1 ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8
9 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
10 #include "llvm/ADT/DenseMap.h"
11 #include "llvm/ADT/STLExtras.h"
12 #include "llvm/ADT/Sequence.h"
13 #include "llvm/ADT/SetVector.h"
14 #include "llvm/ADT/SmallPtrSet.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/ADT/Twine.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/CFG.h"
20 #include "llvm/Analysis/CodeMetrics.h"
21 #include "llvm/Analysis/GuardUtils.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/LoopAnalysisManager.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/LoopIterator.h"
26 #include "llvm/Analysis/LoopPass.h"
27 #include "llvm/Analysis/MemorySSA.h"
28 #include "llvm/Analysis/MemorySSAUpdater.h"
29 #include "llvm/Analysis/MustExecute.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Constant.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/InstrTypes.h"
36 #include "llvm/IR/Instruction.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/Use.h"
41 #include "llvm/IR/Value.h"
42 #include "llvm/InitializePasses.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/Casting.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/GenericDomTree.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Cloning.h"
53 #include "llvm/Transforms/Utils/Local.h"
54 #include "llvm/Transforms/Utils/LoopUtils.h"
55 #include "llvm/Transforms/Utils/ValueMapper.h"
56 #include <algorithm>
57 #include <cassert>
58 #include <iterator>
59 #include <numeric>
60 #include <utility>
61
62 #define DEBUG_TYPE "simple-loop-unswitch"
63
64 using namespace llvm;
65
66 STATISTIC(NumBranches, "Number of branches unswitched");
67 STATISTIC(NumSwitches, "Number of switches unswitched");
68 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
69 STATISTIC(NumTrivial, "Number of unswitches that are trivial");
70 STATISTIC(
71 NumCostMultiplierSkipped,
72 "Number of unswitch candidates that had their cost multiplier skipped");
73
74 static cl::opt<bool> EnableNonTrivialUnswitch(
75 "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
76 cl::desc("Forcibly enables non-trivial loop unswitching rather than "
77 "following the configuration passed into the pass."));
78
79 static cl::opt<int>
80 UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
81 cl::desc("The cost threshold for unswitching a loop."));
82
83 static cl::opt<bool> EnableUnswitchCostMultiplier(
84 "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden,
85 cl::desc("Enable unswitch cost multiplier that prohibits exponential "
86 "explosion in nontrivial unswitch."));
87 static cl::opt<int> UnswitchSiblingsToplevelDiv(
88 "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden,
89 cl::desc("Toplevel siblings divisor for cost multiplier."));
90 static cl::opt<int> UnswitchNumInitialUnscaledCandidates(
91 "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden,
92 cl::desc("Number of unswitch candidates that are ignored when calculating "
93 "cost multiplier."));
94 static cl::opt<bool> UnswitchGuards(
95 "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
96 cl::desc("If enabled, simple loop unswitching will also consider "
97 "llvm.experimental.guard intrinsics as unswitch candidates."));
98 static cl::opt<bool> DropNonTrivialImplicitNullChecks(
99 "simple-loop-unswitch-drop-non-trivial-implicit-null-checks",
100 cl::init(false), cl::Hidden,
101 cl::desc("If enabled, drop make.implicit metadata in unswitched implicit "
102 "null checks to save time analyzing if we can keep it."));
103
104 /// Collect all of the loop invariant input values transitively used by the
105 /// homogeneous instruction graph from a given root.
106 ///
107 /// This essentially walks from a root recursively through loop variant operands
108 /// which have the exact same opcode and finds all inputs which are loop
109 /// invariant. For some operations these can be re-associated and unswitched out
110 /// of the loop entirely.
111 static TinyPtrVector<Value *>
collectHomogenousInstGraphLoopInvariants(Loop & L,Instruction & Root,LoopInfo & LI)112 collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root,
113 LoopInfo &LI) {
114 assert(!L.isLoopInvariant(&Root) &&
115 "Only need to walk the graph if root itself is not invariant.");
116 TinyPtrVector<Value *> Invariants;
117
118 // Build a worklist and recurse through operators collecting invariants.
119 SmallVector<Instruction *, 4> Worklist;
120 SmallPtrSet<Instruction *, 8> Visited;
121 Worklist.push_back(&Root);
122 Visited.insert(&Root);
123 do {
124 Instruction &I = *Worklist.pop_back_val();
125 for (Value *OpV : I.operand_values()) {
126 // Skip constants as unswitching isn't interesting for them.
127 if (isa<Constant>(OpV))
128 continue;
129
130 // Add it to our result if loop invariant.
131 if (L.isLoopInvariant(OpV)) {
132 Invariants.push_back(OpV);
133 continue;
134 }
135
136 // If not an instruction with the same opcode, nothing we can do.
137 Instruction *OpI = dyn_cast<Instruction>(OpV);
138 if (!OpI || OpI->getOpcode() != Root.getOpcode())
139 continue;
140
141 // Visit this operand.
142 if (Visited.insert(OpI).second)
143 Worklist.push_back(OpI);
144 }
145 } while (!Worklist.empty());
146
147 return Invariants;
148 }
149
replaceLoopInvariantUses(Loop & L,Value * Invariant,Constant & Replacement)150 static void replaceLoopInvariantUses(Loop &L, Value *Invariant,
151 Constant &Replacement) {
152 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");
153
154 // Replace uses of LIC in the loop with the given constant.
155 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end(); UI != UE;) {
156 // Grab the use and walk past it so we can clobber it in the use list.
157 Use *U = &*UI++;
158 Instruction *UserI = dyn_cast<Instruction>(U->getUser());
159
160 // Replace this use within the loop body.
161 if (UserI && L.contains(UserI))
162 U->set(&Replacement);
163 }
164 }
165
166 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial
167 /// incoming values along this edge.
areLoopExitPHIsLoopInvariant(Loop & L,BasicBlock & ExitingBB,BasicBlock & ExitBB)168 static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB,
169 BasicBlock &ExitBB) {
170 for (Instruction &I : ExitBB) {
171 auto *PN = dyn_cast<PHINode>(&I);
172 if (!PN)
173 // No more PHIs to check.
174 return true;
175
176 // If the incoming value for this edge isn't loop invariant the unswitch
177 // won't be trivial.
178 if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))
179 return false;
180 }
181 llvm_unreachable("Basic blocks should never be empty!");
182 }
183
184 /// Insert code to test a set of loop invariant values, and conditionally branch
185 /// on them.
buildPartialUnswitchConditionalBranch(BasicBlock & BB,ArrayRef<Value * > Invariants,bool Direction,BasicBlock & UnswitchedSucc,BasicBlock & NormalSucc)186 static void buildPartialUnswitchConditionalBranch(BasicBlock &BB,
187 ArrayRef<Value *> Invariants,
188 bool Direction,
189 BasicBlock &UnswitchedSucc,
190 BasicBlock &NormalSucc) {
191 IRBuilder<> IRB(&BB);
192
193 Value *Cond = Direction ? IRB.CreateOr(Invariants) :
194 IRB.CreateAnd(Invariants);
195 IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
196 Direction ? &NormalSucc : &UnswitchedSucc);
197 }
198
199 /// Rewrite the PHI nodes in an unswitched loop exit basic block.
200 ///
201 /// Requires that the loop exit and unswitched basic block are the same, and
202 /// that the exiting block was a unique predecessor of that block. Rewrites the
203 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
204 /// PHI nodes from the old preheader that now contains the unswitched
205 /// terminator.
rewritePHINodesForUnswitchedExitBlock(BasicBlock & UnswitchedBB,BasicBlock & OldExitingBB,BasicBlock & OldPH)206 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
207 BasicBlock &OldExitingBB,
208 BasicBlock &OldPH) {
209 for (PHINode &PN : UnswitchedBB.phis()) {
210 // When the loop exit is directly unswitched we just need to update the
211 // incoming basic block. We loop to handle weird cases with repeated
212 // incoming blocks, but expect to typically only have one operand here.
213 for (auto i : seq<int>(0, PN.getNumOperands())) {
214 assert(PN.getIncomingBlock(i) == &OldExitingBB &&
215 "Found incoming block different from unique predecessor!");
216 PN.setIncomingBlock(i, &OldPH);
217 }
218 }
219 }
220
221 /// Rewrite the PHI nodes in the loop exit basic block and the split off
222 /// unswitched block.
223 ///
224 /// Because the exit block remains an exit from the loop, this rewrites the
225 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
226 /// nodes into the unswitched basic block to select between the value in the
227 /// old preheader and the loop exit.
rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock & ExitBB,BasicBlock & UnswitchedBB,BasicBlock & OldExitingBB,BasicBlock & OldPH,bool FullUnswitch)228 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
229 BasicBlock &UnswitchedBB,
230 BasicBlock &OldExitingBB,
231 BasicBlock &OldPH,
232 bool FullUnswitch) {
233 assert(&ExitBB != &UnswitchedBB &&
234 "Must have different loop exit and unswitched blocks!");
235 Instruction *InsertPt = &*UnswitchedBB.begin();
236 for (PHINode &PN : ExitBB.phis()) {
237 auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,
238 PN.getName() + ".split", InsertPt);
239
240 // Walk backwards over the old PHI node's inputs to minimize the cost of
241 // removing each one. We have to do this weird loop manually so that we
242 // create the same number of new incoming edges in the new PHI as we expect
243 // each case-based edge to be included in the unswitched switch in some
244 // cases.
245 // FIXME: This is really, really gross. It would be much cleaner if LLVM
246 // allowed us to create a single entry for a predecessor block without
247 // having separate entries for each "edge" even though these edges are
248 // required to produce identical results.
249 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
250 if (PN.getIncomingBlock(i) != &OldExitingBB)
251 continue;
252
253 Value *Incoming = PN.getIncomingValue(i);
254 if (FullUnswitch)
255 // No more edge from the old exiting block to the exit block.
256 PN.removeIncomingValue(i);
257
258 NewPN->addIncoming(Incoming, &OldPH);
259 }
260
261 // Now replace the old PHI with the new one and wire the old one in as an
262 // input to the new one.
263 PN.replaceAllUsesWith(NewPN);
264 NewPN->addIncoming(&PN, &ExitBB);
265 }
266 }
267
268 /// Hoist the current loop up to the innermost loop containing a remaining exit.
269 ///
270 /// Because we've removed an exit from the loop, we may have changed the set of
271 /// loops reachable and need to move the current loop up the loop nest or even
272 /// to an entirely separate nest.
hoistLoopToNewParent(Loop & L,BasicBlock & Preheader,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU,ScalarEvolution * SE)273 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
274 DominatorTree &DT, LoopInfo &LI,
275 MemorySSAUpdater *MSSAU, ScalarEvolution *SE) {
276 // If the loop is already at the top level, we can't hoist it anywhere.
277 Loop *OldParentL = L.getParentLoop();
278 if (!OldParentL)
279 return;
280
281 SmallVector<BasicBlock *, 4> Exits;
282 L.getExitBlocks(Exits);
283 Loop *NewParentL = nullptr;
284 for (auto *ExitBB : Exits)
285 if (Loop *ExitL = LI.getLoopFor(ExitBB))
286 if (!NewParentL || NewParentL->contains(ExitL))
287 NewParentL = ExitL;
288
289 if (NewParentL == OldParentL)
290 return;
291
292 // The new parent loop (if different) should always contain the old one.
293 if (NewParentL)
294 assert(NewParentL->contains(OldParentL) &&
295 "Can only hoist this loop up the nest!");
296
297 // The preheader will need to move with the body of this loop. However,
298 // because it isn't in this loop we also need to update the primary loop map.
299 assert(OldParentL == LI.getLoopFor(&Preheader) &&
300 "Parent loop of this loop should contain this loop's preheader!");
301 LI.changeLoopFor(&Preheader, NewParentL);
302
303 // Remove this loop from its old parent.
304 OldParentL->removeChildLoop(&L);
305
306 // Add the loop either to the new parent or as a top-level loop.
307 if (NewParentL)
308 NewParentL->addChildLoop(&L);
309 else
310 LI.addTopLevelLoop(&L);
311
312 // Remove this loops blocks from the old parent and every other loop up the
313 // nest until reaching the new parent. Also update all of these
314 // no-longer-containing loops to reflect the nesting change.
315 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
316 OldContainingL = OldContainingL->getParentLoop()) {
317 llvm::erase_if(OldContainingL->getBlocksVector(),
318 [&](const BasicBlock *BB) {
319 return BB == &Preheader || L.contains(BB);
320 });
321
322 OldContainingL->getBlocksSet().erase(&Preheader);
323 for (BasicBlock *BB : L.blocks())
324 OldContainingL->getBlocksSet().erase(BB);
325
326 // Because we just hoisted a loop out of this one, we have essentially
327 // created new exit paths from it. That means we need to form LCSSA PHI
328 // nodes for values used in the no-longer-nested loop.
329 formLCSSA(*OldContainingL, DT, &LI, SE);
330
331 // We shouldn't need to form dedicated exits because the exit introduced
332 // here is the (just split by unswitching) preheader. However, after trivial
333 // unswitching it is possible to get new non-dedicated exits out of parent
334 // loop so let's conservatively form dedicated exit blocks and figure out
335 // if we can optimize later.
336 formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU,
337 /*PreserveLCSSA*/ true);
338 }
339 }
340
341 // Return the top-most loop containing ExitBB and having ExitBB as exiting block
342 // or the loop containing ExitBB, if there is no parent loop containing ExitBB
343 // as exiting block.
getTopMostExitingLoop(BasicBlock * ExitBB,LoopInfo & LI)344 static Loop *getTopMostExitingLoop(BasicBlock *ExitBB, LoopInfo &LI) {
345 Loop *TopMost = LI.getLoopFor(ExitBB);
346 Loop *Current = TopMost;
347 while (Current) {
348 if (Current->isLoopExiting(ExitBB))
349 TopMost = Current;
350 Current = Current->getParentLoop();
351 }
352 return TopMost;
353 }
354
355 /// Unswitch a trivial branch if the condition is loop invariant.
356 ///
357 /// This routine should only be called when loop code leading to the branch has
358 /// been validated as trivial (no side effects). This routine checks if the
359 /// condition is invariant and one of the successors is a loop exit. This
360 /// allows us to unswitch without duplicating the loop, making it trivial.
361 ///
362 /// If this routine fails to unswitch the branch it returns false.
363 ///
364 /// If the branch can be unswitched, this routine splits the preheader and
365 /// hoists the branch above that split. Preserves loop simplified form
366 /// (splitting the exit block as necessary). It simplifies the branch within
367 /// the loop to an unconditional branch but doesn't remove it entirely. Further
368 /// cleanup can be done with some simplify-cfg like pass.
369 ///
370 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
371 /// invalidated by this.
unswitchTrivialBranch(Loop & L,BranchInst & BI,DominatorTree & DT,LoopInfo & LI,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)372 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
373 LoopInfo &LI, ScalarEvolution *SE,
374 MemorySSAUpdater *MSSAU) {
375 assert(BI.isConditional() && "Can only unswitch a conditional branch!");
376 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n");
377
378 // The loop invariant values that we want to unswitch.
379 TinyPtrVector<Value *> Invariants;
380
381 // When true, we're fully unswitching the branch rather than just unswitching
382 // some input conditions to the branch.
