1//===----------------- LoopRotationUtils.cpp -----------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file provides utilities to convert a loop into a loop with bottom test.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/Transforms/Utils/LoopRotationUtils.h"
14#include "llvm/ADT/Statistic.h"
15#include "llvm/Analysis/AssumptionCache.h"
16#include "llvm/Analysis/CodeMetrics.h"
17#include "llvm/Analysis/DomTreeUpdater.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/LoopInfo.h"
20#include "llvm/Analysis/MemorySSA.h"
21#include "llvm/Analysis/MemorySSAUpdater.h"
22#include "llvm/Analysis/ScalarEvolution.h"
23#include "llvm/Analysis/ValueTracking.h"
24#include "llvm/IR/CFG.h"
25#include "llvm/IR/DebugInfo.h"
26#include "llvm/IR/Dominators.h"
27#include "llvm/IR/IntrinsicInst.h"
28#include "llvm/IR/MDBuilder.h"
29#include "llvm/IR/ProfDataUtils.h"
30#include "llvm/Support/CommandLine.h"
31#include "llvm/Support/Debug.h"
32#include "llvm/Support/raw_ostream.h"
33#include "llvm/Transforms/Utils/BasicBlockUtils.h"
34#include "llvm/Transforms/Utils/Cloning.h"
35#include "llvm/Transforms/Utils/Local.h"
36#include "llvm/Transforms/Utils/SSAUpdater.h"
37#include "llvm/Transforms/Utils/ValueMapper.h"
38using namespace llvm;
39
40#define DEBUG_TYPE "loop-rotate"
41
42STATISTIC(NumNotRotatedDueToHeaderSize,
43 "Number of loops not rotated due to the header size");
44STATISTIC(NumInstrsHoisted,
45 "Number of instructions hoisted into loop preheader");
46STATISTIC(NumInstrsDuplicated,
47 "Number of instructions cloned into loop preheader");
48
49// Probability that a rotated loop has zero trip count / is never entered.
50static constexpr uint32_t ZeroTripCountWeights[] = {1, 127};
51
52namespace {
53/// A simple loop rotation transformation.
54class LoopRotate {
55 const unsigned MaxHeaderSize;
56 LoopInfo *LI;
57 const TargetTransformInfo *TTI;
58 AssumptionCache *AC;
59 DominatorTree *DT;
60 ScalarEvolution *SE;
61 MemorySSAUpdater *MSSAU;
62 const SimplifyQuery &SQ;
63 bool RotationOnly;
64 bool IsUtilMode;
65 bool PrepareForLTO;
66
67public:
68 LoopRotate(unsigned MaxHeaderSize, LoopInfo *LI,
69 const TargetTransformInfo *TTI, AssumptionCache *AC,
70 DominatorTree *DT, ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
71 const SimplifyQuery &SQ, bool RotationOnly, bool IsUtilMode,
72 bool PrepareForLTO)
73 : MaxHeaderSize(MaxHeaderSize), LI(LI), TTI(TTI), AC(AC), DT(DT), SE(SE),
74 MSSAU(MSSAU), SQ(SQ), RotationOnly(RotationOnly),
75 IsUtilMode(IsUtilMode), PrepareForLTO(PrepareForLTO) {}
76 bool processLoop(Loop *L);
77
78private:
79 bool rotateLoop(Loop *L, bool SimplifiedLatch);
80 bool simplifyLoopLatch(Loop *L);
81};
82} // end anonymous namespace
83
84/// Insert (K, V) pair into the ValueToValueMap, and verify the key did not
85/// previously exist in the map, and the value was inserted.
86static void InsertNewValueIntoMap(ValueToValueMapTy &VM, Value *K, Value *V) {
87 bool Inserted = VM.insert(KV: {K, V}).second;
88 assert(Inserted);
89 (void)Inserted;
90}
91/// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
92/// old header into the preheader. If there were uses of the values produced by
93/// these instruction that were outside of the loop, we have to insert PHI nodes
94/// to merge the two values. Do this now.
95static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader,
96 BasicBlock *OrigPreheader,
97 ValueToValueMapTy &ValueMap,
98 ScalarEvolution *SE,
99 SmallVectorImpl<PHINode*> *InsertedPHIs) {
100 // Remove PHI node entries that are no longer live.
101 BasicBlock::iterator I, E = OrigHeader->end();
102 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(Val&: I); ++I)
103 PN->removeIncomingValue(BB: OrigPreheader);
104
105 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
106 // as necessary.
107 SSAUpdater SSA(InsertedPHIs);
108 for (I = OrigHeader->begin(); I != E; ++I) {
109 Value *OrigHeaderVal = &*I;
110
111 // If there are no uses of the value (e.g. because it returns void), there
112 // is nothing to rewrite.
113 if (OrigHeaderVal->use_empty())
114 continue;
115
116 Value *OrigPreHeaderVal = ValueMap.lookup(Val: OrigHeaderVal);
117
118 // The value now exits in two versions: the initial value in the preheader
119 // and the loop "next" value in the original header.
120 SSA.Initialize(Ty: OrigHeaderVal->getType(), Name: OrigHeaderVal->getName());
121 // Force re-computation of OrigHeaderVal, as some users now need to use the
122 // new PHI node.
123 if (SE)
124 SE->forgetValue(V: OrigHeaderVal);
125 SSA.AddAvailableValue(BB: OrigHeader, V: OrigHeaderVal);
126 SSA.AddAvailableValue(BB: OrigPreheader, V: OrigPreHeaderVal);
127
128 // Visit each use of the OrigHeader instruction.
