1 ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
11 #include "llvm/ADT/DenseMap.h"
12 #include "llvm/ADT/STLExtras.h"
13 #include "llvm/ADT/Sequence.h"
14 #include "llvm/ADT/SetVector.h"
15 #include "llvm/ADT/SmallPtrSet.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/Twine.h"
19 #include "llvm/Analysis/AssumptionCache.h"
20 #include "llvm/Analysis/CFG.h"
21 #include "llvm/Analysis/CodeMetrics.h"
22 #include "llvm/Analysis/GuardUtils.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LoopAnalysisManager.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/LoopIterator.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/MemorySSA.h"
29 #include "llvm/Analysis/MemorySSAUpdater.h"
30 #include "llvm/Analysis/Utils/Local.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/Constant.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/InstrTypes.h"
37 #include "llvm/IR/Instruction.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Use.h"
41 #include "llvm/IR/Value.h"
42 #include "llvm/Pass.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/ErrorHandling.h"
46 #include "llvm/Support/GenericDomTree.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
49 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 #include "llvm/Transforms/Utils/Cloning.h"
51 #include "llvm/Transforms/Utils/LoopUtils.h"
52 #include "llvm/Transforms/Utils/ValueMapper.h"
59 #define DEBUG_TYPE "simple-loop-unswitch"
63 STATISTIC(NumBranches, "Number of branches unswitched");
64 STATISTIC(NumSwitches, "Number of switches unswitched");
65 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
66 STATISTIC(NumTrivial, "Number of unswitches that are trivial");
68 NumCostMultiplierSkipped,
69 "Number of unswitch candidates that had their cost multiplier skipped");
71 static cl::opt<bool> EnableNonTrivialUnswitch(
72 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
73 cl::desc("Forcibly enables non-trivial loop unswitching rather than "
74 "following the configuration passed into the pass."));
77 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
78 cl::desc("The cost threshold for unswitching a loop."));
80 static cl::opt<bool> EnableUnswitchCostMultiplier(
81 "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden,
82 cl::desc("Enable unswitch cost multiplier that prohibits exponential "
83 "explosion in nontrivial unswitch."));
84 static cl::opt<int> UnswitchSiblingsToplevelDiv(
85 "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden,
86 cl::desc("Toplevel siblings divisor for cost multiplier."));
87 static cl::opt<int> UnswitchNumInitialUnscaledCandidates(
88 "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden,
89 cl::desc("Number of unswitch candidates that are ignored when calculating "
91 static cl::opt<bool> UnswitchGuards(
92 "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
93 cl::desc("If enabled, simple loop unswitching will also consider "
94 "llvm.experimental.guard intrinsics as unswitch candidates."));
96 /// Collect all of the loop invariant input values transitively used by the
97 /// homogeneous instruction graph from a given root.
99 /// This essentially walks from a root recursively through loop variant operands
100 /// which have the exact same opcode and finds all inputs which are loop
101 /// invariant. For some operations these can be re-associated and unswitched out
102 /// of the loop entirely.
103 static TinyPtrVector<Value *>
104 collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root,
106 assert(!L.isLoopInvariant(&Root) &&
107 "Only need to walk the graph if root itself is not invariant.");
108 TinyPtrVector<Value *> Invariants;
110 // Build a worklist and recurse through operators collecting invariants.
111 SmallVector<Instruction *, 4> Worklist;
112 SmallPtrSet<Instruction *, 8> Visited;
113 Worklist.push_back(&Root);
114 Visited.insert(&Root);
116 Instruction &I = *Worklist.pop_back_val();
117 for (Value *OpV : I.operand_values()) {
118 // Skip constants as unswitching isn't interesting for them.
119 if (isa<Constant>(OpV))
122 // Add it to our result if loop invariant.
123 if (L.isLoopInvariant(OpV)) {
124 Invariants.push_back(OpV);
128 // If not an instruction with the same opcode, nothing we can do.
129 Instruction *OpI = dyn_cast<Instruction>(OpV);
130 if (!OpI || OpI->getOpcode() != Root.getOpcode())
133 // Visit this operand.
134 if (Visited.insert(OpI).second)
135 Worklist.push_back(OpI);
137 } while (!Worklist.empty());
142 static void replaceLoopInvariantUses(Loop &L, Value *Invariant,
143 Constant &Replacement) {
144 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");
146 // Replace uses of LIC in the loop with the given constant.
147 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); UI != UE;) {
148 // Grab the use and walk past it so we can clobber it in the use list.
150 Instruction *UserI = dyn_cast<Instruction>(U->getUser());
152 // Replace this use within the loop body.
153 if (UserI && L.contains(UserI))
154 U->set(&Replacement);
158 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial
159 /// incoming values along this edge.
160 static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB,
161 BasicBlock &ExitBB) {
162 for (Instruction &I : ExitBB) {
163 auto *PN = dyn_cast<PHINode>(&I);
165 // No more PHIs to check.
168 // If the incoming value for this edge isn't loop invariant the unswitch
170 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))
173 llvm_unreachable("Basic blocks should never be empty!");
176 /// Insert code to test a set of loop invariant values, and conditionally branch
178 static void buildPartialUnswitchConditionalBranch(BasicBlock &BB,
179 ArrayRef<Value *> Invariants,
181 BasicBlock &UnswitchedSucc,
182 BasicBlock &NormalSucc) {
183 IRBuilder<> IRB(&BB);
184 Value *Cond = Invariants.front();
185 for (Value *Invariant :
186 make_range(std::next(Invariants.begin()), Invariants.end()))
188 Cond = IRB.CreateOr(Cond, Invariant);
190 Cond = IRB.CreateAnd(Cond, Invariant);
192 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
193 Direction ? &NormalSucc : &UnswitchedSucc);
196 /// Rewrite the PHI nodes in an unswitched loop exit basic block.
198 /// Requires that the loop exit and unswitched basic block are the same, and
199 /// that the exiting block was a unique predecessor of that block. Rewrites the
200 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
201 /// PHI nodes from the old preheader that now contains the unswitched
203 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
204 BasicBlock &OldExitingBB,
206 for (PHINode &PN : UnswitchedBB.phis()) {
207 // When the loop exit is directly unswitched we just need to update the
208 // incoming basic block. We loop to handle weird cases with repeated
209 // incoming blocks, but expect to typically only have one operand here.
210 for (auto i : seq<int>(0, PN.getNumOperands())) {
211 assert(PN.getIncomingBlock(i) == &OldExitingBB &&
212 "Found incoming block different from unique predecessor!");
213 PN.setIncomingBlock(i, &OldPH);
218 /// Rewrite the PHI nodes in the loop exit basic block and the split off
219 /// unswitched block.
221 /// Because the exit block remains an exit from the loop, this rewrites the
222 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
223 /// nodes into the unswitched basic block to select between the value in the
224 /// old preheader and the loop exit.
225 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
226 BasicBlock &UnswitchedBB,
227 BasicBlock &OldExitingBB,
230 assert(&ExitBB != &UnswitchedBB &&
231 "Must have different loop exit and unswitched blocks!");
232 Instruction *InsertPt = &*UnswitchedBB.begin();
233 for (PHINode &PN : ExitBB.phis()) {
234 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,
235 PN.getName() + ".split", InsertPt);
237 // Walk backwards over the old PHI node's inputs to minimize the cost of
238 // removing each one. We have to do this weird loop manually so that we
239 // create the same number of new incoming edges in the new PHI as we expect
240 // each case-based edge to be included in the unswitched switch in some
242 // FIXME: This is really, really gross. It would be much cleaner if LLVM
243 // allowed us to create a single entry for a predecessor block without
244 // having separate entries for each "edge" even though these edges are
245 // required to produce identical results.
246 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
247 if (PN.getIncomingBlock(i) != &OldExitingBB)
250 Value *Incoming = PN.getIncomingValue(i);
252 // No more edge from the old exiting block to the exit block.
253 PN.removeIncomingValue(i);
255 NewPN->addIncoming(Incoming, &OldPH);
258 // Now replace the old PHI with the new one and wire the old one in as an
259 // input to the new one.
260 PN.replaceAllUsesWith(NewPN);
261 NewPN->addIncoming(&PN, &ExitBB);
265 /// Hoist the current loop up to the innermost loop containing a remaining exit.
267 /// Because we've removed an exit from the loop, we may have changed the set of
268 /// loops reachable and need to move the current loop up the loop nest or even
269 /// to an entirely separate nest.
270 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
271 DominatorTree &DT, LoopInfo &LI) {
272 // If the loop is already at the top level, we can't hoist it anywhere.
273 Loop *OldParentL = L.getParentLoop();
277 SmallVector<BasicBlock *, 4> Exits;
278 L.getExitBlocks(Exits);
279 Loop *NewParentL = nullptr;
280 for (auto *ExitBB : Exits)
281 if (Loop *ExitL = LI.getLoopFor(ExitBB))
282 if (!NewParentL || NewParentL->contains(ExitL))
285 if (NewParentL == OldParentL)
288 // The new parent loop (if different) should always contain the old one.
290 assert(NewParentL->contains(OldParentL) &&
291 "Can only hoist this loop up the nest!");
293 // The preheader will need to move with the body of this loop. However,
294 // because it isn't in this loop we also need to update the primary loop map.
295 assert(OldParentL == LI.getLoopFor(&Preheader) &&
296 "Parent loop of this loop should contain this loop's preheader!");
297 LI.changeLoopFor(&Preheader, NewParentL);
299 // Remove this loop from its old parent.
300 OldParentL->removeChildLoop(&L);
302 // Add the loop either to the new parent or as a top-level loop.
304 NewParentL->addChildLoop(&L);
306 LI.addTopLevelLoop(&L);
308 // Remove this loops blocks from the old parent and every other loop up the
309 // nest until reaching the new parent. Also update all of these
310 // no-longer-containing loops to reflect the nesting change.
311 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
312 OldContainingL = OldContainingL->getParentLoop()) {
313 llvm::erase_if(OldContainingL->getBlocksVector(),
314 [&](const BasicBlock *BB) {
315 return BB == &Preheader || L.contains(BB);
318 OldContainingL->getBlocksSet().erase(&Preheader);
319 for (BasicBlock *BB : L.blocks())
320 OldContainingL->getBlocksSet().erase(BB);
322 // Because we just hoisted a loop out of this one, we have essentially
323 // created new exit paths from it. That means we need to form LCSSA PHI
324 // nodes for values used in the no-longer-nested loop.
325 formLCSSA(*OldContainingL, DT, &LI, nullptr);
327 // We shouldn't need to form dedicated exits because the exit introduced
328 // here is the (just split by unswitching) preheader. However, after trivial
329 // unswitching it is possible to get new non-dedicated exits out of parent
330 // loop so let's conservatively form dedicated exit blocks and figure out
331 // if we can optimize later.
332 formDedicatedExitBlocks(OldContainingL, &DT, &LI, /*PreserveLCSSA*/ true);
336 /// Unswitch a trivial branch if the condition is loop invariant.
338 /// This routine should only be called when loop code leading to the branch has
339 /// been validated as trivial (no side effects). This routine checks if the
340 /// condition is invariant and one of the successors is a loop exit. This
341 /// allows us to unswitch without duplicating the loop, making it trivial.
343 /// If this routine fails to unswitch the branch it returns false.
345 /// If the branch can be unswitched, this routine splits the preheader and
346 /// hoists the branch above that split. Preserves loop simplified form
347 /// (splitting the exit block as necessary). It simplifies the branch within
348 /// the loop to an unconditional branch but doesn't remove it entirely. Further
349 /// cleanup can be done with some simplify-cfg like pass.
351 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
352 /// invalidated by this.
353 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
354 LoopInfo &LI, ScalarEvolution *SE,
355 MemorySSAUpdater *MSSAU) {
356 assert(BI.isConditional() && "Can only unswitch a conditional branch!");
357 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n");
359 // The loop invariant values that we want to unswitch.
360 TinyPtrVector<Value *> Invariants;
362 // When true, we're fully unswitching the branch rather than just unswitching
363 // some input conditions to the branch.
364 bool FullUnswitch = false;
366 if (L.isLoopInvariant(BI.getCondition())) {
367 Invariants.push_back(BI.getCondition());
370 if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition()))
371 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);
372 if (Invariants.empty())
373 // Couldn't find invariant inputs!
377 // Check that one of the branch's successors exits, and which one.
