1 //===-- LoopUtils.cpp - Loop Utility functions -------------------------===//
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 defines common loop utility functions.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/Transforms/Utils/LoopUtils.h"
14 #include "llvm/ADT/DenseSet.h"
15 #include "llvm/ADT/PriorityWorklist.h"
16 #include "llvm/ADT/ScopeExit.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/BasicAliasAnalysis.h"
22 #include "llvm/Analysis/DomTreeUpdater.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/Analysis/InstSimplifyFolder.h"
25 #include "llvm/Analysis/LoopAccessAnalysis.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/MemorySSA.h"
29 #include "llvm/Analysis/MemorySSAUpdater.h"
30 #include "llvm/Analysis/ScalarEvolution.h"
31 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
32 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
33 #include "llvm/IR/DIBuilder.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/MDBuilder.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/PatternMatch.h"
40 #include "llvm/IR/ProfDataUtils.h"
41 #include "llvm/IR/ValueHandle.h"
42 #include "llvm/InitializePasses.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
48
49 using namespace llvm;
50 using namespace llvm::PatternMatch;
51
52 #define DEBUG_TYPE "loop-utils"
53
54 static const char *LLVMLoopDisableNonforced = "llvm.loop.disable_nonforced";
55 static const char *LLVMLoopDisableLICM = "llvm.licm.disable";
56
formDedicatedExitBlocks(Loop * L,DominatorTree * DT,LoopInfo * LI,MemorySSAUpdater * MSSAU,bool PreserveLCSSA)57 bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI,
58 MemorySSAUpdater *MSSAU,
59 bool PreserveLCSSA) {
60 bool Changed = false;
61
62 // We re-use a vector for the in-loop predecesosrs.
63 SmallVector<BasicBlock *, 4> InLoopPredecessors;
64
65 auto RewriteExit = [&](BasicBlock *BB) {
66 assert(InLoopPredecessors.empty() &&
67 "Must start with an empty predecessors list!");
68 auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); });
69
70 // See if there are any non-loop predecessors of this exit block and
71 // keep track of the in-loop predecessors.
72 bool IsDedicatedExit = true;
73 for (auto *PredBB : predecessors(BB))
74 if (L->contains(PredBB)) {
75 if (isa<IndirectBrInst>(PredBB->getTerminator()))
76 // We cannot rewrite exiting edges from an indirectbr.
77 return false;
78
79 InLoopPredecessors.push_back(PredBB);
80 } else {
81 IsDedicatedExit = false;
82 }
83
84 assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!");
85
86 // Nothing to do if this is already a dedicated exit.
87 if (IsDedicatedExit)
88 return false;
89
90 auto *NewExitBB = SplitBlockPredecessors(
91 BB, InLoopPredecessors, ".loopexit", DT, LI, MSSAU, PreserveLCSSA);
92
93 if (!NewExitBB)
94 LLVM_DEBUG(
95 dbgs() << "WARNING: Can't create a dedicated exit block for loop: "
96 << *L << "\n");
97 else
98 LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
99 << NewExitBB->getName() << "\n");
100 return true;
101 };
102
103 // Walk the exit blocks directly rather than building up a data structure for
104 // them, but only visit each one once.
105 SmallPtrSet<BasicBlock *, 4> Visited;
106 for (auto *BB : L->blocks())
107 for (auto *SuccBB : successors(BB)) {
108 // We're looking for exit blocks so skip in-loop successors.
109 if (L->contains(SuccBB))
110 continue;
111
112 // Visit each exit block exactly once.
113 if (!Visited.insert(SuccBB).second)
114 continue;
115
116 Changed |= RewriteExit(SuccBB);
117 }
118
119 return Changed;
120 }
121
122 /// Returns the instructions that use values defined in the loop.
findDefsUsedOutsideOfLoop(Loop * L)123 SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) {
124 SmallVector<Instruction *, 8> UsedOutside;
125
126 for (auto *Block : L->getBlocks())
127 // FIXME: I believe that this could use copy_if if the Inst reference could
128 // be adapted into a pointer.
129 for (auto &Inst : *Block) {
130 auto Users = Inst.users();
131 if (any_of(Users, [&](User *U) {
132 auto *Use = cast<Instruction>(U);
133 return !L->contains(Use->getParent());
134 }))
135 UsedOutside.push_back(&Inst);
136 }
137
138 return UsedOutside;
139 }
140
getLoopAnalysisUsage(AnalysisUsage & AU)141 void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) {
142 // By definition, all loop passes need the LoopInfo analysis and the
143 // Dominator tree it depends on. Because they all participate in the loop
144 // pass manager, they must also preserve these.
145 AU.addRequired<DominatorTreeWrapperPass>();
146 AU.addPreserved<DominatorTreeWrapperPass>();
147 AU.addRequired<LoopInfoWrapperPass>();
148 AU.addPreserved<LoopInfoWrapperPass>();
149
150 // We must also preserve LoopSimplify and LCSSA. We locally access their IDs
151 // here because users shouldn't directly get them from this header.
152 extern char &LoopSimplifyID;
153 extern char &LCSSAID;
154 AU.addRequiredID(LoopSimplifyID);
155 AU.addPreservedID(LoopSimplifyID);
156 AU.addRequiredID(LCSSAID);
157 AU.addPreservedID(LCSSAID);
158 // This is used in the LPPassManager to perform LCSSA verification on passes
159 // which preserve lcssa form
160 AU.addRequired<LCSSAVerificationPass>();
161 AU.addPreserved<LCSSAVerificationPass>();
162
163 // Loop passes are designed to run inside of a loop pass manager which means
164 // that any function analyses they require must be required by the first loop
165 // pass in the manager (so that it is computed before the loop pass manager
166 // runs) and preserved by all loop pasess in the manager. To make this
167 // reasonably robust, the set needed for most loop passes is maintained here.
168 // If your loop pass requires an analysis not listed here, you will need to
169 // carefully audit the loop pass manager nesting structure that results.
170 AU.addRequired<AAResultsWrapperPass>();
171 AU.addPreserved<AAResultsWrapperPass>();
172 AU.addPreserved<BasicAAWrapperPass>();
173 AU.addPreserved<GlobalsAAWrapperPass>();
174 AU.addPreserved<SCEVAAWrapperPass>();
175 AU.addRequired<ScalarEvolutionWrapperPass>();
176 AU.addPreserved<ScalarEvolutionWrapperPass>();
177 // FIXME: When all loop passes preserve MemorySSA, it can be required and
178 // preserved here instead of the individual handling in each pass.
179 }
180
181 /// Manually defined generic "LoopPass" dependency initialization. This is used
182 /// to initialize the exact set of passes from above in \c
183 /// getLoopAnalysisUsage. It can be used within a loop pass's initialization
184 /// with:
185 ///
186 /// INITIALIZE_PASS_DEPENDENCY(LoopPass)
187 ///
188 /// As-if "LoopPass" were a pass.
initializeLoopPassPass(PassRegistry & Registry)189 void llvm::initializeLoopPassPass(PassRegistry &Registry) {
190 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
191 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
192 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
193 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
194 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
195 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
196 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
197 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
198 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
199 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
200 }
201
202 /// Create MDNode for input string.
createStringMetadata(Loop * TheLoop,StringRef Name,unsigned V)203 static MDNode *createStringMetadata(Loop *TheLoop, StringRef Name, unsigned V) {
204 LLVMContext &Context = TheLoop->getHeader()->getContext();
205 Metadata *MDs[] = {
206 MDString::get(Context, Name),
207 ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Context), V))};
208 return MDNode::get(Context, MDs);
209 }
210
211 /// Set input string into loop metadata by keeping other values intact.
212 /// If the string is already in loop metadata update value if it is
213 /// different.
addStringMetadataToLoop(Loop * TheLoop,const char * StringMD,unsigned V)214 void llvm::addStringMetadataToLoop(Loop *TheLoop, const char *StringMD,
215 unsigned V) {
216 SmallVector<Metadata *, 4> MDs(1);
217 // If the loop already has metadata, retain it.
218 MDNode *LoopID = TheLoop->getLoopID();
219 if (LoopID) {
220 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
221 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
222 // If it is of form key = value, try to parse it.
223 if (Node->getNumOperands() == 2) {
224 MDString *S = dyn_cast<MDString>(Node->getOperand(0));
225 if (S && S->getString() == StringMD) {
226 ConstantInt *IntMD =
227 mdconst::extract_or_null<ConstantInt>(Node->getOperand(1));
228 if (IntMD && IntMD->getSExtValue() == V)
229 // It is already in place. Do nothing.
230 return;
231 // We need to update the value, so just skip it here and it will
232 // be added after copying other existed nodes.
233 continue;
234 }
235 }
236 MDs.push_back(Node);
237 }
238 }
239 // Add new metadata.
240 MDs.push_back(createStringMetadata(TheLoop, StringMD, V));
241 // Replace current metadata node with new one.
242 LLVMContext &Context = TheLoop->getHeader()->getContext();
243 MDNode *NewLoopID = MDNode::get(Context, MDs);
244 // Set operand 0 to refer to the loop id itself.
245 NewLoopID->replaceOperandWith(0, NewLoopID);
246 TheLoop->setLoopID(NewLoopID);
247 }
248
249 std::optional<ElementCount>
getOptionalElementCountLoopAttribute(const Loop * TheLoop)250 llvm::getOptionalElementCountLoopAttribute(const Loop *TheLoop) {
251 std::optional<int> Width =
252 getOptionalIntLoopAttribute(TheLoop, "llvm.loop.vectorize.width");
253
254 if (Width) {
255 std::optional<int> IsScalable = getOptionalIntLoopAttribute(
256 TheLoop, "llvm.loop.vectorize.scalable.enable");
257 return ElementCount::get(*Width, IsScalable.value_or(false));
258 }
259
260 return std::nullopt;
261 }
262
makeFollowupLoopID(MDNode * OrigLoopID,ArrayRef<StringRef> FollowupOptions,const char * InheritOptionsExceptPrefix,bool AlwaysNew)263 std::optional<MDNode *> llvm::makeFollowupLoopID(
264 MDNode *OrigLoopID, ArrayRef<StringRef> FollowupOptions,
265 const char *InheritOptionsExceptPrefix, bool AlwaysNew) {
266 if (!OrigLoopID) {
267 if (AlwaysNew)
268 return nullptr;
269 return std::nullopt;
270 }
271
272 assert(OrigLoopID->getOperand(0) == OrigLoopID);
273
274 bool InheritAllAttrs = !InheritOptionsExceptPrefix;
275 bool InheritSomeAttrs =
276 InheritOptionsExceptPrefix && InheritOptionsExceptPrefix[0] != '\0';
277 SmallVector<Metadata *, 8> MDs;
278 MDs.push_back(nullptr);
279
280 bool Changed = false;
281 if (InheritAllAttrs || InheritSomeAttrs) {
282 for (const MDOperand &Existing : drop_begin(OrigLoopID->operands())) {
283 MDNode *Op = cast<MDNode>(Existing.get());
284
285 auto InheritThisAttribute = [InheritSomeAttrs,
286 InheritOptionsExceptPrefix](MDNode *Op) {
287 if (!InheritSomeAttrs)
288 return false;
289
290 // Skip malformatted attribute metadata nodes.
291 if (Op->getNumOperands() == 0)
292 return true;
293 Metadata *NameMD = Op->getOperand(0).get();
294 if (!isa<MDString>(NameMD))
295 return true;
296 StringRef AttrName = cast<MDString>(NameMD)->getString();
297
298 // Do not inherit excluded attributes.
299 return !AttrName.starts_with(InheritOptionsExceptPrefix);
300 };
301
302 if (InheritThisAttribute(Op))
303 MDs.push_back(Op);
304 else
305 Changed = true;
306 }
307 } else {
308 // Modified if we dropped at least one attribute.
309 Changed = OrigLoopID->getNumOperands() > 1;
310 }
311
312 bool HasAnyFollowup = false;
313 for (StringRef OptionName : FollowupOptions) {
314 MDNode *FollowupNode = findOptionMDForLoopID(OrigLoopID, OptionName);
315 if (!FollowupNode)
316 continue;
317
318 HasAnyFollowup = true;
319 for (const MDOperand &Option : drop_begin(FollowupNode->operands())) {
320 MDs.push_back(Option.get());
321 Changed = true;
322 }
323 }
324
325 // Attributes of the followup loop not specified explicity, so signal to the
326 // transformation pass to add suitable attributes.