129 for (Use &U : llvm::make_early_inc_range(Range: OrigHeaderVal->uses())) {
130 // SSAUpdater can't handle a non-PHI use in the same block as an
131 // earlier def. We can easily handle those cases manually.
132 Instruction *UserInst = cast<Instruction>(Val: U.getUser());
133 if (!isa<PHINode>(Val: UserInst)) {
134 BasicBlock *UserBB = UserInst->getParent();
135
136 // The original users in the OrigHeader are already using the
137 // original definitions.
138 if (UserBB == OrigHeader)
139 continue;
140
141 // Users in the OrigPreHeader need to use the value to which the
142 // original definitions are mapped.
143 if (UserBB == OrigPreheader) {
144 U = OrigPreHeaderVal;
145 continue;
146 }
147 }
148
149 // Anything else can be handled by SSAUpdater.
150 SSA.RewriteUse(U);
151 }
152
153 // Replace MetadataAsValue(ValueAsMetadata(OrigHeaderVal)) uses in debug
154 // intrinsics.
155 SmallVector<DbgVariableRecord *, 1> DbgVariableRecords;
156 llvm::findDbgValues(V: OrigHeaderVal, DbgVariableRecords);
157
158 for (DbgVariableRecord *DVR : DbgVariableRecords) {
159 // The original users in the OrigHeader are already using the original
160 // definitions.
161 BasicBlock *UserBB = DVR->getMarker()->getParent();
162 if (UserBB == OrigHeader)
163 continue;
164
165 // Users in the OrigPreHeader need to use the value to which the
166 // original definitions are mapped and anything else can be handled by
167 // the SSAUpdater. To avoid adding PHINodes, check if the value is
168 // available in UserBB, if not substitute poison.
169 Value *NewVal;
170 if (UserBB == OrigPreheader)
171 NewVal = OrigPreHeaderVal;
172 else if (SSA.HasValueForBlock(BB: UserBB))
173 NewVal = SSA.GetValueInMiddleOfBlock(BB: UserBB);
174 else
175 NewVal = PoisonValue::get(T: OrigHeaderVal->getType());
176 DVR->replaceVariableLocationOp(OldValue: OrigHeaderVal, NewValue: NewVal);
177 }
178 }
179}
180
181// Assuming both header and latch are exiting, look for a phi which is only
182// used outside the loop (via a LCSSA phi) in the exit from the header.
183// This means that rotating the loop can remove the phi.
184static bool profitableToRotateLoopExitingLatch(Loop *L) {
185 BasicBlock *Header = L->getHeader();
186 BranchInst *BI = dyn_cast<BranchInst>(Val: Header->getTerminator());
187 assert(BI && BI->isConditional() && "need header with conditional exit");
188 BasicBlock *HeaderExit = BI->getSuccessor(i: 0);
189 if (L->contains(BB: HeaderExit))
190 HeaderExit = BI->getSuccessor(i: 1);
191
192 for (auto &Phi : Header->phis()) {
193 // Look for uses of this phi in the loop/via exits other than the header.
194 if (llvm::any_of(Range: Phi.users(), P: [HeaderExit](const User *U) {
195 return cast<Instruction>(Val: U)->getParent() != HeaderExit;
196 }))
197 continue;
198 return true;
199 }
200 return false;
201}
202
203static void updateBranchWeights(BranchInst &PreHeaderBI, BranchInst &LoopBI,
204 bool HasConditionalPreHeader,
205 bool SuccsSwapped) {
206 MDNode *WeightMD = getBranchWeightMDNode(I: PreHeaderBI);
207 if (WeightMD == nullptr)
208 return;
209
210 // LoopBI should currently be a clone of PreHeaderBI with the same
211 // metadata. But we double check to make sure we don't have a degenerate case
212 // where instsimplify changed the instructions.
213 if (WeightMD != getBranchWeightMDNode(I: LoopBI))
214 return;
215
216 SmallVector<uint32_t, 2> Weights;
217 extractFromBranchWeightMD32(ProfileData: WeightMD, Weights);
218 if (Weights.size() != 2)
219 return;
220 uint32_t OrigLoopExitWeight = Weights[0];
221 uint32_t OrigLoopBackedgeWeight = Weights[1];
222
223 if (SuccsSwapped)
224 std::swap(a&: OrigLoopExitWeight, b&: OrigLoopBackedgeWeight);
225
226 // Update branch weights. Consider the following edge-counts:
227 //
228 // | |-------- |
229 // V V | V
230 // Br i1 ... | Br i1 ...
231 // | | | | |
232 // x| y| | becomes: | y0| |-----
233 // V V | | V V |
234 // Exit Loop | | Loop |
235 // | | | Br i1 ... |
236 // ----- | | | |
237 // x0| x1| y1 | |
238 // V V ----
239 // Exit
240 //
241 // The following must hold:
242 // - x == x0 + x1 # counts to "exit" must stay the same.
243 // - y0 == x - x0 == x1 # how often loop was entered at all.
244 // - y1 == y - y0 # How often loop was repeated (after first iter.).
245 //
246 // We cannot generally deduce how often we had a zero-trip count loop so we
247 // have to make a guess for how to distribute x among the new x0 and x1.
248
249 uint32_t ExitWeight0; // aka x0
250 uint32_t ExitWeight1; // aka x1
251 uint32_t EnterWeight; // aka y0
252 uint32_t LoopBackWeight; // aka y1
253 if (OrigLoopExitWeight > 0 && OrigLoopBackedgeWeight > 0) {
254 ExitWeight0 = 0;
255 if (HasConditionalPreHeader) {
256 // Here we cannot know how many 0-trip count loops we have, so we guess:
257 if (OrigLoopBackedgeWeight >= OrigLoopExitWeight) {
258 // If the loop count is bigger than the exit count then we set
259 // probabilities as if 0-trip count nearly never happens.