378 bool ExitDirection = true;
379 int LoopExitSuccIdx = 0;
380 auto *LoopExitBB = BI.getSuccessor(0);
381 if (L.contains(LoopExitBB)) {
382 ExitDirection = false;
384 LoopExitBB = BI.getSuccessor(1);
385 if (L.contains(LoopExitBB))
388 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
389 auto *ParentBB = BI.getParent();
390 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB))
393 // When unswitching only part of the branch's condition, we need the exit
394 // block to be reached directly from the partially unswitched input. This can
395 // be done when the exit block is along the true edge and the branch condition
396 // is a graph of `or` operations, or the exit block is along the false edge
397 // and the condition is a graph of `and` operations.
400 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::Or)
403 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::And)
409 dbgs() << " unswitching trivial invariant conditions for: " << BI
411 for (Value *Invariant : Invariants) {
412 dbgs() << " " << *Invariant << " == true";
413 if (Invariant != Invariants.back())
419 // If we have scalar evolutions, we need to invalidate them including this
420 // loop and the loop containing the exit block.
422 if (Loop *ExitL = LI.getLoopFor(LoopExitBB))
423 SE->forgetLoop(ExitL);
425 // Forget the entire nest as this exits the entire nest.
426 SE->forgetTopmostLoop(&L);
429 if (MSSAU && VerifyMemorySSA)
430 MSSAU->getMemorySSA()->verifyMemorySSA();
432 // Split the preheader, so that we know that there is a safe place to insert
433 // the conditional branch. We will change the preheader to have a conditional
434 // branch on LoopCond.
435 BasicBlock *OldPH = L.getLoopPreheader();
436 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
438 // Now that we have a place to insert the conditional branch, create a place
439 // to branch to: this is the exit block out of the loop that we are
440 // unswitching. We need to split this if there are other loop predecessors.
441 // Because the loop is in simplified form, *any* other predecessor is enough.
442 BasicBlock *UnswitchedBB;
443 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
444 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
445 "A branch's parent isn't a predecessor!");
446 UnswitchedBB = LoopExitBB;
449 SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI, MSSAU);
452 if (MSSAU && VerifyMemorySSA)
453 MSSAU->getMemorySSA()->verifyMemorySSA();
455 // Actually move the invariant uses into the unswitched position. If possible,
456 // we do this by moving the instructions, but when doing partial unswitching
457 // we do it by building a new merge of the values in the unswitched position.
458 OldPH->getTerminator()->eraseFromParent();
460 // If fully unswitching, we can use the existing branch instruction.
461 // Splice it into the old PH to gate reaching the new preheader and re-point
463 OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(),
466 // Temporarily clone the terminator, to make MSSA update cheaper by
467 // separating "insert edge" updates from "remove edge" ones.
468 ParentBB->getInstList().push_back(BI.clone());
470 // Create a new unconditional branch that will continue the loop as a new
472 BranchInst::Create(ContinueBB, ParentBB);
474 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
475 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
477 // Only unswitching a subset of inputs to the condition, so we will need to
478 // build a new branch that merges the invariant inputs.
480 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
482 "Must have an `or` of `i1`s for the condition!");
484 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
486 "Must have an `and` of `i1`s for the condition!");
487 buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection,
488 *UnswitchedBB, *NewPH);
491 // Update the dominator tree with the added edge.
492 DT.insertEdge(OldPH, UnswitchedBB);
494 // After the dominator tree was updated with the added edge, update MemorySSA
497 SmallVector<CFGUpdate, 1> Updates;
498 Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
499 MSSAU->applyInsertUpdates(Updates, DT);
502 // Finish updating dominator tree and memory ssa for full unswitch.
505 // Remove the cloned branch instruction.
506 ParentBB->getTerminator()->eraseFromParent();
507 // Create unconditional branch now.
508 BranchInst::Create(ContinueBB, ParentBB);
509 MSSAU->removeEdge(ParentBB, LoopExitBB);
511 DT.deleteEdge(ParentBB, LoopExitBB);
514 if (MSSAU && VerifyMemorySSA)
515 MSSAU->getMemorySSA()->verifyMemorySSA();
517 // Rewrite the relevant PHI nodes.
518 if (UnswitchedBB == LoopExitBB)
519 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
521 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
522 *ParentBB, *OldPH, FullUnswitch);
524 // The constant we can replace all of our invariants with inside the loop
525 // body. If any of the invariants have a value other than this the loop won't
527 ConstantInt *Replacement = ExitDirection
528 ? ConstantInt::getFalse(BI.getContext())
529 : ConstantInt::getTrue(BI.getContext());
531 // Since this is an i1 condition we can also trivially replace uses of it
532 // within the loop with a constant.
533 for (Value *Invariant : Invariants)
534 replaceLoopInvariantUses(L, Invariant, *Replacement);
536 // If this was full unswitching, we may have changed the nesting relationship
537 // for this loop so hoist it to its correct parent if needed.
539 hoistLoopToNewParent(L, *NewPH, DT, LI);
541 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n");
547 /// Unswitch a trivial switch if the condition is loop invariant.
549 /// This routine should only be called when loop code leading to the switch has
550 /// been validated as trivial (no side effects). This routine checks if the
551 /// condition is invariant and that at least one of the successors is a loop
552 /// exit. This allows us to unswitch without duplicating the loop, making it
555 /// If this routine fails to unswitch the switch it returns false.
557 /// If the switch can be unswitched, this routine splits the preheader and
558 /// copies the switch above that split. If the default case is one of the
559 /// exiting cases, it copies the non-exiting cases and points them at the new
560 /// preheader. If the default case is not exiting, it copies the exiting cases
561 /// and points the default at the preheader. It preserves loop simplified form
562 /// (splitting the exit blocks as necessary). It simplifies the switch within
563 /// the loop by removing now-dead cases. If the default case is one of those
564 /// unswitched, it replaces its destination with a new basic block containing
565 /// only unreachable. Such basic blocks, while technically loop exits, are not
566 /// considered for unswitching so this is a stable transform and the same
567 /// switch will not be revisited. If after unswitching there is only a single
568 /// in-loop successor, the switch is further simplified to an unconditional
569 /// branch. Still more cleanup can be done with some simplify-cfg like pass.
571 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
572 /// invalidated by this.
573 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
574 LoopInfo &LI, ScalarEvolution *SE,
575 MemorySSAUpdater *MSSAU) {
576 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n");
577 Value *LoopCond = SI.getCondition();
579 // If this isn't switching on an invariant condition, we can't unswitch it.
580 if (!L.isLoopInvariant(LoopCond))
583 auto *ParentBB = SI.getParent();
585 SmallVector<int, 4> ExitCaseIndices;
586 for (auto Case : SI.cases()) {
587 auto *SuccBB = Case.getCaseSuccessor();
588 if (!L.contains(SuccBB) &&
589 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SuccBB))
590 ExitCaseIndices.push_back(Case.getCaseIndex());
592 BasicBlock *DefaultExitBB = nullptr;
593 if (!L.contains(SI.getDefaultDest()) &&
594 areLoopExitPHIsLoopInvariant(L, *ParentBB, *SI.getDefaultDest()) &&
595 !isa<UnreachableInst>(SI.getDefaultDest()->getTerminator()))
596 DefaultExitBB = SI.getDefaultDest();
597 else if (ExitCaseIndices.empty())
600 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n");
602 if (MSSAU && VerifyMemorySSA)
603 MSSAU->getMemorySSA()->verifyMemorySSA();
605 // We may need to invalidate SCEVs for the outermost loop reached by any of
610 // Clear out the default destination temporarily to allow accurate
611 // predecessor lists to be examined below.
612 SI.setDefaultDest(nullptr);
613 // Check the loop containing this exit.
614 Loop *ExitL = LI.getLoopFor(DefaultExitBB);
615 if (!ExitL || ExitL->contains(OuterL))
619 // Store the exit cases into a separate data structure and remove them from
621 SmallVector<std::pair<ConstantInt *, BasicBlock *>, 4> ExitCases;
622 ExitCases.reserve(ExitCaseIndices.size());
623 // We walk the case indices backwards so that we remove the last case first
624 // and don't disrupt the earlier indices.
625 for (unsigned Index : reverse(ExitCaseIndices)) {
626 auto CaseI = SI.case_begin() + Index;
627 // Compute the outer loop from this exit.
628 Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor());
629 if (!ExitL || ExitL->contains(OuterL))
631 // Save the value of this case.
632 ExitCases.push_back({CaseI->getCaseValue(), CaseI->getCaseSuccessor()});
633 // Delete the unswitched cases.
634 SI.removeCase(CaseI);
639 SE->forgetLoop(OuterL);
641 SE->forgetTopmostLoop(&L);
644 // Check if after this all of the remaining cases point at the same
646 BasicBlock *CommonSuccBB = nullptr;
647 if (SI.getNumCases() > 0 &&
648 std::all_of(std::next(SI.case_begin()), SI.case_end(),
649 [&SI](const SwitchInst::CaseHandle &Case) {
650 return Case.getCaseSuccessor() ==
651 SI.case_begin()->getCaseSuccessor();
653 CommonSuccBB = SI.case_begin()->getCaseSuccessor();
654 if (!DefaultExitBB) {
655 // If we're not unswitching the default, we need it to match any cases to
656 // have a common successor or if we have no cases it is the common
658 if (SI.getNumCases() == 0)
659 CommonSuccBB = SI.getDefaultDest();
660 else if (SI.getDefaultDest() != CommonSuccBB)
661 CommonSuccBB = nullptr;
664 // Split the preheader, so that we know that there is a safe place to insert
666 BasicBlock *OldPH = L.getLoopPreheader();
667 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
668 OldPH->getTerminator()->eraseFromParent();
670 // Now add the unswitched switch.
671 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
673 // Rewrite the IR for the unswitched basic blocks. This requires two steps.
674 // First, we split any exit blocks with remaining in-loop predecessors. Then
675 // we update the PHIs in one of two ways depending on if there was a split.
676 // We walk in reverse so that we split in the same order as the cases
677 // appeared. This is purely for convenience of reading the resulting IR, but
678 // it doesn't cost anything really.
679 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
680 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
681 // Handle the default exit if necessary.
682 // FIXME: It'd be great if we could merge this with the loop below but LLVM's
683 // ranges aren't quite powerful enough yet.
685 if (pred_empty(DefaultExitBB)) {
686 UnswitchedExitBBs.insert(DefaultExitBB);
687 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
690 SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI, MSSAU);
691 rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,
693 /*FullUnswitch*/ true);
694 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
697 // Note that we must use a reference in the for loop so that we update the
699 for (auto &CasePair : reverse(ExitCases)) {
700 // Grab a reference to the exit block in the pair so that we can update it.
701 BasicBlock *ExitBB = CasePair.second;
703 // If this case is the last edge into the exit block, we can simply reuse it
704 // as it will no longer be a loop exit. No mapping necessary.
705 if (pred_empty(ExitBB)) {
706 // Only rewrite once.
707 if (UnswitchedExitBBs.insert(ExitBB).second)
708 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
712 // Otherwise we need to split the exit block so that we retain an exit
713 // block from the loop and a target for the unswitched condition.
714 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
716 // If this is the first time we see this, do the split and remember it.
717 SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
718 rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,
720 /*FullUnswitch*/ true);
722 // Update the case pair to point to the split block.
723 CasePair.second = SplitExitBB;
726 // Now add the unswitched cases. We do this in reverse order as we built them
728 for (auto CasePair : reverse(ExitCases)) {
729 ConstantInt *CaseVal = CasePair.first;
730 BasicBlock *UnswitchedBB = CasePair.second;
732 NewSI->addCase(CaseVal, UnswitchedBB);
735 // If the default was unswitched, re-point it and add explicit cases for
736 // entering the loop.
738 NewSI->setDefaultDest(DefaultExitBB);
740 // We removed all the exit cases, so we just copy the cases to the
741 // unswitched switch.
742 for (auto Case : SI.cases())
743 NewSI->addCase(Case.getCaseValue(), NewPH);
746 // If we ended up with a common successor for every path through the switch
747 // after unswitching, rewrite it to an unconditional branch to make it easy
748 // to recognize. Otherwise we potentially have to recognize the default case
749 // pointing at unreachable and other complexity.
751 BasicBlock *BB = SI.getParent();
752 // We may have had multiple edges to this common successor block, so remove
753 // them as predecessors. We skip the first one, either the default or the
754 // actual first case.
755 bool SkippedFirst = DefaultExitBB == nullptr;
756 for (auto Case : SI.cases()) {
757 assert(Case.getCaseSuccessor() == CommonSuccBB &&
758 "Non-common successor!");
764 CommonSuccBB->removePredecessor(BB,
765 /*DontDeleteUselessPHIs*/ true);
767 // Now nuke the switch and replace it with a direct branch.
