards) {
2505  auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2506  Intrinsic::getName(Intrinsic::experimental_guard));
2507  if (GuardDecl && !GuardDecl->use_empty())
2508  CollectGuards = true;
2509  }
2510 
2511  for (auto *BB : L.blocks()) {
2512  if (LI.getLoopFor(BB) != &L)
2513  continue;
2514 
2515  if (CollectGuards)
2516  for (auto &I : *BB)
2517  if (isGuard(&I)) {
2518  auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0);
2519  // TODO: Support AND, OR conditions and partial unswitching.
2520  if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
2521  UnswitchCandidates.push_back({&I, {Cond}});
2522  }
2523 
2524  if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
2525  // We can only consider fully loop-invariant switch conditions as we need
2526  // to completely eliminate the switch after unswitching.
2527  if (!isa<Constant>(SI->getCondition()) &&
2528  L.isLoopInvariant(SI->getCondition()))
2529  UnswitchCandidates.push_back({SI, {SI->getCondition()}});
2530  continue;
2531  }
2532 
2533  auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
2534  if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) ||
2535  BI->getSuccessor(0) == BI->getSuccessor(1))
2536  continue;
2537 
2538  if (L.isLoopInvariant(BI->getCondition())) {
2539  UnswitchCandidates.push_back({BI, {BI->getCondition()}});
2540  continue;
2541  }
2542 
2543  Instruction &CondI = *cast<Instruction>(BI->getCondition());
2544  if (CondI.getOpcode() != Instruction::And &&
2545  CondI.getOpcode() != Instruction::Or)
2546  continue;
2547 
2548  TinyPtrVector<Value *> Invariants =
2550  if (Invariants.empty())
2551  continue;
2552 
2553  UnswitchCandidates.push_back({BI, std::move(Invariants)});
2554  }
2555 
2556  // If we didn't find any candidates, we're done.
2557  if (UnswitchCandidates.empty())
2558  return false;
2559 
2560  // Check if there are irreducible CFG cycles in this loop. If so, we cannot
2561  // easily unswitch non-trivial edges out of the loop. Doing so might turn the
2562  // irreducible control flow into reducible control flow and introduce new
2563  // loops "out of thin air". If we ever discover important use cases for doing
2564  // this, we can add support to loop unswitch, but it is a lot of complexity
2565  // for what seems little or no real world benefit.
2566  LoopBlocksRPO RPOT(&L);
2567  RPOT.perform(&LI);
2568  if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
2569  return false;
2570 
2571  SmallVector<BasicBlock *, 4> ExitBlocks;
2572  L.getUniqueExitBlocks(ExitBlocks);
2573 
2574  // We cannot unswitch if exit blocks contain a cleanuppad instruction as we
2575  // don't know how to split those exit blocks.
2576  // FIXME: We should teach SplitBlock to handle this and remove this
2577  // restriction.
2578  for (auto *ExitBB : ExitBlocks)
2579  if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI())) {
2580  dbgs() << "Cannot unswitch because of cleanuppad in exit block\n";
2581  return false;
2582  }
2583 
2584  LLVM_DEBUG(
2585  dbgs() << "Considering " << UnswitchCandidates.size()
2586  << " non-trivial loop invariant conditions for unswitching.\n");
2587 
2588  // Given that unswitching these terminators will require duplicating parts of
2589  // the loop, so we need to be able to model that cost. Compute the ephemeral
2590  // values and set up a data structure to hold per-BB costs. We cache each
2591  // block's cost so that we don't recompute this when considering different
2592  // subsets of the loop for duplication during unswitching.
2594  CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);
2596 
2597  // Compute the cost of each block, as well as the total loop cost. Also, bail
2598  // out if we see instructions which are incompatible with loop unswitching
2599  // (convergent, noduplicate, or cross-basic-block tokens).
2600  // FIXME: We might be able to safely handle some of these in non-duplicated
2601  // regions.
2602  int LoopCost = 0;
2603  for (auto *BB : L.blocks()) {
2604  int Cost = 0;
2605  for (auto &I : *BB) {
2606  if (EphValues.count(&I))
2607  continue;
2608 
2609  if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
2610  return false;
2611  if (auto CS = CallSite(&I))
2612  if (CS.isConvergent() || CS.cannotDuplicate())
2613  return false;
2614 
2615  Cost += TTI.getUserCost(&I);
2616  }
2617  assert(Cost >= 0 && "Must not have negative costs!");
2618  LoopCost += Cost;
2619  assert(LoopCost >= 0 && "Must not have negative loop costs!");
2620  BBCostMap[BB] = Cost;
2621  }
2622  LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n");
2623 
2624  // Now we find the best candidate by searching for the one with the following
2625  // properties in order:
2626  //
2627  // 1) An unswitching cost below the threshold
2628  // 2) The smallest number of duplicated unswitch candidates (to avoid
2629  // creating redundant subsequent unswitching)
2630  // 3) The smallest cost after unswitching.
2631  //
2632  // We prioritize reducing fanout of unswitch candidates provided the cost
2633  // remains below the threshold because this has a multiplicative effect.
2634  //
2635  // This requires memoizing each dominator subtree to avoid redundant work.
2636  //
2637  // FIXME: Need to actually do the number of candidates part above.