260 ExitWeight0 = ZeroTripCountWeights[0];
261 // Scale up counts if necessary so we can match `ZeroTripCountWeights`
262 // for the `ExitWeight0`:`ExitWeight1` (aka `x0`:`x1` ratio`) ratio.
263 while (OrigLoopExitWeight < ZeroTripCountWeights[1] + ExitWeight0) {
264 // ... but don't overflow.
265 uint32_t const HighBit = uint32_t{1} << (sizeof(uint32_t) * 8 - 1);
266 if ((OrigLoopBackedgeWeight & HighBit) != 0 ||
267 (OrigLoopExitWeight & HighBit) != 0)
268 break;
269 OrigLoopBackedgeWeight <<= 1;
270 OrigLoopExitWeight <<= 1;
271 }
272 } else {
273 // If there's a higher exit-count than backedge-count then we set
274 // probabilities as if there are only 0-trip and 1-trip cases.
275 ExitWeight0 = OrigLoopExitWeight - OrigLoopBackedgeWeight;
276 }
277 } else {
278 // Theoretically, if the loop body must be executed at least once, the
279 // backedge count must be not less than exit count. However the branch
280 // weight collected by sampling-based PGO may be not very accurate due to
281 // sampling. Therefore this workaround is required here to avoid underflow
282 // of unsigned in following update of branch weight.
283 if (OrigLoopExitWeight > OrigLoopBackedgeWeight)
284 OrigLoopBackedgeWeight = OrigLoopExitWeight;
285 }
286 assert(OrigLoopExitWeight >= ExitWeight0 && "Bad branch weight");
287 ExitWeight1 = OrigLoopExitWeight - ExitWeight0;
288 EnterWeight = ExitWeight1;
289 assert(OrigLoopBackedgeWeight >= EnterWeight && "Bad branch weight");
290 LoopBackWeight = OrigLoopBackedgeWeight - EnterWeight;
291 } else if (OrigLoopExitWeight == 0) {
292 if (OrigLoopBackedgeWeight == 0) {
293 // degenerate case... keep everything zero...
294 ExitWeight0 = 0;
295 ExitWeight1 = 0;
296 EnterWeight = 0;
297 LoopBackWeight = 0;
298 } else {
299 // Special case "LoopExitWeight == 0" weights which behaves like an
300 // endless where we don't want loop-enttry (y0) to be the same as
301 // loop-exit (x1).
302 ExitWeight0 = 0;
303 ExitWeight1 = 0;
304 EnterWeight = 1;
305 LoopBackWeight = OrigLoopBackedgeWeight;
306 }
307 } else {
308 // loop is never entered.
309 assert(OrigLoopBackedgeWeight == 0 && "remaining case is backedge zero");
310 ExitWeight0 = 1;
311 ExitWeight1 = 1;
312 EnterWeight = 0;
313 LoopBackWeight = 0;
314 }
315
316 const uint32_t LoopBIWeights[] = {
317 SuccsSwapped ? LoopBackWeight : ExitWeight1,
318 SuccsSwapped ? ExitWeight1 : LoopBackWeight,
319 };
320 setBranchWeights(I&: LoopBI, Weights: LoopBIWeights, /*IsExpected=*/false);
321 if (HasConditionalPreHeader) {
322 const uint32_t PreHeaderBIWeights[] = {
323 SuccsSwapped ? EnterWeight : ExitWeight0,
324 SuccsSwapped ? ExitWeight0 : EnterWeight,
325 };
326 setBranchWeights(I&: PreHeaderBI, Weights: PreHeaderBIWeights, /*IsExpected=*/false);
327 }
328}
329
330/// Rotate loop LP. Return true if the loop is rotated.
331///
332/// \param SimplifiedLatch is true if the latch was just folded into the final
333/// loop exit. In this case we may want to rotate even though the new latch is
334/// now an exiting branch. This rotation would have happened had the latch not
335/// been simplified. However, if SimplifiedLatch is false, then we avoid
336/// rotating loops in which the latch exits to avoid excessive or endless
337/// rotation. LoopRotate should be repeatable and converge to a canonical
338/// form. This property is satisfied because simplifying the loop latch can only
339/// happen once across multiple invocations of the LoopRotate pass.
340bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
341 // If the loop has only one block then there is not much to rotate.
342 if (L->getBlocks().size() == 1)
343 return false;
344
345 bool Rotated = false;
346 BasicBlock *OrigHeader = L->getHeader();
347 BasicBlock *OrigLatch = L->getLoopLatch();
348
349 BranchInst *BI = dyn_cast<BranchInst>(Val: OrigHeader->getTerminator());
350 if (!BI || BI->isUnconditional())
351 return Rotated;
352
353 // If the loop header is not one of the loop exiting blocks then
354 // either this loop is already rotated or it is not
355 // suitable for loop rotation transformations.
356 if (!L->isLoopExiting(BB: OrigHeader))
357 return Rotated;
358
359 // If the loop latch already contains a branch that leaves the loop then the
360 // loop is already rotated.
361 if (!OrigLatch)
362 return Rotated;
363
364 // Rotate if the loop latch was just simplified. Or if it makes the loop exit
365 // count computable. Or if we think it will be profitable.