327 if (!AlwaysNew && !HasAnyFollowup)
328 return std::nullopt;
329
330 // If no attributes were added or remove, the previous loop Id can be reused.
331 if (!AlwaysNew && !Changed)
332 return OrigLoopID;
333
334 // No attributes is equivalent to having no !llvm.loop metadata at all.
335 if (MDs.size() == 1)
336 return nullptr;
337
338 // Build the new loop ID.
339 MDTuple *FollowupLoopID = MDNode::get(OrigLoopID->getContext(), MDs);
340 FollowupLoopID->replaceOperandWith(0, FollowupLoopID);
341 return FollowupLoopID;
342 }
343
hasDisableAllTransformsHint(const Loop * L)344 bool llvm::hasDisableAllTransformsHint(const Loop *L) {
345 return getBooleanLoopAttribute(L, LLVMLoopDisableNonforced);
346 }
347
hasDisableLICMTransformsHint(const Loop * L)348 bool llvm::hasDisableLICMTransformsHint(const Loop *L) {
349 return getBooleanLoopAttribute(L, LLVMLoopDisableLICM);
350 }
351
hasUnrollTransformation(const Loop * L)352 TransformationMode llvm::hasUnrollTransformation(const Loop *L) {
353 if (getBooleanLoopAttribute(L, "llvm.loop.unroll.disable"))
354 return TM_SuppressedByUser;
355
356 std::optional<int> Count =
357 getOptionalIntLoopAttribute(L, "llvm.loop.unroll.count");
358 if (Count)
359 return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
360
361 if (getBooleanLoopAttribute(L, "llvm.loop.unroll.enable"))
362 return TM_ForcedByUser;
363
364 if (getBooleanLoopAttribute(L, "llvm.loop.unroll.full"))
365 return TM_ForcedByUser;
366
367 if (hasDisableAllTransformsHint(L))
368 return TM_Disable;
369
370 return TM_Unspecified;
371 }
372
hasUnrollAndJamTransformation(const Loop * L)373 TransformationMode llvm::hasUnrollAndJamTransformation(const Loop *L) {
374 if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.disable"))
375 return TM_SuppressedByUser;
376
377 std::optional<int> Count =
378 getOptionalIntLoopAttribute(L, "llvm.loop.unroll_and_jam.count");
379 if (Count)
380 return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
381
382 if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.enable"))
383 return TM_ForcedByUser;
384
385 if (hasDisableAllTransformsHint(L))
386 return TM_Disable;
387
388 return TM_Unspecified;
389 }
390
hasVectorizeTransformation(const Loop * L)391 TransformationMode llvm::hasVectorizeTransformation(const Loop *L) {
392 std::optional<bool> Enable =
393 getOptionalBoolLoopAttribute(L, "llvm.loop.vectorize.enable");
394
395 if (Enable == false)
396 return TM_SuppressedByUser;
397
398 std::optional<ElementCount> VectorizeWidth =
399 getOptionalElementCountLoopAttribute(L);
400 std::optional<int> InterleaveCount =
401 getOptionalIntLoopAttribute(L, "llvm.loop.interleave.count");
402
403 // 'Forcing' vector width and interleave count to one effectively disables
404 // this tranformation.
405 if (Enable == true && VectorizeWidth && VectorizeWidth->isScalar() &&
406 InterleaveCount == 1)
407 return TM_SuppressedByUser;
408
409 if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized"))
410 return TM_Disable;
411
412 if (Enable == true)
413 return TM_ForcedByUser;
414
415 if ((VectorizeWidth && VectorizeWidth->isScalar()) && InterleaveCount == 1)
416 return TM_Disable;
417
418 if ((VectorizeWidth && VectorizeWidth->isVector()) || InterleaveCount > 1)
419 return TM_Enable;
420
421 if (hasDisableAllTransformsHint(L))
422 return TM_Disable;
423
424 return TM_Unspecified;
425 }
426
hasDistributeTransformation(const Loop * L)427 TransformationMode llvm::hasDistributeTransformation(const Loop *L) {
428 if (getBooleanLoopAttribute(L, "llvm.loop.distribute.enable"))
429 return TM_ForcedByUser;
430
431 if (hasDisableAllTransformsHint(L))
432 return TM_Disable;
433
434 return TM_Unspecified;
435 }
436
hasLICMVersioningTransformation(const Loop * L)437 TransformationMode llvm::hasLICMVersioningTransformation(const Loop *L) {
438 if (getBooleanLoopAttribute(L, "llvm.loop.licm_versioning.disable"))
439 return TM_SuppressedByUser;
440
441 if (hasDisableAllTransformsHint(L))
442 return TM_Disable;
443
444 return TM_Unspecified;
445 }
446
447 /// Does a BFS from a given node to all of its children inside a given loop.
448 /// The returned vector of nodes includes the starting point.
449 SmallVector<DomTreeNode *, 16>
collectChildrenInLoop(DomTreeNode * N,const Loop * CurLoop)450 llvm::collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop) {
451 SmallVector<DomTreeNode *, 16> Worklist;
452 auto AddRegionToWorklist = [&](DomTreeNode *DTN) {
453 // Only include subregions in the top level loop.
454 BasicBlock *BB = DTN->getBlock();
455 if (CurLoop->contains(BB))
456 Worklist.push_back(DTN);
457 };
458
459 AddRegionToWorklist(N);
460
461 for (size_t I = 0; I < Worklist.size(); I++) {
462 for (DomTreeNode *Child : Worklist[I]->children())
463 AddRegionToWorklist(Child);
464 }
465
466 return Worklist;
467 }
468
isAlmostDeadIV(PHINode * PN,BasicBlock * LatchBlock,Value * Cond)469 bool llvm::isAlmostDeadIV(PHINode *PN, BasicBlock *LatchBlock, Value *Cond) {
470 int LatchIdx = PN->getBasicBlockIndex(LatchBlock);
471 assert(LatchIdx != -1 && "LatchBlock is not a case in this PHINode");
472 Value *IncV = PN->getIncomingValue(LatchIdx);
473
474 for (User *U : PN->users())
475 if (U != Cond && U != IncV) return false;
476
477 for (User *U : IncV->users())
478 if (U != Cond && U != PN) return false;
479 return true;
480 }
481
482
deleteDeadLoop(Loop * L,DominatorTree * DT,ScalarEvolution * SE,LoopInfo * LI,MemorySSA * MSSA)483 void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE,
484 LoopInfo *LI, MemorySSA *MSSA) {
485 assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!");
486 auto *Preheader = L->getLoopPreheader();
487 assert(Preheader && "Preheader should exist!");
488
489 std::unique_ptr<MemorySSAUpdater> MSSAU;
490 if (MSSA)
491 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
492
493 // Now that we know the removal is safe, remove the loop by changing the
494 // branch from the preheader to go to the single exit block.
495 //
496 // Because we're deleting a large chunk of code at once, the sequence in which
497 // we remove things is very important to avoid invalidation issues.
498
499 // Tell ScalarEvolution that the loop is deleted. Do this before
500 // deleting the loop so that ScalarEvolution can look at the loop
501 // to determine what it needs to clean up.
502 if (SE) {
503 SE->forgetLoop(L);
504 SE->forgetBlockAndLoopDispositions();
505 }
506
507 Instruction *OldTerm = Preheader->getTerminator();
508 assert(!OldTerm->mayHaveSideEffects() &&
509 "Preheader must end with a side-effect-free terminator");
510 assert(OldTerm->getNumSuccessors() == 1 &&
511 "Preheader must have a single successor");
512 // Connect the preheader to the exit block. Keep the old edge to the header
513 // around to perform the dominator tree update in two separate steps
514 // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge
515 // preheader -> header.
516 //
517 //
518 // 0. Preheader 1. Preheader 2. Preheader
519 // | | | |
520 // V | V |
521 // Header <--\ | Header <--\ | Header <--\
522 // | | | | | | | | | | |
523 // | V | | | V | | | V |
524 // | Body --/ | | Body --/ | | Body --/
525 // V V V V V
526 // Exit Exit Exit
527 //
528 // By doing this is two separate steps we can perform the dominator tree
529 // update without using the batch update API.
530 //
531 // Even when the loop is never executed, we cannot remove the edge from the
532 // source block to the exit block. Consider the case where the unexecuted loop
533 // branches back to an outer loop. If we deleted the loop and removed the edge
534 // coming to this inner loop, this will break the outer loop structure (by
535 // deleting the backedge of the outer loop). If the outer loop is indeed a
536 // non-loop, it will be deleted in a future iteration of loop deletion pass.
537 IRBuilder<> Builder(OldTerm);
538
539 auto *ExitBlock = L->getUniqueExitBlock();
540 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
541 if (ExitBlock) {
542 assert(ExitBlock && "Should have a unique exit block!");
543 assert(L->hasDedicatedExits() && "Loop should have dedicated exits!");
544
545 Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock);
546 // Remove the old branch. The conditional branch becomes a new terminator.
547 OldTerm->eraseFromParent();
548
549 // Rewrite phis in the exit block to get their inputs from the Preheader
550 // instead of the exiting block.
551 for (PHINode &P : ExitBlock->phis()) {
552 // Set the zero'th element of Phi to be from the preheader and remove all
553 // other incoming values. Given the loop has dedicated exits, all other
554 // incoming values must be from the exiting blocks.
555 int PredIndex = 0;
556 P.setIncomingBlock(PredIndex, Preheader);
557 // Removes all incoming values from all other exiting blocks (including
558 // duplicate values from an exiting block).
559 // Nuke all entries except the zero'th entry which is the preheader entry.
560 P.removeIncomingValueIf([](unsigned Idx) { return Idx != 0; },
561 /* DeletePHIIfEmpty */ false);
562
563 assert((P.getNumIncomingValues() == 1 &&
564 P.getIncomingBlock(PredIndex) == Preheader) &&
565 "Should have exactly one value and that's from the preheader!");
566 }
567
568 if (DT) {
569 DTU.applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}});
570 if (MSSA) {
571 MSSAU->applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}},
572 *DT);
573 if (VerifyMemorySSA)
574 MSSA->verifyMemorySSA();
575 }
576 }
577
578 // Disconnect the loop body by branching directly to its exit.
579 Builder.SetInsertPoint(Preheader->getTerminator());
580 Builder.CreateBr(ExitBlock);
581 // Remove the old branch.
582 Preheader->getTerminator()->eraseFromParent();
583 } else {
584 assert(L->hasNoExitBlocks() &&
585 "Loop should have either zero or one exit blocks.");
586
587 Builder.SetInsertPoint(OldTerm);
588 Builder.CreateUnreachable();
589 Preheader->getTerminator()->eraseFromParent();
590 }
591
592 if (DT) {
593 DTU.applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}});
594 if (MSSA) {
595 MSSAU->applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}},
596 *DT);
597 SmallSetVector<BasicBlock *, 8> DeadBlockSet(L->block_begin(),
598 L->block_end());
599 MSSAU->removeBlocks(DeadBlockSet);
600 if (VerifyMemorySSA)
601 MSSA->verifyMemorySSA();
602 }
603 }
604
605 // Use a map to unique and a vector to guarantee deterministic ordering.
606 llvm::SmallDenseSet<DebugVariable, 4> DeadDebugSet;
607 llvm::SmallVector<DbgVariableIntrinsic *, 4> DeadDebugInst;
608 llvm::SmallVector<DbgVariableRecord *, 4> DeadDbgVariableRecords;
609
610 if (ExitBlock) {
611 // Given LCSSA form is satisfied, we should not have users of instructions
612 // within the dead loop outside of the loop. However, LCSSA doesn't take
613 // unreachable uses into account. We handle them here.
614 // We could do it after drop all references (in this case all users in the
615 // loop will be already eliminated and we have less work to do but according
616 // to API doc of User::dropAllReferences only valid operation after dropping
617 // references, is deletion. So let's substitute all usages of
618 // instruction from the loop with poison value of corresponding type first.