2639  // Given a terminator which might be unswitched, computes the non-duplicated
2640  // cost for that terminator.
2641  auto ComputeUnswitchedCost = [&](Instruction &TI, bool FullUnswitch) {
2642  BasicBlock &BB = *TI.getParent();
2644 
2645  int Cost = LoopCost;
2646  for (BasicBlock *SuccBB : successors(&BB)) {
2647  // Don't count successors more than once.
2648  if (!Visited.insert(SuccBB).second)
2649  continue;
2650 
2651  // If this is a partial unswitch candidate, then it must be a conditional
2652  // branch with a condition of either `or` or `and`. In that case, one of
2653  // the successors is necessarily duplicated, so don't even try to remove
2654  // its cost.
2655  if (!FullUnswitch) {
2656  auto &BI = cast<BranchInst>(TI);
2657  if (cast<Instruction>(BI.getCondition())->getOpcode() ==
2658  Instruction::And) {
2659  if (SuccBB == BI.getSuccessor(1))
2660  continue;
2661  } else {
2662  assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
2663  Instruction::Or &&
2664  "Only `and` and `or` conditions can result in a partial "
2665  "unswitch!");
2666  if (SuccBB == BI.getSuccessor(0))
2667  continue;
2668  }
2669  }
2670 
2671  // This successor's domtree will not need to be duplicated after
2672  // unswitching if the edge to the successor dominates it (and thus the
2673  // entire tree). This essentially means there is no other path into this
2674  // subtree and so it will end up live in only one clone of the loop.
2675  if (SuccBB->getUniquePredecessor() ||
2676  llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2677  return PredBB == &BB || DT.dominates(SuccBB, PredBB);
2678  })) {
2679  Cost -= computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);
2680  assert(Cost >= 0 &&
2681  "Non-duplicated cost should never exceed total loop cost!");
2682  }
2683  }
2684 
2685  // Now scale the cost by the number of unique successors minus one. We
2686  // subtract one because there is already at least one copy of the entire
2687  // loop. This is computing the new cost of unswitching a condition.
2688  // Note that guards always have 2 unique successors that are implicit and
2689  // will be materialized if we decide to unswitch it.
2690  int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
2691  assert(SuccessorsCount > 1 &&
2692  "Cannot unswitch a condition without multiple distinct successors!");
2693  return Cost * (SuccessorsCount - 1);
2694  };
2695  Instruction *BestUnswitchTI = nullptr;
2696  int BestUnswitchCost;
2697  ArrayRef<Value *> BestUnswitchInvariants;
2698  for (auto &TerminatorAndInvariants : UnswitchCandidates) {
2699  Instruction &TI = *TerminatorAndInvariants.first;
2700  ArrayRef<Value *> Invariants = TerminatorAndInvariants.second;
2701  BranchInst *BI = dyn_cast<BranchInst>(&TI);
2702  int CandidateCost = ComputeUnswitchedCost(
2703  TI, /*FullUnswitch*/ !BI || (Invariants.size() == 1 &&
2704  Invariants[0] == BI->getCondition()));
2705  // Calculate cost multiplier which is a tool to limit potentially
2706  // exponential behavior of loop-unswitch.
2708  int CostMultiplier =
2709  calculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
2710  assert(
2711  (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
2712  "cost multiplier needs to be in the range of 1..UnswitchThreshold");
2713  CandidateCost *= CostMultiplier;
2714  LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
2715  << " (multiplier: " << CostMultiplier << ")"
2716  << " for unswitch candidate: " << TI << "\n");
2717  } else {
2718  LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
2719  << " for unswitch candidate: " << TI << "\n");
2720  }
2721 
2722  if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) {
2723  BestUnswitchTI = &TI;
2724  BestUnswitchCost = CandidateCost;
2725  BestUnswitchInvariants = Invariants;
2726  }
2727  }
2728 
2729  if (BestUnswitchCost >= UnswitchThreshold) {
2730  LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: "
2731  << BestUnswitchCost << "\n");
2732  return false;
2733  }
2734 
2735  // If the best candidate is a guard, turn it into a branch.
2736  if (isGuard(BestUnswitchTI))
2737  BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L,
2738  ExitBlocks, DT, LI, MSSAU);
2739 
2740  LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = "
2741  << BestUnswitchCost << ") terminator: " << *BestUnswitchTI
2742  << "\n");
2743  unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants,
2744  ExitBlocks, DT, LI, AC, UnswitchCB, SE, MSSAU);
2745  return true;
2746 }
2747 
2748 /// Unswitch control flow predicated on loop invariant conditions.
2749 ///
2750 /// This first hoists all branches or switches which are trivial (IE, do not
2751 /// require duplicating any part of the loop) out of the loop body. It then
2752 /// looks at other loop invariant control flows and tries to unswitch those as
2753 /// well by cloning the loop if the result is small enough.
2754 ///
2755 /// The `DT`, `LI`, `AC`, `TTI` parameters are required analyses that are also
2756 /// updated based on the unswitch.
2757 /// The `MSSA` analysis is also updated if valid (i.e. its use is enabled).
2758 ///
2759 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
2760 /// true, we will attempt to do non-trivial unswitching as well as trivial
2761 /// unswitching.