768 SI.eraseFromParent();
769 BranchInst::Create(CommonSuccBB, BB);
770 } else if (DefaultExitBB) {
771 assert(SI.getNumCases() > 0 &&
772 "If we had no cases we'd have a common successor!");
773 // Move the last case to the default successor. This is valid as if the
774 // default got unswitched it cannot be reached. This has the advantage of
775 // being simple and keeping the number of edges from this switch to
776 // successors the same, and avoiding any PHI update complexity.
777 auto LastCaseI = std::prev(SI.case_end());
778 SI.setDefaultDest(LastCaseI->getCaseSuccessor());
779 SI.removeCase(LastCaseI);
782 // Walk the unswitched exit blocks and the unswitched split blocks and update
783 // the dominator tree based on the CFG edits. While we are walking unordered
784 // containers here, the API for applyUpdates takes an unordered list of
785 // updates and requires them to not contain duplicates.
786 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
787 for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
788 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});
789 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});
791 for (auto SplitUnswitchedPair : SplitExitBBMap) {
792 auto *UnswitchedBB = SplitUnswitchedPair.second;
793 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedBB});
794 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedBB});
796 DT.applyUpdates(DTUpdates);
799 MSSAU->applyUpdates(DTUpdates, DT);
801 MSSAU->getMemorySSA()->verifyMemorySSA();
804 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
806 // We may have changed the nesting relationship for this loop so hoist it to
807 // its correct parent if needed.
808 hoistLoopToNewParent(L, *NewPH, DT, LI);
812 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n");
816 /// This routine scans the loop to find a branch or switch which occurs before
817 /// any side effects occur. These can potentially be unswitched without
818 /// duplicating the loop. If a branch or switch is successfully unswitched the
819 /// scanning continues to see if subsequent branches or switches have become
820 /// trivial. Once all trivial candidates have been unswitched, this routine
823 /// The return value indicates whether anything was unswitched (and therefore
826 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
827 /// invalidated by this.
828 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
829 LoopInfo &LI, ScalarEvolution *SE,
830 MemorySSAUpdater *MSSAU) {
831 bool Changed = false;
833 // If loop header has only one reachable successor we should keep looking for
834 // trivial condition candidates in the successor as well. An alternative is
835 // to constant fold conditions and merge successors into loop header (then we
836 // only need to check header's terminator). The reason for not doing this in
837 // LoopUnswitch pass is that it could potentially break LoopPassManager's
838 // invariants. Folding dead branches could either eliminate the current loop
839 // or make other loops unreachable. LCSSA form might also not be preserved
840 // after deleting branches. The following code keeps traversing loop header's
841 // successors until it finds the trivial condition candidate (condition that
842 // is not a constant). Since unswitching generates branches with constant
843 // conditions, this scenario could be very common in practice.
844 BasicBlock *CurrentBB = L.getHeader();
845 SmallPtrSet<BasicBlock *, 8> Visited;
846 Visited.insert(CurrentBB);
848 // Check if there are any side-effecting instructions (e.g. stores, calls,
849 // volatile loads) in the part of the loop that the code *would* execute
850 // without unswitching.
851 if (llvm::any_of(*CurrentBB,
852 [](Instruction &I) { return I.mayHaveSideEffects(); }))
855 Instruction *CurrentTerm = CurrentBB->getTerminator();
857 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
858 // Don't bother trying to unswitch past a switch with a constant
859 // condition. This should be removed prior to running this pass by
861 if (isa<Constant>(SI->getCondition()))
864 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU))
865 // Couldn't unswitch this one so we're done.
868 // Mark that we managed to unswitch something.
871 // If unswitching turned the terminator into an unconditional branch then
872 // we can continue. The unswitching logic specifically works to fold any
873 // cases it can into an unconditional branch to make it easier to
875 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
876 if (!BI || BI->isConditional())
879 CurrentBB = BI->getSuccessor(0);
883 auto *BI = dyn_cast<BranchInst>(CurrentTerm);
885 // We do not understand other terminator instructions.
888 // Don't bother trying to unswitch past an unconditional branch or a branch
889 // with a constant value. These should be removed by simplify-cfg prior to
890 // running this pass.
891 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
894 // Found a trivial condition candidate: non-foldable conditional branch. If
895 // we fail to unswitch this, we can't do anything else that is trivial.
896 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU))
899 // Mark that we managed to unswitch something.
902 // If we only unswitched some of the conditions feeding the branch, we won't
903 // have collapsed it to a single successor.
904 BI = cast<BranchInst>(CurrentBB->getTerminator());
905 if (BI->isConditional())
908 // Follow the newly unconditional branch into its successor.
909 CurrentBB = BI->getSuccessor(0);
911 // When continuing, if we exit the loop or reach a previous visited block,
912 // then we can not reach any trivial condition candidates (unfoldable
913 // branch instructions or switch instructions) and no unswitch can happen.
914 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
919 /// Build the cloned blocks for an unswitched copy of the given loop.
921 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
922 /// after the split block (`SplitBB`) that will be used to select between the
923 /// cloned and original loop.
925 /// This routine handles cloning all of the necessary loop blocks and exit
926 /// blocks including rewriting their instructions and the relevant PHI nodes.
927 /// Any loop blocks or exit blocks which are dominated by a different successor
928 /// than the one for this clone of the loop blocks can be trivially skipped. We
929 /// use the `DominatingSucc` map to determine whether a block satisfies that
930 /// property with a simple map lookup.
932 /// It also correctly creates the unconditional branch in the cloned
933 /// unswitched parent block to only point at the unswitched successor.
935 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit
936 /// block splitting is correctly reflected in `LoopInfo`, essentially all of
937 /// the cloned blocks (and their loops) are left without full `LoopInfo`
938 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
939 /// blocks to them but doesn't create the cloned `DominatorTree` structure and
940 /// instead the caller must recompute an accurate DT. It *does* correctly
941 /// update the `AssumptionCache` provided in `AC`.
942 static BasicBlock *buildClonedLoopBlocks(
943 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
944 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
945 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
946 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
947 ValueToValueMapTy &VMap,
948 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
949 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
950 SmallVector<BasicBlock *, 4> NewBlocks;
951 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());
953 // We will need to clone a bunch of blocks, wrap up the clone operation in
955 auto CloneBlock = [&](BasicBlock *OldBB) {
956 // Clone the basic block and insert it before the new preheader.
957 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());
958 NewBB->moveBefore(LoopPH);
960 // Record this block and the mapping.
961 NewBlocks.push_back(NewBB);
967 // We skip cloning blocks when they have a dominating succ that is not the
968 // succ we are cloning for.
969 auto SkipBlock = [&](BasicBlock *BB) {
970 auto It = DominatingSucc.find(BB);
971 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
974 // First, clone the preheader.
975 auto *ClonedPH = CloneBlock(LoopPH);
977 // Then clone all the loop blocks, skipping the ones that aren't necessary.
978 for (auto *LoopBB : L.blocks())
979 if (!SkipBlock(LoopBB))
982 // Split all the loop exit edges so that when we clone the exit blocks, if
983 // any of the exit blocks are *also* a preheader for some other loop, we
984 // don't create multiple predecessors entering the loop header.
985 for (auto *ExitBB : ExitBlocks) {
986 if (SkipBlock(ExitBB))
989 // When we are going to clone an exit, we don't need to clone all the
990 // instructions in the exit block and we want to ensure we have an easy
991 // place to merge the CFG, so split the exit first. This is always safe to
992 // do because there cannot be any non-loop predecessors of a loop exit in
993 // loop simplified form.
994 auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
996 // Rearrange the names to make it easier to write test cases by having the
997 // exit block carry the suffix rather than the merge block carrying the
999 MergeBB->takeName(ExitBB);
1000 ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1002 // Now clone the original exit block.
1003 auto *ClonedExitBB = CloneBlock(ExitBB);
1004 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1005 "Exit block should have been split to have one successor!");
1006 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1007 "Cloned exit block has the wrong successor!");
1009 // Remap any cloned instructions and create a merge phi node for them.
1010 for (auto ZippedInsts : llvm::zip_first(
1011 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),
1012 llvm::make_range(ClonedExitBB->begin(),
1013 std::prev(ClonedExitBB->end())))) {
1014 Instruction &I = std::get<0>(ZippedInsts);
1015 Instruction &ClonedI = std::get<1>(ZippedInsts);
1017 // The only instructions in the exit block should be PHI nodes and
1018 // potentially a landing pad.
1020 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1021 "Bad instruction in exit block!");
1022 // We should have a value map between the instruction and its clone.
1023 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1026 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi",
1027 &*MergeBB->getFirstInsertionPt());
1028 I.replaceAllUsesWith(MergePN);
1029 MergePN->addIncoming(&I, ExitBB);
1030 MergePN->addIncoming(&ClonedI, ClonedExitBB);
1034 // Rewrite the instructions in the cloned blocks to refer to the instructions
1035 // in the cloned blocks. We have to do this as a second pass so that we have
1036 // everything available. Also, we have inserted new instructions which may
1037 // include assume intrinsics, so we update the assumption cache while
1039 for (auto *ClonedBB : NewBlocks)
1040 for (Instruction &I : *ClonedBB) {
1041 RemapInstruction(&I, VMap,
1042 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1043 if (auto *II = dyn_cast<IntrinsicInst>(&I))
1044 if (II->getIntrinsicID() == Intrinsic::assume)
1045 AC.registerAssumption(II);
1048 // Update any PHI nodes in the cloned successors of the skipped blocks to not
1049 // have spurious incoming values.
1050 for (auto *LoopBB : L.blocks())
1051 if (SkipBlock(LoopBB))
1052 for (auto *SuccBB : successors(LoopBB))
1053 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))
1054 for (PHINode &PN : ClonedSuccBB->phis())
1055 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);
1057 // Remove the cloned parent as a predecessor of any successor we ended up
1058 // cloning other than the unswitched one.
1059 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));
1060 for (auto *SuccBB : successors(ParentBB)) {
1061 if (SuccBB == UnswitchedSuccBB)
1064 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));
1068 ClonedSuccBB->removePredecessor(ClonedParentBB,
1069 /*DontDeleteUselessPHIs*/ true);
1072 // Replace the cloned branch with an unconditional branch to the cloned
1073 // unswitched successor.
1074 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));
1075 ClonedParentBB->getTerminator()->eraseFromParent();
1076 BranchInst::Create(ClonedSuccBB, ClonedParentBB);
1078 // If there are duplicate entries in the PHI nodes because of multiple edges
1079 // to the unswitched successor, we need to nuke all but one as we replaced it
1080 // with a direct branch.
1081 for (PHINode &PN : ClonedSuccBB->phis()) {
1083 // Loop over the incoming operands backwards so we can easily delete as we
1084 // go without invalidating the index.
1085 for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1086 if (PN.getIncomingBlock(i) != ClonedParentBB)
1092 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);
1096 // Record the domtree updates for the new blocks.
1097 SmallPtrSet<BasicBlock *, 4> SuccSet;
1098 for (auto *ClonedBB : NewBlocks) {
1099 for (auto *SuccBB : successors(ClonedBB))
1100 if (SuccSet.insert(SuccBB).second)
1101 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});
1108 /// Recursively clone the specified loop and all of its children.
1110 /// The target parent loop for the clone should be provided, or can be null if
1111 /// the clone is a top-level loop. While cloning, all the blocks are mapped
1112 /// with the provided value map. The entire original loop must be present in
1113 /// the value map. The cloned loop is returned.
1114 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1115 const ValueToValueMapTy &VMap, LoopInfo &LI) {
1116 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1117 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1118 ClonedL.reserveBlocks(OrigL.getNumBlocks());
1119 for (auto *BB : OrigL.blocks()) {
1120 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));
1121 ClonedL.addBlockEntry(ClonedBB);
1122 if (LI.getLoopFor(BB) == &OrigL)
1123 LI.changeLoopFor(ClonedBB, &ClonedL);
1127 // We specially handle the first loop because it may get cloned into
1128 // a different parent and because we most commonly are cloning leaf loops.
1129 Loop *ClonedRootL = LI.AllocateLoop();
1131 RootParentL->addChildLoop(ClonedRootL);
1133 LI.addTopLevelLoop(ClonedRootL);
1134 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1136 if (OrigRootL.empty())
1139 // If we have a nest, we can quickly clone the entire loop nest using an
1140 // iterative approach because it is a tree. We keep the cloned parent in the
1141 // data structure to avoid repeatedly querying through a map to find it.