619 for (auto *Block : L->blocks())
620 for (Instruction &I : *Block) {
621 auto *Poison = PoisonValue::get(I.getType());
622 for (Use &U : llvm::make_early_inc_range(I.uses())) {
623 if (auto *Usr = dyn_cast<Instruction>(U.getUser()))
624 if (L->contains(Usr->getParent()))
625 continue;
626 // If we have a DT then we can check that uses outside a loop only in
627 // unreachable block.
628 if (DT)
629 assert(!DT->isReachableFromEntry(U) &&
630 "Unexpected user in reachable block");
631 U.set(Poison);
632 }
633
634 // RemoveDIs: do the same as below for DbgVariableRecords.
635 if (Block->IsNewDbgInfoFormat) {
636 for (DbgVariableRecord &DVR : llvm::make_early_inc_range(
637 filterDbgVars(I.getDbgRecordRange()))) {
638 DebugVariable Key(DVR.getVariable(), DVR.getExpression(),
639 DVR.getDebugLoc().get());
640 if (!DeadDebugSet.insert(Key).second)
641 continue;
642 // Unlinks the DVR from it's container, for later insertion.
643 DVR.removeFromParent();
644 DeadDbgVariableRecords.push_back(&DVR);
645 }
646 }
647
648 // For one of each variable encountered, preserve a debug intrinsic (set
649 // to Poison) and transfer it to the loop exit. This terminates any
650 // variable locations that were set during the loop.
651 auto *DVI = dyn_cast<DbgVariableIntrinsic>(&I);
652 if (!DVI)
653 continue;
654 if (!DeadDebugSet.insert(DebugVariable(DVI)).second)
655 continue;
656 DeadDebugInst.push_back(DVI);
657 }
658
659 // After the loop has been deleted all the values defined and modified
660 // inside the loop are going to be unavailable. Values computed in the
661 // loop will have been deleted, automatically causing their debug uses
662 // be be replaced with undef. Loop invariant values will still be available.
663 // Move dbg.values out the loop so that earlier location ranges are still
664 // terminated and loop invariant assignments are preserved.
665 DIBuilder DIB(*ExitBlock->getModule());
666 BasicBlock::iterator InsertDbgValueBefore =
667 ExitBlock->getFirstInsertionPt();
668 assert(InsertDbgValueBefore != ExitBlock->end() &&
669 "There should be a non-PHI instruction in exit block, else these "
670 "instructions will have no parent.");
671
672 for (auto *DVI : DeadDebugInst)
673 DVI->moveBefore(*ExitBlock, InsertDbgValueBefore);
674
675 // Due to the "head" bit in BasicBlock::iterator, we're going to insert
676 // each DbgVariableRecord right at the start of the block, wheras dbg.values
677 // would be repeatedly inserted before the first instruction. To replicate
678 // this behaviour, do it backwards.
679 for (DbgVariableRecord *DVR : llvm::reverse(DeadDbgVariableRecords))
680 ExitBlock->insertDbgRecordBefore(DVR, InsertDbgValueBefore);
681 }
682
683 // Remove the block from the reference counting scheme, so that we can
684 // delete it freely later.
685 for (auto *Block : L->blocks())
686 Block->dropAllReferences();
687
688 if (MSSA && VerifyMemorySSA)
689 MSSA->verifyMemorySSA();
690
691 if (LI) {
692 // Erase the instructions and the blocks without having to worry
693 // about ordering because we already dropped the references.
694 // NOTE: This iteration is safe because erasing the block does not remove
695 // its entry from the loop's block list. We do that in the next section.
696 for (BasicBlock *BB : L->blocks())
697 BB->eraseFromParent();
698
699 // Finally, the blocks from loopinfo. This has to happen late because
700 // otherwise our loop iterators won't work.
701
702 SmallPtrSet<BasicBlock *, 8> blocks;
703 blocks.insert(L->block_begin(), L->block_end());
704 for (BasicBlock *BB : blocks)
705 LI->removeBlock(BB);
706
707 // The last step is to update LoopInfo now that we've eliminated this loop.
708 // Note: LoopInfo::erase remove the given loop and relink its subloops with
709 // its parent. While removeLoop/removeChildLoop remove the given loop but
710 // not relink its subloops, which is what we want.
711 if (Loop *ParentLoop = L->getParentLoop()) {
712 Loop::iterator I = find(*ParentLoop, L);
713 assert(I != ParentLoop->end() && "Couldn't find loop");
714 ParentLoop->removeChildLoop(I);
715 } else {
716 Loop::iterator I = find(*LI, L);
717 assert(I != LI->end() && "Couldn't find loop");
718 LI->removeLoop(I);
719 }
720 LI->destroy(L);
721 }
722 }
723
breakLoopBackedge(Loop * L,DominatorTree & DT,ScalarEvolution & SE,LoopInfo & LI,MemorySSA * MSSA)724 void llvm::breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE,
725 LoopInfo &LI, MemorySSA *MSSA) {
726 auto *Latch = L->getLoopLatch();
727 assert(Latch && "multiple latches not yet supported");
728 auto *Header = L->getHeader();
729 Loop *OutermostLoop = L->getOutermostLoop();
730
731 SE.forgetLoop(L);
732 SE.forgetBlockAndLoopDispositions();
733
734 std::unique_ptr<MemorySSAUpdater> MSSAU;
735 if (MSSA)
736 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
737
738 // Update the CFG and domtree. We chose to special case a couple of
739 // of common cases for code quality and test readability reasons.
740 [&]() -> void {
741 if (auto *BI = dyn_cast<BranchInst>(Latch->getTerminator())) {
742 if (!BI->isConditional()) {
743 DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
744 (void)changeToUnreachable(BI, /*PreserveLCSSA*/ true, &DTU,
745 MSSAU.get());
746 return;
747 }
748
749 // Conditional latch/exit - note that latch can be shared by inner
750 // and outer loop so the other target doesn't need to an exit
751 if (L->isLoopExiting(Latch)) {
752 // TODO: Generalize ConstantFoldTerminator so that it can be used
753 // here without invalidating LCSSA or MemorySSA. (Tricky case for
754 // LCSSA: header is an exit block of a preceeding sibling loop w/o
755 // dedicated exits.)
756 const unsigned ExitIdx = L->contains(BI->getSuccessor(0)) ? 1 : 0;
757 BasicBlock *ExitBB = BI->getSuccessor(ExitIdx);
758
759 DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
760 Header->removePredecessor(Latch, true);
761
762 IRBuilder<> Builder(BI);
763 auto *NewBI = Builder.CreateBr(ExitBB);
764 // Transfer the metadata to the new branch instruction (minus the
765 // loop info since this is no longer a loop)
766 NewBI->copyMetadata(*BI, {LLVMContext::MD_dbg,
767 LLVMContext::MD_annotation});
768
769 BI->eraseFromParent();
770 DTU.applyUpdates({{DominatorTree::Delete, Latch, Header}});
771 if (MSSA)
772 MSSAU->applyUpdates({{DominatorTree::Delete, Latch, Header}}, DT);
773 return;
774 }
775 }
776
777 // General case. By splitting the backedge, and then explicitly making it
778 // unreachable we gracefully handle corner cases such as switch and invoke
779 // termiantors.
780 auto *BackedgeBB = SplitEdge(Latch, Header, &DT, &LI, MSSAU.get());
781
782 DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
783 (void)changeToUnreachable(BackedgeBB->getTerminator(),
784 /*PreserveLCSSA*/ true, &DTU, MSSAU.get());
785 }();
786
787 // Erase (and destroy) this loop instance. Handles relinking sub-loops
788 // and blocks within the loop as needed.
789 LI.erase(L);
790
791 // If the loop we broke had a parent, then changeToUnreachable might have
792 // caused a block to be removed from the parent loop (see loop_nest_lcssa
793 // test case in zero-btc.ll for an example), thus changing the parent's
794 // exit blocks. If that happened, we need to rebuild LCSSA on the outermost
795 // loop which might have a had a block removed.
796 if (OutermostLoop != L)
797 formLCSSARecursively(*OutermostLoop, DT, &LI, &SE);
798 }
799
800
801 /// Checks if \p L has an exiting latch branch. There may also be other
802 /// exiting blocks. Returns branch instruction terminating the loop
803 /// latch if above check is successful, nullptr otherwise.
getExpectedExitLoopLatchBranch(Loop * L)804 static BranchInst *getExpectedExitLoopLatchBranch(Loop *L) {
805 BasicBlock *Latch = L->getLoopLatch();
806 if (!Latch)
807 return nullptr;
808
809 BranchInst *LatchBR = dyn_cast<BranchInst>(Latch->getTerminator());
810 if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(Latch))
811 return nullptr;
812
813 assert((LatchBR->getSuccessor(0) == L->getHeader() ||
814 LatchBR->getSuccessor(1) == L->getHeader()) &&
815 "At least one edge out of the latch must go to the header");
816
817 return LatchBR;
818 }
819
820 /// Return the estimated trip count for any exiting branch which dominates
821 /// the loop latch.
getEstimatedTripCount(BranchInst * ExitingBranch,Loop * L,uint64_t & OrigExitWeight)822 static std::optional<uint64_t> getEstimatedTripCount(BranchInst *ExitingBranch,
823 Loop *L,
824 uint64_t &OrigExitWeight) {
825 // To estimate the number of times the loop body was executed, we want to
826 // know the number of times the backedge was taken, vs. the number of times
827 // we exited the loop.
828 uint64_t LoopWeight, ExitWeight;
829 if (!extractBranchWeights(*ExitingBranch, LoopWeight, ExitWeight))
830 return std::nullopt;
831
832 if (L->contains(ExitingBranch->getSuccessor(1)))
833 std::swap(LoopWeight, ExitWeight);
834
835 if (!ExitWeight)
836 // Don't have a way to return predicated infinite
837 return std::nullopt;
838
839 OrigExitWeight = ExitWeight;
840
841 // Estimated exit count is a ratio of the loop weight by the weight of the
842 // edge exiting the loop, rounded to nearest.
843 uint64_t ExitCount = llvm::divideNearest(LoopWeight, ExitWeight);
844 // Estimated trip count is one plus estimated exit count.
845 return ExitCount + 1;
846 }
847
848 std::optional<unsigned>
getLoopEstimatedTripCount(Loop * L,unsigned * EstimatedLoopInvocationWeight)849 llvm::getLoopEstimatedTripCount(Loop *L,
850 unsigned *EstimatedLoopInvocationWeight) {
851 // Currently we take the estimate exit count only from the loop latch,
852 // ignoring other exiting blocks. This can overestimate the trip count
853 // if we exit through another exit, but can never underestimate it.
854 // TODO: incorporate information from other exits
855 if (BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L)) {
856 uint64_t ExitWeight;
857 if (std::optional<uint64_t> EstTripCount =
858 getEstimatedTripCount(LatchBranch, L, ExitWeight)) {
859 if (EstimatedLoopInvocationWeight)
860 *EstimatedLoopInvocationWeight = ExitWeight;
861 return *EstTripCount;
862 }
863 }
864 return std::nullopt;
865 }
866
setLoopEstimatedTripCount(Loop * L,unsigned EstimatedTripCount,unsigned EstimatedloopInvocationWeight)867 bool llvm::setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount,
868 unsigned EstimatedloopInvocationWeight) {
869 // At the moment, we currently support changing the estimate trip count of
870 // the latch branch only. We could extend this API to manipulate estimated
871 // trip counts for any exit.
872 BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L);
873 if (!LatchBranch)
874 return false;
875
876 // Calculate taken and exit weights.
877 unsigned LatchExitWeight = 0;
878 unsigned BackedgeTakenWeight = 0;
879
880 if (EstimatedTripCount > 0) {
881 LatchExitWeight = EstimatedloopInvocationWeight;
882 BackedgeTakenWeight = (EstimatedTripCount - 1) * LatchExitWeight;
883 }
884
885 // Make a swap if back edge is taken when condition is "false".
886 if (LatchBranch->getSuccessor(0) != L->getHeader())
887 std::swap(BackedgeTakenWeight, LatchExitWeight);
888
889 MDBuilder MDB(LatchBranch->getContext());
890
891 // Set/Update profile metadata.