2762 ///
2763 /// The `UnswitchCB` callback provided will be run after unswitching is
2764 /// complete, with the first parameter set to `true` if the provided loop
2765 /// remains a loop, and a list of new sibling loops created.
2766 ///
2767 /// If `SE` is non-null, we will update that analysis based on the unswitching
2768 /// done.
2769 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
2771  bool NonTrivial,
2772  function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2773  ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
2774  assert(L.isRecursivelyLCSSAForm(DT, LI) &&
2775  "Loops must be in LCSSA form before unswitching.");
2776  bool Changed = false;
2777 
2778  // Must be in loop simplified form: we need a preheader and dedicated exits.
2779  if (!L.isLoopSimplifyForm())
2780  return false;
2781 
2782  // Try trivial unswitch first before loop over other basic blocks in the loop.
2783  if (unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
2784  // If we unswitched successfully we will want to clean up the loop before
2785  // processing it further so just mark it as unswitched and return.
2786  UnswitchCB(/*CurrentLoopValid*/ true, {});
2787  return true;
2788  }
2789 
2790  // If we're not doing non-trivial unswitching, we're done. We both accept
2791  // a parameter but also check a local flag that can be used for testing
2792  // a debugging.
2793  if (!NonTrivial && !EnableNonTrivialUnswitch)
2794  return false;
2795 
2796  // For non-trivial unswitching, because it often creates new loops, we rely on
2797  // the pass manager to iterate on the loops rather than trying to immediately
2798  // reach a fixed point. There is no substantial advantage to iterating
2799  // internally, and if any of the new loops are simplified enough to contain
2800  // trivial unswitching we want to prefer those.
2801 
2802  // Try to unswitch the best invariant condition. We prefer this full unswitch to
2803  // a partial unswitch when possible below the threshold.
2804  if (unswitchBestCondition(L, DT, LI, AC, TTI, UnswitchCB, SE, MSSAU))
2805  return true;
2806 
2807  // No other opportunities to unswitch.
2808  return Changed;
2809 }
2810 
2813  LPMUpdater &U) {
2814  Function &F = *L.getHeader()->getParent();
2815  (void)F;
2816 
2817  LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
2818  << "\n");
2819 
2820  // Save the current loop name in a variable so that we can report it even
2821  // after it has been deleted.
2822  std::string LoopName = L.getName();
2823 
2824  auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid,
2825  ArrayRef<Loop *> NewLoops) {
2826  // If we did a non-trivial unswitch, we have added new (cloned) loops.
2827  if (!NewLoops.empty())
2828  U.addSiblingLoops(NewLoops);
2829 
2830  // If the current loop remains valid, we should revisit it to catch any
2831  // other unswitch opportunities. Otherwise, we need to mark it as deleted.
2832  if (CurrentLoopValid)
2833  U.revisitCurrentLoop();
2834  else
2835  U.markLoopAsDeleted(L, LoopName);
2836  };
2837 
2839  if (AR.MSSA) {
2840  MSSAU = MemorySSAUpdater(AR.MSSA);
2841  if (VerifyMemorySSA)
2842  AR.MSSA->verifyMemorySSA();
2843  }
2844  if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.TTI, NonTrivial, UnswitchCB,
2845  &AR.SE, MSSAU.hasValue() ? MSSAU.getPointer() : nullptr))
2846  return PreservedAnalyses::all();
2847 
2848  if (AR.MSSA && VerifyMemorySSA)
2849  AR.MSSA->verifyMemorySSA();
2850 
2851  // Historically this pass has had issues with the dominator tree so verify it
2852  // in asserts builds.
2855 }
2856 
2857 namespace {
2858 
2859 class SimpleLoopUnswitchLegacyPass : public LoopPass {
2860  bool NonTrivial;
2861 
2862 public:
2863  static char ID; // Pass ID, replacement for typeid
2864 
2865  explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false)
2866  : LoopPass(ID), NonTrivial(NonTrivial) {
2869  }
2870 
2871  bool runOnLoop(Loop *L, LPPassManager &LPM) override;
2872 
2873  void getAnalysisUsage(AnalysisUsage &AU) const override {
2879  }
2881  }
2882 };
2883 
2884 } // end anonymous namespace
2885 
2886 bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
2887  if (skipLoop(L))
2888  return false;
2889 
2890  Function &F = *L->getHeader()->getParent();
2891 
2892  LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L
2893  << "\n");
2894 
2895  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2896  auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2897  auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2898  auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2899  MemorySSA *MSSA = nullptr;
2902  MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
2903  MSSAU = MemorySSAUpdater(MSSA);
2904  }
2905 
2906  auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
2907  auto *SE = SEWP ? &SEWP->getSE() : nullptr;
2908 
2909  auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid,
2910  ArrayRef<Loop *> NewLoops) {
2911  // If we did a non-trivial unswitch, we have added new (cloned) loops.
2912  for (auto *NewL : NewLoops)
2913  LPM.addLoop(*NewL);
2914 
2915  // If the current loop remains valid, re-add it to the queue. This is
2916  // a little wasteful as we'll finish processing the current loop as well,
2917  // but it is the best we can do in the old PM.