366 if (L->isLoopExiting(BB: OrigLatch) && !SimplifiedLatch && IsUtilMode == false &&
367 !profitableToRotateLoopExitingLatch(L))
368 return Rotated;
369
370 // Check size of original header and reject loop if it is very big or we can't
371 // duplicate blocks inside it.
372 {
373 SmallPtrSet<const Value *, 32> EphValues;
374 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
375
376 CodeMetrics Metrics;
377 Metrics.analyzeBasicBlock(BB: OrigHeader, TTI: *TTI, EphValues, PrepareForLTO);
378 if (Metrics.notDuplicatable) {
379 LLVM_DEBUG(
380 dbgs() << "LoopRotation: NOT rotating - contains non-duplicatable"
381 << " instructions: ";
382 L->dump());
383 return Rotated;
384 }
385 if (Metrics.Convergence != ConvergenceKind::None) {
386 LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains convergent "
387 "instructions: ";
388 L->dump());
389 return Rotated;
390 }
391 if (!Metrics.NumInsts.isValid()) {
392 LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains instructions"
393 " with invalid cost: ";
394 L->dump());
395 return Rotated;
396 }
397 if (Metrics.NumInsts > MaxHeaderSize) {
398 LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains "
399 << Metrics.NumInsts
400 << " instructions, which is more than the threshold ("
401 << MaxHeaderSize << " instructions): ";
402 L->dump());
403 ++NumNotRotatedDueToHeaderSize;
404 return Rotated;
405 }
406
407 // When preparing for LTO, avoid rotating loops with calls that could be
408 // inlined during the LTO stage.
409 if (PrepareForLTO && Metrics.NumInlineCandidates > 0)
410 return Rotated;
411 }
412
413 // Now, this loop is suitable for rotation.
414 BasicBlock *OrigPreheader = L->getLoopPreheader();
415
416 // If the loop could not be converted to canonical form, it must have an
417 // indirectbr in it, just give up.
418 if (!OrigPreheader || !L->hasDedicatedExits())
419 return Rotated;
420
421 // Anything ScalarEvolution may know about this loop or the PHI nodes
422 // in its header will soon be invalidated. We should also invalidate
423 // all outer loops because insertion and deletion of blocks that happens
424 // during the rotation may violate invariants related to backedge taken
425 // infos in them.
426 if (SE) {
427 SE->forgetTopmostLoop(L);
428 // We may hoist some instructions out of loop. In case if they were cached
429 // as "loop variant" or "loop computable", these caches must be dropped.
430 // We also may fold basic blocks, so cached block dispositions also need
431 // to be dropped.
432 SE->forgetBlockAndLoopDispositions();
433 }
434
435 LLVM_DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
436 if (MSSAU && VerifyMemorySSA)
437 MSSAU->getMemorySSA()->verifyMemorySSA();
438
439 // Find new Loop header. NewHeader is a Header's one and only successor
440 // that is inside loop. Header's other successor is outside the
441 // loop. Otherwise loop is not suitable for rotation.
442 BasicBlock *Exit = BI->getSuccessor(i: 0);
443 BasicBlock *NewHeader = BI->getSuccessor(i: 1);
444 bool BISuccsSwapped = L->contains(BB: Exit);
445 if (BISuccsSwapped)
446 std::swap(a&: Exit, b&: NewHeader);
447 assert(NewHeader && "Unable to determine new loop header");
448 assert(L->contains(NewHeader) && !L->contains(Exit) &&
449 "Unable to determine loop header and exit blocks");
450
451 // This code assumes that the new header has exactly one predecessor.
452 // Remove any single-entry PHI nodes in it.
453 assert(NewHeader->getSinglePredecessor() &&
454 "New header doesn't have one pred!");
455 FoldSingleEntryPHINodes(BB: NewHeader);
456
457 // Begin by walking OrigHeader and populating ValueMap with an entry for
458 // each Instruction.
459 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
460 ValueToValueMapTy ValueMap, ValueMapMSSA;
461
462 // For PHI nodes, the value available in OldPreHeader is just the
463 // incoming value from OldPreHeader.
464 for (; PHINode *PN = dyn_cast<PHINode>(Val&: I); ++I)
465 InsertNewValueIntoMap(VM&: ValueMap, K: PN,
466 V: PN->getIncomingValueForBlock(BB: OrigPreheader));
467
468 // For the rest of the instructions, either hoist to the OrigPreheader if
469 // possible or create a clone in the OldPreHeader if not.
470 Instruction *LoopEntryBranch = OrigPreheader->getTerminator();
471
472 // Record all debug records preceding LoopEntryBranch to avoid
473 // duplication.
474 using DbgHash =
475 std::pair<std::pair<hash_code, DILocalVariable *>, DIExpression *>;
476 auto makeHash = [](const DbgVariableRecord *D) -> DbgHash {
477 auto VarLocOps = D->location_ops();
478 return {{hash_combine_range(R&: VarLocOps), D->getVariable()},
479 D->getExpression()};
480 };
481
482 SmallDenseSet<DbgHash, 8> DbgRecords;
483 // Build DbgVariableRecord hashes for DbgVariableRecords attached to the
484 // terminator.
485 for (const DbgVariableRecord &DVR :
486 filterDbgVars(R: OrigPreheader->getTerminator()->getDbgRecordRange()))
487 DbgRecords.insert(V: makeHash(&DVR));
488
489 // Remember the local noalias scope declarations in the header. After the
490 // rotation, they must be duplicated and the scope must be cloned. This
491 // avoids unwanted interaction across iterations.