383 bool FullUnswitch = false;
384
385 if (L.isLoopInvariant(BI.getCondition())) {
386 Invariants.push_back(BI.getCondition());
387 FullUnswitch = true;
388 } else {
389 if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition()))
390 Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);
391 if (Invariants.empty())
392 // Couldn't find invariant inputs!
393 return false;
394 }
395
396 // Check that one of the branch's successors exits, and which one.
397 bool ExitDirection = true;
398 int LoopExitSuccIdx = 0;
399 auto *LoopExitBB = BI.getSuccessor(0);
400 if (L.contains(LoopExitBB)) {
401 ExitDirection = false;
402 LoopExitSuccIdx = 1;
403 LoopExitBB = BI.getSuccessor(1);
404 if (L.contains(LoopExitBB))
405 return false;
406 }
407 auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
408 auto *ParentBB = BI.getParent();
409 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB))
410 return false;
411
412 // When unswitching only part of the branch's condition, we need the exit
413 // block to be reached directly from the partially unswitched input. This can
414 // be done when the exit block is along the true edge and the branch condition
415 // is a graph of `or` operations, or the exit block is along the false edge
416 // and the condition is a graph of `and` operations.
417 if (!FullUnswitch) {
418 if (ExitDirection) {
419 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::Or)
420 return false;
421 } else {
422 if (cast<Instruction>(BI.getCondition())->getOpcode() != Instruction::And)
423 return false;
424 }
425 }
426
427 LLVM_DEBUG({
428 dbgs() << " unswitching trivial invariant conditions for: " << BI
429 << "\n";
430 for (Value *Invariant : Invariants) {
431 dbgs() << " " << *Invariant << " == true";
432 if (Invariant != Invariants.back())
433 dbgs() << " ||";
434 dbgs() << "\n";
435 }
436 });
437
438 // If we have scalar evolutions, we need to invalidate them including this
439 // loop, the loop containing the exit block and the topmost parent loop
440 // exiting via LoopExitBB.
441 if (SE) {
442 if (Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI))
443 SE->forgetLoop(ExitL);
444 else
445 // Forget the entire nest as this exits the entire nest.
446 SE->forgetTopmostLoop(&L);
447 }
448
449 if (MSSAU && VerifyMemorySSA)
450 MSSAU->getMemorySSA()->verifyMemorySSA();
451
452 // Split the preheader, so that we know that there is a safe place to insert
453 // the conditional branch. We will change the preheader to have a conditional
454 // branch on LoopCond.
455 BasicBlock *OldPH = L.getLoopPreheader();
456 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
457
458 // Now that we have a place to insert the conditional branch, create a place
459 // to branch to: this is the exit block out of the loop that we are
460 // unswitching. We need to split this if there are other loop predecessors.
461 // Because the loop is in simplified form, *any* other predecessor is enough.
462 BasicBlock *UnswitchedBB;
463 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
464 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
465 "A branch's parent isn't a predecessor!");
466 UnswitchedBB = LoopExitBB;
467 } else {
468 UnswitchedBB =
469 SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI, MSSAU);
470 }
471
472 if (MSSAU && VerifyMemorySSA)
473 MSSAU->getMemorySSA()->verifyMemorySSA();
474
475 // Actually move the invariant uses into the unswitched position. If possible,
476 // we do this by moving the instructions, but when doing partial unswitching
477 // we do it by building a new merge of the values in the unswitched position.
478 OldPH->getTerminator()->eraseFromParent();
479 if (FullUnswitch) {
480 // If fully unswitching, we can use the existing branch instruction.
481 // Splice it into the old PH to gate reaching the new preheader and re-point
482 // its successors.
483 OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(),
484 BI);
485 if (MSSAU) {
486 // Temporarily clone the terminator, to make MSSA update cheaper by
487 // separating "insert edge" updates from "remove edge" ones.
488 ParentBB->getInstList().push_back(BI.clone());
489 } else {
490 // Create a new unconditional branch that will continue the loop as a new
491 // terminator.
492 BranchInst::Create(ContinueBB, ParentBB);
493 }
494 BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
495 BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
496 } else {
497 // Only unswitching a subset of inputs to the condition, so we will need to
498 // build a new branch that merges the invariant inputs.
499 if (ExitDirection)
500 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
501 Instruction::Or &&
502 "Must have an `or` of `i1`s for the condition!");
503 else
504 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
505 Instruction::And &&
506 "Must have an `and` of `i1`s for the condition!");
507 buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection,
508 *UnswitchedBB, *NewPH);
509 }
510
511 // Update the dominator tree with the added edge.
512 DT.insertEdge(OldPH, UnswitchedBB);
513
514 // After the dominator tree was updated with the added edge, update MemorySSA
515 // if available.
516 if (MSSAU) {
517 SmallVector<CFGUpdate, 1> Updates;
518 Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
519 MSSAU->applyInsertUpdates(Updates, DT);
520 }
521
522 // Finish updating dominator tree and memory ssa for full unswitch.
523 if (FullUnswitch) {
524 if (MSSAU) {
525 // Remove the cloned branch instruction.
526 ParentBB->getTerminator()->eraseFromParent();
527 // Create unconditional branch now.
528 BranchInst::Create(ContinueBB, ParentBB);
529 MSSAU->removeEdge(ParentBB, LoopExitBB);
530 }
531 DT.deleteEdge(ParentBB, LoopExitBB);
532 }
533
534 if (MSSAU && VerifyMemorySSA)
535 MSSAU->getMemorySSA()->verifyMemorySSA();
536
537 // Rewrite the relevant PHI nodes.
538 if (UnswitchedBB == LoopExitBB)
539 rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
540 else
541 rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
542 *ParentBB, *OldPH, FullUnswitch);
543
544 // The constant we can replace all of our invariants with inside the loop
545 // body. If any of the invariants have a value other than this the loop won't
546 // be entered.
547 ConstantInt *Replacement = ExitDirection
548 ? ConstantInt::getFalse(BI.getContext())
549 : ConstantInt::getTrue(BI.getContext());
550
551 // Since this is an i1 condition we can also trivially replace uses of it
552 // within the loop with a constant.
553 for (Value *Invariant : Invariants)
554 replaceLoopInvariantUses(L, Invariant, *Replacement);
555
556 // If this was full unswitching, we may have changed the nesting relationship
557 // for this loop so hoist it to its correct parent if needed.
558 if (FullUnswitch)
559 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
560
561 if (MSSAU && VerifyMemorySSA)
562 MSSAU->getMemorySSA()->verifyMemorySSA();
563
564 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n");
565 ++NumTrivial;
566 ++NumBranches;
567 return true;
568 }
569
570 /// Unswitch a trivial switch if the condition is loop invariant.
571 ///
572 /// This routine should only be called when loop code leading to the switch has
573 /// been validated as trivial (no side effects). This routine checks if the
574 /// condition is invariant and that at least one of the successors is a loop
575 /// exit. This allows us to unswitch without duplicating the loop, making it
576 /// trivial.
577 ///
578 /// If this routine fails to unswitch the switch it returns false.
579 ///
580 /// If the switch can be unswitched, this routine splits the preheader and
581 /// copies the switch above that split. If the default case is one of the
582 /// exiting cases, it copies the non-exiting cases and points them at the new
583 /// preheader. If the default case is not exiting, it copies the exiting cases
584 /// and points the default at the preheader. It preserves loop simplified form
585 /// (splitting the exit blocks as necessary). It simplifies the switch within
586 /// the loop by removing now-dead cases. If the default case is one of those
587 /// unswitched, it replaces its destination with a new basic block containing
588 /// only unreachable. Such basic blocks, while technically loop exits, are not
589 /// considered for unswitching so this is a stable transform and the same
590 /// switch will not be revisited. If after unswitching there is only a single
591 /// in-loop successor, the switch is further simplified to an unconditional
592 /// branch. Still more cleanup can be done with some simplify-cfg like pass.
593 ///
594 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
595 /// invalidated by this.
unswitchTrivialSwitch(Loop & L,SwitchInst & SI,DominatorTree & DT,LoopInfo & LI,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)596 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
597 LoopInfo &LI, ScalarEvolution *SE,
598 MemorySSAUpdater *MSSAU) {
599 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n");
600 Value *LoopCond = SI.getCondition();
601
602 // If this isn't switching on an invariant condition, we can't unswitch it.
603 if (!L.isLoopInvariant(LoopCond))
604 return false;
605
606 auto *ParentBB = SI.getParent();
607
608 // The same check must be used both for the default and the exit cases. We
609 // should never leave edges from the switch instruction to a basic block that
610 // we are unswitching, hence the condition used to determine the default case
611 // needs to also be used to populate ExitCaseIndices, which is then used to
612 // remove cases from the switch.
613 auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) {
614 // BBToCheck is not an exit block if it is inside loop L.
615 if (L.contains(&BBToCheck))
616 return false;
617 // BBToCheck is not trivial to unswitch if its phis aren't loop invariant.
618 if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck))
619 return false;
620 // We do not unswitch a block that only has an unreachable statement, as
621 // it's possible this is a previously unswitched block. Only unswitch if
622 // either the terminator is not unreachable, or, if it is, it's not the only
623 // instruction in the block.
624 auto *TI = BBToCheck.getTerminator();
625 bool isUnreachable = isa<UnreachableInst>(TI);
626 return !isUnreachable ||
627 (isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI));
628 };
629
630 SmallVector<int, 4> ExitCaseIndices;
631 for (auto Case : SI.cases())
632 if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor()))
633 ExitCaseIndices.push_back(Case.getCaseIndex());
634 BasicBlock *DefaultExitBB = nullptr;
635 SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight =
636 SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0);
637 if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) {
638 DefaultExitBB = SI.getDefaultDest();
639 } else if (ExitCaseIndices.empty())
640 return false;
641
642 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n");
643
644 if (MSSAU && VerifyMemorySSA)
645 MSSAU->getMemorySSA()->verifyMemorySSA();
646
647 // We may need to invalidate SCEVs for the outermost loop reached by any of
648 // the exits.
649 Loop *OuterL = &L;
650
651 if (DefaultExitBB) {
652 // Clear out the default destination temporarily to allow accurate
653 // predecessor lists to be examined below.
654 SI.setDefaultDest(nullptr);
655 // Check the loop containing this exit.
656 Loop *ExitL = LI.getLoopFor(DefaultExitBB);
657 if (!ExitL || ExitL->contains(OuterL))
658 OuterL = ExitL;
659 }
660
661 // Store the exit cases into a separate data structure and remove them from
662 // the switch.
663 SmallVector<std::tuple<ConstantInt *, BasicBlock *,
664 SwitchInstProfUpdateWrapper::CaseWeightOpt>,
665 4> ExitCases;
666 ExitCases.reserve(ExitCaseIndices.size());
667 SwitchInstProfUpdateWrapper SIW(SI);
668 // We walk the case indices backwards so that we remove the last case first
669 // and don't disrupt the earlier indices.
670 for (unsigned Index : reverse(ExitCaseIndices)) {
671 auto CaseI = SI.case_begin() + Index;
672 // Compute the outer loop from this exit.
673 Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor());
674 if (!ExitL || ExitL->contains(OuterL))
675 OuterL = ExitL;
676 // Save the value of this case.
677 auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex());
678 ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W);
679 // Delete the unswitched cases.
680 SIW.removeCase(CaseI);
681 }
682
683 if (SE) {
684 if (OuterL)
685 SE->forgetLoop(OuterL);
686 else
687 SE->forgetTopmostLoop(&L);
688 }
689
690 // Check if after this all of the remaining cases point at the same
691 // successor.
692 BasicBlock *CommonSuccBB = nullptr;
693 if (SI.getNumCases() > 0 &&
694 std::all_of(std::next(SI.case_begin()), SI.case_end(),
695 [&SI](const SwitchInst::CaseHandle &Case) {
696 return Case.getCaseSuccessor() ==
697 SI.case_begin()->getCaseSuccessor();
698 }))
699 CommonSuccBB = SI.case_begin()->getCaseSuccessor();
700 if (!DefaultExitBB) {
701 // If we're not unswitching the default, we need it to match any cases to
702 // have a common successor or if we have no cases it is the common
703 // successor.
704 if (SI.getNumCases() == 0)
705 CommonSuccBB = SI.getDefaultDest();
706 else if (SI.getDefaultDest() != CommonSuccBB)
707 CommonSuccBB = nullptr;
708 }
709
710 // Split the preheader, so that we know that there is a safe place to insert
711 // the switch.
712 BasicBlock *OldPH = L.getLoopPreheader();
713 BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
714 OldPH->getTerminator()->eraseFromParent();
715
716 // Now add the unswitched switch.
717 auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
718 SwitchInstProfUpdateWrapper NewSIW(*NewSI);
719
720 // Rewrite the IR for the unswitched basic blocks. This requires two steps.
721 // First, we split any exit blocks with remaining in-loop predecessors. Then
722 // we update the PHIs in one of two ways depending on if there was a split.
723 // We walk in reverse so that we split in the same order as the cases
724 // appeared. This is purely for convenience of reading the resulting IR, but
725 // it doesn't cost anything really.
726 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
727 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
728 // Handle the default exit if necessary.
729 // FIXME: It'd be great if we could merge this with the loop below but LLVM's
730 // ranges aren't quite powerful enough yet.
731 if (DefaultExitBB) {
732 if (pred_empty(DefaultExitBB)) {
733 UnswitchedExitBBs.insert(DefaultExitBB);
734 rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
735 } else {
736 auto *SplitBB =
737 SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI, MSSAU);
738 rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,
739 *ParentBB, *OldPH,
740 /*FullUnswitch*/ true);
741 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
742 }
743 }
744 // Note that we must use a reference in the for loop so that we update the
745 // container.
746 for (auto &ExitCase : reverse(ExitCases)) {
747 // Grab a reference to the exit block in the pair so that we can update it.
748 BasicBlock *ExitBB = std::get<1>(ExitCase);
749
750 // If this case is the last edge into the exit block, we can simply reuse it
751 // as it will no longer be a loop exit. No mapping necessary.
752 if (pred_empty(ExitBB)) {
753 // Only rewrite once.
754 if (UnswitchedExitBBs.insert(ExitBB).second)
755 rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
756 continue;
757 }
758
759 // Otherwise we need to split the exit block so that we retain an exit
760 // block from the loop and a target for the unswitched condition.
761 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
762 if (!SplitExitBB) {
763 // If this is the first time we see this, do the split and remember it.
764 SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
765 rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,
766 *ParentBB, *OldPH,
767 /*FullUnswitch*/ true);
768 }
769 // Update the case pair to point to the split block.
770 std::get<1>(ExitCase) = SplitExitBB;
771 }
772
773 // Now add the unswitched cases. We do this in reverse order as we built them
774 // in reverse order.
775 for (auto &ExitCase : reverse(ExitCases)) {
776 ConstantInt *CaseVal = std::get<0>(ExitCase);
777 BasicBlock *UnswitchedBB = std::get<1>(ExitCase);
778
779 NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase));
780 }
781
782 // If the default was unswitched, re-point it and add explicit cases for
783 // entering the loop.
784 if (DefaultExitBB) {
785 NewSIW->setDefaultDest(DefaultExitBB);
786 NewSIW.setSuccessorWeight(0, DefaultCaseWeight);
787
788 // We removed all the exit cases, so we just copy the cases to the
789 // unswitched switch.
790 for (const auto &Case : SI.cases())
791 NewSIW.addCase(Case.getCaseValue(), NewPH,
792 SIW.getSuccessorWeight(Case.getSuccessorIndex()));
793 } else if (DefaultCaseWeight) {
794 // We have to set branch weight of the default case.