2918  if (CurrentLoopValid)
2919  LPM.addLoop(*L);
2920  else
2921  LPM.markLoopAsDeleted(*L);
2922  };
2923 
2924  if (MSSA && VerifyMemorySSA)
2925  MSSA->verifyMemorySSA();
2926 
2927  bool Changed = unswitchLoop(*L, DT, LI, AC, TTI, NonTrivial, UnswitchCB, SE,
2928  MSSAU.hasValue() ? MSSAU.getPointer() : nullptr);
2929 
2930  if (MSSA && VerifyMemorySSA)
2931  MSSA->verifyMemorySSA();
2932 
2933  // If anything was unswitched, also clear any cached information about this
2934  // loop.
2935  LPM.deleteSimpleAnalysisLoop(L);
2936 
2937  // Historically this pass has had issues with the dominator tree so verify it
2938  // in asserts builds.
2940 
2941  return Changed;
2942 }
2943 
2945 INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
2946  "Simple unswitch loops", false, false)
2953 INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
2954  "Simple unswitch loops", false, false)
2955 
2957  return new SimpleLoopUnswitchLegacyPass(NonTrivial);
2958 }
Pass interface - Implemented by all &#39;passes&#39;.
Definition: Pass.h:80
const T & front() const
front - Get the first element.
Definition: ArrayRef.h:151
static void collectEphemeralValues(const Loop *L, AssumptionCache *AC, SmallPtrSetImpl< const Value *> &EphValues)
Collect a loop&#39;s ephemeral values (those used only by an assume or similar intrinsics in the loop)...
Definition: CodeMetrics.cpp:71
static cl::opt< int > UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden, cl::desc("The cost threshold for unswitching a loop."))
void destroy(LoopT *L)
Destroy a loop that has been removed from the LoopInfo nest.
Definition: LoopInfo.h:787
unsigned getNumCases() const
Return the number of &#39;cases&#39; in this switch instruction, excluding the default case.
static cl::opt< bool > UnswitchGuards("simple-loop-unswitch-guards", cl::init(true), cl::Hidden, cl::desc("If enabled, simple loop unswitching will also consider " "llvm.experimental.guard intrinsics as unswitch candidates."))
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
static BasicBlock * buildClonedLoopBlocks(Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB, ArrayRef< BasicBlock *> ExitBlocks, BasicBlock *ParentBB, BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB, const SmallDenseMap< BasicBlock *, BasicBlock *, 16 > &DominatingSucc, ValueToValueMapTy &VMap, SmallVectorImpl< DominatorTree::UpdateType > &DTUpdates, AssumptionCache &AC, DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU)
Build the cloned blocks for an unswitched copy of the given loop.
use_iterator use_end()
Definition: Value.h:346
This routine provides some synthesis utilities to produce sequences of values.
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:584
BranchInst * CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, MDNode *BranchWeights=nullptr, MDNode *Unpredictable=nullptr)
Create a conditional &#39;br Cond, TrueDest, FalseDest&#39; instruction.
Definition: IRBuilder.h:853
CaseIt case_end()
Returns a read/write iterator that points one past the last in the SwitchInst.
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
void removePredecessor(BasicBlock *Pred, bool DontDeleteUselessPHIs=false)
Notify the BasicBlock that the predecessor Pred is no longer able to reach it.
Definition: BasicBlock.cpp:301
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
Definition: LoopInfoImpl.h:224
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
static cl::opt< bool > EnableUnswitchCostMultiplier("enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden, cl::desc("Enable unswitch cost multiplier that prohibits exponential " "explosion in nontrivial unswitch."))
static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, LoopInfo &LI, ScalarEvolution *SE, MemorySSAUpdater *MSSAU)
This routine scans the loop to find a branch or switch which occurs before any side effects occur...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
iterator_range< CaseIt > cases()
Iteration adapter for range-for loops.
std::vector< BlockT * > & getBlocksVector()
Return a direct, mutable handle to the blocks vector so that we can mutate it efficiently with techni...
Definition: LoopInfo.h:169
PreservedAnalyses getLoopPassPreservedAnalyses()
Returns the minimum set of Analyses that all loop passes must preserve.
bool VerifyMemorySSA
Enables verification of MemorySSA.
Definition: MemorySSA.cpp:82
This class represents lattice values for constants.
Definition: AllocatorList.h:23
void swapSuccessors()
Swap the successors of this branch instruction.
size_type size() const
Determine the number of elements in the SetVector.
Definition: SetVector.h:77
void updateExitBlocksForClonedLoop(ArrayRef< BasicBlock *> ExitBlocks, const ValueToValueMapTy &VMap, DominatorTree &DT)
Update phi nodes in exit block successors following cloning.
iterator begin() const
Definition: ArrayRef.h:136
void removeDuplicatePhiEdgesBetween(BasicBlock *From, BasicBlock *To)
Update the MemoryPhi in To to have a single incoming edge from From, following a CFG change that repl...
simple loop unswitch
bool isRecursivelyLCSSAForm(DominatorTree &DT, const LoopInfo &LI) const
Return true if this Loop and all inner subloops are in LCSSA form.
Definition: LoopInfo.cpp:183
void applyInsertUpdates(ArrayRef< CFGUpdate > Updates, DominatorTree &DT)
Apply CFG insert updates, analogous with the DT edge updates.
unsigned getLoopDepth(const BlockT *BB) const
Return the loop nesting level of the specified block.