1142 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1143 // Build up the loops to clone in reverse order as we'll clone them from the
1145 for (Loop *ChildL : llvm::reverse(OrigRootL))
1146 LoopsToClone.push_back({ClonedRootL, ChildL});
1148 Loop *ClonedParentL, *L;
1149 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();
1150 Loop *ClonedL = LI.AllocateLoop();
1151 ClonedParentL->addChildLoop(ClonedL);
1152 AddClonedBlocksToLoop(*L, *ClonedL);
1153 for (Loop *ChildL : llvm::reverse(*L))
1154 LoopsToClone.push_back({ClonedL, ChildL});
1155 } while (!LoopsToClone.empty());
1160 /// Build the cloned loops of an original loop from unswitching.
1162 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1163 /// operation. We need to re-verify that there even is a loop (as the backedge
1164 /// may not have been cloned), and even if there are remaining backedges the
1165 /// backedge set may be different. However, we know that each child loop is
1166 /// undisturbed, we only need to find where to place each child loop within
1167 /// either any parent loop or within a cloned version of the original loop.
1169 /// Because child loops may end up cloned outside of any cloned version of the
1170 /// original loop, multiple cloned sibling loops may be created. All of them
1171 /// are returned so that the newly introduced loop nest roots can be
1173 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1174 const ValueToValueMapTy &VMap, LoopInfo &LI,
1175 SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1176 Loop *ClonedL = nullptr;
1178 auto *OrigPH = OrigL.getLoopPreheader();
1179 auto *OrigHeader = OrigL.getHeader();
1181 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));
1182 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));
1184 // We need to know the loops of the cloned exit blocks to even compute the
1185 // accurate parent loop. If we only clone exits to some parent of the
1186 // original parent, we want to clone into that outer loop. We also keep track
1187 // of the loops that our cloned exit blocks participate in.
1188 Loop *ParentL = nullptr;
1189 SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1190 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1191 ClonedExitsInLoops.reserve(ExitBlocks.size());
1192 for (auto *ExitBB : ExitBlocks)
1193 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))
1194 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1195 ExitLoopMap[ClonedExitBB] = ExitL;
1196 ClonedExitsInLoops.push_back(ClonedExitBB);
1197 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1200 assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1201 ParentL->contains(OrigL.getParentLoop())) &&
1202 "The computed parent loop should always contain (or be) the parent of "
1203 "the original loop.");
1205 // We build the set of blocks dominated by the cloned header from the set of
1206 // cloned blocks out of the original loop. While not all of these will
1207 // necessarily be in the cloned loop, it is enough to establish that they
1208 // aren't in unreachable cycles, etc.
1209 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1210 for (auto *BB : OrigL.blocks())
1211 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))
1212 ClonedLoopBlocks.insert(ClonedBB);
1214 // Rebuild the set of blocks that will end up in the cloned loop. We may have
1215 // skipped cloning some region of this loop which can in turn skip some of
1216 // the backedges so we have to rebuild the blocks in the loop based on the
1217 // backedges that remain after cloning.
1218 SmallVector<BasicBlock *, 16> Worklist;
1219 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1220 for (auto *Pred : predecessors(ClonedHeader)) {
1221 // The only possible non-loop header predecessor is the preheader because
1222 // we know we cloned the loop in simplified form.
1223 if (Pred == ClonedPH)
1226 // Because the loop was in simplified form, the only non-loop predecessor
1227 // should be the preheader.
1228 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1229 "header other than the preheader "
1230 "that is not part of the loop!");
1232 // Insert this block into the loop set and on the first visit (and if it
1233 // isn't the header we're currently walking) put it into the worklist to
1235 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)
1236 Worklist.push_back(Pred);
1239 // If we had any backedges then there *is* a cloned loop. Put the header into
1240 // the loop set and then walk the worklist backwards to find all the blocks
1241 // that remain within the loop after cloning.
1242 if (!BlocksInClonedLoop.empty()) {
1243 BlocksInClonedLoop.insert(ClonedHeader);
1245 while (!Worklist.empty()) {
1246 BasicBlock *BB = Worklist.pop_back_val();
1247 assert(BlocksInClonedLoop.count(BB) &&
1248 "Didn't put block into the loop set!");
1250 // Insert any predecessors that are in the possible set into the cloned
1251 // set, and if the insert is successful, add them to the worklist. Note
1252 // that we filter on the blocks that are definitely reachable via the
1253 // backedge to the loop header so we may prune out dead code within the
1255 for (auto *Pred : predecessors(BB))
1256 if (ClonedLoopBlocks.count(Pred) &&
1257 BlocksInClonedLoop.insert(Pred).second)
1258 Worklist.push_back(Pred);
1261 ClonedL = LI.AllocateLoop();
1263 ParentL->addBasicBlockToLoop(ClonedPH, LI);
1264 ParentL->addChildLoop(ClonedL);
1266 LI.addTopLevelLoop(ClonedL);
1268 NonChildClonedLoops.push_back(ClonedL);
1270 ClonedL->reserveBlocks(BlocksInClonedLoop.size());
1271 // We don't want to just add the cloned loop blocks based on how we
1272 // discovered them. The original order of blocks was carefully built in
1273 // a way that doesn't rely on predecessor ordering. Rather than re-invent
1274 // that logic, we just re-walk the original blocks (and those of the child
1275 // loops) and filter them as we add them into the cloned loop.
1276 for (auto *BB : OrigL.blocks()) {
1277 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));
1278 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))
1281 // Directly add the blocks that are only in this loop.
1282 if (LI.getLoopFor(BB) == &OrigL) {
1283 ClonedL->addBasicBlockToLoop(ClonedBB, LI);
1287 // We want to manually add it to this loop and parents.
1288 // Registering it with LoopInfo will happen when we clone the top
1289 // loop for this block.
1290 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1291 PL->addBlockEntry(ClonedBB);
1294 // Now add each child loop whose header remains within the cloned loop. All
1295 // of the blocks within the loop must satisfy the same constraints as the
1296 // header so once we pass the header checks we can just clone the entire
1298 for (Loop *ChildL : OrigL) {
1299 auto *ClonedChildHeader =
1300 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1301 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))
1305 // We should never have a cloned child loop header but fail to have
1306 // all of the blocks for that child loop.
1307 for (auto *ChildLoopBB : ChildL->blocks())
1308 assert(BlocksInClonedLoop.count(
1309 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1310 "Child cloned loop has a header within the cloned outer "
1311 "loop but not all of its blocks!");
1314 cloneLoopNest(*ChildL, ClonedL, VMap, LI);
1318 // Now that we've handled all the components of the original loop that were
1319 // cloned into a new loop, we still need to handle anything from the original
1320 // loop that wasn't in a cloned loop.
1322 // Figure out what blocks are left to place within any loop nest containing
1323 // the unswitched loop. If we never formed a loop, the cloned PH is one of
1325 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1326 if (BlocksInClonedLoop.empty())
1327 UnloopedBlockSet.insert(ClonedPH);
1328 for (auto *ClonedBB : ClonedLoopBlocks)
1329 if (!BlocksInClonedLoop.count(ClonedBB))
1330 UnloopedBlockSet.insert(ClonedBB);
1332 // Copy the cloned exits and sort them in ascending loop depth, we'll work
1333 // backwards across these to process them inside out. The order shouldn't
1334 // matter as we're just trying to build up the map from inside-out; we use
1335 // the map in a more stably ordered way below.
1336 auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1337 llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1338 return ExitLoopMap.lookup(LHS)->getLoopDepth() <
1339 ExitLoopMap.lookup(RHS)->getLoopDepth();
1342 // Populate the existing ExitLoopMap with everything reachable from each
1343 // exit, starting from the inner most exit.
1344 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1345 assert(Worklist.empty() && "Didn't clear worklist!");
1347 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1348 Loop *ExitL = ExitLoopMap.lookup(ExitBB);
1350 // Walk the CFG back until we hit the cloned PH adding everything reachable
1351 // and in the unlooped set to this exit block's loop.
1352 Worklist.push_back(ExitBB);
1354 BasicBlock *BB = Worklist.pop_back_val();
1355 // We can stop recursing at the cloned preheader (if we get there).
1359 for (BasicBlock *PredBB : predecessors(BB)) {
1360 // If this pred has already been moved to our set or is part of some
1361 // (inner) loop, no update needed.
1362 if (!UnloopedBlockSet.erase(PredBB)) {
1364 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1365 "Predecessor not mapped to a loop!");
1369 // We just insert into the loop set here. We'll add these blocks to the
1370 // exit loop after we build up the set in an order that doesn't rely on
1371 // predecessor order (which in turn relies on use list order).
1372 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;
1374 assert(Inserted && "Should only visit an unlooped block once!");
1376 // And recurse through to its predecessors.
1377 Worklist.push_back(PredBB);
1379 } while (!Worklist.empty());
1382 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned
1383 // blocks to their outer loops, walk the cloned blocks and the cloned exits
1384 // in their original order adding them to the correct loop.
1386 // We need a stable insertion order. We use the order of the original loop
1387 // order and map into the correct parent loop.
1388 for (auto *BB : llvm::concat<BasicBlock *const>(
1389 makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))
1390 if (Loop *OuterL = ExitLoopMap.lookup(BB))
1391 OuterL->addBasicBlockToLoop(BB, LI);
1394 for (auto &BBAndL : ExitLoopMap) {
1395 auto *BB = BBAndL.first;
1396 auto *OuterL = BBAndL.second;
1397 assert(LI.getLoopFor(BB) == OuterL &&
1398 "Failed to put all blocks into outer loops!");
1402 // Now that all the blocks are placed into the correct containing loop in the
1403 // absence of child loops, find all the potentially cloned child loops and
1404 // clone them into whatever outer loop we placed their header into.
1405 for (Loop *ChildL : OrigL) {
1406 auto *ClonedChildHeader =
1407 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1408 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))
1412 for (auto *ChildLoopBB : ChildL->blocks())
1413 assert(VMap.count(ChildLoopBB) &&
1414 "Cloned a child loop header but not all of that loops blocks!");
1417 NonChildClonedLoops.push_back(cloneLoopNest(
1418 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));
1423 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1424 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1425 DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1426 // Find all the dead clones, and remove them from their successors.
1427 SmallVector<BasicBlock *, 16> DeadBlocks;
1428 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))
1429 for (auto &VMap : VMaps)
1430 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))
1431 if (!DT.isReachableFromEntry(ClonedBB)) {
1432 for (BasicBlock *SuccBB : successors(ClonedBB))
1433 SuccBB->removePredecessor(ClonedBB);
1434 DeadBlocks.push_back(ClonedBB);
1437 // Remove all MemorySSA in the dead blocks
1439 SmallPtrSet<BasicBlock *, 16> DeadBlockSet(DeadBlocks.begin(),
1441 MSSAU->removeBlocks(DeadBlockSet);
1444 // Drop any remaining references to break cycles.
1445 for (BasicBlock *BB : DeadBlocks)
1446 BB->dropAllReferences();
1447 // Erase them from the IR.
1448 for (BasicBlock *BB : DeadBlocks)
1449 BB->eraseFromParent();
1452 static void deleteDeadBlocksFromLoop(Loop &L,
1453 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1454 DominatorTree &DT, LoopInfo &LI,
1455 MemorySSAUpdater *MSSAU) {
1456 // Find all the dead blocks tied to this loop, and remove them from their
1458 SmallPtrSet<BasicBlock *, 16> DeadBlockSet;
1460 // Start with loop/exit blocks and get a transitive closure of reachable dead
1462 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1464 DeathCandidates.append(L.blocks().begin(), L.blocks().end());
1465 while (!DeathCandidates.empty()) {
1466 auto *BB = DeathCandidates.pop_back_val();
1467 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {
1468 for (BasicBlock *SuccBB : successors(BB)) {
1469 SuccBB->removePredecessor(BB);
1470 DeathCandidates.push_back(SuccBB);
1472 DeadBlockSet.insert(BB);
1476 // Remove all MemorySSA in the dead blocks
1478 MSSAU->removeBlocks(DeadBlockSet);
1480 // Filter out the dead blocks from the exit blocks list so that it can be
1481 // used in the caller.
1482 llvm::erase_if(ExitBlocks,
1483 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1485 // Walk from this loop up through its parents removing all of the dead blocks.
1486 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1487 for (auto *BB : DeadBlockSet)
1488 ParentL->getBlocksSet().erase(BB);
1489 llvm::erase_if(ParentL->getBlocksVector(),
1490 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1493 // Now delete the dead child loops. This raw delete will clear them
1495 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {
1496 if (!DeadBlockSet.count(ChildL->getHeader()))
1499 assert(llvm::all_of(ChildL->blocks(),
1500 [&](BasicBlock *ChildBB) {
1501 return DeadBlockSet.count(ChildBB);
1503 "If the child loop header is dead all blocks in the child loop must "
1504 "be dead as well!");
1509 // Remove the loop mappings for the dead blocks and drop all the references
1510 // from these blocks to others to handle cyclic references as we start
1511 // deleting the blocks themselves.