795 uint64_t SW = *DefaultCaseWeight;
796 for (const auto &Case : SI.cases()) {
797 auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex());
798 assert(W &&
799 "case weight must be defined as default case weight is defined");
800 SW += *W;
801 }
802 NewSIW.setSuccessorWeight(0, SW);
803 }
804
805 // If we ended up with a common successor for every path through the switch
806 // after unswitching, rewrite it to an unconditional branch to make it easy
807 // to recognize. Otherwise we potentially have to recognize the default case
808 // pointing at unreachable and other complexity.
809 if (CommonSuccBB) {
810 BasicBlock *BB = SI.getParent();
811 // We may have had multiple edges to this common successor block, so remove
812 // them as predecessors. We skip the first one, either the default or the
813 // actual first case.
814 bool SkippedFirst = DefaultExitBB == nullptr;
815 for (auto Case : SI.cases()) {
816 assert(Case.getCaseSuccessor() == CommonSuccBB &&
817 "Non-common successor!");
818 (void)Case;
819 if (!SkippedFirst) {
820 SkippedFirst = true;
821 continue;
822 }
823 CommonSuccBB->removePredecessor(BB,
824 /*KeepOneInputPHIs*/ true);
825 }
826 // Now nuke the switch and replace it with a direct branch.
827 SIW.eraseFromParent();
828 BranchInst::Create(CommonSuccBB, BB);
829 } else if (DefaultExitBB) {
830 assert(SI.getNumCases() > 0 &&
831 "If we had no cases we'd have a common successor!");
832 // Move the last case to the default successor. This is valid as if the
833 // default got unswitched it cannot be reached. This has the advantage of
834 // being simple and keeping the number of edges from this switch to
835 // successors the same, and avoiding any PHI update complexity.
836 auto LastCaseI = std::prev(SI.case_end());
837
838 SI.setDefaultDest(LastCaseI->getCaseSuccessor());
839 SIW.setSuccessorWeight(
840 0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex()));
841 SIW.removeCase(LastCaseI);
842 }
843
844 // Walk the unswitched exit blocks and the unswitched split blocks and update
845 // the dominator tree based on the CFG edits. While we are walking unordered
846 // containers here, the API for applyUpdates takes an unordered list of
847 // updates and requires them to not contain duplicates.
848 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
849 for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
850 DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});
851 DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});
852 }
853 for (auto SplitUnswitchedPair : SplitExitBBMap) {
854 DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first});
855 DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second});
856 }
857 DT.applyUpdates(DTUpdates);
858
859 if (MSSAU) {
860 MSSAU->applyUpdates(DTUpdates, DT);
861 if (VerifyMemorySSA)
862 MSSAU->getMemorySSA()->verifyMemorySSA();
863 }
864
865 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
866
867 // We may have changed the nesting relationship for this loop so hoist it to
868 // its correct parent if needed.
869 hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
870
871 if (MSSAU && VerifyMemorySSA)
872 MSSAU->getMemorySSA()->verifyMemorySSA();
873
874 ++NumTrivial;
875 ++NumSwitches;
876 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n");
877 return true;
878 }
879
880 /// This routine scans the loop to find a branch or switch which occurs before
881 /// any side effects occur. These can potentially be unswitched without
882 /// duplicating the loop. If a branch or switch is successfully unswitched the
883 /// scanning continues to see if subsequent branches or switches have become
884 /// trivial. Once all trivial candidates have been unswitched, this routine
885 /// returns.
886 ///
887 /// The return value indicates whether anything was unswitched (and therefore
888 /// changed).
889 ///
890 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
891 /// invalidated by this.
unswitchAllTrivialConditions(Loop & L,DominatorTree & DT,LoopInfo & LI,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)892 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
893 LoopInfo &LI, ScalarEvolution *SE,
894 MemorySSAUpdater *MSSAU) {
895 bool Changed = false;
896
897 // If loop header has only one reachable successor we should keep looking for
898 // trivial condition candidates in the successor as well. An alternative is
899 // to constant fold conditions and merge successors into loop header (then we
900 // only need to check header's terminator). The reason for not doing this in
901 // LoopUnswitch pass is that it could potentially break LoopPassManager's
902 // invariants. Folding dead branches could either eliminate the current loop
903 // or make other loops unreachable. LCSSA form might also not be preserved
904 // after deleting branches. The following code keeps traversing loop header's
905 // successors until it finds the trivial condition candidate (condition that
906 // is not a constant). Since unswitching generates branches with constant
907 // conditions, this scenario could be very common in practice.
908 BasicBlock *CurrentBB = L.getHeader();
909 SmallPtrSet<BasicBlock *, 8> Visited;
910 Visited.insert(CurrentBB);
911 do {
912 // Check if there are any side-effecting instructions (e.g. stores, calls,
913 // volatile loads) in the part of the loop that the code *would* execute
914 // without unswitching.
915 if (MSSAU) // Possible early exit with MSSA
916 if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB))
917 if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end()))
918 return Changed;
919 if (llvm::any_of(*CurrentBB,
920 [](Instruction &I) { return I.mayHaveSideEffects(); }))
921 return Changed;
922
923 Instruction *CurrentTerm = CurrentBB->getTerminator();
924
925 if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
926 // Don't bother trying to unswitch past a switch with a constant
927 // condition. This should be removed prior to running this pass by
928 // simplify-cfg.
929 if (isa<Constant>(SI->getCondition()))
930 return Changed;
931
932 if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU))
933 // Couldn't unswitch this one so we're done.
934 return Changed;
935
936 // Mark that we managed to unswitch something.
937 Changed = true;
938
939 // If unswitching turned the terminator into an unconditional branch then
940 // we can continue. The unswitching logic specifically works to fold any
941 // cases it can into an unconditional branch to make it easier to
942 // recognize here.
943 auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
944 if (!BI || BI->isConditional())
945 return Changed;
946
947 CurrentBB = BI->getSuccessor(0);
948 continue;
949 }
950
951 auto *BI = dyn_cast<BranchInst>(CurrentTerm);
952 if (!BI)
953 // We do not understand other terminator instructions.
954 return Changed;
955
956 // Don't bother trying to unswitch past an unconditional branch or a branch
957 // with a constant value. These should be removed by simplify-cfg prior to
958 // running this pass.
959 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
960 return Changed;
961
962 // Found a trivial condition candidate: non-foldable conditional branch. If
963 // we fail to unswitch this, we can't do anything else that is trivial.
964 if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU))
965 return Changed;
966
967 // Mark that we managed to unswitch something.
968 Changed = true;
969
970 // If we only unswitched some of the conditions feeding the branch, we won't
971 // have collapsed it to a single successor.
972 BI = cast<BranchInst>(CurrentBB->getTerminator());
973 if (BI->isConditional())
974 return Changed;
975
976 // Follow the newly unconditional branch into its successor.
977 CurrentBB = BI->getSuccessor(0);
978
979 // When continuing, if we exit the loop or reach a previous visited block,
980 // then we can not reach any trivial condition candidates (unfoldable
981 // branch instructions or switch instructions) and no unswitch can happen.
982 } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
983
984 return Changed;
985 }
986
987 /// Build the cloned blocks for an unswitched copy of the given loop.
988 ///
989 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
990 /// after the split block (`SplitBB`) that will be used to select between the
991 /// cloned and original loop.
992 ///
993 /// This routine handles cloning all of the necessary loop blocks and exit
994 /// blocks including rewriting their instructions and the relevant PHI nodes.
995 /// Any loop blocks or exit blocks which are dominated by a different successor
996 /// than the one for this clone of the loop blocks can be trivially skipped. We
997 /// use the `DominatingSucc` map to determine whether a block satisfies that
998 /// property with a simple map lookup.
999 ///
1000 /// It also correctly creates the unconditional branch in the cloned
1001 /// unswitched parent block to only point at the unswitched successor.
1002 ///
1003 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit
1004 /// block splitting is correctly reflected in `LoopInfo`, essentially all of
1005 /// the cloned blocks (and their loops) are left without full `LoopInfo`
1006 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
1007 /// blocks to them but doesn't create the cloned `DominatorTree` structure and
1008 /// instead the caller must recompute an accurate DT. It *does* correctly
1009 /// update the `AssumptionCache` provided in `AC`.
buildClonedLoopBlocks(Loop & L,BasicBlock * LoopPH,BasicBlock * SplitBB,ArrayRef<BasicBlock * > ExitBlocks,BasicBlock * ParentBB,BasicBlock * UnswitchedSuccBB,BasicBlock * ContinueSuccBB,const SmallDenseMap<BasicBlock *,BasicBlock *,16> & DominatingSucc,ValueToValueMapTy & VMap,SmallVectorImpl<DominatorTree::UpdateType> & DTUpdates,AssumptionCache & AC,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU)1010 static BasicBlock *buildClonedLoopBlocks(
1011 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
1012 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
1013 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
1014 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
1015 ValueToValueMapTy &VMap,
1016 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
1017 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
1018 SmallVector<BasicBlock *, 4> NewBlocks;
1019 NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());
1020
1021 // We will need to clone a bunch of blocks, wrap up the clone operation in
1022 // a helper.
1023 auto CloneBlock = [&](BasicBlock *OldBB) {
1024 // Clone the basic block and insert it before the new preheader.
1025 BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());
1026 NewBB->moveBefore(LoopPH);
1027
1028 // Record this block and the mapping.
1029 NewBlocks.push_back(NewBB);
1030 VMap[OldBB] = NewBB;
1031
1032 return NewBB;
1033 };
1034
1035 // We skip cloning blocks when they have a dominating succ that is not the
1036 // succ we are cloning for.
1037 auto SkipBlock = [&](BasicBlock *BB) {
1038 auto It = DominatingSucc.find(BB);
1039 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
1040 };
1041
1042 // First, clone the preheader.
1043 auto *ClonedPH = CloneBlock(LoopPH);
1044
1045 // Then clone all the loop blocks, skipping the ones that aren't necessary.
1046 for (auto *LoopBB : L.blocks())
1047 if (!SkipBlock(LoopBB))
1048 CloneBlock(LoopBB);
1049
1050 // Split all the loop exit edges so that when we clone the exit blocks, if
1051 // any of the exit blocks are *also* a preheader for some other loop, we
1052 // don't create multiple predecessors entering the loop header.
1053 for (auto *ExitBB : ExitBlocks) {
1054 if (SkipBlock(ExitBB))
1055 continue;
1056
1057 // When we are going to clone an exit, we don't need to clone all the
1058 // instructions in the exit block and we want to ensure we have an easy
1059 // place to merge the CFG, so split the exit first. This is always safe to
1060 // do because there cannot be any non-loop predecessors of a loop exit in
1061 // loop simplified form.
1062 auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
1063
1064 // Rearrange the names to make it easier to write test cases by having the
1065 // exit block carry the suffix rather than the merge block carrying the
1066 // suffix.
1067 MergeBB->takeName(ExitBB);
1068 ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1069
1070 // Now clone the original exit block.
1071 auto *ClonedExitBB = CloneBlock(ExitBB);
1072 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1073 "Exit block should have been split to have one successor!");
1074 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1075 "Cloned exit block has the wrong successor!");
1076
1077 // Remap any cloned instructions and create a merge phi node for them.
1078 for (auto ZippedInsts : llvm::zip_first(
1079 llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),
1080 llvm::make_range(ClonedExitBB->begin(),
1081 std::prev(ClonedExitBB->end())))) {
1082 Instruction &I = std::get<0>(ZippedInsts);
1083 Instruction &ClonedI = std::get<1>(ZippedInsts);
1084
1085 // The only instructions in the exit block should be PHI nodes and
1086 // potentially a landing pad.
1087 assert(
1088 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1089 "Bad instruction in exit block!");
1090 // We should have a value map between the instruction and its clone.
1091 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1092
1093 auto *MergePN =
1094 PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi",
1095 &*MergeBB->getFirstInsertionPt());
1096 I.replaceAllUsesWith(MergePN);
1097 MergePN->addIncoming(&I, ExitBB);
1098 MergePN->addIncoming(&ClonedI, ClonedExitBB);
1099 }
1100 }
1101
1102 // Rewrite the instructions in the cloned blocks to refer to the instructions
1103 // in the cloned blocks. We have to do this as a second pass so that we have
1104 // everything available. Also, we have inserted new instructions which may
1105 // include assume intrinsics, so we update the assumption cache while
1106 // processing this.
1107 for (auto *ClonedBB : NewBlocks)
1108 for (Instruction &I : *ClonedBB) {
1109 RemapInstruction(&I, VMap,
1110 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1111 if (auto *II = dyn_cast<IntrinsicInst>(&I))
1112 if (II->getIntrinsicID() == Intrinsic::assume)
1113 AC.registerAssumption(II);
1114 }
1115
1116 // Update any PHI nodes in the cloned successors of the skipped blocks to not
1117 // have spurious incoming values.
1118 for (auto *LoopBB : L.blocks())
1119 if (SkipBlock(LoopBB))
1120 for (auto *SuccBB : successors(LoopBB))
1121 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))
1122 for (PHINode &PN : ClonedSuccBB->phis())
1123 PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);
1124
1125 // Remove the cloned parent as a predecessor of any successor we ended up
1126 // cloning other than the unswitched one.
1127 auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));
1128 for (auto *SuccBB : successors(ParentBB)) {
1129 if (SuccBB == UnswitchedSuccBB)
1130 continue;
1131
1132 auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));
1133 if (!ClonedSuccBB)
1134 continue;
1135
1136 ClonedSuccBB->removePredecessor(ClonedParentBB,
1137 /*KeepOneInputPHIs*/ true);
1138 }
1139
1140 // Replace the cloned branch with an unconditional branch to the cloned
1141 // unswitched successor.
1142 auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));
1143 Instruction *ClonedTerminator = ClonedParentBB->getTerminator();
1144 // Trivial Simplification. If Terminator is a conditional branch and
1145 // condition becomes dead - erase it.
1146 Value *ClonedConditionToErase = nullptr;
1147 if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator))
1148 ClonedConditionToErase = BI->getCondition();
1149 else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator))
1150 ClonedConditionToErase = SI->getCondition();
1151
1152 ClonedTerminator->eraseFromParent();
1153 BranchInst::Create(ClonedSuccBB, ClonedParentBB);
1154
1155 if (ClonedConditionToErase)
1156 RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr,
1157 MSSAU);
1158
1159 // If there are duplicate entries in the PHI nodes because of multiple edges
1160 // to the unswitched successor, we need to nuke all but one as we replaced it
1161 // with a direct branch.
1162 for (PHINode &PN : ClonedSuccBB->phis()) {
1163 bool Found = false;
1164 // Loop over the incoming operands backwards so we can easily delete as we
1165 // go without invalidating the index.
1166 for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1167 if (PN.getIncomingBlock(i) != ClonedParentBB)
1168 continue;
1169 if (!Found) {
1170 Found = true;
1171 continue;
1172 }
1173 PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);
1174 }
1175 }
1176
1177 // Record the domtree updates for the new blocks.
1178 SmallPtrSet<BasicBlock *, 4> SuccSet;
1179 for (auto *ClonedBB : NewBlocks) {
1180 for (auto *SuccBB : successors(ClonedBB))
1181 if (SuccSet.insert(SuccBB).second)
1182 DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});
1183 SuccSet.clear();
1184 }
1185
1186 return ClonedPH;
1187 }
1188
1189 /// Recursively clone the specified loop and all of its children.