Definition: LoopInfo.h:696
unsigned getLoopDepth() const
Return the nesting level of this loop.
Definition: LoopInfo.h:91
void reserveBlocks(unsigned size)
interface to do reserve() for Blocks
Definition: LoopInfo.h:371
LoopT * removeChildLoop(iterator I)
This removes the specified child from being a subloop of this loop.
Definition: LoopInfo.h:339
The main scalar evolution driver.
void removeBlocks(const SmallPtrSetImpl< BasicBlock *> &DeadBlocks)
Remove all MemoryAcceses in a set of BasicBlocks about to be deleted.
BlockT * getLoopPreheader() const
If there is a preheader for this loop, return it.
Definition: LoopInfoImpl.h:173
An immutable pass that tracks lazily created AssumptionCache objects.
Value * getCondition() const
CaseIt case_begin()
Returns a read/write iterator that points to the first case in the SwitchInst.
An efficient, type-erasing, non-owning reference to a callable.
Definition: STLExtras.h:116
A cache of @llvm.assume calls within a function.
Represents a read-write access to memory, whether it is a must-alias, or a may-alias.
Definition: MemorySSA.h:372
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:704
auto count_if(R &&Range, UnaryPredicate P) -> typename std::iterator_traits< decltype(adl_begin(Range))>::difference_type
Wrapper function around std::count_if to count the number of times an element satisfying a given pred...
Definition: STLExtras.h:1259
unsigned second
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1185
std::vector< LoopT * > & getSubLoopsVector()
Definition: LoopInfo.h:134
BasicBlock * getSuccessor(unsigned i) const
STATISTIC(NumFunctions, "Total number of functions")
void setArgOperand(unsigned i, Value *v)
Definition: InstrTypes.h:1139
The adaptor from a function pass to a loop pass computes these analyses and makes them available to t...
F(f)
static int computeDomSubtreeCost(DomTreeNode &N, const SmallDenseMap< BasicBlock *, int, 4 > &BBCostMap, SmallDenseMap< DomTreeNode *, int, 4 > &DTCostMap)
Recursively compute the cost of a dominator subtree based on the per-block cost map provided...
Value * getCondition() const
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.cpp:137
This defines the Use class.
static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB, BasicBlock &ExitBB)
Check that all the LCSSA PHI nodes in the loop exit block have trivial incoming values along this edg...
void reserve(size_type N)
Definition: SmallVector.h:374
TinyPtrVector - This class is specialized for cases where there are normally 0 or 1 element in a vect...
Definition: TinyPtrVector.h:30
bool isReachableFromEntry(const Use &U) const
Provide an overload for a Use.
Definition: Dominators.cpp:299
PreservedAnalyses run(Loop &L, LoopAnalysisManager &AM, LoopStandardAnalysisResults &AR, LPMUpdater &U)
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
Definition: SmallPtrSet.h:343
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:268
static void deleteDeadBlocksFromLoop(Loop &L, SmallVectorImpl< BasicBlock *> &ExitBlocks, DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU)
BasicBlock * SplitEdge(BasicBlock *From, BasicBlock *To, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr)
Split the edge connecting specified block.
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1134
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:221
void removeEdge(BasicBlock *From, BasicBlock *To)
Update the MemoryPhi in To following an edge deletion between From and To.
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
void initializeSimpleLoopUnswitchLegacyPassPass(PassRegistry &)
Legacy analysis pass which computes MemorySSA.
Definition: MemorySSA.h:955
MemorySSA * getMemorySSA() const
Get handle on MemorySSA.
StringRef getName(ID id)
Return the LLVM name for an intrinsic, such as "llvm.ppc.altivec.lvx".
Definition: Function.cpp:625
bool verify(VerificationLevel VL=VerificationLevel::Full) const
verify - checks if the tree is correct.
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:80
static constexpr UpdateKind Delete
LoopT * getLoopFor(const BlockT *BB) const
Return the inner most loop that BB lives in.
Definition: LoopInfo.h:689
static void unswitchNontrivialInvariants(Loop &L, Instruction &TI, ArrayRef< Value *> Invariants, SmallVectorImpl< BasicBlock *> &ExitBlocks, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, function_ref< void(bool, ArrayRef< Loop *>)> UnswitchCB, ScalarEvolution *SE, MemorySSAUpdater *MSSAU)
SmallPtrSetImpl< const BlockT * > & getBlocksSet()
Return a direct, mutable handle to the blocks set so that we can mutate it efficiently.
Definition: LoopInfo.h:175
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:450
void deleteSimpleAnalysisLoop(Loop *L)
Invoke deleteAnalysisLoop hook for all passes that implement simple analysis interface.
Definition: LoopPass.cpp:118
A Use represents the edge between a Value definition and its users.
Definition: Use.h:55
static Optional< unsigned > getOpcode(ArrayRef< VPValue *> Values)
Returns the opcode of Values or ~0 if they do not all agree.
Definition: VPlanSLP.cpp:196
Encapsulates MemorySSA, including all data associated with memory accesses.
Definition: MemorySSA.h:700
void applyUpdates(ArrayRef< CFGUpdate > Updates, DominatorTree &DT)
Apply CFG updates, analogous with the DT edge updates.