1512 for (auto *BB : DeadBlockSet) {
1513 // Check that the dominator tree has already been updated.
1514 assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1515 LI.changeLoopFor(BB, nullptr);
1516 BB->dropAllReferences();
1519 // Actually delete the blocks now that they've been fully unhooked from the
1521 for (auto *BB : DeadBlockSet)
1522 BB->eraseFromParent();
1525 /// Recompute the set of blocks in a loop after unswitching.
1527 /// This walks from the original headers predecessors to rebuild the loop. We
1528 /// take advantage of the fact that new blocks can't have been added, and so we
1529 /// filter by the original loop's blocks. This also handles potentially
1530 /// unreachable code that we don't want to explore but might be found examining
1531 /// the predecessors of the header.
1533 /// If the original loop is no longer a loop, this will return an empty set. If
1534 /// it remains a loop, all the blocks within it will be added to the set
1535 /// (including those blocks in inner loops).
1536 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1538 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1540 auto *PH = L.getLoopPreheader();
1541 auto *Header = L.getHeader();
1543 // A worklist to use while walking backwards from the header.
1544 SmallVector<BasicBlock *, 16> Worklist;
1546 // First walk the predecessors of the header to find the backedges. This will
1547 // form the basis of our walk.
1548 for (auto *Pred : predecessors(Header)) {
1549 // Skip the preheader.
1553 // Because the loop was in simplified form, the only non-loop predecessor
1554 // is the preheader.
1555 assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1556 "than the preheader that is not part of the "
1559 // Insert this block into the loop set and on the first visit and, if it
1560 // isn't the header we're currently walking, put it into the worklist to
1562 if (LoopBlockSet.insert(Pred).second && Pred != Header)
1563 Worklist.push_back(Pred);
1566 // If no backedges were found, we're done.
1567 if (LoopBlockSet.empty())
1568 return LoopBlockSet;
1570 // We found backedges, recurse through them to identify the loop blocks.
1571 while (!Worklist.empty()) {
1572 BasicBlock *BB = Worklist.pop_back_val();
1573 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1575 // No need to walk past the header.
1579 // Because we know the inner loop structure remains valid we can use the
1580 // loop structure to jump immediately across the entire nested loop.
1581 // Further, because it is in loop simplified form, we can directly jump
1582 // to its preheader afterward.
1583 if (Loop *InnerL = LI.getLoopFor(BB))
1585 assert(L.contains(InnerL) &&
1586 "Should not reach a loop *outside* this loop!");
1587 // The preheader is the only possible predecessor of the loop so
1588 // insert it into the set and check whether it was already handled.
1589 auto *InnerPH = InnerL->getLoopPreheader();
1590 assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1591 "but not contain the inner loop "
1593 if (!LoopBlockSet.insert(InnerPH).second)
1594 // The only way to reach the preheader is through the loop body
1595 // itself so if it has been visited the loop is already handled.
1598 // Insert all of the blocks (other than those already present) into
1599 // the loop set. We expect at least the block that led us to find the
1600 // inner loop to be in the block set, but we may also have other loop
1601 // blocks if they were already enqueued as predecessors of some other
1602 // outer loop block.
1603 for (auto *InnerBB : InnerL->blocks()) {
1604 if (InnerBB == BB) {
1605 assert(LoopBlockSet.count(InnerBB) &&
1606 "Block should already be in the set!");
1610 LoopBlockSet.insert(InnerBB);
1613 // Add the preheader to the worklist so we will continue past the
1615 Worklist.push_back(InnerPH);
1619 // Insert any predecessors that were in the original loop into the new
1620 // set, and if the insert is successful, add them to the worklist.
1621 for (auto *Pred : predecessors(BB))
1622 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)
1623 Worklist.push_back(Pred);
1626 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1628 // We've found all the blocks participating in the loop, return our completed
1630 return LoopBlockSet;
1633 /// Rebuild a loop after unswitching removes some subset of blocks and edges.
1635 /// The removal may have removed some child loops entirely but cannot have
1636 /// disturbed any remaining child loops. However, they may need to be hoisted
1637 /// to the parent loop (or to be top-level loops). The original loop may be
1638 /// completely removed.
1640 /// The sibling loops resulting from this update are returned. If the original
1641 /// loop remains a valid loop, it will be the first entry in this list with all
1642 /// of the newly sibling loops following it.
1644 /// Returns true if the loop remains a loop after unswitching, and false if it
1645 /// is no longer a loop after unswitching (and should not continue to be
1647 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1649 SmallVectorImpl<Loop *> &HoistedLoops) {
1650 auto *PH = L.getLoopPreheader();
1652 // Compute the actual parent loop from the exit blocks. Because we may have
1653 // pruned some exits the loop may be different from the original parent.
1654 Loop *ParentL = nullptr;
1655 SmallVector<Loop *, 4> ExitLoops;
1656 SmallVector<BasicBlock *, 4> ExitsInLoops;
1657 ExitsInLoops.reserve(ExitBlocks.size());
1658 for (auto *ExitBB : ExitBlocks)
1659 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1660 ExitLoops.push_back(ExitL);
1661 ExitsInLoops.push_back(ExitBB);
1662 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1666 // Recompute the blocks participating in this loop. This may be empty if it
1667 // is no longer a loop.
1668 auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1670 // If we still have a loop, we need to re-set the loop's parent as the exit
1671 // block set changing may have moved it within the loop nest. Note that this
1672 // can only happen when this loop has a parent as it can only hoist the loop
1674 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1675 // Remove this loop's (original) blocks from all of the intervening loops.
1676 for (Loop *IL = L.getParentLoop(); IL != ParentL;
1677 IL = IL->getParentLoop()) {
1678 IL->getBlocksSet().erase(PH);
1679 for (auto *BB : L.blocks())
1680 IL->getBlocksSet().erase(BB);
1681 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {
1682 return BB == PH || L.contains(BB);
1686 LI.changeLoopFor(PH, ParentL);
1687 L.getParentLoop()->removeChildLoop(&L);
1689 ParentL->addChildLoop(&L);
1691 LI.addTopLevelLoop(&L);
1694 // Now we update all the blocks which are no longer within the loop.
1695 auto &Blocks = L.getBlocksVector();
1697 LoopBlockSet.empty()
1699 : std::stable_partition(
1700 Blocks.begin(), Blocks.end(),
1701 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });
1703 // Before we erase the list of unlooped blocks, build a set of them.
1704 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1705 if (LoopBlockSet.empty())
1706 UnloopedBlocks.insert(PH);
1708 // Now erase these blocks from the loop.
1709 for (auto *BB : make_range(BlocksSplitI, Blocks.end()))
1710 L.getBlocksSet().erase(BB);
1711 Blocks.erase(BlocksSplitI, Blocks.end());
1713 // Sort the exits in ascending loop depth, we'll work backwards across these
1714 // to process them inside out.
1715 std::stable_sort(ExitsInLoops.begin(), ExitsInLoops.end(),
1716 [&](BasicBlock *LHS, BasicBlock *RHS) {
1717 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);
1720 // We'll build up a set for each exit loop.
1721 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1722 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1724 auto RemoveUnloopedBlocksFromLoop =
1725 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1726 for (auto *BB : UnloopedBlocks)
1727 L.getBlocksSet().erase(BB);
1728 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {
1729 return UnloopedBlocks.count(BB);
1733 SmallVector<BasicBlock *, 16> Worklist;
1734 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1735 assert(Worklist.empty() && "Didn't clear worklist!");
1736 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
1738 // Grab the next exit block, in decreasing loop depth order.
1739 BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1740 Loop &ExitL = *LI.getLoopFor(ExitBB);
1741 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
1743 // Erase all of the unlooped blocks from the loops between the previous
1744 // exit loop and this exit loop. This works because the ExitInLoops list is
1745 // sorted in increasing order of loop depth and thus we visit loops in
1746 // decreasing order of loop depth.
1747 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
1748 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1750 // Walk the CFG back until we hit the cloned PH adding everything reachable
1751 // and in the unlooped set to this exit block's loop.
1752 Worklist.push_back(ExitBB);
1754 BasicBlock *BB = Worklist.pop_back_val();
1755 // We can stop recursing at the cloned preheader (if we get there).
1759 for (BasicBlock *PredBB : predecessors(BB)) {
1760 // If this pred has already been moved to our set or is part of some
1761 // (inner) loop, no update needed.
1762 if (!UnloopedBlocks.erase(PredBB)) {
1763 assert((NewExitLoopBlocks.count(PredBB) ||
1764 ExitL.contains(LI.getLoopFor(PredBB))) &&
1765 "Predecessor not in a nested loop (or already visited)!");
1769 // We just insert into the loop set here. We'll add these blocks to the
1770 // exit loop after we build up the set in a deterministic order rather
1771 // than the predecessor-influenced visit order.
1772 bool Inserted = NewExitLoopBlocks.insert(PredBB).second;
1774 assert(Inserted && "Should only visit an unlooped block once!");
1776 // And recurse through to its predecessors.
1777 Worklist.push_back(PredBB);
1779 } while (!Worklist.empty());
1781 // If blocks in this exit loop were directly part of the original loop (as
1782 // opposed to a child loop) update the map to point to this exit loop. This
1783 // just updates a map and so the fact that the order is unstable is fine.
1784 for (auto *BB : NewExitLoopBlocks)
1785 if (Loop *BBL = LI.getLoopFor(BB))
1786 if (BBL == &L || !L.contains(BBL))
1787 LI.changeLoopFor(BB, &ExitL);
1789 // We will remove the remaining unlooped blocks from this loop in the next
1790 // iteration or below.
1791 NewExitLoopBlocks.clear();
1794 // Any remaining unlooped blocks are no longer part of any loop unless they
1795 // are part of some child loop.
1796 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
1797 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1798 for (auto *BB : UnloopedBlocks)
1799 if (Loop *BBL = LI.getLoopFor(BB))
1800 if (BBL == &L || !L.contains(BBL))
1801 LI.changeLoopFor(BB, nullptr);
1803 // Sink all the child loops whose headers are no longer in the loop set to
1804 // the parent (or to be top level loops). We reach into the loop and directly
1805 // update its subloop vector to make this batch update efficient.
1806 auto &SubLoops = L.getSubLoopsVector();
1807 auto SubLoopsSplitI =
1808 LoopBlockSet.empty()
1810 : std::stable_partition(
1811 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {
1812 return LoopBlockSet.count(SubL->getHeader());
1814 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {
1815 HoistedLoops.push_back(HoistedL);
1816 HoistedL->setParentLoop(nullptr);
1818 // To compute the new parent of this hoisted loop we look at where we
1819 // placed the preheader above. We can't lookup the header itself because we
1820 // retained the mapping from the header to the hoisted loop. But the
1821 // preheader and header should have the exact same new parent computed
1822 // based on the set of exit blocks from the original loop as the preheader
1823 // is a predecessor of the header and so reached in the reverse walk. And
1824 // because the loops were all in simplified form the preheader of the
1825 // hoisted loop can't be part of some *other* loop.
1826 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))
1827 NewParentL->addChildLoop(HoistedL);
1829 LI.addTopLevelLoop(HoistedL);
1831 SubLoops.erase(SubLoopsSplitI, SubLoops.end());
1833 // Actually delete the loop if nothing remained within it.
1834 if (Blocks.empty()) {
1835 assert(SubLoops.empty() &&
1836 "Failed to remove all subloops from the original loop!");
1837 if (Loop *ParentL = L.getParentLoop())
1838 ParentL->removeChildLoop(llvm::find(*ParentL, &L));
1840 LI.removeLoop(llvm::find(LI, &L));
1848 /// Helper to visit a dominator subtree, invoking a callable on each node.
1850 /// Returning false at any point will stop walking past that node of the tree.
1851 template <typename CallableT>
1852 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
1853 SmallVector<DomTreeNode *, 4> DomWorklist;
1854 DomWorklist.push_back(DT[BB]);
1856 SmallPtrSet<DomTreeNode *, 4> Visited;
1857 Visited.insert(DT[BB]);
1860 DomTreeNode *N = DomWorklist.pop_back_val();
1863 if (!Callable(N->getBlock()))
1866 // Accumulate the child nodes.