1190 ///
1191 /// The target parent loop for the clone should be provided, or can be null if
1192 /// the clone is a top-level loop. While cloning, all the blocks are mapped
1193 /// with the provided value map. The entire original loop must be present in
1194 /// the value map. The cloned loop is returned.
cloneLoopNest(Loop & OrigRootL,Loop * RootParentL,const ValueToValueMapTy & VMap,LoopInfo & LI)1195 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1196 const ValueToValueMapTy &VMap, LoopInfo &LI) {
1197 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1198 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1199 ClonedL.reserveBlocks(OrigL.getNumBlocks());
1200 for (auto *BB : OrigL.blocks()) {
1201 auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));
1202 ClonedL.addBlockEntry(ClonedBB);
1203 if (LI.getLoopFor(BB) == &OrigL)
1204 LI.changeLoopFor(ClonedBB, &ClonedL);
1205 }
1206 };
1207
1208 // We specially handle the first loop because it may get cloned into
1209 // a different parent and because we most commonly are cloning leaf loops.
1210 Loop *ClonedRootL = LI.AllocateLoop();
1211 if (RootParentL)
1212 RootParentL->addChildLoop(ClonedRootL);
1213 else
1214 LI.addTopLevelLoop(ClonedRootL);
1215 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1216
1217 if (OrigRootL.isInnermost())
1218 return ClonedRootL;
1219
1220 // If we have a nest, we can quickly clone the entire loop nest using an
1221 // iterative approach because it is a tree. We keep the cloned parent in the
1222 // data structure to avoid repeatedly querying through a map to find it.
1223 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1224 // Build up the loops to clone in reverse order as we'll clone them from the
1225 // back.
1226 for (Loop *ChildL : llvm::reverse(OrigRootL))
1227 LoopsToClone.push_back({ClonedRootL, ChildL});
1228 do {
1229 Loop *ClonedParentL, *L;
1230 std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();
1231 Loop *ClonedL = LI.AllocateLoop();
1232 ClonedParentL->addChildLoop(ClonedL);
1233 AddClonedBlocksToLoop(*L, *ClonedL);
1234 for (Loop *ChildL : llvm::reverse(*L))
1235 LoopsToClone.push_back({ClonedL, ChildL});
1236 } while (!LoopsToClone.empty());
1237
1238 return ClonedRootL;
1239 }
1240
1241 /// Build the cloned loops of an original loop from unswitching.
1242 ///
1243 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1244 /// operation. We need to re-verify that there even is a loop (as the backedge
1245 /// may not have been cloned), and even if there are remaining backedges the
1246 /// backedge set may be different. However, we know that each child loop is
1247 /// undisturbed, we only need to find where to place each child loop within
1248 /// either any parent loop or within a cloned version of the original loop.
1249 ///
1250 /// Because child loops may end up cloned outside of any cloned version of the
1251 /// original loop, multiple cloned sibling loops may be created. All of them
1252 /// are returned so that the newly introduced loop nest roots can be
1253 /// identified.
buildClonedLoops(Loop & OrigL,ArrayRef<BasicBlock * > ExitBlocks,const ValueToValueMapTy & VMap,LoopInfo & LI,SmallVectorImpl<Loop * > & NonChildClonedLoops)1254 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1255 const ValueToValueMapTy &VMap, LoopInfo &LI,
1256 SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1257 Loop *ClonedL = nullptr;
1258
1259 auto *OrigPH = OrigL.getLoopPreheader();
1260 auto *OrigHeader = OrigL.getHeader();
1261
1262 auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));
1263 auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));
1264
1265 // We need to know the loops of the cloned exit blocks to even compute the
1266 // accurate parent loop. If we only clone exits to some parent of the
1267 // original parent, we want to clone into that outer loop. We also keep track
1268 // of the loops that our cloned exit blocks participate in.
1269 Loop *ParentL = nullptr;
1270 SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1271 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1272 ClonedExitsInLoops.reserve(ExitBlocks.size());
1273 for (auto *ExitBB : ExitBlocks)
1274 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))
1275 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1276 ExitLoopMap[ClonedExitBB] = ExitL;
1277 ClonedExitsInLoops.push_back(ClonedExitBB);
1278 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1279 ParentL = ExitL;
1280 }
1281 assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1282 ParentL->contains(OrigL.getParentLoop())) &&
1283 "The computed parent loop should always contain (or be) the parent of "
1284 "the original loop.");
1285
1286 // We build the set of blocks dominated by the cloned header from the set of
1287 // cloned blocks out of the original loop. While not all of these will
1288 // necessarily be in the cloned loop, it is enough to establish that they
1289 // aren't in unreachable cycles, etc.
1290 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1291 for (auto *BB : OrigL.blocks())
1292 if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))
1293 ClonedLoopBlocks.insert(ClonedBB);
1294
1295 // Rebuild the set of blocks that will end up in the cloned loop. We may have
1296 // skipped cloning some region of this loop which can in turn skip some of
1297 // the backedges so we have to rebuild the blocks in the loop based on the
1298 // backedges that remain after cloning.
1299 SmallVector<BasicBlock *, 16> Worklist;
1300 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1301 for (auto *Pred : predecessors(ClonedHeader)) {
1302 // The only possible non-loop header predecessor is the preheader because
1303 // we know we cloned the loop in simplified form.
1304 if (Pred == ClonedPH)
1305 continue;
1306
1307 // Because the loop was in simplified form, the only non-loop predecessor
1308 // should be the preheader.
1309 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1310 "header other than the preheader "
1311 "that is not part of the loop!");
1312
1313 // Insert this block into the loop set and on the first visit (and if it
1314 // isn't the header we're currently walking) put it into the worklist to
1315 // recurse through.
1316 if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)
1317 Worklist.push_back(Pred);
1318 }
1319
1320 // If we had any backedges then there *is* a cloned loop. Put the header into
1321 // the loop set and then walk the worklist backwards to find all the blocks
1322 // that remain within the loop after cloning.
1323 if (!BlocksInClonedLoop.empty()) {
1324 BlocksInClonedLoop.insert(ClonedHeader);
1325
1326 while (!Worklist.empty()) {
1327 BasicBlock *BB = Worklist.pop_back_val();
1328 assert(BlocksInClonedLoop.count(BB) &&
1329 "Didn't put block into the loop set!");
1330
1331 // Insert any predecessors that are in the possible set into the cloned
1332 // set, and if the insert is successful, add them to the worklist. Note
1333 // that we filter on the blocks that are definitely reachable via the
1334 // backedge to the loop header so we may prune out dead code within the
1335 // cloned loop.
1336 for (auto *Pred : predecessors(BB))
1337 if (ClonedLoopBlocks.count(Pred) &&
1338 BlocksInClonedLoop.insert(Pred).second)
1339 Worklist.push_back(Pred);
1340 }
1341
1342 ClonedL = LI.AllocateLoop();
1343 if (ParentL) {
1344 ParentL->addBasicBlockToLoop(ClonedPH, LI);
1345 ParentL->addChildLoop(ClonedL);
1346 } else {
1347 LI.addTopLevelLoop(ClonedL);
1348 }
1349 NonChildClonedLoops.push_back(ClonedL);
1350
1351 ClonedL->reserveBlocks(BlocksInClonedLoop.size());
1352 // We don't want to just add the cloned loop blocks based on how we
1353 // discovered them. The original order of blocks was carefully built in
1354 // a way that doesn't rely on predecessor ordering. Rather than re-invent
1355 // that logic, we just re-walk the original blocks (and those of the child
1356 // loops) and filter them as we add them into the cloned loop.
1357 for (auto *BB : OrigL.blocks()) {
1358 auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));
1359 if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))
1360 continue;
1361
1362 // Directly add the blocks that are only in this loop.
1363 if (LI.getLoopFor(BB) == &OrigL) {
1364 ClonedL->addBasicBlockToLoop(ClonedBB, LI);
1365 continue;
1366 }
1367
1368 // We want to manually add it to this loop and parents.
1369 // Registering it with LoopInfo will happen when we clone the top
1370 // loop for this block.
1371 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1372 PL->addBlockEntry(ClonedBB);
1373 }
1374
1375 // Now add each child loop whose header remains within the cloned loop. All
1376 // of the blocks within the loop must satisfy the same constraints as the
1377 // header so once we pass the header checks we can just clone the entire
1378 // child loop nest.
1379 for (Loop *ChildL : OrigL) {
1380 auto *ClonedChildHeader =
1381 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1382 if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))
1383 continue;
1384
1385 #ifndef NDEBUG
1386 // We should never have a cloned child loop header but fail to have
1387 // all of the blocks for that child loop.
1388 for (auto *ChildLoopBB : ChildL->blocks())
1389 assert(BlocksInClonedLoop.count(
1390 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1391 "Child cloned loop has a header within the cloned outer "
1392 "loop but not all of its blocks!");
1393 #endif
1394
1395 cloneLoopNest(*ChildL, ClonedL, VMap, LI);
1396 }
1397 }
1398
1399 // Now that we've handled all the components of the original loop that were
1400 // cloned into a new loop, we still need to handle anything from the original
1401 // loop that wasn't in a cloned loop.
1402
1403 // Figure out what blocks are left to place within any loop nest containing
1404 // the unswitched loop. If we never formed a loop, the cloned PH is one of
1405 // them.
1406 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1407 if (BlocksInClonedLoop.empty())
1408 UnloopedBlockSet.insert(ClonedPH);
1409 for (auto *ClonedBB : ClonedLoopBlocks)
1410 if (!BlocksInClonedLoop.count(ClonedBB))
1411 UnloopedBlockSet.insert(ClonedBB);
1412
1413 // Copy the cloned exits and sort them in ascending loop depth, we'll work
1414 // backwards across these to process them inside out. The order shouldn't
1415 // matter as we're just trying to build up the map from inside-out; we use
1416 // the map in a more stably ordered way below.
1417 auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1418 llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1419 return ExitLoopMap.lookup(LHS)->getLoopDepth() <
1420 ExitLoopMap.lookup(RHS)->getLoopDepth();
1421 });
1422
1423 // Populate the existing ExitLoopMap with everything reachable from each
1424 // exit, starting from the inner most exit.
1425 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1426 assert(Worklist.empty() && "Didn't clear worklist!");
1427
1428 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1429 Loop *ExitL = ExitLoopMap.lookup(ExitBB);
1430
1431 // Walk the CFG back until we hit the cloned PH adding everything reachable
1432 // and in the unlooped set to this exit block's loop.
1433 Worklist.push_back(ExitBB);
1434 do {
1435 BasicBlock *BB = Worklist.pop_back_val();
1436 // We can stop recursing at the cloned preheader (if we get there).
1437 if (BB == ClonedPH)
1438 continue;
1439
1440 for (BasicBlock *PredBB : predecessors(BB)) {
1441 // If this pred has already been moved to our set or is part of some
1442 // (inner) loop, no update needed.
1443 if (!UnloopedBlockSet.erase(PredBB)) {
1444 assert(
1445 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1446 "Predecessor not mapped to a loop!");
1447 continue;
1448 }
1449
1450 // We just insert into the loop set here. We'll add these blocks to the
1451 // exit loop after we build up the set in an order that doesn't rely on
1452 // predecessor order (which in turn relies on use list order).
1453 bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;
1454 (void)Inserted;
1455 assert(Inserted && "Should only visit an unlooped block once!");
1456
1457 // And recurse through to its predecessors.
1458 Worklist.push_back(PredBB);
1459 }
1460 } while (!Worklist.empty());
1461 }
1462
1463 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned
1464 // blocks to their outer loops, walk the cloned blocks and the cloned exits
1465 // in their original order adding them to the correct loop.
1466
1467 // We need a stable insertion order. We use the order of the original loop
1468 // order and map into the correct parent loop.
1469 for (auto *BB : llvm::concat<BasicBlock *const>(
1470 makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))
1471 if (Loop *OuterL = ExitLoopMap.lookup(BB))
1472 OuterL->addBasicBlockToLoop(BB, LI);
1473
1474 #ifndef NDEBUG
1475 for (auto &BBAndL : ExitLoopMap) {
1476 auto *BB = BBAndL.first;
1477 auto *OuterL = BBAndL.second;
1478 assert(LI.getLoopFor(BB) == OuterL &&
1479 "Failed to put all blocks into outer loops!");
1480 }
1481 #endif
1482
1483 // Now that all the blocks are placed into the correct containing loop in the
1484 // absence of child loops, find all the potentially cloned child loops and
1485 // clone them into whatever outer loop we placed their header into.
1486 for (Loop *ChildL : OrigL) {
1487 auto *ClonedChildHeader =
1488 cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1489 if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))
1490 continue;
1491
1492 #ifndef NDEBUG
1493 for (auto *ChildLoopBB : ChildL->blocks())
1494 assert(VMap.count(ChildLoopBB) &&
1495 "Cloned a child loop header but not all of that loops blocks!");
1496 #endif
1497
1498 NonChildClonedLoops.push_back(cloneLoopNest(
1499 *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));
1500 }
1501 }
1502
1503 static void
deleteDeadClonedBlocks(Loop & L,ArrayRef<BasicBlock * > ExitBlocks,ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,DominatorTree & DT,MemorySSAUpdater * MSSAU)1504 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1505 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1506 DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1507 // Find all the dead clones, and remove them from their successors.
1508 SmallVector<BasicBlock *, 16> DeadBlocks;
1509 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))
1510 for (auto &VMap : VMaps)
1511 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))
1512 if (!DT.isReachableFromEntry(ClonedBB)) {
1513 for (BasicBlock *SuccBB : successors(ClonedBB))
1514 SuccBB->removePredecessor(ClonedBB);
1515 DeadBlocks.push_back(ClonedBB);
1516 }
1517
1518 // Remove all MemorySSA in the dead blocks
1519 if (MSSAU) {
1520 SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(),
1521 DeadBlocks.end());
1522 MSSAU->removeBlocks(DeadBlockSet);
1523 }
1524
1525 // Drop any remaining references to break cycles.
1526 for (BasicBlock *BB : DeadBlocks)
1527 BB->dropAllReferences();
1528 // Erase them from the IR.
1529 for (BasicBlock *BB : DeadBlocks)
1530 BB->eraseFromParent();
1531 }
1532
deleteDeadBlocksFromLoop(Loop & L,SmallVectorImpl<BasicBlock * > & ExitBlocks,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU)1533 static void deleteDeadBlocksFromLoop(Loop &L,
1534 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1535 DominatorTree &DT, LoopInfo &LI,
1536 MemorySSAUpdater *MSSAU) {
1537 // Find all the dead blocks tied to this loop, and remove them from their
1538 // successors.
1539 SmallSetVector<BasicBlock *, 8> DeadBlockSet;
1540
1541 // Start with loop/exit blocks and get a transitive closure of reachable dead
1542 // blocks.
1543 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1544 ExitBlocks.end());
1545 DeathCandidates.append(L.blocks().begin(), L.blocks().end());
1546 while (!DeathCandidates.empty()) {
1547 auto *BB = DeathCandidates.pop_back_val();
1548 if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {
1549 for (BasicBlock *SuccBB : successors(BB)) {
1550 SuccBB->removePredecessor(BB);
1551 DeathCandidates.push_back(SuccBB);
1552 }
1553 DeadBlockSet.insert(BB);
1554 }
1555 }
1556
1557 // Remove all MemorySSA in the dead blocks
1558 if (MSSAU)
1559 MSSAU->removeBlocks(DeadBlockSet);
1560
1561 // Filter out the dead blocks from the exit blocks list so that it can be
1562 // used in the caller.