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:742
void deleteEdge(NodeT *From, NodeT *To)
Inform the dominator tree about a CFG edge deletion and update the tree.
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:284
BlockT * getHeader() const
Definition: LoopInfo.h:99
void getExitBlocks(SmallVectorImpl< BlockT *> &ExitBlocks) const
Return all of the successor blocks of this loop.
Definition: LoopInfoImpl.h:62
auto reverse(ContainerTy &&C, typename std::enable_if< has_rbegin< ContainerTy >::value >::type *=nullptr) -> decltype(make_range(C.rbegin(), C.rend()))
Definition: STLExtras.h:266
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Fast - This calling convention attempts to make calls as fast as possible (e.g.
Definition: CallingConv.h:42
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:141
MemoryUseOrDef * getMemoryAccess(const Instruction *I) const
Given a memory Mod/Ref&#39;ing instruction, get the MemorySSA access associated with it.
Definition: MemorySSA.h:712
void addBasicBlockToLoop(BlockT *NewBB, LoopInfoBase< BlockT, LoopT > &LI)
This method is used by other analyses to update loop information.
Definition: LoopInfoImpl.h:250
void addTopLevelLoop(LoopT *New)
This adds the specified loop to the collection of top-level loops.
Definition: LoopInfo.h:740
This header provides classes for managing per-loop analyses.
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:428
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:170
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:290
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
void insertEdge(NodeT *From, NodeT *To)
Inform the dominator tree about a CFG edge insertion and update the tree.
static SmallPtrSet< const BasicBlock *, 16 > recomputeLoopBlockSet(Loop &L, LoopInfo &LI)
Recompute the set of blocks in a loop after unswitching.
If this flag is set, the remapper knows that only local values within a function (such as an instruct...
Definition: ValueMapper.h:72
size_type count(const key_type &key) const
Count the number of elements of a given key in the SetVector.
Definition: SetVector.h:210
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1181
void applyUpdates(ArrayRef< UpdateType > Updates)
Inform the dominator tree about a sequence of CFG edge insertions and deletions and perform a batch u...
NodeT * getBlock() const
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:422
ValueMap< const Value *, WeakTrackingVH > ValueToValueMapTy
static void buildClonedLoops(Loop &OrigL, ArrayRef< BasicBlock *> ExitBlocks, const ValueToValueMapTy &VMap, LoopInfo &LI, SmallVectorImpl< Loop *> &NonChildClonedLoops)
Build the cloned loops of an original loop from unswitching.
Wrapper pass for TargetTransformInfo.
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
static constexpr UpdateKind Insert
static TinyPtrVector< Value * > collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root, LoopInfo &LI)
Collect all of the loop invariant input values transitively used by the homogeneous instruction graph...
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
void push_back(EltTy NewVal)
Conditional or Unconditional Branch instruction.
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:148
This is an important base class in LLVM.
Definition: Constant.h:41
LLVM_NODISCARD bool empty() const
Definition: SmallPtrSet.h:91
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:128
This file contains the declarations for the subclasses of Constant, which represent the different fla...
iterator end() const
Definition: LoopInfo.h:665
INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch", "Simple unswitch loops", false, false) INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass
const Instruction & front() const
Definition: BasicBlock.h:280
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:370
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:561
BasicBlock * getDefaultDest() const
void moveAllAfterSpliceBlocks(BasicBlock *From, BasicBlock *To, Instruction *Start)
From block was spliced into From and To.
Represent the analysis usage information of a pass.
void splice(iterator where, iplist_impl &L2)
Definition: ilist.h:328
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:1192
static Loop * cloneLoopNest(Loop &OrigRootL, Loop *RootParentL, const ValueToValueMapTy &VMap, LoopInfo &LI)
Recursively clone the specified loop and all of its children.
const T * getPointer() const
Definition: Optional.h:154
bool empty() const
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:381
bool pred_empty(const BasicBlock *BB)
Definition: CFG.h:116
void updateForClonedLoop(const LoopBlocksRPO &LoopBlocks, ArrayRef< BasicBlock *> ExitBlocks, const ValueToValueMapTy &VM, bool IgnoreIncomingWithNoClones=false)
Update MemorySSA after a loop was cloned, given the blocks in RPO order, the exit blocks and a 1:1 ma...
CaseIt removeCase(CaseIt I)
This method removes the specified case and its successor from the switch instruction.
detail::zippy< detail::zip_first, T, U, Args... > zip_first(T &&t, U &&u, Args &&... args)
zip iterator that, for the sake of efficiency, assumes the first iteratee to be the shortest...
Definition: STLExtras.h:669
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
size_t size() const
Definition: SmallVector.h:52
DomTreeNodeBase< NodeT > * getNode(const NodeT *BB) const
getNode - return the (Post)DominatorTree node for the specified basic block.
auto find(R &&Range, const T &Val) -> decltype(adl_begin(Range))
Provide wrappers to std::find which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1206
static BranchInst * turnGuardIntoBranch(IntrinsicInst *GI, Loop &L, SmallVectorImpl< BasicBlock *> &ExitBlocks, DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU)
Turns a llvm.experimental.guard intrinsic into implicit control flow branch, making the following rep...