492 SmallVector<NoAliasScopeDeclInst *, 6> NoAliasDeclInstructions;
493 for (Instruction &I : *OrigHeader)
494 if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(Val: &I))
495 NoAliasDeclInstructions.push_back(Elt: Decl);
496
497 Module *M = OrigHeader->getModule();
498
499 // Track the next DbgRecord to clone. If we have a sequence where an
500 // instruction is hoisted instead of being cloned:
501 // DbgRecord blah
502 // %foo = add i32 0, 0
503 // DbgRecord xyzzy
504 // %bar = call i32 @foobar()
505 // where %foo is hoisted, then the DbgRecord "blah" will be seen twice, once
506 // attached to %foo, then when %foo his hoisted it will "fall down" onto the
507 // function call:
508 // DbgRecord blah
509 // DbgRecord xyzzy
510 // %bar = call i32 @foobar()
511 // causing it to appear attached to the call too.
512 //
513 // To avoid this, cloneDebugInfoFrom takes an optional "start cloning from
514 // here" position to account for this behaviour. We point it at any
515 // DbgRecords on the next instruction, here labelled xyzzy, before we hoist
516 // %foo. Later, we only only clone DbgRecords from that position (xyzzy)
517 // onwards, which avoids cloning DbgRecord "blah" multiple times. (Stored as
518 // a range because it gives us a natural way of testing whether
519 // there were DbgRecords on the next instruction before we hoisted things).
520 iterator_range<DbgRecord::self_iterator> NextDbgInsts =
521 (I != E) ? I->getDbgRecordRange() : DbgMarker::getEmptyDbgRecordRange();
522
523 while (I != E) {
524 Instruction *Inst = &*I++;
525
526 // If the instruction's operands are invariant and it doesn't read or write
527 // memory, then it is safe to hoist. Doing this doesn't change the order of
528 // execution in the preheader, but does prevent the instruction from
529 // executing in each iteration of the loop. This means it is safe to hoist
530 // something that might trap, but isn't safe to hoist something that reads
531 // memory (without proving that the loop doesn't write).
532 if (L->hasLoopInvariantOperands(I: Inst) && !Inst->mayReadFromMemory() &&
533 !Inst->mayWriteToMemory() && !Inst->isTerminator() &&
534 !isa<AllocaInst>(Val: Inst) &&
535 // It is not safe to hoist the value of these instructions in
536 // coroutines, as the addresses of otherwise eligible variables (e.g.
537 // thread-local variables and errno) may change if the coroutine is
538 // resumed in a different thread.Therefore, we disable this
539 // optimization for correctness. However, this may block other correct
540 // optimizations.
541 // FIXME: This should be reverted once we have a better model for
542 // memory access in coroutines.
543 !Inst->getFunction()->isPresplitCoroutine()) {
544
545 if (!NextDbgInsts.empty()) {
546 auto DbgValueRange =
547 LoopEntryBranch->cloneDebugInfoFrom(From: Inst, FromHere: NextDbgInsts.begin());
548 RemapDbgRecordRange(M, Range: DbgValueRange, VM&: ValueMap,
549 Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
550 // Erase anything we've seen before.
551 for (DbgVariableRecord &DVR :
552 make_early_inc_range(Range: filterDbgVars(R: DbgValueRange)))
553 if (DbgRecords.count(V: makeHash(&DVR)))
554 DVR.eraseFromParent();
555 }
556
557 NextDbgInsts = I->getDbgRecordRange();
558
559 Inst->moveBefore(InsertPos: LoopEntryBranch->getIterator());
560
561 ++NumInstrsHoisted;
562 continue;
563 }
564
565 // Otherwise, create a duplicate of the instruction.
566 Instruction *C = Inst->clone();
567 if (const DebugLoc &DL = C->getDebugLoc())
568 mapAtomInstance(DL, VMap&: ValueMap);
569
570 C->insertBefore(InsertPos: LoopEntryBranch->getIterator());
571
572 ++NumInstrsDuplicated;
573
574 if (!NextDbgInsts.empty()) {
575 auto Range = C->cloneDebugInfoFrom(From: Inst, FromHere: NextDbgInsts.begin());
576 RemapDbgRecordRange(M, Range, VM&: ValueMap,
577 Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
578 NextDbgInsts = DbgMarker::getEmptyDbgRecordRange();
579 // Erase anything we've seen before.
580 for (DbgVariableRecord &DVR : make_early_inc_range(Range: filterDbgVars(R: Range)))
581 if (DbgRecords.count(V: makeHash(&DVR)))
582 DVR.eraseFromParent();
583 }
584
585 // Eagerly remap the operands of the instruction.
586 RemapInstruction(I: C, VM&: ValueMap,
587 Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
588
589 // With the operands remapped, see if the instruction constant folds or is
590 // otherwise simplifyable. This commonly occurs because the entry from PHI
591 // nodes allows icmps and other instructions to fold.
592 Value *V = simplifyInstruction(I: C, Q: SQ);
593 if (V && LI->replacementPreservesLCSSAForm(From: C, To: V)) {
594 // If so, then delete the temporary instruction and stick the folded value
595 // in the map.
596 InsertNewValueIntoMap(VM&: ValueMap, K: Inst, V);
597 if (!C->mayHaveSideEffects()) {
598 C->eraseFromParent();
599 C = nullptr;
600 }
601 } else {
602 InsertNewValueIntoMap(VM&: ValueMap, K: Inst, V: C);
603 }
604 if (C) {
605 // Otherwise, stick the new instruction into the new block!