892 LatchBranch->setMetadata(
893 LLVMContext::MD_prof,
894 MDB.createBranchWeights(BackedgeTakenWeight, LatchExitWeight));
895
896 return true;
897 }
898
hasIterationCountInvariantInParent(Loop * InnerLoop,ScalarEvolution & SE)899 bool llvm::hasIterationCountInvariantInParent(Loop *InnerLoop,
900 ScalarEvolution &SE) {
901 Loop *OuterL = InnerLoop->getParentLoop();
902 if (!OuterL)
903 return true;
904
905 // Get the backedge taken count for the inner loop
906 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
907 const SCEV *InnerLoopBECountSC = SE.getExitCount(InnerLoop, InnerLoopLatch);
908 if (isa<SCEVCouldNotCompute>(InnerLoopBECountSC) ||
909 !InnerLoopBECountSC->getType()->isIntegerTy())
910 return false;
911
912 // Get whether count is invariant to the outer loop
913 ScalarEvolution::LoopDisposition LD =
914 SE.getLoopDisposition(InnerLoopBECountSC, OuterL);
915 if (LD != ScalarEvolution::LoopInvariant)
916 return false;
917
918 return true;
919 }
920
getReductionIntrinsicID(RecurKind RK)921 constexpr Intrinsic::ID llvm::getReductionIntrinsicID(RecurKind RK) {
922 switch (RK) {
923 default:
924 llvm_unreachable("Unexpected recurrence kind");
925 case RecurKind::Add:
926 return Intrinsic::vector_reduce_add;
927 case RecurKind::Mul:
928 return Intrinsic::vector_reduce_mul;
929 case RecurKind::And:
930 return Intrinsic::vector_reduce_and;
931 case RecurKind::Or:
932 return Intrinsic::vector_reduce_or;
933 case RecurKind::Xor:
934 return Intrinsic::vector_reduce_xor;
935 case RecurKind::FMulAdd:
936 case RecurKind::FAdd:
937 return Intrinsic::vector_reduce_fadd;
938 case RecurKind::FMul:
939 return Intrinsic::vector_reduce_fmul;
940 case RecurKind::SMax:
941 return Intrinsic::vector_reduce_smax;
942 case RecurKind::SMin:
943 return Intrinsic::vector_reduce_smin;
944 case RecurKind::UMax:
945 return Intrinsic::vector_reduce_umax;
946 case RecurKind::UMin:
947 return Intrinsic::vector_reduce_umin;
948 case RecurKind::FMax:
949 return Intrinsic::vector_reduce_fmax;
950 case RecurKind::FMin:
951 return Intrinsic::vector_reduce_fmin;
952 case RecurKind::FMaximum:
953 return Intrinsic::vector_reduce_fmaximum;
954 case RecurKind::FMinimum:
955 return Intrinsic::vector_reduce_fminimum;
956 }
957 }
958
getArithmeticReductionInstruction(Intrinsic::ID RdxID)959 unsigned llvm::getArithmeticReductionInstruction(Intrinsic::ID RdxID) {
960 switch (RdxID) {
961 case Intrinsic::vector_reduce_fadd:
962 return Instruction::FAdd;
963 case Intrinsic::vector_reduce_fmul:
964 return Instruction::FMul;
965 case Intrinsic::vector_reduce_add:
966 return Instruction::Add;
967 case Intrinsic::vector_reduce_mul:
968 return Instruction::Mul;
969 case Intrinsic::vector_reduce_and:
970 return Instruction::And;
971 case Intrinsic::vector_reduce_or:
972 return Instruction::Or;
973 case Intrinsic::vector_reduce_xor:
974 return Instruction::Xor;
975 case Intrinsic::vector_reduce_smax:
976 case Intrinsic::vector_reduce_smin:
977 case Intrinsic::vector_reduce_umax:
978 case Intrinsic::vector_reduce_umin:
979 return Instruction::ICmp;
980 case Intrinsic::vector_reduce_fmax:
981 case Intrinsic::vector_reduce_fmin:
982 return Instruction::FCmp;
983 default:
984 llvm_unreachable("Unexpected ID");
985 }
986 }
987
getMinMaxReductionIntrinsicOp(Intrinsic::ID RdxID)988 Intrinsic::ID llvm::getMinMaxReductionIntrinsicOp(Intrinsic::ID RdxID) {
989 switch (RdxID) {
990 default:
991 llvm_unreachable("Unknown min/max recurrence kind");
992 case Intrinsic::vector_reduce_umin:
993 return Intrinsic::umin;
994 case Intrinsic::vector_reduce_umax:
995 return Intrinsic::umax;
996 case Intrinsic::vector_reduce_smin:
997 return Intrinsic::smin;
998 case Intrinsic::vector_reduce_smax:
999 return Intrinsic::smax;
1000 case Intrinsic::vector_reduce_fmin:
1001 return Intrinsic::minnum;
1002 case Intrinsic::vector_reduce_fmax:
1003 return Intrinsic::maxnum;
1004 case Intrinsic::vector_reduce_fminimum:
1005 return Intrinsic::minimum;
1006 case Intrinsic::vector_reduce_fmaximum:
1007 return Intrinsic::maximum;
1008 }
1009 }
1010
getMinMaxReductionIntrinsicOp(RecurKind RK)1011 Intrinsic::ID llvm::getMinMaxReductionIntrinsicOp(RecurKind RK) {
1012 switch (RK) {
1013 default:
1014 llvm_unreachable("Unknown min/max recurrence kind");
1015 case RecurKind::UMin:
1016 return Intrinsic::umin;
1017 case RecurKind::UMax:
1018 return Intrinsic::umax;
1019 case RecurKind::SMin:
1020 return Intrinsic::smin;
1021 case RecurKind::SMax:
1022 return Intrinsic::smax;
1023 case RecurKind::FMin:
1024 return Intrinsic::minnum;
1025 case RecurKind::FMax:
1026 return Intrinsic::maxnum;
1027 case RecurKind::FMinimum:
1028 return Intrinsic::minimum;
1029 case RecurKind::FMaximum:
1030 return Intrinsic::maximum;
1031 }
1032 }
1033
getMinMaxReductionRecurKind(Intrinsic::ID RdxID)1034 RecurKind llvm::getMinMaxReductionRecurKind(Intrinsic::ID RdxID) {
1035 switch (RdxID) {
1036 case Intrinsic::vector_reduce_smax:
1037 return RecurKind::SMax;
1038 case Intrinsic::vector_reduce_smin:
1039 return RecurKind::SMin;
1040 case Intrinsic::vector_reduce_umax:
1041 return RecurKind::UMax;
1042 case Intrinsic::vector_reduce_umin:
1043 return RecurKind::UMin;
1044 case Intrinsic::vector_reduce_fmax:
1045 return RecurKind::FMax;
1046 case Intrinsic::vector_reduce_fmin:
1047 return RecurKind::FMin;
1048 default:
1049 return RecurKind::None;
1050 }
1051 }
1052
getMinMaxReductionPredicate(RecurKind RK)1053 CmpInst::Predicate llvm::getMinMaxReductionPredicate(RecurKind RK) {
1054 switch (RK) {
1055 default:
1056 llvm_unreachable("Unknown min/max recurrence kind");
1057 case RecurKind::UMin:
1058 return CmpInst::ICMP_ULT;
1059 case RecurKind::UMax:
1060 return CmpInst::ICMP_UGT;
1061 case RecurKind::SMin:
1062 return CmpInst::ICMP_SLT;
1063 case RecurKind::SMax:
1064 return CmpInst::ICMP_SGT;
1065 case RecurKind::FMin:
1066 return CmpInst::FCMP_OLT;
1067 case RecurKind::FMax:
1068 return CmpInst::FCMP_OGT;
1069 // We do not add FMinimum/FMaximum recurrence kind here since there is no
1070 // equivalent predicate which compares signed zeroes according to the
1071 // semantics of the intrinsics (llvm.minimum/maximum).
1072 }
1073 }
1074
createMinMaxOp(IRBuilderBase & Builder,RecurKind RK,Value * Left,Value * Right)1075 Value *llvm::createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left,
1076 Value *Right) {
1077 Type *Ty = Left->getType();
1078 if (Ty->isIntOrIntVectorTy() ||
1079 (RK == RecurKind::FMinimum || RK == RecurKind::FMaximum)) {
1080 // TODO: Add float minnum/maxnum support when FMF nnan is set.
1081 Intrinsic::ID Id = getMinMaxReductionIntrinsicOp(RK);
1082 return Builder.CreateIntrinsic(Ty, Id, {Left, Right}, nullptr,
1083 "rdx.minmax");
1084 }
1085 CmpInst::Predicate Pred = getMinMaxReductionPredicate(RK);
1086 Value *Cmp = Builder.CreateCmp(Pred, Left, Right, "rdx.minmax.cmp");
1087 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1088 return Select;
1089 }
1090
1091 // Helper to generate an ordered reduction.
getOrderedReduction(IRBuilderBase & Builder,Value * Acc,Value * Src,unsigned Op,RecurKind RdxKind)1092 Value *llvm::getOrderedReduction(IRBuilderBase &Builder, Value *Acc, Value *Src,
1093 unsigned Op, RecurKind RdxKind) {
1094 unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
1095
1096 // Extract and apply reduction ops in ascending order:
1097 // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1]
1098 Value *Result = Acc;
1099 for (unsigned ExtractIdx = 0; ExtractIdx != VF; ++ExtractIdx) {
1100 Value *Ext =
1101 Builder.CreateExtractElement(Src, Builder.getInt32(ExtractIdx));
1102
1103 if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
1104 Result = Builder.CreateBinOp((Instruction::BinaryOps)Op, Result, Ext,
1105 "bin.rdx");
1106 } else {
1107 assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
1108 "Invalid min/max");
1109 Result = createMinMaxOp(Builder, RdxKind, Result, Ext);
1110 }
1111 }
1112
1113 return Result;
1114 }
1115
1116 // Helper to generate a log2 shuffle reduction.
getShuffleReduction(IRBuilderBase & Builder,Value * Src,unsigned Op,TargetTransformInfo::ReductionShuffle RS,RecurKind RdxKind)1117 Value *llvm::getShuffleReduction(IRBuilderBase &Builder, Value *Src,
1118 unsigned Op,
1119 TargetTransformInfo::ReductionShuffle RS,
1120 RecurKind RdxKind) {
1121 unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
1122 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1123 // and vector ops, reducing the set of values being computed by half each
1124 // round.
1125 assert(isPowerOf2_32(VF) &&
1126 "Reduction emission only supported for pow2 vectors!");
1127 // Note: fast-math-flags flags are controlled by the builder configuration
1128 // and are assumed to apply to all generated arithmetic instructions. Other
1129 // poison generating flags (nsw/nuw/inbounds/inrange/exact) are not part
1130 // of the builder configuration, and since they're not passed explicitly,
1131 // will never be relevant here. Note that it would be generally unsound to
1132 // propagate these from an intrinsic call to the expansion anyways as we/
1133 // change the order of operations.
1134 auto BuildShuffledOp = [&Builder, &Op,
1135 &RdxKind](SmallVectorImpl<int> &ShuffleMask,
1136 Value *&TmpVec) -> void {
1137 Value *Shuf = Builder.CreateShuffleVector(TmpVec, ShuffleMask, "rdx.shuf");
1138 if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
1139 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
1140 "bin.rdx");
1141 } else {
1142 assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
1143 "Invalid min/max");
1144 TmpVec = createMinMaxOp(Builder, RdxKind, TmpVec, Shuf);
1145 }
1146 };
1147
1148 Value *TmpVec = Src;
1149 if (TargetTransformInfo::ReductionShuffle::Pairwise == RS) {
1150 SmallVector<int, 32> ShuffleMask(VF);
1151 for (unsigned stride = 1; stride < VF; stride <<= 1) {
1152 // Initialise the mask with undef.
1153 std::fill(ShuffleMask.begin(), ShuffleMask.end(), -1);
1154 for (unsigned j = 0; j < VF; j += stride << 1) {
1155 ShuffleMask[j] = j + stride;
1156 }
1157 BuildShuffledOp(ShuffleMask, TmpVec);
1158 }
1159 } else {
1160 SmallVector<int, 32> ShuffleMask(VF);
1161 for (unsigned i = VF; i != 1; i >>= 1) {
1162 // Move the upper half of the vector to the lower half.