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
static void buildPartialUnswitchConditionalBranch(BasicBlock &BB, ArrayRef< Value *> Invariants, bool Direction, BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc)
Insert code to test a set of loop invariant values, and conditionally branch on them.
void moveToPlace(MemoryUseOrDef *What, BasicBlock *BB, MemorySSA::InsertionPlace Where)
bool isLoopInvariant(const Value *V) const
Return true if the specified value is loop invariant.
Definition: LoopInfo.cpp:57
void setSuccessor(unsigned idx, BasicBlock *NewSucc)
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:1115
bool formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution *SE)
Put loop into LCSSA form.
Definition: LCSSA.cpp:304
size_type size() const
Definition: SmallPtrSet.h:92
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
This class provides an interface for updating the loop pass manager based on mutations to the loop ne...
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:333
bool contains(const LoopT *L) const
Return true if the specified loop is contained within in this loop.
Definition: LoopInfo.h:109
A SetVector that performs no allocations if smaller than a certain size.
Definition: SetVector.h:297
static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, LoopInfo &LI, ScalarEvolution *SE, MemorySSAUpdater *MSSAU)
Unswitch a trivial switch if the condition is loop invariant.
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:417
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
void verifyMemorySSA() const
Verify that MemorySSA is self consistent (IE definitions dominate all uses, uses appear in the right ...
Definition: MemorySSA.cpp:1775
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
bool erase(PtrType Ptr)
erase - If the set contains the specified pointer, remove it and return true, otherwise return false...
Definition: SmallPtrSet.h:377
static void replaceLoopInvariantUses(Loop &L, Value *Invariant, Constant &Replacement)
iterator end()
Definition: BasicBlock.h:270
static SwitchInst * Create(Value *Value, BasicBlock *Default, unsigned NumCases, Instruction *InsertBefore=nullptr)
static cl::opt< bool > EnableNonTrivialUnswitch("enable-nontrivial-unswitch", cl::init(false), cl::Hidden, cl::desc("Forcibly enables non-trivial loop unswitching rather than " "following the configuration passed into the pass."))
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:845
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:248
static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, BasicBlock &OldExitingBB, BasicBlock &OldPH)
Rewrite the PHI nodes in an unswitched loop exit basic block.
iterator end() const
Definition: ArrayRef.h:137
LoopT * removeLoop(iterator I)
This removes the specified top-level loop from this loop info object.
Definition: LoopInfo.h:710
LoopT * AllocateLoop(ArgsTy &&... Args)
Definition: LoopInfo.h:653
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:379
simple loop Simple unswitch loops
Pass * createSimpleLoopUnswitchLegacyPass(bool NonTrivial=false)
Create the legacy pass object for the simple loop unswitcher.
static BranchInst * Create(BasicBlock *IfTrue, Instruction *InsertBefore=nullptr)
bool isConditional() const
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:124
void markLoopAsDeleted(Loop &L)
Definition: LoopPass.cpp:142
static void deleteDeadClonedBlocks(Loop &L, ArrayRef< BasicBlock *> ExitBlocks, ArrayRef< std::unique_ptr< ValueToValueMapTy >> VMaps, DominatorTree &DT, MemorySSAUpdater *MSSAU)
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:577
void erase_if(Container &C, UnaryPredicate P)
Provide a container algorithm similar to C++ Library Fundamentals v2&#39;s erase_if which is equivalent t...
Definition: STLExtras.h:1329
bool isGuard(const User *U)
Returns true iff U has semantics of a guard.
Definition: GuardUtils.cpp:17
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
unsigned Log2_32(uint32_t Value)
Return the floor log base 2 of the specified value, -1 if the value is zero.
Definition: MathExtras.h:538
Function * getFunction(StringRef Name) const
Look up the specified function in the module symbol table.
Definition: Module.cpp:175
iterator begin() const
Definition: LoopInfo.h:664
BasicBlock * CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap, const Twine &NameSuffix="", Function *F=nullptr, ClonedCodeInfo *CodeInfo=nullptr, DebugInfoFinder *DIFinder=nullptr)
Return a copy of the specified basic block, but without embedding the block into a particular functio...
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:250
static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader, DominatorTree &DT, LoopInfo &LI)
Hoist the current loop up to the innermost loop containing a remaining exit.
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:392
If this flag is set, the remapper ignores missing function-local entries (Argument, Instruction, BasicBlock) that are not in the value map.
Definition: ValueMapper.h:90
Instruction * SplitBlockAndInsertIfThen(Value *Cond, Instruction *SplitBefore, bool Unreachable, MDNode *BranchWeights=nullptr, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr)
Split the containing block at the specified instruction - everything before SplitBefore stays in the ...
static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, TargetTransformInfo &TTI, function_ref< void(bool, ArrayRef< Loop *>)> UnswitchCB, ScalarEvolution *SE, MemorySSAUpdater *MSSAU)
LoopT * getParentLoop() const
Definition: LoopInfo.h:100
use_iterator use_begin()
Definition: Value.h:338
unsigned getNumBlocks() const
Get the number of blocks in this loop in constant time.
Definition: LoopInfo.h:162
bool hasValue() const
Definition: Optional.h:166
bool isLoopSimplifyForm() const
Return true if the Loop is in the form that the LoopSimplify form transforms loops to...