606 C->setName(Inst->getName());
607
608 if (auto *II = dyn_cast<AssumeInst>(Val: C))
609 AC->registerAssumption(CI: II);
610 // MemorySSA cares whether the cloned instruction was inserted or not, and
611 // not whether it can be remapped to a simplified value.
612 if (MSSAU)
613 InsertNewValueIntoMap(VM&: ValueMapMSSA, K: Inst, V: C);
614 }
615 }
616
617 if (!NoAliasDeclInstructions.empty()) {
618 // There are noalias scope declarations:
619 // (general):
620 // Original: OrigPre { OrigHeader NewHeader ... Latch }
621 // after: (OrigPre+OrigHeader') { NewHeader ... Latch OrigHeader }
622 //
623 // with D: llvm.experimental.noalias.scope.decl,
624 // U: !noalias or !alias.scope depending on D
625 // ... { D U1 U2 } can transform into:
626 // (0) : ... { D U1 U2 } // no relevant rotation for this part
627 // (1) : ... D' { U1 U2 D } // D is part of OrigHeader
628 // (2) : ... D' U1' { U2 D U1 } // D, U1 are part of OrigHeader
629 //
630 // We now want to transform:
631 // (1) -> : ... D' { D U1 U2 D'' }
632 // (2) -> : ... D' U1' { D U2 D'' U1'' }
633 // D: original llvm.experimental.noalias.scope.decl
634 // D', U1': duplicate with replaced scopes
635 // D'', U1'': different duplicate with replaced scopes
636 // This ensures a safe fallback to 'may_alias' introduced by the rotate,
637 // as U1'' and U1' scopes will not be compatible wrt to the local restrict
638
639 // Clone the llvm.experimental.noalias.decl again for the NewHeader.
640 BasicBlock::iterator NewHeaderInsertionPoint =
641 NewHeader->getFirstNonPHIIt();
642 for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions) {
643 LLVM_DEBUG(dbgs() << " Cloning llvm.experimental.noalias.scope.decl:"
644 << *NAD << "\n");
645 Instruction *NewNAD = NAD->clone();
646 NewNAD->insertBefore(BB&: *NewHeader, InsertPos: NewHeaderInsertionPoint);
647 }
648
649 // Scopes must now be duplicated, once for OrigHeader and once for
650 // OrigPreHeader'.
651 {
652 auto &Context = NewHeader->getContext();
653
654 SmallVector<MDNode *, 8> NoAliasDeclScopes;
655 for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions)
656 NoAliasDeclScopes.push_back(Elt: NAD->getScopeList());
657
658 LLVM_DEBUG(dbgs() << " Updating OrigHeader scopes\n");
659 cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, NewBlocks: {OrigHeader}, Context,
660 Ext: "h.rot");
661 LLVM_DEBUG(OrigHeader->dump());
662
663 // Keep the compile time impact low by only adapting the inserted block
664 // of instructions in the OrigPreHeader. This might result in slightly
665 // more aliasing between these instructions and those that were already
666 // present, but it will be much faster when the original PreHeader is
667 // large.
668 LLVM_DEBUG(dbgs() << " Updating part of OrigPreheader scopes\n");
669 auto *FirstDecl =
670 cast<Instruction>(Val&: ValueMap[*NoAliasDeclInstructions.begin()]);
671 auto *LastInst = &OrigPreheader->back();
672 cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, IStart: FirstDecl, IEnd: LastInst,
673 Context, Ext: "pre.rot");
674 LLVM_DEBUG(OrigPreheader->dump());
675
676 LLVM_DEBUG(dbgs() << " Updated NewHeader:\n");
677 LLVM_DEBUG(NewHeader->dump());
678 }
679 }
680
681 // Along with all the other instructions, we just cloned OrigHeader's
682 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
683 // successors by duplicating their incoming values for OrigHeader.
684 for (BasicBlock *SuccBB : successors(BB: OrigHeader))
685 for (BasicBlock::iterator BI = SuccBB->begin();
686 PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI)
687 PN->addIncoming(V: PN->getIncomingValueForBlock(BB: OrigHeader), BB: OrigPreheader);
688
689 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
690 // OrigPreHeader's old terminator (the original branch into the loop), and
691 // remove the corresponding incoming values from the PHI nodes in OrigHeader.
692 LoopEntryBranch->eraseFromParent();
693 OrigPreheader->flushTerminatorDbgRecords();
694
695 // Update MemorySSA before the rewrite call below changes the 1:1
696 // instruction:cloned_instruction_or_value mapping.
697 if (MSSAU) {
698 InsertNewValueIntoMap(VM&: ValueMapMSSA, K: OrigHeader, V: OrigPreheader);
699 MSSAU->updateForClonedBlockIntoPred(BB: OrigHeader, P1: OrigPreheader,
700 VM: ValueMapMSSA);
701 }
702
703 SmallVector<PHINode *, 2> InsertedPHIs;
704 // If there were any uses of instructions in the duplicated block outside the
705 // loop, update them, inserting PHI nodes as required
706 RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap, SE,
707 InsertedPHIs: &InsertedPHIs);
708
709 // Attach debug records to the new phis if that phi uses a value that
710 // previously had debug metadata attached. This keeps the debug info
711 // up-to-date in the loop body.
712 if (!InsertedPHIs.empty())
713 insertDebugValuesForPHIs(BB: OrigHeader, InsertedPHIs);
714
715 // NewHeader is now the header of the loop.
716 L->moveToHeader(BB: NewHeader);
717 assert(L->getHeader() == NewHeader && "Latch block is our new header");
718
719 // Inform DT about changes to the CFG.