1163 for (unsigned j = 0; j != i / 2; ++j)
1164 ShuffleMask[j] = i / 2 + j;
1165
1166 // Fill the rest of the mask with undef.
1167 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(), -1);
1168 BuildShuffledOp(ShuffleMask, TmpVec);
1169 }
1170 }
1171 // The result is in the first element of the vector.
1172 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1173 }
1174
createAnyOfTargetReduction(IRBuilderBase & Builder,Value * Src,const RecurrenceDescriptor & Desc,PHINode * OrigPhi)1175 Value *llvm::createAnyOfTargetReduction(IRBuilderBase &Builder, Value *Src,
1176 const RecurrenceDescriptor &Desc,
1177 PHINode *OrigPhi) {
1178 assert(
1179 RecurrenceDescriptor::isAnyOfRecurrenceKind(Desc.getRecurrenceKind()) &&
1180 "Unexpected reduction kind");
1181 Value *InitVal = Desc.getRecurrenceStartValue();
1182 Value *NewVal = nullptr;
1183
1184 // First use the original phi to determine the new value we're trying to
1185 // select from in the loop.
1186 SelectInst *SI = nullptr;
1187 for (auto *U : OrigPhi->users()) {
1188 if ((SI = dyn_cast<SelectInst>(U)))
1189 break;
1190 }
1191 assert(SI && "One user of the original phi should be a select");
1192
1193 if (SI->getTrueValue() == OrigPhi)
1194 NewVal = SI->getFalseValue();
1195 else {
1196 assert(SI->getFalseValue() == OrigPhi &&
1197 "At least one input to the select should be the original Phi");
1198 NewVal = SI->getTrueValue();
1199 }
1200
1201 // If any predicate is true it means that we want to select the new value.
1202 Value *AnyOf =
1203 Src->getType()->isVectorTy() ? Builder.CreateOrReduce(Src) : Src;
1204 // The compares in the loop may yield poison, which propagates through the
1205 // bitwise ORs. Freeze it here before the condition is used.
1206 AnyOf = Builder.CreateFreeze(AnyOf);
1207 return Builder.CreateSelect(AnyOf, NewVal, InitVal, "rdx.select");
1208 }
1209
createSimpleTargetReduction(IRBuilderBase & Builder,Value * Src,RecurKind RdxKind)1210 Value *llvm::createSimpleTargetReduction(IRBuilderBase &Builder, Value *Src,
1211 RecurKind RdxKind) {
1212 auto *SrcVecEltTy = cast<VectorType>(Src->getType())->getElementType();
1213 switch (RdxKind) {
1214 case RecurKind::Add:
1215 return Builder.CreateAddReduce(Src);
1216 case RecurKind::Mul:
1217 return Builder.CreateMulReduce(Src);
1218 case RecurKind::And:
1219 return Builder.CreateAndReduce(Src);
1220 case RecurKind::Or:
1221 return Builder.CreateOrReduce(Src);
1222 case RecurKind::Xor:
1223 return Builder.CreateXorReduce(Src);
1224 case RecurKind::FMulAdd:
1225 case RecurKind::FAdd:
1226 return Builder.CreateFAddReduce(ConstantFP::getNegativeZero(SrcVecEltTy),
1227 Src);
1228 case RecurKind::FMul:
1229 return Builder.CreateFMulReduce(ConstantFP::get(SrcVecEltTy, 1.0), Src);
1230 case RecurKind::SMax:
1231 return Builder.CreateIntMaxReduce(Src, true);
1232 case RecurKind::SMin:
1233 return Builder.CreateIntMinReduce(Src, true);
1234 case RecurKind::UMax:
1235 return Builder.CreateIntMaxReduce(Src, false);
1236 case RecurKind::UMin:
1237 return Builder.CreateIntMinReduce(Src, false);
1238 case RecurKind::FMax:
1239 return Builder.CreateFPMaxReduce(Src);
1240 case RecurKind::FMin:
1241 return Builder.CreateFPMinReduce(Src);
1242 case RecurKind::FMinimum:
1243 return Builder.CreateFPMinimumReduce(Src);
1244 case RecurKind::FMaximum:
1245 return Builder.CreateFPMaximumReduce(Src);
1246 default:
1247 llvm_unreachable("Unhandled opcode");
1248 }
1249 }
1250
createSimpleTargetReduction(VectorBuilder & VBuilder,Value * Src,const RecurrenceDescriptor & Desc)1251 Value *llvm::createSimpleTargetReduction(VectorBuilder &VBuilder, Value *Src,
1252 const RecurrenceDescriptor &Desc) {
1253 RecurKind Kind = Desc.getRecurrenceKind();
1254 assert(!RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) &&
1255 "AnyOf reduction is not supported.");
1256 Intrinsic::ID Id = getReductionIntrinsicID(Kind);
1257 auto *SrcTy = cast<VectorType>(Src->getType());
1258 Type *SrcEltTy = SrcTy->getElementType();
1259 Value *Iden =
1260 Desc.getRecurrenceIdentity(Kind, SrcEltTy, Desc.getFastMathFlags());
1261 Value *Ops[] = {Iden, Src};
1262 return VBuilder.createSimpleTargetReduction(Id, SrcTy, Ops);
1263 }
1264
createTargetReduction(IRBuilderBase & B,const RecurrenceDescriptor & Desc,Value * Src,PHINode * OrigPhi)1265 Value *llvm::createTargetReduction(IRBuilderBase &B,
1266 const RecurrenceDescriptor &Desc, Value *Src,
1267 PHINode *OrigPhi) {
1268 // TODO: Support in-order reductions based on the recurrence descriptor.
1269 // All ops in the reduction inherit fast-math-flags from the recurrence
1270 // descriptor.
1271 IRBuilderBase::FastMathFlagGuard FMFGuard(B);
1272 B.setFastMathFlags(Desc.getFastMathFlags());
1273
1274 RecurKind RK = Desc.getRecurrenceKind();
1275 if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RK))
1276 return createAnyOfTargetReduction(B, Src, Desc, OrigPhi);
1277
1278 return createSimpleTargetReduction(B, Src, RK);
1279 }
1280
createOrderedReduction(IRBuilderBase & B,const RecurrenceDescriptor & Desc,Value * Src,Value * Start)1281 Value *llvm::createOrderedReduction(IRBuilderBase &B,
1282 const RecurrenceDescriptor &Desc,
1283 Value *Src, Value *Start) {
1284 assert((Desc.getRecurrenceKind() == RecurKind::FAdd ||
1285 Desc.getRecurrenceKind() == RecurKind::FMulAdd) &&
1286 "Unexpected reduction kind");
1287 assert(Src->getType()->isVectorTy() && "Expected a vector type");
1288 assert(!Start->getType()->isVectorTy() && "Expected a scalar type");
1289
1290 return B.CreateFAddReduce(Start, Src);
1291 }
1292
createOrderedReduction(VectorBuilder & VBuilder,const RecurrenceDescriptor & Desc,Value * Src,Value * Start)1293 Value *llvm::createOrderedReduction(VectorBuilder &VBuilder,
1294 const RecurrenceDescriptor &Desc,
1295 Value *Src, Value *Start) {
1296 assert((Desc.getRecurrenceKind() == RecurKind::FAdd ||
1297 Desc.getRecurrenceKind() == RecurKind::FMulAdd) &&
1298 "Unexpected reduction kind");
1299 assert(Src->getType()->isVectorTy() && "Expected a vector type");
1300 assert(!Start->getType()->isVectorTy() && "Expected a scalar type");
1301
1302 Intrinsic::ID Id = getReductionIntrinsicID(RecurKind::FAdd);
1303 auto *SrcTy = cast<VectorType>(Src->getType());
1304 Value *Ops[] = {Start, Src};
1305 return VBuilder.createSimpleTargetReduction(Id, SrcTy, Ops);
1306 }
1307
propagateIRFlags(Value * I,ArrayRef<Value * > VL,Value * OpValue,bool IncludeWrapFlags)1308 void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue,
1309 bool IncludeWrapFlags) {
1310 auto *VecOp = dyn_cast<Instruction>(I);
1311 if (!VecOp)
1312 return;
1313 auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0])
1314 : dyn_cast<Instruction>(OpValue);
1315 if (!Intersection)
1316 return;
1317 const unsigned Opcode = Intersection->getOpcode();
1318 VecOp->copyIRFlags(Intersection, IncludeWrapFlags);
1319 for (auto *V : VL) {
1320 auto *Instr = dyn_cast<Instruction>(V);
1321 if (!Instr)
1322 continue;
1323 if (OpValue == nullptr || Opcode == Instr->getOpcode())
1324 VecOp->andIRFlags(V);
1325 }
1326 }
1327
isKnownNegativeInLoop(const SCEV * S,const Loop * L,ScalarEvolution & SE)1328 bool llvm::isKnownNegativeInLoop(const SCEV *S, const Loop *L,
1329 ScalarEvolution &SE) {
1330 const SCEV *Zero = SE.getZero(S->getType());
1331 return SE.isAvailableAtLoopEntry(S, L) &&
1332 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, S, Zero);
1333 }
1334
isKnownNonNegativeInLoop(const SCEV * S,const Loop * L,ScalarEvolution & SE)1335 bool llvm::isKnownNonNegativeInLoop(const SCEV *S, const Loop *L,
1336 ScalarEvolution &SE) {
1337 const SCEV *Zero = SE.getZero(S->getType());
1338 return SE.isAvailableAtLoopEntry(S, L) &&
1339 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGE, S, Zero);
1340 }
1341
isKnownPositiveInLoop(const SCEV * S,const Loop * L,ScalarEvolution & SE)1342 bool llvm::isKnownPositiveInLoop(const SCEV *S, const Loop *L,
1343 ScalarEvolution &SE) {
1344 const SCEV *Zero = SE.getZero(S->getType());
1345 return SE.isAvailableAtLoopEntry(S, L) &&
1346 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGT, S, Zero);
1347 }
1348
isKnownNonPositiveInLoop(const SCEV * S,const Loop * L,ScalarEvolution & SE)1349 bool llvm::isKnownNonPositiveInLoop(const SCEV *S, const Loop *L,
1350 ScalarEvolution &SE) {
1351 const SCEV *Zero = SE.getZero(S->getType());
1352 return SE.isAvailableAtLoopEntry(S, L) &&
1353 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLE, S, Zero);
1354 }
1355
cannotBeMinInLoop(const SCEV * S,const Loop * L,ScalarEvolution & SE,bool Signed)1356 bool llvm::cannotBeMinInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1357 bool Signed) {
1358 unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1359 APInt Min = Signed ? APInt::getSignedMinValue(BitWidth) :
1360 APInt::getMinValue(BitWidth);
1361 auto Predicate = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1362 return SE.isAvailableAtLoopEntry(S, L) &&
1363 SE.isLoopEntryGuardedByCond(L, Predicate, S,
1364 SE.getConstant(Min));
1365 }
1366
cannotBeMaxInLoop(const SCEV * S,const Loop * L,ScalarEvolution & SE,bool Signed)1367 bool llvm::cannotBeMaxInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1368 bool Signed) {
1369 unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1370 APInt Max = Signed ? APInt::getSignedMaxValue(BitWidth) :
1371 APInt::getMaxValue(BitWidth);
1372 auto Predicate = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1373 return SE.isAvailableAtLoopEntry(S, L) &&
1374 SE.isLoopEntryGuardedByCond(L, Predicate, S,
1375 SE.getConstant(Max));
1376 }
1377
1378 //===----------------------------------------------------------------------===//
1379 // rewriteLoopExitValues - Optimize IV users outside the loop.
1380 // As a side effect, reduces the amount of IV processing within the loop.
1381 //===----------------------------------------------------------------------===//
1382
hasHardUserWithinLoop(const Loop * L,const Instruction * I)1383 static bool hasHardUserWithinLoop(const Loop *L, const Instruction *I) {
1384 SmallPtrSet<const Instruction *, 8> Visited;
1385 SmallVector<const Instruction *, 8> WorkList;
1386 Visited.insert(I);
1387 WorkList.push_back(I);
1388 while (!WorkList.empty()) {
1389 const Instruction *Curr = WorkList.pop_back_val();
1390 // This use is outside the loop, nothing to do.