1563 llvm::erase_if(ExitBlocks,
1564 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1565
1566 // Walk from this loop up through its parents removing all of the dead blocks.
1567 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1568 for (auto *BB : DeadBlockSet)
1569 ParentL->getBlocksSet().erase(BB);
1570 llvm::erase_if(ParentL->getBlocksVector(),
1571 [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1572 }
1573
1574 // Now delete the dead child loops. This raw delete will clear them
1575 // recursively.
1576 llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {
1577 if (!DeadBlockSet.count(ChildL->getHeader()))
1578 return false;
1579
1580 assert(llvm::all_of(ChildL->blocks(),
1581 [&](BasicBlock *ChildBB) {
1582 return DeadBlockSet.count(ChildBB);
1583 }) &&
1584 "If the child loop header is dead all blocks in the child loop must "
1585 "be dead as well!");
1586 LI.destroy(ChildL);
1587 return true;
1588 });
1589
1590 // Remove the loop mappings for the dead blocks and drop all the references
1591 // from these blocks to others to handle cyclic references as we start
1592 // deleting the blocks themselves.
1593 for (auto *BB : DeadBlockSet) {
1594 // Check that the dominator tree has already been updated.
1595 assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1596 LI.changeLoopFor(BB, nullptr);
1597 // Drop all uses of the instructions to make sure we won't have dangling
1598 // uses in other blocks.
1599 for (auto &I : *BB)
1600 if (!I.use_empty())
1601 I.replaceAllUsesWith(UndefValue::get(I.getType()));
1602 BB->dropAllReferences();
1603 }
1604
1605 // Actually delete the blocks now that they've been fully unhooked from the
1606 // IR.
1607 for (auto *BB : DeadBlockSet)
1608 BB->eraseFromParent();
1609 }
1610
1611 /// Recompute the set of blocks in a loop after unswitching.
1612 ///
1613 /// This walks from the original headers predecessors to rebuild the loop. We
1614 /// take advantage of the fact that new blocks can't have been added, and so we
1615 /// filter by the original loop's blocks. This also handles potentially
1616 /// unreachable code that we don't want to explore but might be found examining
1617 /// the predecessors of the header.
1618 ///
1619 /// If the original loop is no longer a loop, this will return an empty set. If
1620 /// it remains a loop, all the blocks within it will be added to the set
1621 /// (including those blocks in inner loops).
recomputeLoopBlockSet(Loop & L,LoopInfo & LI)1622 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1623 LoopInfo &LI) {
1624 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1625
1626 auto *PH = L.getLoopPreheader();
1627 auto *Header = L.getHeader();
1628
1629 // A worklist to use while walking backwards from the header.
1630 SmallVector<BasicBlock *, 16> Worklist;
1631
1632 // First walk the predecessors of the header to find the backedges. This will
1633 // form the basis of our walk.
1634 for (auto *Pred : predecessors(Header)) {
1635 // Skip the preheader.
1636 if (Pred == PH)
1637 continue;
1638
1639 // Because the loop was in simplified form, the only non-loop predecessor
1640 // is the preheader.
1641 assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1642 "than the preheader that is not part of the "
1643 "loop!");
1644
1645 // Insert this block into the loop set and on the first visit and, if it
1646 // isn't the header we're currently walking, put it into the worklist to
1647 // recurse through.
1648 if (LoopBlockSet.insert(Pred).second && Pred != Header)
1649 Worklist.push_back(Pred);
1650 }
1651
1652 // If no backedges were found, we're done.
1653 if (LoopBlockSet.empty())
1654 return LoopBlockSet;
1655
1656 // We found backedges, recurse through them to identify the loop blocks.
1657 while (!Worklist.empty()) {
1658 BasicBlock *BB = Worklist.pop_back_val();
1659 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1660
1661 // No need to walk past the header.
1662 if (BB == Header)
1663 continue;
1664
1665 // Because we know the inner loop structure remains valid we can use the
1666 // loop structure to jump immediately across the entire nested loop.
1667 // Further, because it is in loop simplified form, we can directly jump
1668 // to its preheader afterward.
1669 if (Loop *InnerL = LI.getLoopFor(BB))
1670 if (InnerL != &L) {
1671 assert(L.contains(InnerL) &&
1672 "Should not reach a loop *outside* this loop!");
1673 // The preheader is the only possible predecessor of the loop so
1674 // insert it into the set and check whether it was already handled.
1675 auto *InnerPH = InnerL->getLoopPreheader();
1676 assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1677 "but not contain the inner loop "
1678 "preheader!");
1679 if (!LoopBlockSet.insert(InnerPH).second)
1680 // The only way to reach the preheader is through the loop body
1681 // itself so if it has been visited the loop is already handled.
1682 continue;
1683
1684 // Insert all of the blocks (other than those already present) into
1685 // the loop set. We expect at least the block that led us to find the
1686 // inner loop to be in the block set, but we may also have other loop
1687 // blocks if they were already enqueued as predecessors of some other
1688 // outer loop block.
1689 for (auto *InnerBB : InnerL->blocks()) {
1690 if (InnerBB == BB) {
1691 assert(LoopBlockSet.count(InnerBB) &&
1692 "Block should already be in the set!");
1693 continue;
1694 }
1695
1696 LoopBlockSet.insert(InnerBB);
1697 }
1698
1699 // Add the preheader to the worklist so we will continue past the
1700 // loop body.
1701 Worklist.push_back(InnerPH);
1702 continue;
1703 }
1704
1705 // Insert any predecessors that were in the original loop into the new
1706 // set, and if the insert is successful, add them to the worklist.
1707 for (auto *Pred : predecessors(BB))
1708 if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)
1709 Worklist.push_back(Pred);
1710 }
1711
1712 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1713
1714 // We've found all the blocks participating in the loop, return our completed
1715 // set.
1716 return LoopBlockSet;
1717 }
1718
1719 /// Rebuild a loop after unswitching removes some subset of blocks and edges.
1720 ///
1721 /// The removal may have removed some child loops entirely but cannot have
1722 /// disturbed any remaining child loops. However, they may need to be hoisted
1723 /// to the parent loop (or to be top-level loops). The original loop may be
1724 /// completely removed.
1725 ///
1726 /// The sibling loops resulting from this update are returned. If the original
1727 /// loop remains a valid loop, it will be the first entry in this list with all
1728 /// of the newly sibling loops following it.
1729 ///
1730 /// Returns true if the loop remains a loop after unswitching, and false if it
1731 /// is no longer a loop after unswitching (and should not continue to be
1732 /// referenced).
rebuildLoopAfterUnswitch(Loop & L,ArrayRef<BasicBlock * > ExitBlocks,LoopInfo & LI,SmallVectorImpl<Loop * > & HoistedLoops)1733 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1734 LoopInfo &LI,
1735 SmallVectorImpl<Loop *> &HoistedLoops) {
1736 auto *PH = L.getLoopPreheader();
1737
1738 // Compute the actual parent loop from the exit blocks. Because we may have
1739 // pruned some exits the loop may be different from the original parent.
1740 Loop *ParentL = nullptr;
1741 SmallVector<Loop *, 4> ExitLoops;
1742 SmallVector<BasicBlock *, 4> ExitsInLoops;
1743 ExitsInLoops.reserve(ExitBlocks.size());
1744 for (auto *ExitBB : ExitBlocks)
1745 if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1746 ExitLoops.push_back(ExitL);
1747 ExitsInLoops.push_back(ExitBB);
1748 if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1749 ParentL = ExitL;
1750 }
1751
1752 // Recompute the blocks participating in this loop. This may be empty if it
1753 // is no longer a loop.
1754 auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1755
1756 // If we still have a loop, we need to re-set the loop's parent as the exit
1757 // block set changing may have moved it within the loop nest. Note that this
1758 // can only happen when this loop has a parent as it can only hoist the loop
1759 // *up* the nest.
1760 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1761 // Remove this loop's (original) blocks from all of the intervening loops.
1762 for (Loop *IL = L.getParentLoop(); IL != ParentL;
1763 IL = IL->getParentLoop()) {
1764 IL->getBlocksSet().erase(PH);
1765 for (auto *BB : L.blocks())
1766 IL->getBlocksSet().erase(BB);
1767 llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {
1768 return BB == PH || L.contains(BB);
1769 });
1770 }
1771
1772 LI.changeLoopFor(PH, ParentL);
1773 L.getParentLoop()->removeChildLoop(&L);
1774 if (ParentL)
1775 ParentL->addChildLoop(&L);
1776 else
1777 LI.addTopLevelLoop(&L);
1778 }
1779
1780 // Now we update all the blocks which are no longer within the loop.
1781 auto &Blocks = L.getBlocksVector();
1782 auto BlocksSplitI =
1783 LoopBlockSet.empty()
1784 ? Blocks.begin()
1785 : std::stable_partition(
1786 Blocks.begin(), Blocks.end(),
1787 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });
1788
1789 // Before we erase the list of unlooped blocks, build a set of them.
1790 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1791 if (LoopBlockSet.empty())
1792 UnloopedBlocks.insert(PH);
1793
1794 // Now erase these blocks from the loop.
1795 for (auto *BB : make_range(BlocksSplitI, Blocks.end()))
1796 L.getBlocksSet().erase(BB);
1797 Blocks.erase(BlocksSplitI, Blocks.end());
1798
1799 // Sort the exits in ascending loop depth, we'll work backwards across these
1800 // to process them inside out.
1801 llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1802 return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);
1803 });
1804
1805 // We'll build up a set for each exit loop.
1806 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1807 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1808
1809 auto RemoveUnloopedBlocksFromLoop =
1810 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1811 for (auto *BB : UnloopedBlocks)
1812 L.getBlocksSet().erase(BB);
1813 llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {
1814 return UnloopedBlocks.count(BB);
1815 });
1816 };
1817
1818 SmallVector<BasicBlock *, 16> Worklist;
1819 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1820 assert(Worklist.empty() && "Didn't clear worklist!");
1821 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
1822
1823 // Grab the next exit block, in decreasing loop depth order.
1824 BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1825 Loop &ExitL = *LI.getLoopFor(ExitBB);
1826 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
1827
1828 // Erase all of the unlooped blocks from the loops between the previous
1829 // exit loop and this exit loop. This works because the ExitInLoops list is
1830 // sorted in increasing order of loop depth and thus we visit loops in
1831 // decreasing order of loop depth.
1832 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
1833 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1834
1835 // Walk the CFG back until we hit the cloned PH adding everything reachable
1836 // and in the unlooped set to this exit block's loop.
1837 Worklist.push_back(ExitBB);
1838 do {
1839 BasicBlock *BB = Worklist.pop_back_val();
1840 // We can stop recursing at the cloned preheader (if we get there).
1841 if (BB == PH)
1842 continue;
1843
1844 for (BasicBlock *PredBB : predecessors(BB)) {
1845 // If this pred has already been moved to our set or is part of some
1846 // (inner) loop, no update needed.
1847 if (!UnloopedBlocks.erase(PredBB)) {
1848 assert((NewExitLoopBlocks.count(PredBB) ||
1849 ExitL.contains(LI.getLoopFor(PredBB))) &&
1850 "Predecessor not in a nested loop (or already visited)!");
1851 continue;
1852 }
1853
1854 // We just insert into the loop set here. We'll add these blocks to the
1855 // exit loop after we build up the set in a deterministic order rather
1856 // than the predecessor-influenced visit order.
1857 bool Inserted = NewExitLoopBlocks.insert(PredBB).second;
1858 (void)Inserted;
1859 assert(Inserted && "Should only visit an unlooped block once!");
1860
1861 // And recurse through to its predecessors.
1862 Worklist.push_back(PredBB);
1863 }
1864 } while (!Worklist.empty());
1865
1866 // If blocks in this exit loop were directly part of the original loop (as
1867 // opposed to a child loop) update the map to point to this exit loop. This
1868 // just updates a map and so the fact that the order is unstable is fine.
1869 for (auto *BB : NewExitLoopBlocks)
1870 if (Loop *BBL = LI.getLoopFor(BB))
1871 if (BBL == &L || !L.contains(BBL))
1872 LI.changeLoopFor(BB, &ExitL);
1873
1874 // We will remove the remaining unlooped blocks from this loop in the next
1875 // iteration or below.
1876 NewExitLoopBlocks.clear();
1877 }
1878
1879 // Any remaining unlooped blocks are no longer part of any loop unless they
1880 // are part of some child loop.
1881 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
1882 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1883 for (auto *BB : UnloopedBlocks)
1884 if (Loop *BBL = LI.getLoopFor(BB))
1885 if (BBL == &L || !L.contains(BBL))
1886 LI.changeLoopFor(BB, nullptr);
1887
1888 // Sink all the child loops whose headers are no longer in the loop set to
1889 // the parent (or to be top level loops). We reach into the loop and directly
1890 // update its subloop vector to make this batch update efficient.
1891 auto &SubLoops = L.getSubLoopsVector();
1892 auto SubLoopsSplitI =
1893 LoopBlockSet.empty()
1894 ? SubLoops.begin()
1895 : std::stable_partition(
1896 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {
1897 return LoopBlockSet.count(SubL->getHeader());
1898 });
1899 for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {
1900 HoistedLoops.push_back(HoistedL);
1901 HoistedL->setParentLoop(nullptr);
1902
1903 // To compute the new parent of this hoisted loop we look at where we
1904 // placed the preheader above. We can't lookup the header itself because we
1905 // retained the mapping from the header to the hoisted loop. But the
1906 // preheader and header should have the exact same new parent computed
1907 // based on the set of exit blocks from the original loop as the preheader
1908 // is a predecessor of the header and so reached in the reverse walk. And
1909 // because the loops were all in simplified form the preheader of the
1910 // hoisted loop can't be part of some *other* loop.
1911 if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))
1912 NewParentL->addChildLoop(HoistedL);
1913 else
1914 LI.addTopLevelLoop(HoistedL);
1915 }
1916 SubLoops.erase(SubLoopsSplitI, SubLoops.end());
1917
1918 // Actually delete the loop if nothing remained within it.
1919 if (Blocks.empty()) {
1920 assert(SubLoops.empty() &&
1921 "Failed to remove all subloops from the original loop!");
1922 if (Loop *ParentL = L.getParentLoop())
1923 ParentL->removeChildLoop(llvm::find(*ParentL, &L));
1924 else
1925 LI.removeLoop(llvm::find(LI, &L));
1926 LI.destroy(&L);
1927 return false;
1928 }
1929
1930 return true;
1931 }
1932
1933 /// Helper to visit a dominator subtree, invoking a callable on each node.
1934 ///
1935 /// Returning false at any point will stop walking past that node of the tree.
1936 template <typename CallableT>
visitDomSubTree(DominatorTree & DT,BasicBlock * BB,CallableT Callable)1937 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
1938 SmallVector<DomTreeNode *, 4> DomWorklist;
1939 DomWorklist.push_back(DT[BB]);
1940 #ifndef NDEBUG
1941 SmallPtrSet<DomTreeNode *, 4> Visited;
1942 Visited.insert(DT[BB]);
1943 #endif
1944 do {
1945 DomTreeNode *N = DomWorklist.pop_back_val();
1946
1947 // Visit this node.