720 if (DT) {
721 // The OrigPreheader branches to the NewHeader and Exit now. Then, inform
722 // the DT about the removed edge to the OrigHeader (that got removed).
723 SmallVector<DominatorTree::UpdateType, 3> Updates = {
724 {DominatorTree::Insert, OrigPreheader, Exit},
725 {DominatorTree::Insert, OrigPreheader, NewHeader},
726 {DominatorTree::Delete, OrigPreheader, OrigHeader}};
727
728 if (MSSAU) {
729 MSSAU->applyUpdates(Updates, DT&: *DT, /*UpdateDT=*/UpdateDTFirst: true);
730 if (VerifyMemorySSA)
731 MSSAU->getMemorySSA()->verifyMemorySSA();
732 } else {
733 DT->applyUpdates(Updates);
734 }
735 }
736
737 // At this point, we've finished our major CFG changes. As part of cloning
738 // the loop into the preheader we've simplified instructions and the
739 // duplicated conditional branch may now be branching on a constant. If it is
740 // branching on a constant and if that constant means that we enter the loop,
741 // then we fold away the cond branch to an uncond branch. This simplifies the
742 // loop in cases important for nested loops, and it also means we don't have
743 // to split as many edges.
744 BranchInst *PHBI = cast<BranchInst>(Val: OrigPreheader->getTerminator());
745 assert(PHBI->isConditional() && "Should be clone of BI condbr!");
746 const Value *Cond = PHBI->getCondition();
747 const bool HasConditionalPreHeader =
748 !isa<ConstantInt>(Val: Cond) ||
749 PHBI->getSuccessor(i: cast<ConstantInt>(Val: Cond)->isZero()) != NewHeader;
750
751 updateBranchWeights(PreHeaderBI&: *PHBI, LoopBI&: *BI, HasConditionalPreHeader, SuccsSwapped: BISuccsSwapped);
752
753 if (HasConditionalPreHeader) {
754 // The conditional branch can't be folded, handle the general case.
755 // Split edges as necessary to preserve LoopSimplify form.
756
757 // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
758 // thus is not a preheader anymore.
759 // Split the edge to form a real preheader.
760 BasicBlock *NewPH = SplitCriticalEdge(
761 Src: OrigPreheader, Dst: NewHeader,
762 Options: CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
763 NewPH->setName(NewHeader->getName() + ".lr.ph");
764
765 // Preserve canonical loop form, which means that 'Exit' should have only
766 // one predecessor. Note that Exit could be an exit block for multiple
767 // nested loops, causing both of the edges to now be critical and need to
768 // be split.
769 SmallVector<BasicBlock *, 4> ExitPreds(predecessors(BB: Exit));
770 bool SplitLatchEdge = false;
771 for (BasicBlock *ExitPred : ExitPreds) {
772 // We only need to split loop exit edges.
773 Loop *PredLoop = LI->getLoopFor(BB: ExitPred);
774 if (!PredLoop || PredLoop->contains(BB: Exit) ||
775 isa<IndirectBrInst>(Val: ExitPred->getTerminator()))
776 continue;
777 SplitLatchEdge |= L->getLoopLatch() == ExitPred;
778 BasicBlock *ExitSplit = SplitCriticalEdge(
779 Src: ExitPred, Dst: Exit,
780 Options: CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
781 ExitSplit->moveBefore(MovePos: Exit);
782 }
783 assert(SplitLatchEdge &&
784 "Despite splitting all preds, failed to split latch exit?");
785 (void)SplitLatchEdge;
786 } else {
787 // We can fold the conditional branch in the preheader, this makes things
788 // simpler. The first step is to remove the extra edge to the Exit block.
789 Exit->removePredecessor(Pred: OrigPreheader, KeepOneInputPHIs: true /*preserve LCSSA*/);
790 BranchInst *NewBI = BranchInst::Create(IfTrue: NewHeader, InsertBefore: PHBI->getIterator());
791 NewBI->setDebugLoc(PHBI->getDebugLoc());
792 PHBI->eraseFromParent();
793
794 // With our CFG finalized, update DomTree if it is available.
795 if (DT)
796 DT->deleteEdge(From: OrigPreheader, To: Exit);
797
798 // Update MSSA too, if available.
799 if (MSSAU)
800 MSSAU->removeEdge(From: OrigPreheader, To: Exit);
801 }
802
803 assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
804 assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
805
806 if (MSSAU && VerifyMemorySSA)
807 MSSAU->getMemorySSA()->verifyMemorySSA();
808
809 // Now that the CFG and DomTree are in a consistent state again, try to merge
810 // the OrigHeader block into OrigLatch. This will succeed if they are
811 // connected by an unconditional branch. This is just a cleanup so the
812 // emitted code isn't too gross in this common case.
813 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
814 BasicBlock *PredBB = OrigHeader->getUniquePredecessor();
815 bool DidMerge = MergeBlockIntoPredecessor(BB: OrigHeader, DTU: &DTU, LI, MSSAU);
816 if (DidMerge)
817 RemoveRedundantDbgInstrs(BB: PredBB);
818
819 if (MSSAU && VerifyMemorySSA)
820 MSSAU->getMemorySSA()->verifyMemorySSA();
821
822 LLVM_DEBUG(dbgs() << "LoopRotation: into "; L->dump());
823
824 return true;
825}
826
827/// Determine whether the instructions in this range may be safely and cheaply
828/// speculated. This is not an important enough situation to develop complex
829/// heuristics. We handle a single arithmetic instruction along with any type
830/// conversions.