1391 if (!L->contains(Curr))
1392 continue;
1393 // Do we assume it is a "hard" use which will not be eliminated easily?
1394 if (Curr->mayHaveSideEffects())
1395 return true;
1396 // Otherwise, add all its users to worklist.
1397 for (const auto *U : Curr->users()) {
1398 auto *UI = cast<Instruction>(U);
1399 if (Visited.insert(UI).second)
1400 WorkList.push_back(UI);
1401 }
1402 }
1403 return false;
1404 }
1405
1406 // Collect information about PHI nodes which can be transformed in
1407 // rewriteLoopExitValues.
1408 struct RewritePhi {
1409 PHINode *PN; // For which PHI node is this replacement?
1410 unsigned Ith; // For which incoming value?
1411 const SCEV *ExpansionSCEV; // The SCEV of the incoming value we are rewriting.
1412 Instruction *ExpansionPoint; // Where we'd like to expand that SCEV?
1413 bool HighCost; // Is this expansion a high-cost?
1414
RewritePhiRewritePhi1415 RewritePhi(PHINode *P, unsigned I, const SCEV *Val, Instruction *ExpansionPt,
1416 bool H)
1417 : PN(P), Ith(I), ExpansionSCEV(Val), ExpansionPoint(ExpansionPt),
1418 HighCost(H) {}
1419 };
1420
1421 // Check whether it is possible to delete the loop after rewriting exit
1422 // value. If it is possible, ignore ReplaceExitValue and do rewriting
1423 // aggressively.
canLoopBeDeleted(Loop * L,SmallVector<RewritePhi,8> & RewritePhiSet)1424 static bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
1425 BasicBlock *Preheader = L->getLoopPreheader();
1426 // If there is no preheader, the loop will not be deleted.
1427 if (!Preheader)
1428 return false;
1429
1430 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
1431 // We obviate multiple ExitingBlocks case for simplicity.
1432 // TODO: If we see testcase with multiple ExitingBlocks can be deleted
1433 // after exit value rewriting, we can enhance the logic here.
1434 SmallVector<BasicBlock *, 4> ExitingBlocks;
1435 L->getExitingBlocks(ExitingBlocks);
1436 SmallVector<BasicBlock *, 8> ExitBlocks;
1437 L->getUniqueExitBlocks(ExitBlocks);
1438 if (ExitBlocks.size() != 1 || ExitingBlocks.size() != 1)
1439 return false;
1440
1441 BasicBlock *ExitBlock = ExitBlocks[0];
1442 BasicBlock::iterator BI = ExitBlock->begin();
1443 while (PHINode *P = dyn_cast<PHINode>(BI)) {
1444 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
1445
1446 // If the Incoming value of P is found in RewritePhiSet, we know it
1447 // could be rewritten to use a loop invariant value in transformation
1448 // phase later. Skip it in the loop invariant check below.
1449 bool found = false;
1450 for (const RewritePhi &Phi : RewritePhiSet) {
1451 unsigned i = Phi.Ith;
1452 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
1453 found = true;
1454 break;
1455 }
1456 }
1457
1458 Instruction *I;
1459 if (!found && (I = dyn_cast<Instruction>(Incoming)))
1460 if (!L->hasLoopInvariantOperands(I))
1461 return false;
1462
1463 ++BI;
1464 }
1465
1466 for (auto *BB : L->blocks())
1467 if (llvm::any_of(*BB, [](Instruction &I) {
1468 return I.mayHaveSideEffects();
1469 }))
1470 return false;
1471
1472 return true;
1473 }
1474
1475 /// Checks if it is safe to call InductionDescriptor::isInductionPHI for \p Phi,
1476 /// and returns true if this Phi is an induction phi in the loop. When
1477 /// isInductionPHI returns true, \p ID will be also be set by isInductionPHI.
checkIsIndPhi(PHINode * Phi,Loop * L,ScalarEvolution * SE,InductionDescriptor & ID)1478 static bool checkIsIndPhi(PHINode *Phi, Loop *L, ScalarEvolution *SE,
1479 InductionDescriptor &ID) {
1480 if (!Phi)
1481 return false;
1482 if (!L->getLoopPreheader())
1483 return false;
1484 if (Phi->getParent() != L->getHeader())
1485 return false;
1486 return InductionDescriptor::isInductionPHI(Phi, L, SE, ID);
1487 }
1488
rewriteLoopExitValues(Loop * L,LoopInfo * LI,TargetLibraryInfo * TLI,ScalarEvolution * SE,const TargetTransformInfo * TTI,SCEVExpander & Rewriter,DominatorTree * DT,ReplaceExitVal ReplaceExitValue,SmallVector<WeakTrackingVH,16> & DeadInsts)1489 int llvm::rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI,
1490 ScalarEvolution *SE,
1491 const TargetTransformInfo *TTI,
1492 SCEVExpander &Rewriter, DominatorTree *DT,
1493 ReplaceExitVal ReplaceExitValue,
1494 SmallVector<WeakTrackingVH, 16> &DeadInsts) {
1495 // Check a pre-condition.
1496 assert(L->isRecursivelyLCSSAForm(*DT, *LI) &&
1497 "Indvars did not preserve LCSSA!");
1498
1499 SmallVector<BasicBlock*, 8> ExitBlocks;
1500 L->getUniqueExitBlocks(ExitBlocks);
1501
1502 SmallVector<RewritePhi, 8> RewritePhiSet;
1503 // Find all values that are computed inside the loop, but used outside of it.
1504 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
1505 // the exit blocks of the loop to find them.
1506 for (BasicBlock *ExitBB : ExitBlocks) {
1507 // If there are no PHI nodes in this exit block, then no values defined
1508 // inside the loop are used on this path, skip it.
1509 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
1510 if (!PN) continue;
1511
1512 unsigned NumPreds = PN->getNumIncomingValues();
1513
1514 // Iterate over all of the PHI nodes.
1515 BasicBlock::iterator BBI = ExitBB->begin();
1516 while ((PN = dyn_cast<PHINode>(BBI++))) {
1517 if (PN->use_empty())
1518 continue; // dead use, don't replace it
1519
1520 if (!SE->isSCEVable(PN->getType()))
1521 continue;
1522
1523 // Iterate over all of the values in all the PHI nodes.
1524 for (unsigned i = 0; i != NumPreds; ++i) {
1525 // If the value being merged in is not integer or is not defined
1526 // in the loop, skip it.
1527 Value *InVal = PN->getIncomingValue(i);
1528 if (!isa<Instruction>(InVal))
1529 continue;
1530
1531 // If this pred is for a subloop, not L itself, skip it.
1532 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
1533 continue; // The Block is in a subloop, skip it.
1534
1535 // Check that InVal is defined in the loop.
1536 Instruction *Inst = cast<Instruction>(InVal);
1537 if (!L->contains(Inst))
1538 continue;
1539
1540 // Find exit values which are induction variables in the loop, and are
1541 // unused in the loop, with the only use being the exit block PhiNode,
1542 // and the induction variable update binary operator.
1543 // The exit value can be replaced with the final value when it is cheap
1544 // to do so.
1545 if (ReplaceExitValue == UnusedIndVarInLoop) {
1546 InductionDescriptor ID;
1547 PHINode *IndPhi = dyn_cast<PHINode>(Inst);
1548 if (IndPhi) {
1549 if (!checkIsIndPhi(IndPhi, L, SE, ID))
1550 continue;
1551 // This is an induction PHI. Check that the only users are PHI
1552 // nodes, and induction variable update binary operators.
1553 if (llvm::any_of(Inst->users(), [&](User *U) {
1554 if (!isa<PHINode>(U) && !isa<BinaryOperator>(U))
1555 return true;
1556 BinaryOperator *B = dyn_cast<BinaryOperator>(U);
1557 if (B && B != ID.getInductionBinOp())
1558 return true;
1559 return false;
1560 }))
1561 continue;
1562 } else {
1563 // If it is not an induction phi, it must be an induction update
1564 // binary operator with an induction phi user.
1565 BinaryOperator *B = dyn_cast<BinaryOperator>(Inst);
1566 if (!B)
1567 continue;
1568 if (llvm::any_of(Inst->users(), [&](User *U) {
1569 PHINode *Phi = dyn_cast<PHINode>(U);
1570 if (Phi != PN && !checkIsIndPhi(Phi, L, SE, ID))
1571 return true;
1572 return false;
1573 }))
1574 continue;
1575 if (B != ID.getInductionBinOp())
1576 continue;
1577 }
1578 }
1579
1580 // Okay, this instruction has a user outside of the current loop
1581 // and varies predictably *inside* the loop. Evaluate the value it
1582 // contains when the loop exits, if possible. We prefer to start with
1583 // expressions which are true for all exits (so as to maximize
1584 // expression reuse by the SCEVExpander), but resort to per-exit
1585 // evaluation if that fails.
1586 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
1587 if (isa<SCEVCouldNotCompute>(ExitValue) ||
1588 !SE->isLoopInvariant(ExitValue, L) ||
1589 !Rewriter.isSafeToExpand(ExitValue)) {
1590 // TODO: This should probably be sunk into SCEV in some way; maybe a
1591 // getSCEVForExit(SCEV*, L, ExitingBB)? It can be generalized for
1592 // most SCEV expressions and other recurrence types (e.g. shift
1593 // recurrences). Is there existing code we can reuse?
1594 const SCEV *ExitCount = SE->getExitCount(L, PN->getIncomingBlock(i));
1595 if (isa<SCEVCouldNotCompute>(ExitCount))
1596 continue;
1597 if (auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Inst)))
1598 if (AddRec->getLoop() == L)
1599 ExitValue = AddRec->evaluateAtIteration(ExitCount, *SE);
1600 if (isa<SCEVCouldNotCompute>(ExitValue) ||
1601 !SE->isLoopInvariant(ExitValue, L) ||
1602 !Rewriter.isSafeToExpand(ExitValue))
1603 continue;
1604 }
1605
1606 // Computing the value outside of the loop brings no benefit if it is
1607 // definitely used inside the loop in a way which can not be optimized
1608 // away. Avoid doing so unless we know we have a value which computes
1609 // the ExitValue already. TODO: This should be merged into SCEV
1610 // expander to leverage its knowledge of existing expressions.
1611 if (ReplaceExitValue != AlwaysRepl && !isa<SCEVConstant>(ExitValue) &&
1612 !isa<SCEVUnknown>(ExitValue) && hasHardUserWithinLoop(L, Inst))
1613 continue;
1614
1615 // Check if expansions of this SCEV would count as being high cost.
1616 bool HighCost = Rewriter.isHighCostExpansion(
1617 ExitValue, L, SCEVCheapExpansionBudget, TTI, Inst);
1618
1619 // Note that we must not perform expansions until after
1620 // we query *all* the costs, because if we perform temporary expansion
1621 // inbetween, one that we might not intend to keep, said expansion
1622 // *may* affect cost calculation of the next SCEV's we'll query,
1623 // and next SCEV may errneously get smaller cost.
1624
1625 // Collect all the candidate PHINodes to be rewritten.
1626 Instruction *InsertPt =
1627 (isa<PHINode>(Inst) || isa<LandingPadInst>(Inst)) ?
1628 &*Inst->getParent()->getFirstInsertionPt() : Inst;
1629 RewritePhiSet.emplace_back(PN, i, ExitValue, InsertPt, HighCost);
1630 }
1631 }
1632 }
1633
1634 // TODO: evaluate whether it is beneficial to change how we calculate
1635 // high-cost: if we have SCEV 'A' which we know we will expand, should we
1636 // calculate the cost of other SCEV's after expanding SCEV 'A', thus
1637 // potentially giving cost bonus to those other SCEV's?
1638
1639 bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet);
1640 int NumReplaced = 0;
1641
1642 // Transformation.
1643 for (const RewritePhi &Phi : RewritePhiSet) {
1644 PHINode *PN = Phi.PN;
1645
1646 // Only do the rewrite when the ExitValue can be expanded cheaply.
1647 // If LoopCanBeDel is true, rewrite exit value aggressively.