1948 if (!Callable(N->getBlock()))
1949 continue;
1950
1951 // Accumulate the child nodes.
1952 for (DomTreeNode *ChildN : *N) {
1953 assert(Visited.insert(ChildN).second &&
1954 "Cannot visit a node twice when walking a tree!");
1955 DomWorklist.push_back(ChildN);
1956 }
1957 } while (!DomWorklist.empty());
1958 }
1959
unswitchNontrivialInvariants(Loop & L,Instruction & TI,ArrayRef<Value * > Invariants,SmallVectorImpl<BasicBlock * > & ExitBlocks,DominatorTree & DT,LoopInfo & LI,AssumptionCache & AC,function_ref<void (bool,ArrayRef<Loop * >)> UnswitchCB,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)1960 static void unswitchNontrivialInvariants(
1961 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
1962 SmallVectorImpl<BasicBlock *> &ExitBlocks, DominatorTree &DT, LoopInfo &LI,
1963 AssumptionCache &AC, function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
1964 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
1965 auto *ParentBB = TI.getParent();
1966 BranchInst *BI = dyn_cast<BranchInst>(&TI);
1967 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
1968
1969 // We can only unswitch switches, conditional branches with an invariant
1970 // condition, or combining invariant conditions with an instruction.
1971 assert((SI || (BI && BI->isConditional())) &&
1972 "Can only unswitch switches and conditional branch!");
1973 bool FullUnswitch = SI || BI->getCondition() == Invariants[0];
1974 if (FullUnswitch)
1975 assert(Invariants.size() == 1 &&
1976 "Cannot have other invariants with full unswitching!");
1977 else
1978 assert(isa<Instruction>(BI->getCondition()) &&
1979 "Partial unswitching requires an instruction as the condition!");
1980
1981 if (MSSAU && VerifyMemorySSA)
1982 MSSAU->getMemorySSA()->verifyMemorySSA();
1983
1984 // Constant and BBs tracking the cloned and continuing successor. When we are
1985 // unswitching the entire condition, this can just be trivially chosen to
1986 // unswitch towards `true`. However, when we are unswitching a set of
1987 // invariants combined with `and` or `or`, the combining operation determines
1988 // the best direction to unswitch: we want to unswitch the direction that will
1989 // collapse the branch.
1990 bool Direction = true;
1991 int ClonedSucc = 0;
1992 if (!FullUnswitch) {
1993 if (cast<Instruction>(BI->getCondition())->getOpcode() != Instruction::Or) {
1994 assert(cast<Instruction>(BI->getCondition())->getOpcode() ==
1995 Instruction::And &&
1996 "Only `or` and `and` instructions can combine invariants being "
1997 "unswitched.");
1998 Direction = false;
1999 ClonedSucc = 1;
2000 }
2001 }
2002
2003 BasicBlock *RetainedSuccBB =
2004 BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();
2005 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
2006 if (BI)
2007 UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));
2008 else
2009 for (auto Case : SI->cases())
2010 if (Case.getCaseSuccessor() != RetainedSuccBB)
2011 UnswitchedSuccBBs.insert(Case.getCaseSuccessor());
2012
2013 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
2014 "Should not unswitch the same successor we are retaining!");
2015
2016 // The branch should be in this exact loop. Any inner loop's invariant branch
2017 // should be handled by unswitching that inner loop. The caller of this
2018 // routine should filter out any candidates that remain (but were skipped for
2019 // whatever reason).
2020 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
2021
2022 // Compute the parent loop now before we start hacking on things.
2023 Loop *ParentL = L.getParentLoop();
2024 // Get blocks in RPO order for MSSA update, before changing the CFG.
2025 LoopBlocksRPO LBRPO(&L);
2026 if (MSSAU)
2027 LBRPO.perform(&LI);
2028
2029 // Compute the outer-most loop containing one of our exit blocks. This is the
2030 // furthest up our loopnest which can be mutated, which we will use below to
2031 // update things.
2032 Loop *OuterExitL = &L;
2033 for (auto *ExitBB : ExitBlocks) {
2034 Loop *NewOuterExitL = LI.getLoopFor(ExitBB);
2035 if (!NewOuterExitL) {
2036 // We exited the entire nest with this block, so we're done.
2037 OuterExitL = nullptr;
2038 break;
2039 }
2040 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))
2041 OuterExitL = NewOuterExitL;
2042 }
2043
2044 // At this point, we're definitely going to unswitch something so invalidate
2045 // any cached information in ScalarEvolution for the outer most loop
2046 // containing an exit block and all nested loops.
2047 if (SE) {
2048 if (OuterExitL)
2049 SE->forgetLoop(OuterExitL);
2050 else
2051 SE->forgetTopmostLoop(&L);
2052 }
2053
2054 // If the edge from this terminator to a successor dominates that successor,
2055 // store a map from each block in its dominator subtree to it. This lets us
2056 // tell when cloning for a particular successor if a block is dominated by
2057 // some *other* successor with a single data structure. We use this to
2058 // significantly reduce cloning.
2059 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
2060 for (auto *SuccBB : llvm::concat<BasicBlock *const>(
2061 makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs))
2062 if (SuccBB->getUniquePredecessor() ||
2063 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2064 return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);
2065 }))
2066 visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {
2067 DominatingSucc[BB] = SuccBB;
2068 return true;
2069 });
2070
2071 // Split the preheader, so that we know that there is a safe place to insert
2072 // the conditional branch. We will change the preheader to have a conditional
2073 // branch on LoopCond. The original preheader will become the split point
2074 // between the unswitched versions, and we will have a new preheader for the
2075 // original loop.
2076 BasicBlock *SplitBB = L.getLoopPreheader();
2077 BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU);
2078
2079 // Keep track of the dominator tree updates needed.
2080 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2081
2082 // Clone the loop for each unswitched successor.
2083 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
2084 VMaps.reserve(UnswitchedSuccBBs.size());
2085 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
2086 for (auto *SuccBB : UnswitchedSuccBBs) {
2087 VMaps.emplace_back(new ValueToValueMapTy());
2088 ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2089 L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,
2090 DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU);
2091 }
2092
2093 // Drop metadata if we may break its semantics by moving this instr into the
2094 // split block.
2095 if (TI.getMetadata(LLVMContext::MD_make_implicit)) {
2096 if (DropNonTrivialImplicitNullChecks)
2097 // Do not spend time trying to understand if we can keep it, just drop it
2098 // to save compile time.
2099 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2100 else {
2101 // It is only legal to preserve make.implicit metadata if we are
2102 // guaranteed no reach implicit null check after following this branch.
2103 ICFLoopSafetyInfo SafetyInfo;
2104 SafetyInfo.computeLoopSafetyInfo(&L);
2105 if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))
2106 TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2107 }
2108 }
2109
2110 // The stitching of the branched code back together depends on whether we're
2111 // doing full unswitching or not with the exception that we always want to
2112 // nuke the initial terminator placed in the split block.
2113 SplitBB->getTerminator()->eraseFromParent();
2114 if (FullUnswitch) {
2115 // Splice the terminator from the original loop and rewrite its
2116 // successors.
2117 SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI);
2118
2119 // Keep a clone of the terminator for MSSA updates.
2120 Instruction *NewTI = TI.clone();
2121 ParentBB->getInstList().push_back(NewTI);
2122
2123 // First wire up the moved terminator to the preheaders.
2124 if (BI) {
2125 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2126 BI->setSuccessor(ClonedSucc, ClonedPH);
2127 BI->setSuccessor(1 - ClonedSucc, LoopPH);
2128 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2129 } else {
2130 assert(SI && "Must either be a branch or switch!");
2131
2132 // Walk the cases and directly update their successors.
2133 assert(SI->getDefaultDest() == RetainedSuccBB &&
2134 "Not retaining default successor!");
2135 SI->setDefaultDest(LoopPH);
2136 for (auto &Case : SI->cases())
2137 if (Case.getCaseSuccessor() == RetainedSuccBB)
2138 Case.setSuccessor(LoopPH);
2139 else
2140 Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);
2141
2142 // We need to use the set to populate domtree updates as even when there
2143 // are multiple cases pointing at the same successor we only want to
2144 // remove and insert one edge in the domtree.
2145 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2146 DTUpdates.push_back(
2147 {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});
2148 }
2149
2150 if (MSSAU) {
2151 DT.applyUpdates(DTUpdates);
2152 DTUpdates.clear();
2153
2154 // Remove all but one edge to the retained block and all unswitched
2155 // blocks. This is to avoid having duplicate entries in the cloned Phis,
2156 // when we know we only keep a single edge for each case.
2157 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB);
2158 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2159 MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB);
2160
2161 for (auto &VMap : VMaps)
2162 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2163 /*IgnoreIncomingWithNoClones=*/true);
2164 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2165
2166 // Remove all edges to unswitched blocks.
2167 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2168 MSSAU->removeEdge(ParentBB, SuccBB);
2169 }
2170
2171 // Now unhook the successor relationship as we'll be replacing
2172 // the terminator with a direct branch. This is much simpler for branches
2173 // than switches so we handle those first.
2174 if (BI) {
2175 // Remove the parent as a predecessor of the unswitched successor.
2176 assert(UnswitchedSuccBBs.size() == 1 &&
2177 "Only one possible unswitched block for a branch!");
2178 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2179 UnswitchedSuccBB->removePredecessor(ParentBB,
2180 /*KeepOneInputPHIs*/ true);
2181 DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2182 } else {
2183 // Note that we actually want to remove the parent block as a predecessor
2184 // of *every* case successor. The case successor is either unswitched,
2185 // completely eliminating an edge from the parent to that successor, or it
2186 // is a duplicate edge to the retained successor as the retained successor
2187 // is always the default successor and as we'll replace this with a direct
2188 // branch we no longer need the duplicate entries in the PHI nodes.
2189 SwitchInst *NewSI = cast<SwitchInst>(NewTI);
2190 assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2191 "Not retaining default successor!");
2192 for (auto &Case : NewSI->cases())
2193 Case.getCaseSuccessor()->removePredecessor(
2194 ParentBB,
2195 /*KeepOneInputPHIs*/ true);
2196
2197 // We need to use the set to populate domtree updates as even when there
2198 // are multiple cases pointing at the same successor we only want to
2199 // remove and insert one edge in the domtree.
2200 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2201 DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});
2202 }
2203
2204 // After MSSAU update, remove the cloned terminator instruction NewTI.
2205 ParentBB->getTerminator()->eraseFromParent();
2206
2207 // Create a new unconditional branch to the continuing block (as opposed to
2208 // the one cloned).
2209 BranchInst::Create(RetainedSuccBB, ParentBB);
2210 } else {
2211 assert(BI && "Only branches have partial unswitching.");
2212 assert(UnswitchedSuccBBs.size() == 1 &&
2213 "Only one possible unswitched block for a branch!");
2214 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2215 // When doing a partial unswitch, we have to do a bit more work to build up
2216 // the branch in the split block.
2217 buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction,
2218 *ClonedPH, *LoopPH);
2219 DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2220
2221 if (MSSAU) {
2222 DT.applyUpdates(DTUpdates);
2223 DTUpdates.clear();
2224
2225 // Perform MSSA cloning updates.
2226 for (auto &VMap : VMaps)
2227 MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2228 /*IgnoreIncomingWithNoClones=*/true);
2229 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2230 }
2231 }
2232
2233 // Apply the updates accumulated above to get an up-to-date dominator tree.
2234 DT.applyUpdates(DTUpdates);
2235
2236 // Now that we have an accurate dominator tree, first delete the dead cloned
2237 // blocks so that we can accurately build any cloned loops. It is important to
2238 // not delete the blocks from the original loop yet because we still want to
2239 // reference the original loop to understand the cloned loop's structure.
2240 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2241
2242 // Build the cloned loop structure itself. This may be substantially
2243 // different from the original structure due to the simplified CFG. This also
2244 // handles inserting all the cloned blocks into the correct loops.
2245 SmallVector<Loop *, 4> NonChildClonedLoops;
2246 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2247 buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);
2248
2249 // Now that our cloned loops have been built, we can update the original loop.
2250 // First we delete the dead blocks from it and then we rebuild the loop
2251 // structure taking these deletions into account.
2252 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU);
2253
2254 if (MSSAU && VerifyMemorySSA)
2255 MSSAU->getMemorySSA()->verifyMemorySSA();
2256
2257 SmallVector<Loop *, 4> HoistedLoops;
2258 bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops);
2259
2260 if (MSSAU && VerifyMemorySSA)
2261 MSSAU->getMemorySSA()->verifyMemorySSA();
2262
2263 // This transformation has a high risk of corrupting the dominator tree, and
2264 // the below steps to rebuild loop structures will result in hard to debug
2265 // errors in that case so verify that the dominator tree is sane first.
2266 // FIXME: Remove this when the bugs stop showing up and rely on existing
2267 // verification steps.
2268 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2269
2270 if (BI) {
2271 // If we unswitched a branch which collapses the condition to a known
2272 // constant we want to replace all the uses of the invariants within both
2273 // the original and cloned blocks. We do this here so that we can use the
2274 // now updated dominator tree to identify which side the users are on.
2275 assert(UnswitchedSuccBBs.size() == 1 &&
2276 "Only one possible unswitched block for a branch!");
2277 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2278
2279 // When considering multiple partially-unswitched invariants
2280 // we cant just go replace them with constants in both branches.
2281 //
2282 // For 'AND' we infer that true branch ("continue") means true
2283 // for each invariant operand.
2284 // For 'OR' we can infer that false branch ("continue") means false
2285 // for each invariant operand.
2286 // So it happens that for multiple-partial case we dont replace
2287 // in the unswitched branch.
2288 bool ReplaceUnswitched = FullUnswitch || (Invariants.size() == 1);
2289
2290 ConstantInt *UnswitchedReplacement =
2291 Direction ? ConstantInt::getTrue(BI->getContext())
2292 : ConstantInt::getFalse(BI->getContext());
2293 ConstantInt *ContinueReplacement =
2294 Direction ? ConstantInt::getFalse(BI->getContext())
2295 : ConstantInt::getTrue(BI->getContext());
2296 for (Value *Invariant : Invariants)
2297 for (auto UI = Invariant->use_begin(), UE = Invariant->use_end();
2298 UI != UE;) {
2299 // Grab the use and walk past it so we can clobber it in the use list.
2300 Use *U = &*UI++;
2301 Instruction *UserI = dyn_cast<Instruction>(U->getUser());
2302 if (!UserI)
2303 continue;
2304
2305 // Replace it with the 'continue' side if in the main loop body, and the
2306 // unswitched if in the cloned blocks.
2307 if (DT.dominates(LoopPH, UserI->getParent()))
2308 U->set(ContinueReplacement);
2309 else if (ReplaceUnswitched &&
2310 DT.dominates(ClonedPH, UserI->getParent()))
2311 U->set(UnswitchedReplacement);
2312 }
2313 }
2314
2315 // We can change which blocks are exit blocks of all the cloned sibling
2316 // loops, the current loop, and any parent loops which shared exit blocks
2317 // with the current loop. As a consequence, we need to re-form LCSSA for
2318 // them. But we shouldn't need to re-form LCSSA for any child loops.
2319 // FIXME: This could be made more efficient by tracking which exit blocks are
2320 // new, and focusing on them, but that isn't likely to be necessary.