1867 for (DomTreeNode *ChildN : *N) {
1868 assert(Visited.insert(ChildN).second &&
1869 "Cannot visit a node twice when walking a tree!");
1870 DomWorklist.push_back(ChildN);
1872 } while (!DomWorklist.empty());
1875 static void unswitchNontrivialInvariants(
1876 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
1877 SmallVectorImpl<BasicBlock *> &ExitBlocks, DominatorTree &DT, LoopInfo &LI,
1878 AssumptionCache &AC, function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
1879 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
1880 auto *ParentBB = TI.getParent();
1881 BranchInst *BI = dyn_cast<BranchInst>(&TI);
1882 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
1884 // We can only unswitch switches, conditional branches with an invariant
1885 // condition, or combining invariant conditions with an instruction.
1886 assert((SI || BI->isConditional()) &&
1887 "Can only unswitch switches and conditional branch!");
1888 bool FullUnswitch = SI || BI->getCondition() == Invariants[0];
1890 assert(Invariants.size() == 1 &&
1891 "Cannot have other invariants with full unswitching!");
1893 assert(isa<Instruction>(BI->getCondition()) &&
1894 "Partial unswitching requires an instruction as the condition!");
1896 if (MSSAU && VerifyMemorySSA)
1897 MSSAU->getMemorySSA()->verifyMemorySSA();
1899 // Constant and BBs tracking the cloned and continuing successor. When we are
1900 // unswitching the entire condition, this can just be trivially chosen to
1901 // unswitch towards `true`. However, when we are unswitching a set of
1902 // invariants combined with `and` or `or`, the combining operation determines
1903 // the best direction to unswitch: we want to unswitch the direction that will
1904 // collapse the branch.
1905 bool Direction = true;
1907 if (!FullUnswitch) {
1908 if (cast<Instruction>(BI->getCondition())->getOpcode() != Instruction::Or) {
1909 assert(cast<Instruction>(BI->getCondition())->getOpcode() ==
1911 "Only `or` and `and` instructions can combine invariants being "
1918 BasicBlock *RetainedSuccBB =
1919 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();
1920 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
1922 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));
1924 for (auto Case : SI->cases())
1925 if (Case.getCaseSuccessor() != RetainedSuccBB)
1926 UnswitchedSuccBBs.insert(Case.getCaseSuccessor());
1928 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
1929 "Should not unswitch the same successor we are retaining!");
1931 // The branch should be in this exact loop. Any inner loop's invariant branch
1932 // should be handled by unswitching that inner loop. The caller of this
1933 // routine should filter out any candidates that remain (but were skipped for
1934 // whatever reason).
1935 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
1937 // Compute the parent loop now before we start hacking on things.
1938 Loop *ParentL = L.getParentLoop();
1939 // Get blocks in RPO order for MSSA update, before changing the CFG.
1940 LoopBlocksRPO LBRPO(&L);
1944 // Compute the outer-most loop containing one of our exit blocks. This is the
1945 // furthest up our loopnest which can be mutated, which we will use below to
1947 Loop *OuterExitL = &L;
1948 for (auto *ExitBB : ExitBlocks) {
1949 Loop *NewOuterExitL = LI.getLoopFor(ExitBB);
1950 if (!NewOuterExitL) {
1951 // We exited the entire nest with this block, so we're done.
1952 OuterExitL = nullptr;
1955 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))
1956 OuterExitL = NewOuterExitL;
1959 // At this point, we're definitely going to unswitch something so invalidate
1960 // any cached information in ScalarEvolution for the outer most loop
1961 // containing an exit block and all nested loops.
1964 SE->forgetLoop(OuterExitL);
1966 SE->forgetTopmostLoop(&L);
1969 // If the edge from this terminator to a successor dominates that successor,
1970 // store a map from each block in its dominator subtree to it. This lets us
1971 // tell when cloning for a particular successor if a block is dominated by
1972 // some *other* successor with a single data structure. We use this to
1973 // significantly reduce cloning.
1974 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
1975 for (auto *SuccBB : llvm::concat<BasicBlock *const>(
1976 makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs))
1977 if (SuccBB->getUniquePredecessor() ||
1978 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
1979 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);
1981 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {
1982 DominatingSucc[BB] = SuccBB;
1986 // Split the preheader, so that we know that there is a safe place to insert
1987 // the conditional branch. We will change the preheader to have a conditional
1988 // branch on LoopCond. The original preheader will become the split point
1989 // between the unswitched versions, and we will have a new preheader for the
1991 BasicBlock *SplitBB = L.getLoopPreheader();
1992 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU);
1994 // Keep track of the dominator tree updates needed.
1995 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
1997 // Clone the loop for each unswitched successor.
1998 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
1999 VMaps.reserve(UnswitchedSuccBBs.size());
2000 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
2001 for (auto *SuccBB : UnswitchedSuccBBs) {
2002 VMaps.emplace_back(new ValueToValueMapTy());
2003 ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2004 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,
2005 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU);
2008 // The stitching of the branched code back together depends on whether we're
2009 // doing full unswitching or not with the exception that we always want to
2010 // nuke the initial terminator placed in the split block.
2011 SplitBB->getTerminator()->eraseFromParent();
2013 // Splice the terminator from the original loop and rewrite its
2015 SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI);
2017 // Keep a clone of the terminator for MSSA updates.
2018 Instruction *NewTI = TI.clone();
2019 ParentBB->getInstList().push_back(NewTI);
2021 // First wire up the moved terminator to the preheaders.
2023 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2024 BI->setSuccessor(ClonedSucc, ClonedPH);
2025 BI->setSuccessor(1 - ClonedSucc, LoopPH);
2026 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2028 assert(SI && "Must either be a branch or switch!");
2030 // Walk the cases and directly update their successors.
2031 assert(SI->getDefaultDest() == RetainedSuccBB &&
2032 "Not retaining default successor!");
2033 SI->setDefaultDest(LoopPH);
2034 for (auto &Case : SI->cases())
2035 if (Case.getCaseSuccessor() == RetainedSuccBB)
2036 Case.setSuccessor(LoopPH);
2038 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);
2040 // We need to use the set to populate domtree updates as even when there
2041 // are multiple cases pointing at the same successor we only want to
2042 // remove and insert one edge in the domtree.
2043 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2044 DTUpdates.push_back(
2045 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});
2049 DT.applyUpdates(DTUpdates);
2052 // Remove all but one edge to the retained block and all unswitched
2053 // blocks. This is to avoid having duplicate entries in the cloned Phis,
2054 // when we know we only keep a single edge for each case.
2055 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB);
2056 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2057 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB);
2059 for (auto &VMap : VMaps)
2060 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2061 /*IgnoreIncomingWithNoClones=*/true);
2062 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2064 // Remove all edges to unswitched blocks.
2065 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2066 MSSAU->removeEdge(ParentBB, SuccBB);
2069 // Now unhook the successor relationship as we'll be replacing
2070 // the terminator with a direct branch. This is much simpler for branches
2071 // than switches so we handle those first.
2073 // Remove the parent as a predecessor of the unswitched successor.
2074 assert(UnswitchedSuccBBs.size() == 1 &&
2075 "Only one possible unswitched block for a branch!");
2076 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2077 UnswitchedSuccBB->removePredecessor(ParentBB,
2078 /*DontDeleteUselessPHIs*/ true);
2079 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2081 // Note that we actually want to remove the parent block as a predecessor
2082 // of *every* case successor. The case successor is either unswitched,
2083 // completely eliminating an edge from the parent to that successor, or it
2084 // is a duplicate edge to the retained successor as the retained successor
2085 // is always the default successor and as we'll replace this with a direct
2086 // branch we no longer need the duplicate entries in the PHI nodes.
2087 SwitchInst *NewSI = cast<SwitchInst>(NewTI);
2088 assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2089 "Not retaining default successor!");
2090 for (auto &Case : NewSI->cases())
2091 Case.getCaseSuccessor()->removePredecessor(
2093 /*DontDeleteUselessPHIs*/ true);
2095 // We need to use the set to populate domtree updates as even when there
2096 // are multiple cases pointing at the same successor we only want to
2097 // remove and insert one edge in the domtree.
2098 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2099 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});
2102 // After MSSAU update, remove the cloned terminator instruction NewTI.
2103 ParentBB->getTerminator()->eraseFromParent();
2105 // Create a new unconditional branch to the continuing block (as opposed to
2107 BranchInst::Create(RetainedSuccBB, ParentBB);
2109 assert(BI && "Only branches have partial unswitching.");
2110 assert(UnswitchedSuccBBs.size() == 1 &&
2111 "Only one possible unswitched block for a branch!");
2112 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2113 // When doing a partial unswitch, we have to do a bit more work to build up
2114 // the branch in the split block.
2115 buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction,
2116 *ClonedPH, *LoopPH);
2117 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2120 // Apply the updates accumulated above to get an up-to-date dominator tree.
2121 DT.applyUpdates(DTUpdates);
2122 if (!FullUnswitch && MSSAU) {
2123 // Update MSSA for partial unswitch, after DT update.
2124 SmallVector<CFGUpdate, 1> Updates;
2126 {cfg::UpdateKind::Insert, SplitBB, ClonedPHs.begin()->second});
2127 MSSAU->applyInsertUpdates(Updates, DT);
2130 // Now that we have an accurate dominator tree, first delete the dead cloned
2131 // blocks so that we can accurately build any cloned loops. It is important to
2132 // not delete the blocks from the original loop yet because we still want to
2133 // reference the original loop to understand the cloned loop's structure.
2134 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2136 // Build the cloned loop structure itself. This may be substantially
2137 // different from the original structure due to the simplified CFG. This also
2138 // handles inserting all the cloned blocks into the correct loops.
2139 SmallVector<Loop *, 4> NonChildClonedLoops;
2140 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2141 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);
2143 // Now that our cloned loops have been built, we can update the original loop.
2144 // First we delete the dead blocks from it and then we rebuild the loop
2145 // structure taking these deletions into account.
2146 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU);
2148 if (MSSAU && VerifyMemorySSA)
2149 MSSAU->getMemorySSA()->verifyMemorySSA();
2151 SmallVector<Loop *, 4> HoistedLoops;
2152 bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops);
2154 if (MSSAU && VerifyMemorySSA)
2155 MSSAU->getMemorySSA()->verifyMemorySSA();
2157 // This transformation has a high risk of corrupting the dominator tree, and
2158 // the below steps to rebuild loop structures will result in hard to debug
2159 // errors in that case so verify that the dominator tree is sane first.
2160 // FIXME: Remove this when the bugs stop showing up and rely on existing
2161 // verification steps.
2162 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2165 // If we unswitched a branch which collapses the condition to a known
2166 // constant we want to replace all the uses of the invariants within both
2167 // the original and cloned blocks. We do this here so that we can use the
2168 // now updated dominator tree to identify which side the users are on.
2169 assert(UnswitchedSuccBBs.size() == 1 &&
2170 "Only one possible unswitched block for a branch!");
2171 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2173 // When considering multiple partially-unswitched invariants
2174 // we cant just go replace them with constants in both branches.
2176 // For 'AND' we infer that true branch ("continue") means true
2177 // for each invariant operand.
2178 // For 'OR' we can infer that false branch ("continue") means false
2179 // for each invariant operand.
2180 // So it happens that for multiple-partial case we dont replace
2181 // in the unswitched branch.
2182 bool ReplaceUnswitched = FullUnswitch || (Invariants.size() == 1);
2184 ConstantInt *UnswitchedReplacement =
2185 Direction ? ConstantInt::getTrue(BI->getContext())
2186 : ConstantInt::getFalse(BI->getContext());
2187 ConstantInt *ContinueReplacement =
2188 Direction ? ConstantInt::getFalse(BI->getContext())
2189 : ConstantInt::getTrue(BI->getContext());
2190 for (Value *Invariant : Invariants)
2191 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end();
2193 // Grab the use and walk past it so we can clobber it in the use list.
2195 Instruction *UserI = dyn_cast<Instruction>(U->getUser());
2199 // Replace it with the 'continue' side if in the main loop body, and the
2200 // unswitched if in the cloned blocks.
2201 if (DT.dominates(LoopPH, UserI->getParent()))
2202 U->set(ContinueReplacement);
2203 else if (ReplaceUnswitched &&
2204 DT.dominates(ClonedPH, UserI->getParent()))
2205 U->set(UnswitchedReplacement);
2209 // We can change which blocks are exit blocks of all the cloned sibling
2210 // loops, the current loop, and any parent loops which shared exit blocks
2211 // with the current loop. As a consequence, we need to re-form LCSSA for
2212 // them. But we shouldn't need to re-form LCSSA for any child loops.
2213 // FIXME: This could be made more efficient by tracking which exit blocks are
2214 // new, and focusing on them, but that isn't likely to be necessary.