1648 if ((ReplaceExitValue == OnlyCheapRepl ||
1649 ReplaceExitValue == UnusedIndVarInLoop) &&
1650 !LoopCanBeDel && Phi.HighCost)
1651 continue;
1652
1653 Value *ExitVal = Rewriter.expandCodeFor(
1654 Phi.ExpansionSCEV, Phi.PN->getType(), Phi.ExpansionPoint);
1655
1656 LLVM_DEBUG(dbgs() << "rewriteLoopExitValues: AfterLoopVal = " << *ExitVal
1657 << '\n'
1658 << " LoopVal = " << *(Phi.ExpansionPoint) << "\n");
1659
1660 #ifndef NDEBUG
1661 // If we reuse an instruction from a loop which is neither L nor one of
1662 // its containing loops, we end up breaking LCSSA form for this loop by
1663 // creating a new use of its instruction.
1664 if (auto *ExitInsn = dyn_cast<Instruction>(ExitVal))
1665 if (auto *EVL = LI->getLoopFor(ExitInsn->getParent()))
1666 if (EVL != L)
1667 assert(EVL->contains(L) && "LCSSA breach detected!");
1668 #endif
1669
1670 NumReplaced++;
1671 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
1672 PN->setIncomingValue(Phi.Ith, ExitVal);
1673 // It's necessary to tell ScalarEvolution about this explicitly so that
1674 // it can walk the def-use list and forget all SCEVs, as it may not be
1675 // watching the PHI itself. Once the new exit value is in place, there
1676 // may not be a def-use connection between the loop and every instruction
1677 // which got a SCEVAddRecExpr for that loop.
1678 SE->forgetValue(PN);
1679
1680 // If this instruction is dead now, delete it. Don't do it now to avoid
1681 // invalidating iterators.
1682 if (isInstructionTriviallyDead(Inst, TLI))
1683 DeadInsts.push_back(Inst);
1684
1685 // Replace PN with ExitVal if that is legal and does not break LCSSA.
1686 if (PN->getNumIncomingValues() == 1 &&
1687 LI->replacementPreservesLCSSAForm(PN, ExitVal)) {
1688 PN->replaceAllUsesWith(ExitVal);
1689 PN->eraseFromParent();
1690 }
1691 }
1692
1693 // The insertion point instruction may have been deleted; clear it out
1694 // so that the rewriter doesn't trip over it later.
1695 Rewriter.clearInsertPoint();
1696 return NumReplaced;
1697 }
1698
1699 /// Set weights for \p UnrolledLoop and \p RemainderLoop based on weights for
1700 /// \p OrigLoop.
setProfileInfoAfterUnrolling(Loop * OrigLoop,Loop * UnrolledLoop,Loop * RemainderLoop,uint64_t UF)1701 void llvm::setProfileInfoAfterUnrolling(Loop *OrigLoop, Loop *UnrolledLoop,
1702 Loop *RemainderLoop, uint64_t UF) {
1703 assert(UF > 0 && "Zero unrolled factor is not supported");
1704 assert(UnrolledLoop != RemainderLoop &&
1705 "Unrolled and Remainder loops are expected to distinct");
1706
1707 // Get number of iterations in the original scalar loop.
1708 unsigned OrigLoopInvocationWeight = 0;
1709 std::optional<unsigned> OrigAverageTripCount =
1710 getLoopEstimatedTripCount(OrigLoop, &OrigLoopInvocationWeight);
1711 if (!OrigAverageTripCount)
1712 return;
1713
1714 // Calculate number of iterations in unrolled loop.
1715 unsigned UnrolledAverageTripCount = *OrigAverageTripCount / UF;
1716 // Calculate number of iterations for remainder loop.
1717 unsigned RemainderAverageTripCount = *OrigAverageTripCount % UF;
1718
1719 setLoopEstimatedTripCount(UnrolledLoop, UnrolledAverageTripCount,
1720 OrigLoopInvocationWeight);
1721 setLoopEstimatedTripCount(RemainderLoop, RemainderAverageTripCount,
1722 OrigLoopInvocationWeight);
1723 }
1724
1725 /// Utility that implements appending of loops onto a worklist.
1726 /// Loops are added in preorder (analogous for reverse postorder for trees),
1727 /// and the worklist is processed LIFO.
1728 template <typename RangeT>
appendReversedLoopsToWorklist(RangeT && Loops,SmallPriorityWorklist<Loop *,4> & Worklist)1729 void llvm::appendReversedLoopsToWorklist(
1730 RangeT &&Loops, SmallPriorityWorklist<Loop *, 4> &Worklist) {
1731 // We use an internal worklist to build up the preorder traversal without
1732 // recursion.
1733 SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist;
1734
1735 // We walk the initial sequence of loops in reverse because we generally want
1736 // to visit defs before uses and the worklist is LIFO.
1737 for (Loop *RootL : Loops) {
1738 assert(PreOrderLoops.empty() && "Must start with an empty preorder walk.");
1739 assert(PreOrderWorklist.empty() &&
1740 "Must start with an empty preorder walk worklist.");
1741 PreOrderWorklist.push_back(RootL);
1742 do {
1743 Loop *L = PreOrderWorklist.pop_back_val();
1744 PreOrderWorklist.append(L->begin(), L->end());
1745 PreOrderLoops.push_back(L);
1746 } while (!PreOrderWorklist.empty());
1747
1748 Worklist.insert(std::move(PreOrderLoops));
1749 PreOrderLoops.clear();
1750 }
1751 }
1752
1753 template <typename RangeT>
appendLoopsToWorklist(RangeT && Loops,SmallPriorityWorklist<Loop *,4> & Worklist)1754 void llvm::appendLoopsToWorklist(RangeT &&Loops,
1755 SmallPriorityWorklist<Loop *, 4> &Worklist) {
1756 appendReversedLoopsToWorklist(reverse(Loops), Worklist);
1757 }
1758
1759 template void llvm::appendLoopsToWorklist<ArrayRef<Loop *> &>(
1760 ArrayRef<Loop *> &Loops, SmallPriorityWorklist<Loop *, 4> &Worklist);
1761
1762 template void
1763 llvm::appendLoopsToWorklist<Loop &>(Loop &L,
1764 SmallPriorityWorklist<Loop *, 4> &Worklist);
1765
appendLoopsToWorklist(LoopInfo & LI,SmallPriorityWorklist<Loop *,4> & Worklist)1766 void llvm::appendLoopsToWorklist(LoopInfo &LI,
1767 SmallPriorityWorklist<Loop *, 4> &Worklist) {
1768 appendReversedLoopsToWorklist(LI, Worklist);
1769 }
1770
cloneLoop(Loop * L,Loop * PL,ValueToValueMapTy & VM,LoopInfo * LI,LPPassManager * LPM)1771 Loop *llvm::cloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
1772 LoopInfo *LI, LPPassManager *LPM) {
1773 Loop &New = *LI->AllocateLoop();
1774 if (PL)
1775 PL->addChildLoop(&New);
1776 else
1777 LI->addTopLevelLoop(&New);
1778
1779 if (LPM)
1780 LPM->addLoop(New);
1781
1782 // Add all of the blocks in L to the new loop.
1783 for (BasicBlock *BB : L->blocks())
1784 if (LI->getLoopFor(BB) == L)
1785 New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), *LI);
1786
1787 // Add all of the subloops to the new loop.
1788 for (Loop *I : *L)
1789 cloneLoop(I, &New, VM, LI, LPM);
1790
1791 return &New;
1792 }
1793
1794 /// IR Values for the lower and upper bounds of a pointer evolution. We
1795 /// need to use value-handles because SCEV expansion can invalidate previously
1796 /// expanded values. Thus expansion of a pointer can invalidate the bounds for
1797 /// a previous one.
1798 struct PointerBounds {
1799 TrackingVH<Value> Start;
1800 TrackingVH<Value> End;
1801 Value *StrideToCheck;
1802 };
1803
1804 /// Expand code for the lower and upper bound of the pointer group \p CG
1805 /// in \p TheLoop. \return the values for the bounds.
expandBounds(const RuntimeCheckingPtrGroup * CG,Loop * TheLoop,Instruction * Loc,SCEVExpander & Exp,bool HoistRuntimeChecks)1806 static PointerBounds expandBounds(const RuntimeCheckingPtrGroup *CG,
1807 Loop *TheLoop, Instruction *Loc,
1808 SCEVExpander &Exp, bool HoistRuntimeChecks) {
1809 LLVMContext &Ctx = Loc->getContext();
1810 Type *PtrArithTy = PointerType::get(Ctx, CG->AddressSpace);
1811
1812 Value *Start = nullptr, *End = nullptr;
1813 LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1814 const SCEV *Low = CG->Low, *High = CG->High, *Stride = nullptr;
1815
1816 // If the Low and High values are themselves loop-variant, then we may want
1817 // to expand the range to include those covered by the outer loop as well.
1818 // There is a trade-off here with the advantage being that creating checks
1819 // using the expanded range permits the runtime memory checks to be hoisted
1820 // out of the outer loop. This reduces the cost of entering the inner loop,
1821 // which can be significant for low trip counts. The disadvantage is that
1822 // there is a chance we may now never enter the vectorized inner loop,
1823 // whereas using a restricted range check could have allowed us to enter at
1824 // least once. This is why the behaviour is not currently the default and is
1825 // controlled by the parameter 'HoistRuntimeChecks'.
1826 if (HoistRuntimeChecks && TheLoop->getParentLoop() &&
1827 isa<SCEVAddRecExpr>(High) && isa<SCEVAddRecExpr>(Low)) {
1828 auto *HighAR = cast<SCEVAddRecExpr>(High);
1829 auto *LowAR = cast<SCEVAddRecExpr>(Low);
1830 const Loop *OuterLoop = TheLoop->getParentLoop();
1831 ScalarEvolution &SE = *Exp.getSE();
1832 const SCEV *Recur = LowAR->getStepRecurrence(SE);
1833 if (Recur == HighAR->getStepRecurrence(SE) &&
1834 HighAR->getLoop() == OuterLoop && LowAR->getLoop() == OuterLoop) {
1835 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1836 const SCEV *OuterExitCount = SE.getExitCount(OuterLoop, OuterLoopLatch);
1837 if (!isa<SCEVCouldNotCompute>(OuterExitCount) &&
1838 OuterExitCount->getType()->isIntegerTy()) {
1839 const SCEV *NewHigh =
1840 cast<SCEVAddRecExpr>(High)->evaluateAtIteration(OuterExitCount, SE);
1841 if (!isa<SCEVCouldNotCompute>(NewHigh)) {
1842 LLVM_DEBUG(dbgs() << "LAA: Expanded RT check for range to include "
1843 "outer loop in order to permit hoisting\n");
1844 High = NewHigh;
1845 Low = cast<SCEVAddRecExpr>(Low)->getStart();
1846 // If there is a possibility that the stride is negative then we have
1847 // to generate extra checks to ensure the stride is positive.
1848 if (!SE.isKnownNonNegative(
1849 SE.applyLoopGuards(Recur, HighAR->getLoop()))) {
1850 Stride = Recur;
1851 LLVM_DEBUG(dbgs() << "LAA: ... but need to check stride is "
1852 "positive: "
1853 << *Stride << '\n');
1854 }
1855 }
1856 }
1857 }
1858 }
1859
1860 Start = Exp.expandCodeFor(Low, PtrArithTy, Loc);
1861 End = Exp.expandCodeFor(High, PtrArithTy, Loc);
1862 if (CG->NeedsFreeze) {
1863 IRBuilder<> Builder(Loc);
1864 Start = Builder.CreateFreeze(Start, Start->getName() + ".fr");
1865 End = Builder.CreateFreeze(End, End->getName() + ".fr");
1866 }
1867 Value *StrideVal =
1868 Stride ? Exp.expandCodeFor(Stride, Stride->getType(), Loc) : nullptr;
1869 LLVM_DEBUG(dbgs() << "Start: " << *Low << " End: " << *High << "\n");
1870 return {Start, End, StrideVal};
1871 }
1872
1873 /// Turns a collection of checks into a collection of expanded upper and
1874 /// lower bounds for both pointers in the check.