Definition: LoopInfo.cpp:192
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:132
void registerAssumption(CallInst *CI)
Add an @llvm.assume intrinsic to this function&#39;s cache.
void forgetLoop(const Loop *L)
This method should be called by the client when it has changed a loop in a way that may effect Scalar...
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:650
void addChildLoop(LoopT *NewChild)
Add the specified loop to be a child of this loop.
Definition: LoopInfo.h:330
static cl::opt< int > UnswitchSiblingsToplevelDiv("unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden, cl::desc("Toplevel siblings divisor for cost multiplier."))
int getUserCost(const User *U, ArrayRef< const Value *> Operands) const
Estimate the cost of a given IR user when lowered.
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
StringRef getName() const
Definition: LoopInfo.h:588
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:464
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:213
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
SymbolTableList< BasicBlock >::iterator eraseFromParent()
Unlink &#39;this&#39; from the containing function and delete it.
Definition: BasicBlock.cpp:114
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
void addLoop(Loop &L)
Definition: LoopPass.cpp:76
void getLoopAnalysisUsage(AnalysisUsage &AU)
Helper to consistently add the set of standard passes to a loop pass&#39;s AnalysisUsage.
Definition: LoopUtils.cpp:131
iterator_range< value_op_iterator > operand_values()
Definition: User.h:261
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:322
void changeLoopFor(BlockT *BB, LoopT *L)
Change the top-level loop that contains BB to the specified loop.
Definition: LoopInfo.h:721
iterator_range< const_phi_iterator > phis() const
Returns a range that iterates over the phis in the basic block.
Definition: BasicBlock.h:324
size_type count(const_arg_type_t< KeyT > Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: DenseMap.h:171
static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, BasicBlock &UnswitchedBB, BasicBlock &OldExitingBB, BasicBlock &OldPH, bool FullUnswitch)
Rewrite the PHI nodes in the loop exit basic block and the split off unswitched block.
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:211
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1163
Wrapper class to LoopBlocksDFS that provides a standard begin()/end() interface for the DFS reverse p...
Definition: LoopIterator.h:172
bool empty() const
Definition: LoopInfo.h:145
Multiway switch.
size_type count(const KeyT &Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: ValueMap.h:157
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
LLVM Value Representation.
Definition: Value.h:72
static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, TargetTransformInfo &TTI, bool NonTrivial, function_ref< void(bool, ArrayRef< Loop *>)> UnswitchCB, ScalarEvolution *SE, MemorySSAUpdater *MSSAU)
Unswitch control flow predicated on loop invariant conditions.
void setDefaultDest(BasicBlock *DefaultCase)
succ_range successors(Instruction *I)
Definition: CFG.h:259
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:86
BasicBlock * SplitBlock(BasicBlock *Old, Instruction *SplitPt, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr)
Split the specified block at the specified instruction - everything before SplitPt stays in Old and e...
bool formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI, bool PreserveLCSSA)
Ensure that all exit blocks of the loop are dedicated exits.
Definition: LoopUtils.cpp:48
This file exposes an interface to building/using memory SSA to walk memory instructions using a use/d...
The legacy pass manager&#39;s analysis pass to compute loop information.
Definition: LoopInfo.h:969
void getUniqueExitBlocks(SmallVectorImpl< BlockT *> &ExitBlocks) const
Return all unique successor blocks of this loop.
Definition: LoopInfoImpl.h:99
This file defines a set of templates that efficiently compute a dominator tree over a generic graph...
static int calculateUnswitchCostMultiplier(Instruction &TI, Loop &L, LoopInfo &LI, DominatorTree &DT, ArrayRef< std::pair< Instruction *, TinyPtrVector< Value *>>> UnswitchCandidates)
Cost multiplier is a way to limit potentially exponential behavior of loop-unswitch.
A container for analyses that lazily runs them and caches their results.
static cl::opt< int > UnswitchNumInitialUnscaledCandidates("unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden, cl::desc("Number of unswitch candidates that are ignored when calculating " "cost multiplier."))
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
void perform(LoopInfo *LI)
Traverse the loop blocks and store the DFS result.
Definition: LoopIterator.h:180
#define LLVM_DEBUG(X)
Definition: Debug.h:122
iterator_range< block_iterator > blocks() const
Definition: LoopInfo.h:155
static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef< BasicBlock *> ExitBlocks, LoopInfo &LI, SmallVectorImpl< Loop *> &HoistedLoops)
Rebuild a loop after unswitching removes some subset of blocks and edges.
void moveBefore(BasicBlock *MovePos)
Unlink this basic block from its current function and insert it into the function that MovePos lives ...
Definition: BasicBlock.cpp:120
void dropAllReferences()
Cause all subinstructions to "let go" of all the references that said subinstructions are maintaining...
Definition: BasicBlock.cpp:226
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
const BasicBlock * getParent() const
Definition: Instruction.h:66
void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable)
Helper to visit a dominator subtree, invoking a callable on each node.
static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, LoopInfo &LI, ScalarEvolution *SE, MemorySSAUpdater *MSSAU)
Unswitch a trivial branch if the condition is loop invariant.
void forgetTopmostLoop(const Loop *L)
cl::opt< bool > EnableMSSALoopDependency
Enables memory ssa as a dependency for loop passes.