2321 //
2322 // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2323 // loop nest and update every loop that could have had its exits changed. We
2324 // also need to cover any intervening loops. We add all of these loops to
2325 // a list and sort them by loop depth to achieve this without updating
2326 // unnecessary loops.
2327 auto UpdateLoop = [&](Loop &UpdateL) {
2328 #ifndef NDEBUG
2329 UpdateL.verifyLoop();
2330 for (Loop *ChildL : UpdateL) {
2331 ChildL->verifyLoop();
2332 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2333 "Perturbed a child loop's LCSSA form!");
2334 }
2335 #endif
2336 // First build LCSSA for this loop so that we can preserve it when
2337 // forming dedicated exits. We don't want to perturb some other loop's
2338 // LCSSA while doing that CFG edit.
2339 formLCSSA(UpdateL, DT, &LI, SE);
2340
2341 // For loops reached by this loop's original exit blocks we may
2342 // introduced new, non-dedicated exits. At least try to re-form dedicated
2343 // exits for these loops. This may fail if they couldn't have dedicated
2344 // exits to start with.
2345 formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true);
2346 };
2347
2348 // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2349 // and we can do it in any order as they don't nest relative to each other.
2350 //
2351 // Also check if any of the loops we have updated have become top-level loops
2352 // as that will necessitate widening the outer loop scope.
2353 for (Loop *UpdatedL :
2354 llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {
2355 UpdateLoop(*UpdatedL);
2356 if (UpdatedL->isOutermost())
2357 OuterExitL = nullptr;
2358 }
2359 if (IsStillLoop) {
2360 UpdateLoop(L);
2361 if (L.isOutermost())
2362 OuterExitL = nullptr;
2363 }
2364
2365 // If the original loop had exit blocks, walk up through the outer most loop
2366 // of those exit blocks to update LCSSA and form updated dedicated exits.
2367 if (OuterExitL != &L)
2368 for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2369 OuterL = OuterL->getParentLoop())
2370 UpdateLoop(*OuterL);
2371
2372 #ifndef NDEBUG
2373 // Verify the entire loop structure to catch any incorrect updates before we
2374 // progress in the pass pipeline.
2375 LI.verify(DT);
2376 #endif
2377
2378 // Now that we've unswitched something, make callbacks to report the changes.
2379 // For that we need to merge together the updated loops and the cloned loops
2380 // and check whether the original loop survived.
2381 SmallVector<Loop *, 4> SibLoops;
2382 for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
2383 if (UpdatedL->getParentLoop() == ParentL)
2384 SibLoops.push_back(UpdatedL);
2385 UnswitchCB(IsStillLoop, SibLoops);
2386
2387 if (MSSAU && VerifyMemorySSA)
2388 MSSAU->getMemorySSA()->verifyMemorySSA();
2389
2390 if (BI)
2391 ++NumBranches;
2392 else
2393 ++NumSwitches;
2394 }
2395
2396 /// Recursively compute the cost of a dominator subtree based on the per-block
2397 /// cost map provided.
2398 ///
2399 /// The recursive computation is memozied into the provided DT-indexed cost map
2400 /// to allow querying it for most nodes in the domtree without it becoming
2401 /// quadratic.
2402 static int
computeDomSubtreeCost(DomTreeNode & N,const SmallDenseMap<BasicBlock *,int,4> & BBCostMap,SmallDenseMap<DomTreeNode *,int,4> & DTCostMap)2403 computeDomSubtreeCost(DomTreeNode &N,
2404 const SmallDenseMap<BasicBlock *, int, 4> &BBCostMap,
2405 SmallDenseMap<DomTreeNode *, int, 4> &DTCostMap) {
2406 // Don't accumulate cost (or recurse through) blocks not in our block cost
2407 // map and thus not part of the duplication cost being considered.
2408 auto BBCostIt = BBCostMap.find(N.getBlock());
2409 if (BBCostIt == BBCostMap.end())
2410 return 0;
2411
2412 // Lookup this node to see if we already computed its cost.
2413 auto DTCostIt = DTCostMap.find(&N);
2414 if (DTCostIt != DTCostMap.end())
2415 return DTCostIt->second;
2416
2417 // If not, we have to compute it. We can't use insert above and update
2418 // because computing the cost may insert more things into the map.
2419 int Cost = std::accumulate(
2420 N.begin(), N.end(), BBCostIt->second, [&](int Sum, DomTreeNode *ChildN) {
2421 return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
2422 });
2423 bool Inserted = DTCostMap.insert({&N, Cost}).second;
2424 (void)Inserted;
2425 assert(Inserted && "Should not insert a node while visiting children!");
2426 return Cost;
2427 }
2428
2429 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2430 /// making the following replacement:
2431 ///
2432 /// --code before guard--
2433 /// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2434 /// --code after guard--
2435 ///
2436 /// into
2437 ///
2438 /// --code before guard--
2439 /// br i1 %cond, label %guarded, label %deopt
2440 ///
2441 /// guarded:
2442 /// --code after guard--
2443 ///
2444 /// deopt:
2445 /// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2446 /// unreachable
2447 ///
2448 /// It also makes all relevant DT and LI updates, so that all structures are in
2449 /// valid state after this transform.
2450 static BranchInst *
turnGuardIntoBranch(IntrinsicInst * GI,Loop & L,SmallVectorImpl<BasicBlock * > & ExitBlocks,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU)2451 turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2452 SmallVectorImpl<BasicBlock *> &ExitBlocks,
2453 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
2454 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2455 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2456 BasicBlock *CheckBB = GI->getParent();
2457
2458 if (MSSAU && VerifyMemorySSA)
2459 MSSAU->getMemorySSA()->verifyMemorySSA();
2460
2461 // Remove all CheckBB's successors from DomTree. A block can be seen among
2462 // successors more than once, but for DomTree it should be added only once.
2463 SmallPtrSet<BasicBlock *, 4> Successors;
2464 for (auto *Succ : successors(CheckBB))
2465 if (Successors.insert(Succ).second)
2466 DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ});
2467
2468 Instruction *DeoptBlockTerm =
2469 SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true);
2470 BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
2471 // SplitBlockAndInsertIfThen inserts control flow that branches to
2472 // DeoptBlockTerm if the condition is true. We want the opposite.
2473 CheckBI->swapSuccessors();
2474
2475 BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
2476 GuardedBlock->setName("guarded");
2477 CheckBI->getSuccessor(1)->setName("deopt");
2478 BasicBlock *DeoptBlock = CheckBI->getSuccessor(1);
2479
2480 // We now have a new exit block.
2481 ExitBlocks.push_back(CheckBI->getSuccessor(1));
2482
2483 if (MSSAU)
2484 MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI);
2485
2486 GI->moveBefore(DeoptBlockTerm);
2487 GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));
2488
2489 // Add new successors of CheckBB into DomTree.
2490 for (auto *Succ : successors(CheckBB))
2491 DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ});
2492
2493 // Now the blocks that used to be CheckBB's successors are GuardedBlock's
2494 // successors.
2495 for (auto *Succ : Successors)
2496 DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ});
2497
2498 // Make proper changes to DT.
2499 DT.applyUpdates(DTUpdates);
2500 // Inform LI of a new loop block.
2501 L.addBasicBlockToLoop(GuardedBlock, LI);
2502
2503 if (MSSAU) {
2504 MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI));
2505 MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator);
2506 if (VerifyMemorySSA)
2507 MSSAU->getMemorySSA()->verifyMemorySSA();
2508 }
2509
2510 ++NumGuards;
2511 return CheckBI;
2512 }
2513
2514 /// Cost multiplier is a way to limit potentially exponential behavior
2515 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2516 /// candidates available. Also accounting for the number of "sibling" loops with
2517 /// the idea to account for previous unswitches that already happened on this
2518 /// cluster of loops. There was an attempt to keep this formula simple,
2519 /// just enough to limit the worst case behavior. Even if it is not that simple
2520 /// now it is still not an attempt to provide a detailed heuristic size
2521 /// prediction.
2522 ///
2523 /// TODO: Make a proper accounting of "explosion" effect for all kinds of
2524 /// unswitch candidates, making adequate predictions instead of wild guesses.
2525 /// That requires knowing not just the number of "remaining" candidates but
2526 /// also costs of unswitching for each of these candidates.
CalculateUnswitchCostMultiplier(Instruction & TI,Loop & L,LoopInfo & LI,DominatorTree & DT,ArrayRef<std::pair<Instruction *,TinyPtrVector<Value * >>> UnswitchCandidates)2527 static int CalculateUnswitchCostMultiplier(
2528 Instruction &TI, Loop &L, LoopInfo &LI, DominatorTree &DT,
2529 ArrayRef<std::pair<Instruction *, TinyPtrVector<Value *>>>
2530 UnswitchCandidates) {
2531
2532 // Guards and other exiting conditions do not contribute to exponential
2533 // explosion as soon as they dominate the latch (otherwise there might be
2534 // another path to the latch remaining that does not allow to eliminate the
2535 // loop copy on unswitch).
2536 BasicBlock *Latch = L.getLoopLatch();
2537 BasicBlock *CondBlock = TI.getParent();
2538 if (DT.dominates(CondBlock, Latch) &&
2539 (isGuard(&TI) ||
2540 llvm::count_if(successors(&TI), [&L](BasicBlock *SuccBB) {
2541 return L.contains(SuccBB);
2542 }) <= 1)) {
2543 NumCostMultiplierSkipped++;
2544 return 1;
2545 }
2546
2547 auto *ParentL = L.getParentLoop();
2548 int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2549 : std::distance(LI.begin(), LI.end()));
2550 // Count amount of clones that all the candidates might cause during
2551 // unswitching. Branch/guard counts as 1, switch counts as log2 of its cases.
2552 int UnswitchedClones = 0;
2553 for (auto Candidate : UnswitchCandidates) {
2554 Instruction *CI = Candidate.first;
2555 BasicBlock *CondBlock = CI->getParent();
2556 bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch);
2557 if (isGuard(CI)) {
2558 if (!SkipExitingSuccessors)
2559 UnswitchedClones++;
2560 continue;
2561 }
2562 int NonExitingSuccessors = llvm::count_if(
2563 successors(CondBlock), [SkipExitingSuccessors, &L](BasicBlock *SuccBB) {
2564 return !SkipExitingSuccessors || L.contains(SuccBB);
2565 });
2566 UnswitchedClones += Log2_32(NonExitingSuccessors);
2567 }
2568
2569 // Ignore up to the "unscaled candidates" number of unswitch candidates
2570 // when calculating the power-of-two scaling of the cost. The main idea
2571 // with this control is to allow a small number of unswitches to happen
2572 // and rely more on siblings multiplier (see below) when the number
2573 // of candidates is small.
2574 unsigned ClonesPower =
2575 std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0);
2576
2577 // Allowing top-level loops to spread a bit more than nested ones.
2578 int SiblingsMultiplier =
2579 std::max((ParentL ? SiblingsCount
2580 : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2581 1);
2582 // Compute the cost multiplier in a way that won't overflow by saturating
2583 // at an upper bound.
2584 int CostMultiplier;
2585 if (ClonesPower > Log2_32(UnswitchThreshold) ||
2586 SiblingsMultiplier > UnswitchThreshold)
2587 CostMultiplier = UnswitchThreshold;
2588 else
2589 CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower),
2590 (int)UnswitchThreshold);
2591
2592 LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier
2593 << " (siblings " << SiblingsMultiplier << " * clones "
2594 << (1 << ClonesPower) << ")"
2595 << " for unswitch candidate: " << TI << "\n");
2596 return CostMultiplier;
2597 }
2598
2599 static bool
unswitchBestCondition(Loop & L,DominatorTree & DT,LoopInfo & LI,AssumptionCache & AC,TargetTransformInfo & TTI,function_ref<void (bool,ArrayRef<Loop * >)> UnswitchCB,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)2600 unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
2601 AssumptionCache &AC, TargetTransformInfo &TTI,
2602 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2603 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
2604 // Collect all invariant conditions within this loop (as opposed to an inner
2605 // loop which would be handled when visiting that inner loop).
2606 SmallVector<std::pair<Instruction *, TinyPtrVector<Value *>>, 4>
2607 UnswitchCandidates;
2608
2609 // Whether or not we should also collect guards in the loop.
2610 bool CollectGuards = false;
2611 if (UnswitchGuards) {
2612 auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2613 Intrinsic::getName(Intrinsic::experimental_guard));
2614 if (GuardDecl && !GuardDecl->use_empty())
2615 CollectGuards = true;
2616 }
2617
2618 for (auto *BB : L.blocks()) {
2619 if (LI.getLoopFor(BB) != &L)
2620 continue;
2621
2622 if (CollectGuards)
2623 for (auto &I : *BB)
2624 if (isGuard(&I)) {
2625 auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0);
2626 // TODO: Support AND, OR conditions and partial unswitching.
2627 if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
2628 UnswitchCandidates.push_back({&I, {Cond}});
2629 }
2630
2631 if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
2632 // We can only consider fully loop-invariant switch conditions as we need
2633 // to completely eliminate the switch after unswitching.
2634 if (!isa<Constant>(SI->getCondition()) &&
2635 L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor())
2636 UnswitchCandidates.push_back({SI, {SI->getCondition()}});
2637 continue;
2638 }
2639
2640 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
2641 if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) ||
2642 BI->getSuccessor(0) == BI->getSuccessor(1))
2643 continue;
2644
2645 if (L.isLoopInvariant(BI->getCondition())) {
2646 UnswitchCandidates.push_back({BI, {BI->getCondition()}});
2647 continue;
2648 }
2649
2650 Instruction &CondI = *cast<Instruction>(BI->getCondition());
2651 if (CondI.getOpcode() != Instruction::And &&
2652 CondI.getOpcode() != Instruction::Or)
2653 continue;
2654
2655 TinyPtrVector<Value *> Invariants =
2656 collectHomogenousInstGraphLoopInvariants(L, CondI, LI);
2657 if (Invariants.empty())
2658 continue;
2659
2660 UnswitchCandidates.push_back({BI, std::move(Invariants)});
2661 }
2662
2663 // If we didn't find any candidates, we're done.
2664 if (UnswitchCandidates.empty())
2665 return false;
2666
2667 // Check if there are irreducible CFG cycles in this loop. If so, we cannot
2668 // easily unswitch non-trivial edges out of the loop. Doing so might turn the
2669 // irreducible control flow into reducible control flow and introduce new
2670 // loops "out of thin air". If we ever discover important use cases for doing
2671 // this, we can add support to loop unswitch, but it is a lot of complexity
2672 // for what seems little or no real world benefit.
2673 LoopBlocksRPO RPOT(&L);
2674 RPOT.perform(&LI);
2675 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
2676 return false;
2677
2678 SmallVector<BasicBlock *, 4> ExitBlocks;
2679 L.getUniqueExitBlocks(ExitBlocks);
2680
2681 // We cannot unswitch if exit blocks contain a cleanuppad instruction as we
2682 // don't know how to split those exit blocks.
2683 // FIXME: We should teach SplitBlock to handle this and remove this
2684 // restriction.
2685 for (auto *ExitBB : ExitBlocks)
2686 if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI())) {
2687 dbgs() << "Cannot unswitch because of cleanuppad in exit block\n";
2688 return false;
2689 }
2690
2691 LLVM_DEBUG(
2692 dbgs() << "Considering " << UnswitchCandidates.size()
2693 << " non-trivial loop invariant conditions for unswitching.\n");
2694
2695 // Given that unswitching these terminators will require duplicating parts of
2696 // the loop, so we need to be able to model that cost. Compute the ephemeral
2697 // values and set up a data structure to hold per-BB costs. We cache each
2698 // block's cost so that we don't recompute this when considering different
2699 // subsets of the loop for duplication during unswitching.