2216 // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2217 // loop nest and update every loop that could have had its exits changed. We
2218 // also need to cover any intervening loops. We add all of these loops to
2219 // a list and sort them by loop depth to achieve this without updating
2220 // unnecessary loops.
2221 auto UpdateLoop = [&](Loop &UpdateL) {
2223 UpdateL.verifyLoop();
2224 for (Loop *ChildL : UpdateL) {
2225 ChildL->verifyLoop();
2226 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2227 "Perturbed a child loop's LCSSA form!");
2230 // First build LCSSA for this loop so that we can preserve it when
2231 // forming dedicated exits. We don't want to perturb some other loop's
2232 // LCSSA while doing that CFG edit.
2233 formLCSSA(UpdateL, DT, &LI, nullptr);
2235 // For loops reached by this loop's original exit blocks we may
2236 // introduced new, non-dedicated exits. At least try to re-form dedicated
2237 // exits for these loops. This may fail if they couldn't have dedicated
2238 // exits to start with.
2239 formDedicatedExitBlocks(&UpdateL, &DT, &LI, /*PreserveLCSSA*/ true);
2242 // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2243 // and we can do it in any order as they don't nest relative to each other.
2245 // Also check if any of the loops we have updated have become top-level loops
2246 // as that will necessitate widening the outer loop scope.
2247 for (Loop *UpdatedL :
2248 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {
2249 UpdateLoop(*UpdatedL);
2250 if (!UpdatedL->getParentLoop())
2251 OuterExitL = nullptr;
2255 if (!L.getParentLoop())
2256 OuterExitL = nullptr;
2259 // If the original loop had exit blocks, walk up through the outer most loop
2260 // of those exit blocks to update LCSSA and form updated dedicated exits.
2261 if (OuterExitL != &L)
2262 for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2263 OuterL = OuterL->getParentLoop())
2264 UpdateLoop(*OuterL);
2267 // Verify the entire loop structure to catch any incorrect updates before we
2268 // progress in the pass pipeline.
2272 // Now that we've unswitched something, make callbacks to report the changes.
2273 // For that we need to merge together the updated loops and the cloned loops
2274 // and check whether the original loop survived.
2275 SmallVector<Loop *, 4> SibLoops;
2276 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
2277 if (UpdatedL->getParentLoop() == ParentL)
2278 SibLoops.push_back(UpdatedL);
2279 UnswitchCB(IsStillLoop, SibLoops);
2281 if (MSSAU && VerifyMemorySSA)
2282 MSSAU->getMemorySSA()->verifyMemorySSA();
2290 /// Recursively compute the cost of a dominator subtree based on the per-block
2291 /// cost map provided.
2293 /// The recursive computation is memozied into the provided DT-indexed cost map
2294 /// to allow querying it for most nodes in the domtree without it becoming
2297 computeDomSubtreeCost(DomTreeNode &N,
2298 const SmallDenseMap<BasicBlock *, int, 4> &BBCostMap,
2299 SmallDenseMap<DomTreeNode *, int, 4> &DTCostMap) {
2300 // Don't accumulate cost (or recurse through) blocks not in our block cost
2301 // map and thus not part of the duplication cost being considered.
2302 auto BBCostIt = BBCostMap.find(N.getBlock());
2303 if (BBCostIt == BBCostMap.end())
2306 // Lookup this node to see if we already computed its cost.
2307 auto DTCostIt = DTCostMap.find(&N);
2308 if (DTCostIt != DTCostMap.end())
2309 return DTCostIt->second;
2311 // If not, we have to compute it. We can't use insert above and update
2312 // because computing the cost may insert more things into the map.
2313 int Cost = std::accumulate(
2314 N.begin(), N.end(), BBCostIt->second, [&](int Sum, DomTreeNode *ChildN) {
2315 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
2317 bool Inserted = DTCostMap.insert({&N, Cost}).second;
2319 assert(Inserted && "Should not insert a node while visiting children!");
2323 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2324 /// making the following replacement:
2326 /// --code before guard--
2327 /// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2328 /// --code after guard--
2332 /// --code before guard--
2333 /// br i1 %cond, label %guarded, label %deopt
2336 /// --code after guard--
2339 /// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2342 /// It also makes all relevant DT and LI updates, so that all structures are in
2343 /// valid state after this transform.
2345 turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2346 SmallVectorImpl<BasicBlock *> &ExitBlocks,
2347 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
2348 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2349 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2350 BasicBlock *CheckBB = GI->getParent();
2352 if (MSSAU && VerifyMemorySSA)
2353 MSSAU->getMemorySSA()->verifyMemorySSA();
2355 // Remove all CheckBB's successors from DomTree. A block can be seen among
2356 // successors more than once, but for DomTree it should be added only once.
2357 SmallPtrSet<BasicBlock *, 4> Successors;
2358 for (auto *Succ : successors(CheckBB))
2359 if (Successors.insert(Succ).second)
2360 DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ});
2362 Instruction *DeoptBlockTerm =
2363 SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true);
2364 BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
2365 // SplitBlockAndInsertIfThen inserts control flow that branches to
2366 // DeoptBlockTerm if the condition is true. We want the opposite.
2367 CheckBI->swapSuccessors();
2369 BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
2370 GuardedBlock->setName("guarded");
2371 CheckBI->getSuccessor(1)->setName("deopt");
2372 BasicBlock *DeoptBlock = CheckBI->getSuccessor(1);
2374 // We now have a new exit block.
2375 ExitBlocks.push_back(CheckBI->getSuccessor(1));
2378 MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI);
2380 GI->moveBefore(DeoptBlockTerm);
2381 GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));
2383 // Add new successors of CheckBB into DomTree.
2384 for (auto *Succ : successors(CheckBB))
2385 DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ});
2387 // Now the blocks that used to be CheckBB's successors are GuardedBlock's
2389 for (auto *Succ : Successors)
2390 DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ});
2392 // Make proper changes to DT.
2393 DT.applyUpdates(DTUpdates);
2394 // Inform LI of a new loop block.
2395 L.addBasicBlockToLoop(GuardedBlock, LI);
2398 MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI));
2399 MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::End);
2400 if (VerifyMemorySSA)
2401 MSSAU->getMemorySSA()->verifyMemorySSA();
2408 /// Cost multiplier is a way to limit potentially exponential behavior
2409 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2410 /// candidates available. Also accounting for the number of "sibling" loops with
2411 /// the idea to account for previous unswitches that already happened on this
2412 /// cluster of loops. There was an attempt to keep this formula simple,
2413 /// just enough to limit the worst case behavior. Even if it is not that simple
2414 /// now it is still not an attempt to provide a detailed heuristic size
2417 /// TODO: Make a proper accounting of "explosion" effect for all kinds of
2418 /// unswitch candidates, making adequate predictions instead of wild guesses.
2419 /// That requires knowing not just the number of "remaining" candidates but
2420 /// also costs of unswitching for each of these candidates.
2421 static int calculateUnswitchCostMultiplier(
2422 Instruction &TI, Loop &L, LoopInfo &LI, DominatorTree &DT,
2423 ArrayRef<std::pair<Instruction *, TinyPtrVector<Value *>>>
2424 UnswitchCandidates) {
2426 // Guards and other exiting conditions do not contribute to exponential
2427 // explosion as soon as they dominate the latch (otherwise there might be
2428 // another path to the latch remaining that does not allow to eliminate the
2429 // loop copy on unswitch).
2430 BasicBlock *Latch = L.getLoopLatch();
2431 BasicBlock *CondBlock = TI.getParent();
2432 if (DT.dominates(CondBlock, Latch) &&
2434 llvm::count_if(successors(&TI), [&L](BasicBlock *SuccBB) {
2435 return L.contains(SuccBB);
2437 NumCostMultiplierSkipped++;
2441 auto *ParentL = L.getParentLoop();
2442 int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2443 : std::distance(LI.begin(), LI.end()));
2444 // Count amount of clones that all the candidates might cause during
2445 // unswitching. Branch/guard counts as 1, switch counts as log2 of its cases.
2446 int UnswitchedClones = 0;
2447 for (auto Candidate : UnswitchCandidates) {
2448 Instruction *CI = Candidate.first;
2449 BasicBlock *CondBlock = CI->getParent();
2450 bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch);
2452 if (!SkipExitingSuccessors)
2456 int NonExitingSuccessors = llvm::count_if(
2457 successors(CondBlock), [SkipExitingSuccessors, &L](BasicBlock *SuccBB) {
2458 return !SkipExitingSuccessors || L.contains(SuccBB);
2460 UnswitchedClones += Log2_32(NonExitingSuccessors);
2463 // Ignore up to the "unscaled candidates" number of unswitch candidates
2464 // when calculating the power-of-two scaling of the cost. The main idea
2465 // with this control is to allow a small number of unswitches to happen
2466 // and rely more on siblings multiplier (see below) when the number
2467 // of candidates is small.
2468 unsigned ClonesPower =
2469 std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0);
2471 // Allowing top-level loops to spread a bit more than nested ones.
2472 int SiblingsMultiplier =
2473 std::max((ParentL ? SiblingsCount
2474 : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2476 // Compute the cost multiplier in a way that won't overflow by saturating
2477 // at an upper bound.
2479 if (ClonesPower > Log2_32(UnswitchThreshold) ||
2480 SiblingsMultiplier > UnswitchThreshold)
2481 CostMultiplier = UnswitchThreshold;
2483 CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower),
2484 (int)UnswitchThreshold);
2486 LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier
2487 << " (siblings " << SiblingsMultiplier << " * clones "
2488 << (1 << ClonesPower) << ")"
2489 << " for unswitch candidate: " << TI << "\n");
2490 return CostMultiplier;
2494 unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
2495 AssumptionCache &AC, TargetTransformInfo &TTI,
2496 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2497 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
2498 // Collect all invariant conditions within this loop (as opposed to an inner
2499 // loop which would be handled when visiting that inner loop).
2500 SmallVector<std::pair<Instruction *, TinyPtrVector<Value *>>, 4>
2503 // Whether or not we should also collect guards in the loop.
2504 bool CollectGuards = false;
2505 if (UnswitchGuards) {
2506 auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2507 Intrinsic::getName(Intrinsic::experimental_guard));
2508 if (GuardDecl && !GuardDecl->use_empty())
2509 CollectGuards = true;
2512 for (auto *BB : L.blocks()) {
2513 if (LI.getLoopFor(BB) != &L)
2519 auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0);
2520 // TODO: Support AND, OR conditions and partial unswitching.
2521 if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
2522 UnswitchCandidates.push_back({&I, {Cond}});
2525 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
2526 // We can only consider fully loop-invariant switch conditions as we need
2527 // to completely eliminate the switch after unswitching.
2528 if (!isa<Constant>(SI->getCondition()) &&
2529 L.isLoopInvariant(SI->getCondition()))
2530 UnswitchCandidates.push_back({SI, {SI->getCondition()}});
2534 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
2535 if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) ||
2536 BI->getSuccessor(0) == BI->getSuccessor(1))
2539 if (L.isLoopInvariant(BI->getCondition())) {
2540 UnswitchCandidates.push_back({BI, {BI->getCondition()}});
2544 Instruction &CondI = *cast<Instruction>(BI->getCondition());
2545 if (CondI.getOpcode() != Instruction::And &&
2546 CondI.getOpcode() != Instruction::Or)
2549 TinyPtrVector<Value *> Invariants =
2550 collectHomogenousInstGraphLoopInvariants(L, CondI, LI);
2551 if (Invariants.empty())
2554 UnswitchCandidates.push_back({BI, std::move(Invariants)});
2557 // If we didn't find any candidates, we're done.
2558 if (UnswitchCandidates.empty())
2561 // Check if there are irreducible CFG cycles in this loop. If so, we cannot
2562 // easily unswitch non-trivial edges out of the loop. Doing so might turn the
2563 // irreducible control flow into reducible control flow and introduce new
2564 // loops "out of thin air". If we ever discover important use cases for doing
2565 // this, we can add support to loop unswitch, but it is a lot of complexity
2566 // for what seems little or no real world benefit.
2567 LoopBlocksRPO RPOT(&L);
2569 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
2572 SmallVector<BasicBlock *, 4> ExitBlocks;
2573 L.getUniqueExitBlocks(ExitBlocks);
2575 // We cannot unswitch if exit blocks contain a cleanuppad instruction as we
2576 // don't know how to split those exit blocks.