831static bool shouldSpeculateInstrs(BasicBlock::iterator Begin,
832 BasicBlock::iterator End, Loop *L) {
833 bool seenIncrement = false;
834 bool MultiExitLoop = false;
835
836 if (!L->getExitingBlock())
837 MultiExitLoop = true;
838
839 for (BasicBlock::iterator I = Begin; I != End; ++I) {
840
841 if (!isSafeToSpeculativelyExecute(I: &*I))
842 return false;
843
844 switch (I->getOpcode()) {
845 default:
846 return false;
847 case Instruction::GetElementPtr:
848 // GEPs are cheap if all indices are constant.
849 if (!cast<GEPOperator>(Val&: I)->hasAllConstantIndices())
850 return false;
851 // fall-thru to increment case
852 [[fallthrough]];
853 case Instruction::Add:
854 case Instruction::Sub:
855 case Instruction::And:
856 case Instruction::Or:
857 case Instruction::Xor:
858 case Instruction::Shl:
859 case Instruction::LShr:
860 case Instruction::AShr: {
861 Value *IVOpnd =
862 !isa<Constant>(Val: I->getOperand(i: 0))
863 ? I->getOperand(i: 0)
864 : !isa<Constant>(Val: I->getOperand(i: 1)) ? I->getOperand(i: 1) : nullptr;
865 if (!IVOpnd)
866 return false;
867
868 // If increment operand is used outside of the loop, this speculation
869 // could cause extra live range interference.
870 if (MultiExitLoop) {
871 for (User *UseI : IVOpnd->users()) {
872 auto *UserInst = cast<Instruction>(Val: UseI);
873 if (!L->contains(Inst: UserInst))
874 return false;
875 }
876 }
877
878 if (seenIncrement)
879 return false;
880 seenIncrement = true;
881 break;
882 }
883 case Instruction::Trunc:
884 case Instruction::ZExt:
885 case Instruction::SExt:
886 // ignore type conversions
887 break;
888 }
889 }
890 return true;
891}
892
893/// Fold the loop tail into the loop exit by speculating the loop tail
894/// instructions. Typically, this is a single post-increment. In the case of a
895/// simple 2-block loop, hoisting the increment can be much better than
896/// duplicating the entire loop header. In the case of loops with early exits,
897/// rotation will not work anyway, but simplifyLoopLatch will put the loop in
898/// canonical form so downstream passes can handle it.
899///
900/// I don't believe this invalidates SCEV.
901bool LoopRotate::simplifyLoopLatch(Loop *L) {
902 BasicBlock *Latch = L->getLoopLatch();
903 if (!Latch || Latch->hasAddressTaken())
904 return false;
905
906 BranchInst *Jmp = dyn_cast<BranchInst>(Val: Latch->getTerminator());
907 if (!Jmp || !Jmp->isUnconditional())
908 return false;
909
910 BasicBlock *LastExit = Latch->getSinglePredecessor();
911 if (!LastExit || !L->isLoopExiting(BB: LastExit))
912 return false;
913
914 BranchInst *BI = dyn_cast<BranchInst>(Val: LastExit->getTerminator());
915 if (!BI)
916 return false;
917
918 if (!shouldSpeculateInstrs(Begin: Latch->begin(), End: Jmp->getIterator(), L))
919 return false;
920
921 LLVM_DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
922 << LastExit->getName() << "\n");
923
924 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
925 MergeBlockIntoPredecessor(BB: Latch, DTU: &DTU, LI, MSSAU, MemDep: nullptr,
926 /*PredecessorWithTwoSuccessors=*/true);
927
928 if (SE) {
929 // Merging blocks may remove blocks reference in the block disposition cache. Clear the cache.
930 SE->forgetBlockAndLoopDispositions();
931 }
932
933 if (MSSAU && VerifyMemorySSA)
934 MSSAU->getMemorySSA()->verifyMemorySSA();
935
936 return true;
937}
938
939/// Rotate \c L, and return true if any modification was made.
940bool LoopRotate::processLoop(Loop *L) {
941 // Save the loop metadata.
942 MDNode *LoopMD = L->getLoopID();
943
944 bool SimplifiedLatch = false;
945
946 // Simplify the loop latch before attempting to rotate the header
947 // upward. Rotation may not be needed if the loop tail can be folded into the
948 // loop exit.
949 if (!RotationOnly)
950 SimplifiedLatch = simplifyLoopLatch(L);
951
952 bool MadeChange = rotateLoop(L, SimplifiedLatch);
953 assert((!MadeChange || L->isLoopExiting(L->getLoopLatch())) &&
954 "Loop latch should be exiting after loop-rotate.");
955
956 // Restore the loop metadata.
957 // NB! We presume LoopRotation DOESN'T ADD its own metadata.
958 if ((MadeChange || SimplifiedLatch) && LoopMD)
959 L->setLoopID(LoopMD);
960
961 return MadeChange || SimplifiedLatch;
962}
963
964
965/// The utility to convert a loop into a loop with bottom test.
966bool llvm::LoopRotation(Loop *L, LoopInfo *LI, const TargetTransformInfo *TTI,
967 AssumptionCache *AC, DominatorTree *DT,
968 ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
969 const SimplifyQuery &SQ, bool RotationOnly = true,
970 unsigned Threshold = unsigned(-1),
971 bool IsUtilMode = true, bool PrepareForLTO) {
972 LoopRotate LR(Threshold, LI, TTI, AC, DT, SE, MSSAU, SQ, RotationOnly,
973 IsUtilMode, PrepareForLTO);
974 return LR.processLoop(L);
975}
976