2700 SmallPtrSet<const Value *, 4> EphValues;
2701 CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);
2702 SmallDenseMap<BasicBlock *, int, 4> BBCostMap;
2703
2704 // Compute the cost of each block, as well as the total loop cost. Also, bail
2705 // out if we see instructions which are incompatible with loop unswitching
2706 // (convergent, noduplicate, or cross-basic-block tokens).
2707 // FIXME: We might be able to safely handle some of these in non-duplicated
2708 // regions.
2709 TargetTransformInfo::TargetCostKind CostKind =
2710 L.getHeader()->getParent()->hasMinSize()
2711 ? TargetTransformInfo::TCK_CodeSize
2712 : TargetTransformInfo::TCK_SizeAndLatency;
2713 int LoopCost = 0;
2714 for (auto *BB : L.blocks()) {
2715 int Cost = 0;
2716 for (auto &I : *BB) {
2717 if (EphValues.count(&I))
2718 continue;
2719
2720 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
2721 return false;
2722 if (auto *CB = dyn_cast<CallBase>(&I))
2723 if (CB->isConvergent() || CB->cannotDuplicate())
2724 return false;
2725
2726 Cost += TTI.getUserCost(&I, CostKind);
2727 }
2728 assert(Cost >= 0 && "Must not have negative costs!");
2729 LoopCost += Cost;
2730 assert(LoopCost >= 0 && "Must not have negative loop costs!");
2731 BBCostMap[BB] = Cost;
2732 }
2733 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n");
2734
2735 // Now we find the best candidate by searching for the one with the following
2736 // properties in order:
2737 //
2738 // 1) An unswitching cost below the threshold
2739 // 2) The smallest number of duplicated unswitch candidates (to avoid
2740 // creating redundant subsequent unswitching)
2741 // 3) The smallest cost after unswitching.
2742 //
2743 // We prioritize reducing fanout of unswitch candidates provided the cost
2744 // remains below the threshold because this has a multiplicative effect.
2745 //
2746 // This requires memoizing each dominator subtree to avoid redundant work.
2747 //
2748 // FIXME: Need to actually do the number of candidates part above.
2749 SmallDenseMap<DomTreeNode *, int, 4> DTCostMap;
2750 // Given a terminator which might be unswitched, computes the non-duplicated
2751 // cost for that terminator.
2752 auto ComputeUnswitchedCost = [&](Instruction &TI, bool FullUnswitch) {
2753 BasicBlock &BB = *TI.getParent();
2754 SmallPtrSet<BasicBlock *, 4> Visited;
2755
2756 int Cost = LoopCost;
2757 for (BasicBlock *SuccBB : successors(&BB)) {
2758 // Don't count successors more than once.
2759 if (!Visited.insert(SuccBB).second)
2760 continue;
2761
2762 // If this is a partial unswitch candidate, then it must be a conditional
2763 // branch with a condition of either `or` or `and`. In that case, one of
2764 // the successors is necessarily duplicated, so don't even try to remove
2765 // its cost.
2766 if (!FullUnswitch) {
2767 auto &BI = cast<BranchInst>(TI);
2768 if (cast<Instruction>(BI.getCondition())->getOpcode() ==
2769 Instruction::And) {
2770 if (SuccBB == BI.getSuccessor(1))
2771 continue;
2772 } else {
2773 assert(cast<Instruction>(BI.getCondition())->getOpcode() ==
2774 Instruction::Or &&
2775 "Only `and` and `or` conditions can result in a partial "
2776 "unswitch!");
2777 if (SuccBB == BI.getSuccessor(0))
2778 continue;
2779 }
2780 }
2781
2782 // This successor's domtree will not need to be duplicated after
2783 // unswitching if the edge to the successor dominates it (and thus the
2784 // entire tree). This essentially means there is no other path into this
2785 // subtree and so it will end up live in only one clone of the loop.
2786 if (SuccBB->getUniquePredecessor() ||
2787 llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2788 return PredBB == &BB || DT.dominates(SuccBB, PredBB);
2789 })) {
2790 Cost -= computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);
2791 assert(Cost >= 0 &&
2792 "Non-duplicated cost should never exceed total loop cost!");
2793 }
2794 }
2795
2796 // Now scale the cost by the number of unique successors minus one. We
2797 // subtract one because there is already at least one copy of the entire
2798 // loop. This is computing the new cost of unswitching a condition.
2799 // Note that guards always have 2 unique successors that are implicit and
2800 // will be materialized if we decide to unswitch it.
2801 int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
2802 assert(SuccessorsCount > 1 &&
2803 "Cannot unswitch a condition without multiple distinct successors!");
2804 return Cost * (SuccessorsCount - 1);
2805 };
2806 Instruction *BestUnswitchTI = nullptr;
2807 int BestUnswitchCost = 0;
2808 ArrayRef<Value *> BestUnswitchInvariants;
2809 for (auto &TerminatorAndInvariants : UnswitchCandidates) {
2810 Instruction &TI = *TerminatorAndInvariants.first;
2811 ArrayRef<Value *> Invariants = TerminatorAndInvariants.second;
2812 BranchInst *BI = dyn_cast<BranchInst>(&TI);
2813 int CandidateCost = ComputeUnswitchedCost(
2814 TI, /*FullUnswitch*/ !BI || (Invariants.size() == 1 &&
2815 Invariants[0] == BI->getCondition()));
2816 // Calculate cost multiplier which is a tool to limit potentially
2817 // exponential behavior of loop-unswitch.
2818 if (EnableUnswitchCostMultiplier) {
2819 int CostMultiplier =
2820 CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
2821 assert(
2822 (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
2823 "cost multiplier needs to be in the range of 1..UnswitchThreshold");
2824 CandidateCost *= CostMultiplier;
2825 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
2826 << " (multiplier: " << CostMultiplier << ")"
2827 << " for unswitch candidate: " << TI << "\n");
2828 } else {
2829 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
2830 << " for unswitch candidate: " << TI << "\n");
2831 }
2832
2833 if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) {
2834 BestUnswitchTI = &TI;
2835 BestUnswitchCost = CandidateCost;
2836 BestUnswitchInvariants = Invariants;
2837 }
2838 }
2839 assert(BestUnswitchTI && "Failed to find loop unswitch candidate");
2840
2841 if (BestUnswitchCost >= UnswitchThreshold) {
2842 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: "
2843 << BestUnswitchCost << "\n");
2844 return false;
2845 }
2846
2847 // If the best candidate is a guard, turn it into a branch.
2848 if (isGuard(BestUnswitchTI))
2849 BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L,
2850 ExitBlocks, DT, LI, MSSAU);
2851
2852 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = "
2853 << BestUnswitchCost << ") terminator: " << *BestUnswitchTI
2854 << "\n");
2855 unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants,
2856 ExitBlocks, DT, LI, AC, UnswitchCB, SE, MSSAU);
2857 return true;
2858 }
2859
2860 /// Unswitch control flow predicated on loop invariant conditions.
2861 ///
2862 /// This first hoists all branches or switches which are trivial (IE, do not
2863 /// require duplicating any part of the loop) out of the loop body. It then
2864 /// looks at other loop invariant control flows and tries to unswitch those as
2865 /// well by cloning the loop if the result is small enough.
2866 ///
2867 /// The `DT`, `LI`, `AC`, `TTI` parameters are required analyses that are also
2868 /// updated based on the unswitch.
2869 /// The `MSSA` analysis is also updated if valid (i.e. its use is enabled).
2870 ///
2871 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
2872 /// true, we will attempt to do non-trivial unswitching as well as trivial
2873 /// unswitching.
2874 ///
2875 /// The `UnswitchCB` callback provided will be run after unswitching is
2876 /// complete, with the first parameter set to `true` if the provided loop
2877 /// remains a loop, and a list of new sibling loops created.
2878 ///
2879 /// If `SE` is non-null, we will update that analysis based on the unswitching
2880 /// done.
unswitchLoop(Loop & L,DominatorTree & DT,LoopInfo & LI,AssumptionCache & AC,TargetTransformInfo & TTI,bool NonTrivial,function_ref<void (bool,ArrayRef<Loop * >)> UnswitchCB,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)2881 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
2882 AssumptionCache &AC, TargetTransformInfo &TTI,
2883 bool NonTrivial,
2884 function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
2885 ScalarEvolution *SE, MemorySSAUpdater *MSSAU) {
2886 assert(L.isRecursivelyLCSSAForm(DT, LI) &&
2887 "Loops must be in LCSSA form before unswitching.");
2888
2889 // Must be in loop simplified form: we need a preheader and dedicated exits.
2890 if (!L.isLoopSimplifyForm())
2891 return false;
2892
2893 // Try trivial unswitch first before loop over other basic blocks in the loop.
2894 if (unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
2895 // If we unswitched successfully we will want to clean up the loop before
2896 // processing it further so just mark it as unswitched and return.
2897 UnswitchCB(/*CurrentLoopValid*/ true, {});
2898 return true;
2899 }
2900
2901 // If we're not doing non-trivial unswitching, we're done. We both accept
2902 // a parameter but also check a local flag that can be used for testing
2903 // a debugging.
2904 if (!NonTrivial && !EnableNonTrivialUnswitch)
2905 return false;
2906
2907 // For non-trivial unswitching, because it often creates new loops, we rely on
2908 // the pass manager to iterate on the loops rather than trying to immediately
2909 // reach a fixed point. There is no substantial advantage to iterating
2910 // internally, and if any of the new loops are simplified enough to contain
2911 // trivial unswitching we want to prefer those.
2912
2913 // Try to unswitch the best invariant condition. We prefer this full unswitch to
2914 // a partial unswitch when possible below the threshold.
2915 if (unswitchBestCondition(L, DT, LI, AC, TTI, UnswitchCB, SE, MSSAU))
2916 return true;
2917
2918 // No other opportunities to unswitch.
2919 return false;
2920 }
2921
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)2922 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
2923 LoopStandardAnalysisResults &AR,
2924 LPMUpdater &U) {
2925 Function &F = *L.getHeader()->getParent();
2926 (void)F;
2927
2928 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
2929 << "\n");
2930
2931 // Save the current loop name in a variable so that we can report it even
2932 // after it has been deleted.
2933 std::string LoopName = std::string(L.getName());
2934
2935 auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid,
2936 ArrayRef<Loop *> NewLoops) {
2937 // If we did a non-trivial unswitch, we have added new (cloned) loops.
2938 if (!NewLoops.empty())
2939 U.addSiblingLoops(NewLoops);
2940
2941 // If the current loop remains valid, we should revisit it to catch any
2942 // other unswitch opportunities. Otherwise, we need to mark it as deleted.
2943 if (CurrentLoopValid)
2944 U.revisitCurrentLoop();
2945 else
2946 U.markLoopAsDeleted(L, LoopName);
2947 };
2948
2949 Optional<MemorySSAUpdater> MSSAU;
2950 if (AR.MSSA) {
2951 MSSAU = MemorySSAUpdater(AR.MSSA);
2952 if (VerifyMemorySSA)
2953 AR.MSSA->verifyMemorySSA();
2954 }
2955 if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.TTI, NonTrivial, UnswitchCB,
2956 &AR.SE, MSSAU.hasValue() ? MSSAU.getPointer() : nullptr))
2957 return PreservedAnalyses::all();
2958
2959 if (AR.MSSA && VerifyMemorySSA)
2960 AR.MSSA->verifyMemorySSA();
2961
2962 // Historically this pass has had issues with the dominator tree so verify it
2963 // in asserts builds.
2964 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));
2965
2966 auto PA = getLoopPassPreservedAnalyses();
2967 if (AR.MSSA)
2968 PA.preserve<MemorySSAAnalysis>();
2969 return PA;
2970 }
2971
2972 namespace {
2973
2974 class SimpleLoopUnswitchLegacyPass : public LoopPass {
2975 bool NonTrivial;
2976
2977 public:
2978 static char ID; // Pass ID, replacement for typeid
2979
SimpleLoopUnswitchLegacyPass(bool NonTrivial=false)2980 explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false)
2981 : LoopPass(ID), NonTrivial(NonTrivial) {
2982 initializeSimpleLoopUnswitchLegacyPassPass(
2983 *PassRegistry::getPassRegistry());
2984 }
2985
2986 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
2987
getAnalysisUsage(AnalysisUsage & AU) const2988 void getAnalysisUsage(AnalysisUsage &AU) const override {
2989 AU.addRequired<AssumptionCacheTracker>();
2990 AU.addRequired<TargetTransformInfoWrapperPass>();
2991 if (EnableMSSALoopDependency) {
2992 AU.addRequired<MemorySSAWrapperPass>();
2993 AU.addPreserved<MemorySSAWrapperPass>();
2994 }
2995 getLoopAnalysisUsage(AU);
2996 }
2997 };
2998
2999 } // end anonymous namespace
3000
runOnLoop(Loop * L,LPPassManager & LPM)3001 bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
3002 if (skipLoop(L))
3003 return false;
3004
3005 Function &F = *L->getHeader()->getParent();
3006
3007 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L
3008 << "\n");
3009
3010 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3011 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
3012 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3013 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
3014 MemorySSA *MSSA = nullptr;
3015 Optional<MemorySSAUpdater> MSSAU;
3016 if (EnableMSSALoopDependency) {
3017 MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
3018 MSSAU = MemorySSAUpdater(MSSA);
3019 }
3020
3021 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
3022 auto *SE = SEWP ? &SEWP->getSE() : nullptr;
3023
3024 auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid,
3025 ArrayRef<Loop *> NewLoops) {
3026 // If we did a non-trivial unswitch, we have added new (cloned) loops.
3027 for (auto *NewL : NewLoops)
3028 LPM.addLoop(*NewL);
3029
3030 // If the current loop remains valid, re-add it to the queue. This is
3031 // a little wasteful as we'll finish processing the current loop as well,
3032 // but it is the best we can do in the old PM.
3033 if (CurrentLoopValid)
3034 LPM.addLoop(*L);
3035 else
3036 LPM.markLoopAsDeleted(*L);
3037 };
3038
3039 if (MSSA && VerifyMemorySSA)
3040 MSSA->verifyMemorySSA();
3041
3042 bool Changed = unswitchLoop(*L, DT, LI, AC, TTI, NonTrivial, UnswitchCB, SE,
3043 MSSAU.hasValue() ? MSSAU.getPointer() : nullptr);
3044
3045 if (MSSA && VerifyMemorySSA)
3046 MSSA->verifyMemorySSA();
3047
3048 // Historically this pass has had issues with the dominator tree so verify it
3049 // in asserts builds.
3050 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
3051
3052 return Changed;
3053 }
3054
3055 char SimpleLoopUnswitchLegacyPass::ID = 0;
3056 INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
3057 "Simple unswitch loops", false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)3058 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3059 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3060 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
3061 INITIALIZE_PASS_DEPENDENCY(LoopPass)
3062 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
3063 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
3064 INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
3065 "Simple unswitch loops", false, false)
3066
3067 Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) {
3068 return new SimpleLoopUnswitchLegacyPass(NonTrivial);
3069 }
3070