2577 // FIXME: We should teach SplitBlock to handle this and remove this
2579 for (auto *ExitBB : ExitBlocks)
2580 if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI())) {
2581 dbgs() << "Cannot unswitch because of cleanuppad in exit block\n";
2586 dbgs() << "Considering " << UnswitchCandidates.size()
2587 << " non-trivial loop invariant conditions for unswitching.\n");
2589 // Given that unswitching these terminators will require duplicating parts of
2590 // the loop, so we need to be able to model that cost. Compute the ephemeral
2591 // values and set up a data structure to hold per-BB costs. We cache each
2592 // block's cost so that we don't recompute this when considering different
2593 // subsets of the loop for duplication during unswitching.
2594 SmallPtrSet<const Value *, 4> EphValues;
2595 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);
2596 SmallDenseMap<BasicBlock *, int, 4> BBCostMap;
2598 // Compute the cost of each block, as well as the total loop cost. Also, bail
2599 // out if we see instructions which are incompatible with loop unswitching
2600 // (convergent, noduplicate, or cross-basic-block tokens).
2601 // FIXME: We might be able to safely handle some of these in non-duplicated
2604 for (auto *BB : L.blocks()) {
2606 for (auto &I : *BB) {
2607 if (EphValues.count(&I))
2610 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
2612 if (auto CS = CallSite(&I))
2613 if (CS.isConvergent() || CS.cannotDuplicate())
2616 Cost += TTI.getUserCost(&I);
2618 assert(Cost >= 0 && "Must not have negative costs!");
2620 assert(LoopCost >= 0 && "Must not have negative loop costs!");
2621 BBCostMap[BB] = Cost;
2623 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n");
2625 // Now we find the best candidate by searching for the one with the following
2626 // properties in order:
2628 // 1) An unswitching cost below the threshold
2629 // 2) The smallest number of duplicated unswitch candidates (to avoid
2630 // creating redundant subsequent unswitching)
2631 // 3) The smallest cost after unswitching.
2633 // We prioritize reducing fanout of unswitch candidates provided the cost
2634 // remains below the threshold because this has a multiplicative effect.
2636 // This requires memoizing each dominator subtree to avoid redundant work.
2638 // FIXME: Need to actually do the number of candidates part above.
2639 SmallDenseMap<DomTreeNode *, int, 4> DTCostMap;
2640 // Given a terminator which might be unswitched, computes the non-duplicated
2641 // cost for that terminator.
2642 auto ComputeUnswitchedCost = [&](Instruction &TI, bool FullUnswitch) {
2643 BasicBlock &BB = *TI.getParent();
2644 SmallPtrSet<BasicBlock *, 4> Visited;
2646 int Cost = LoopCost;
2647 for (BasicBlock *SuccBB : successors(&BB)) {
2648 // Don't count successors more than once.
2649 if (!Visited.insert(SuccBB).second)
2652 // If this is a partial unswitch candidate, then it must be a conditional
2653 // branch with a condition of either `or` or `and`. In that case, one of
2654 // the successors is necessarily duplicated, so don't even try to remove
2656 if (!FullUnswitch) {
2657 auto &BI = cast<BranchInst>(TI);
2658 if (cast<Instruction>(BI.getCondition())->getOpcode() ==
2660 if (SuccBB == BI.getSuccessor(1))
2663 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
2665 "Only `and` and `or` conditions can result in a partial "
2667 if (SuccBB == BI.getSuccessor(0))
2672 // This successor's domtree will not need to be duplicated after
2673 // unswitching if the edge to the successor dominates it (and thus the
2674 // entire tree). This essentially means there is no other path into this
2675 // subtree and so it will end up live in only one clone of the loop.
2676 if (SuccBB->getUniquePredecessor() ||
2677 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2678 return PredBB == &BB || DT.dominates(SuccBB, PredBB);
2680 Cost -= computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);
2682 "Non-duplicated cost should never exceed total loop cost!");
2686 // Now scale the cost by the number of unique successors minus one. We
2687 // subtract one because there is already at least one copy of the entire
2688 // loop. This is computing the new cost of unswitching a condition.
2689 // Note that guards always have 2 unique successors that are implicit and
2690 // will be materialized if we decide to unswitch it.
2691 int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
2692 assert(SuccessorsCount > 1 &&
2693 "Cannot unswitch a condition without multiple distinct successors!");
2694 return Cost * (SuccessorsCount - 1);
2696 Instruction *BestUnswitchTI = nullptr;
2697 int BestUnswitchCost;
2698 ArrayRef<Value *> BestUnswitchInvariants;
2699 for (auto &TerminatorAndInvariants : UnswitchCandidates) {
2700 Instruction &TI = *TerminatorAndInvariants.first;
2701 ArrayRef<Value *> Invariants = TerminatorAndInvariants.second;
2702 BranchInst *BI = dyn_cast<BranchInst>(&TI);
2703 int CandidateCost = ComputeUnswitchedCost(
2704 TI, /*FullUnswitch*/ !BI || (Invariants.size() == 1 &&
2705 Invariants[0] == BI->getCondition()));
2706 // Calculate cost multiplier which is a tool to limit potentially
2707 // exponential behavior of loop-unswitch.
2708 if (EnableUnswitchCostMultiplier) {
2709 int CostMultiplier =
2710 calculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
2712 (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
2713 "cost multiplier needs to be in the range of 1..UnswitchThreshold");
2714 CandidateCost *= CostMultiplier;
2715 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
2716 << " (multiplier: " << CostMultiplier << ")"
2717 << " for unswitch candidate: " << TI << "\n");
2719 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
2720 << " for unswitch candidate: " << TI << "\n");
2723 if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) {
2724 BestUnswitchTI = &TI;
2725 BestUnswitchCost = CandidateCost;
2726 BestUnswitchInvariants = Invariants;
2730 if (BestUnswitchCost >= UnswitchThreshold) {
2731 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: "
2732 << BestUnswitchCost << "\n");
2736 // If the best candidate is a guard, turn it into a branch.
2737 if (isGuard(BestUnswitchTI))
2738 BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L,
2739 ExitBlocks, DT, LI, MSSAU);
2741 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = "
2742 << BestUnswitchCost << ") terminator: " << *BestUnswitchTI
2744 unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants,
2745 ExitBlocks, DT, LI, AC, UnswitchCB, SE, MSSAU);
2749 /// Unswitch control flow predicated on loop invariant conditions.
2751 /// This first hoists all branches or switches which are trivial (IE, do not
2752 /// require duplicating any part of the loop) out of the loop body. It then
2753 /// looks at other loop invariant control flows and tries to unswitch those as
2754 /// well by cloning the loop if the result is small enough.
2756 /// The `DT`, `LI`, `AC`, `TTI` parameters are required analyses that are also
2757 /// updated based on the unswitch.
2758 /// The `MSSA` analysis is also updated if valid (i.e. its use is enabled).
2760 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
2761 /// true, we will attempt to do non-trivial unswitching as well as trivial
2764 /// The `UnswitchCB` callback provided will be run after unswitching is
2765 /// complete, with the first parameter set to `true` if the provided loop
2766 /// remains a loop, and a list of new sibling loops created.
2768 /// If `SE` is non-null, we will update that analysis based on the unswitching
2770 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
2771 AssumptionCache &AC, TargetTransformInfo &TTI,
2773 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2774 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
2775 assert(L.isRecursivelyLCSSAForm(DT, LI) &&
2776 "Loops must be in LCSSA form before unswitching.");
2777 bool Changed = false;
2779 // Must be in loop simplified form: we need a preheader and dedicated exits.
2780 if (!L.isLoopSimplifyForm())
2783 // Try trivial unswitch first before loop over other basic blocks in the loop.
2784 if (unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
2785 // If we unswitched successfully we will want to clean up the loop before
2786 // processing it further so just mark it as unswitched and return.
2787 UnswitchCB(/*CurrentLoopValid*/ true, {});
2791 // If we're not doing non-trivial unswitching, we're done. We both accept
2792 // a parameter but also check a local flag that can be used for testing
2794 if (!NonTrivial && !EnableNonTrivialUnswitch)
2797 // For non-trivial unswitching, because it often creates new loops, we rely on
2798 // the pass manager to iterate on the loops rather than trying to immediately
2799 // reach a fixed point. There is no substantial advantage to iterating
2800 // internally, and if any of the new loops are simplified enough to contain
2801 // trivial unswitching we want to prefer those.
2803 // Try to unswitch the best invariant condition. We prefer this full unswitch to
2804 // a partial unswitch when possible below the threshold.
2805 if (unswitchBestCondition(L, DT, LI, AC, TTI, UnswitchCB, SE, MSSAU))
2808 // No other opportunities to unswitch.
2812 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
2813 LoopStandardAnalysisResults &AR,
2815 Function &F = *L.getHeader()->getParent();
2818 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
2821 // Save the current loop name in a variable so that we can report it even
2822 // after it has been deleted.
2823 std::string LoopName = L.getName();
2825 auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid,
2826 ArrayRef<Loop *> NewLoops) {
2827 // If we did a non-trivial unswitch, we have added new (cloned) loops.
2828 if (!NewLoops.empty())
2829 U.addSiblingLoops(NewLoops);
2831 // If the current loop remains valid, we should revisit it to catch any
2832 // other unswitch opportunities. Otherwise, we need to mark it as deleted.
2833 if (CurrentLoopValid)
2834 U.revisitCurrentLoop();
2836 U.markLoopAsDeleted(L, LoopName);
2839 Optional<MemorySSAUpdater> MSSAU;
2841 MSSAU = MemorySSAUpdater(AR.MSSA);
2842 if (VerifyMemorySSA)
2843 AR.MSSA->verifyMemorySSA();
2845 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.TTI, NonTrivial, UnswitchCB,
2846 &AR.SE, MSSAU.hasValue() ? MSSAU.getPointer() : nullptr))
2847 return PreservedAnalyses::all();
2849 if (AR.MSSA && VerifyMemorySSA)
2850 AR.MSSA->verifyMemorySSA();
2852 // Historically this pass has had issues with the dominator tree so verify it
2853 // in asserts builds.
2854 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));
2855 return getLoopPassPreservedAnalyses();
2860 class SimpleLoopUnswitchLegacyPass : public LoopPass {
2864 static char ID; // Pass ID, replacement for typeid
2866 explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false)
2867 : LoopPass(ID), NonTrivial(NonTrivial) {
2868 initializeSimpleLoopUnswitchLegacyPassPass(
2869 *PassRegistry::getPassRegistry());
2872 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
2874 void getAnalysisUsage(AnalysisUsage &AU) const override {
2875 AU.addRequired<AssumptionCacheTracker>();
2876 AU.addRequired<TargetTransformInfoWrapperPass>();
2877 if (EnableMSSALoopDependency) {
2878 AU.addRequired<MemorySSAWrapperPass>();
2879 AU.addPreserved<MemorySSAWrapperPass>();
2881 getLoopAnalysisUsage(AU);
2885 } // end anonymous namespace
2887 bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
2891 Function &F = *L->getHeader()->getParent();
2893 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L
2896 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2897 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2898 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2899 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2900 MemorySSA *MSSA = nullptr;
2901 Optional<MemorySSAUpdater> MSSAU;
2902 if (EnableMSSALoopDependency) {
2903 MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
2904 MSSAU = MemorySSAUpdater(MSSA);
2907 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
2908 auto *SE = SEWP ? &SEWP->getSE() : nullptr;
2910 auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid,
2911 ArrayRef<Loop *> NewLoops) {
2912 // If we did a non-trivial unswitch, we have added new (cloned) loops.
2913 for (auto *NewL : NewLoops)
2916 // If the current loop remains valid, re-add it to the queue. This is
2917 // a little wasteful as we'll finish processing the current loop as well,
2918 // but it is the best we can do in the old PM.
2919 if (CurrentLoopValid)
2922 LPM.markLoopAsDeleted(*L);
2925 if (MSSA && VerifyMemorySSA)
2926 MSSA->verifyMemorySSA();
2928 bool Changed = unswitchLoop(*L, DT, LI, AC, TTI, NonTrivial, UnswitchCB, SE,
2929 MSSAU.hasValue() ? MSSAU.getPointer() : nullptr);
2931 if (MSSA && VerifyMemorySSA)
2932 MSSA->verifyMemorySSA();
2934 // If anything was unswitched, also clear any cached information about this
2936 LPM.deleteSimpleAnalysisLoop(L);
2938 // Historically this pass has had issues with the dominator tree so verify it
2939 // in asserts builds.
2940 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2945 char SimpleLoopUnswitchLegacyPass::ID = 0;
2946 INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
2947 "Simple unswitch loops", false, false)
2948 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2949 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2950 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2951 INITIALIZE_PASS_DEPENDENCY(LoopPass)
2952 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
2953 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
2954 INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
2955 "Simple unswitch loops", false, false)
2957 Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) {
2958 return new SimpleLoopUnswitchLegacyPass(NonTrivial);