1875 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4>
expandBounds(const SmallVectorImpl<RuntimePointerCheck> & PointerChecks,Loop * L,Instruction * Loc,SCEVExpander & Exp,bool HoistRuntimeChecks)1876 expandBounds(const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, Loop *L,
1877 Instruction *Loc, SCEVExpander &Exp, bool HoistRuntimeChecks) {
1878 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1879
1880 // Here we're relying on the SCEV Expander's cache to only emit code for the
1881 // same bounds once.
1882 transform(PointerChecks, std::back_inserter(ChecksWithBounds),
1883 [&](const RuntimePointerCheck &Check) {
1884 PointerBounds First = expandBounds(Check.first, L, Loc, Exp,
1885 HoistRuntimeChecks),
1886 Second = expandBounds(Check.second, L, Loc, Exp,
1887 HoistRuntimeChecks);
1888 return std::make_pair(First, Second);
1889 });
1890
1891 return ChecksWithBounds;
1892 }
1893
addRuntimeChecks(Instruction * Loc,Loop * TheLoop,const SmallVectorImpl<RuntimePointerCheck> & PointerChecks,SCEVExpander & Exp,bool HoistRuntimeChecks)1894 Value *llvm::addRuntimeChecks(
1895 Instruction *Loc, Loop *TheLoop,
1896 const SmallVectorImpl<RuntimePointerCheck> &PointerChecks,
1897 SCEVExpander &Exp, bool HoistRuntimeChecks) {
1898 // TODO: Move noalias annotation code from LoopVersioning here and share with LV if possible.
1899 // TODO: Pass RtPtrChecking instead of PointerChecks and SE separately, if possible
1900 auto ExpandedChecks =
1901 expandBounds(PointerChecks, TheLoop, Loc, Exp, HoistRuntimeChecks);
1902
1903 LLVMContext &Ctx = Loc->getContext();
1904 IRBuilder<InstSimplifyFolder> ChkBuilder(Ctx,
1905 Loc->getDataLayout());
1906 ChkBuilder.SetInsertPoint(Loc);
1907 // Our instructions might fold to a constant.
1908 Value *MemoryRuntimeCheck = nullptr;
1909
1910 for (const auto &[A, B] : ExpandedChecks) {
1911 // Check if two pointers (A and B) conflict where conflict is computed as:
1912 // start(A) <= end(B) && start(B) <= end(A)
1913
1914 assert((A.Start->getType()->getPointerAddressSpace() ==
1915 B.End->getType()->getPointerAddressSpace()) &&
1916 (B.Start->getType()->getPointerAddressSpace() ==
1917 A.End->getType()->getPointerAddressSpace()) &&
1918 "Trying to bounds check pointers with different address spaces");
1919
1920 // [A|B].Start points to the first accessed byte under base [A|B].
1921 // [A|B].End points to the last accessed byte, plus one.
1922 // There is no conflict when the intervals are disjoint:
1923 // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
1924 //
1925 // bound0 = (B.Start < A.End)
1926 // bound1 = (A.Start < B.End)
1927 // IsConflict = bound0 & bound1
1928 Value *Cmp0 = ChkBuilder.CreateICmpULT(A.Start, B.End, "bound0");
1929 Value *Cmp1 = ChkBuilder.CreateICmpULT(B.Start, A.End, "bound1");
1930 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1931 if (A.StrideToCheck) {
1932 Value *IsNegativeStride = ChkBuilder.CreateICmpSLT(
1933 A.StrideToCheck, ConstantInt::get(A.StrideToCheck->getType(), 0),
1934 "stride.check");
1935 IsConflict = ChkBuilder.CreateOr(IsConflict, IsNegativeStride);
1936 }
1937 if (B.StrideToCheck) {
1938 Value *IsNegativeStride = ChkBuilder.CreateICmpSLT(
1939 B.StrideToCheck, ConstantInt::get(B.StrideToCheck->getType(), 0),
1940 "stride.check");
1941 IsConflict = ChkBuilder.CreateOr(IsConflict, IsNegativeStride);
1942 }
1943 if (MemoryRuntimeCheck) {
1944 IsConflict =
1945 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1946 }
1947 MemoryRuntimeCheck = IsConflict;
1948 }
1949
1950 return MemoryRuntimeCheck;
1951 }
1952
addDiffRuntimeChecks(Instruction * Loc,ArrayRef<PointerDiffInfo> Checks,SCEVExpander & Expander,function_ref<Value * (IRBuilderBase &,unsigned)> GetVF,unsigned IC)1953 Value *llvm::addDiffRuntimeChecks(
1954 Instruction *Loc, ArrayRef<PointerDiffInfo> Checks, SCEVExpander &Expander,
1955 function_ref<Value *(IRBuilderBase &, unsigned)> GetVF, unsigned IC) {
1956
1957 LLVMContext &Ctx = Loc->getContext();
1958 IRBuilder<InstSimplifyFolder> ChkBuilder(Ctx,
1959 Loc->getDataLayout());
1960 ChkBuilder.SetInsertPoint(Loc);
1961 // Our instructions might fold to a constant.
1962 Value *MemoryRuntimeCheck = nullptr;
1963
1964 auto &SE = *Expander.getSE();
1965 // Map to keep track of created compares, The key is the pair of operands for
1966 // the compare, to allow detecting and re-using redundant compares.
1967 DenseMap<std::pair<Value *, Value *>, Value *> SeenCompares;
1968 for (const auto &[SrcStart, SinkStart, AccessSize, NeedsFreeze] : Checks) {
1969 Type *Ty = SinkStart->getType();
1970 // Compute VF * IC * AccessSize.
1971 auto *VFTimesUFTimesSize =
1972 ChkBuilder.CreateMul(GetVF(ChkBuilder, Ty->getScalarSizeInBits()),
1973 ConstantInt::get(Ty, IC * AccessSize));
1974 Value *Diff =
1975 Expander.expandCodeFor(SE.getMinusSCEV(SinkStart, SrcStart), Ty, Loc);
1976
1977 // Check if the same compare has already been created earlier. In that case,
1978 // there is no need to check it again.
1979 Value *IsConflict = SeenCompares.lookup({Diff, VFTimesUFTimesSize});
1980 if (IsConflict)
1981 continue;
1982
1983 IsConflict =
1984 ChkBuilder.CreateICmpULT(Diff, VFTimesUFTimesSize, "diff.check");
1985 SeenCompares.insert({{Diff, VFTimesUFTimesSize}, IsConflict});
1986 if (NeedsFreeze)
1987 IsConflict =
1988 ChkBuilder.CreateFreeze(IsConflict, IsConflict->getName() + ".fr");
1989 if (MemoryRuntimeCheck) {
1990 IsConflict =
1991 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1992 }
1993 MemoryRuntimeCheck = IsConflict;
1994 }
1995
1996 return MemoryRuntimeCheck;
1997 }
1998
1999 std::optional<IVConditionInfo>
hasPartialIVCondition(const Loop & L,unsigned MSSAThreshold,const MemorySSA & MSSA,AAResults & AA)2000 llvm::hasPartialIVCondition(const Loop &L, unsigned MSSAThreshold,
2001 const MemorySSA &MSSA, AAResults &AA) {
2002 auto *TI = dyn_cast<BranchInst>(L.getHeader()->getTerminator());
2003 if (!TI || !TI->isConditional())
2004 return {};
2005
2006 auto *CondI = dyn_cast<Instruction>(TI->getCondition());
2007 // The case with the condition outside the loop should already be handled
2008 // earlier.
2009 // Allow CmpInst and TruncInsts as they may be users of load instructions
2010 // and have potential for partial unswitching
2011 if (!CondI || !isa<CmpInst, TruncInst>(CondI) || !L.contains(CondI))
2012 return {};
2013
2014 SmallVector<Instruction *> InstToDuplicate;
2015 InstToDuplicate.push_back(CondI);
2016
2017 SmallVector<Value *, 4> WorkList;
2018 WorkList.append(CondI->op_begin(), CondI->op_end());
2019
2020 SmallVector<MemoryAccess *, 4> AccessesToCheck;
2021 SmallVector<MemoryLocation, 4> AccessedLocs;
2022 while (!WorkList.empty()) {
2023 Instruction *I = dyn_cast<Instruction>(WorkList.pop_back_val());
2024 if (!I || !L.contains(I))
2025 continue;
2026
2027 // TODO: support additional instructions.
2028 if (!isa<LoadInst>(I) && !isa<GetElementPtrInst>(I))
2029 return {};
2030
2031 // Do not duplicate volatile and atomic loads.
2032 if (auto *LI = dyn_cast<LoadInst>(I))
2033 if (LI->isVolatile() || LI->isAtomic())
2034 return {};
2035
2036 InstToDuplicate.push_back(I);
2037 if (MemoryAccess *MA = MSSA.getMemoryAccess(I)) {
2038 if (auto *MemUse = dyn_cast_or_null<MemoryUse>(MA)) {
2039 // Queue the defining access to check for alias checks.
2040 AccessesToCheck.push_back(MemUse->getDefiningAccess());
2041 AccessedLocs.push_back(MemoryLocation::get(I));
2042 } else {
2043 // MemoryDefs may clobber the location or may be atomic memory
2044 // operations. Bail out.
2045 return {};
2046 }
2047 }
2048 WorkList.append(I->op_begin(), I->op_end());
2049 }
2050
2051 if (InstToDuplicate.empty())
2052 return {};
2053
2054 SmallVector<BasicBlock *, 4> ExitingBlocks;
2055 L.getExitingBlocks(ExitingBlocks);
2056 auto HasNoClobbersOnPath =
2057 [&L, &AA, &AccessedLocs, &ExitingBlocks, &InstToDuplicate,
2058 MSSAThreshold](BasicBlock *Succ, BasicBlock *Header,
2059 SmallVector<MemoryAccess *, 4> AccessesToCheck)
2060 -> std::optional<IVConditionInfo> {
2061 IVConditionInfo Info;
2062 // First, collect all blocks in the loop that are on a patch from Succ
2063 // to the header.
2064 SmallVector<BasicBlock *, 4> WorkList;
2065 WorkList.push_back(Succ);
2066 WorkList.push_back(Header);
2067 SmallPtrSet<BasicBlock *, 4> Seen;
2068 Seen.insert(Header);
2069 Info.PathIsNoop &=
2070 all_of(*Header, [](Instruction &I) { return !I.mayHaveSideEffects(); });
2071
2072 while (!WorkList.empty()) {
2073 BasicBlock *Current = WorkList.pop_back_val();
2074 if (!L.contains(Current))
2075 continue;
2076 const auto &SeenIns = Seen.insert(Current);
2077 if (!SeenIns.second)
2078 continue;
2079
2080 Info.PathIsNoop &= all_of(
2081 *Current, [](Instruction &I) { return !I.mayHaveSideEffects(); });
2082 WorkList.append(succ_begin(Current), succ_end(Current));
2083 }
2084
2085 // Require at least 2 blocks on a path through the loop. This skips
2086 // paths that directly exit the loop.
2087 if (Seen.size() < 2)
2088 return {};
2089
2090 // Next, check if there are any MemoryDefs that are on the path through
2091 // the loop (in the Seen set) and they may-alias any of the locations in
2092 // AccessedLocs. If that is the case, they may modify the condition and
2093 // partial unswitching is not possible.
2094 SmallPtrSet<MemoryAccess *, 4> SeenAccesses;
2095 while (!AccessesToCheck.empty()) {
2096 MemoryAccess *Current = AccessesToCheck.pop_back_val();
2097 auto SeenI = SeenAccesses.insert(Current);
2098 if (!SeenI.second || !Seen.contains(Current->getBlock()))
2099 continue;
2100
2101 // Bail out if exceeded the threshold.
2102 if (SeenAccesses.size() >= MSSAThreshold)
2103 return {};
2104
2105 // MemoryUse are read-only accesses.
2106 if (isa<MemoryUse>(Current))
2107 continue;
2108
2109 // For a MemoryDef, check if is aliases any of the location feeding
2110 // the original condition.
2111 if (auto *CurrentDef = dyn_cast<MemoryDef>(Current)) {
2112 if (any_of(AccessedLocs, [&AA, CurrentDef](MemoryLocation &Loc) {
2113 return isModSet(
2114 AA.getModRefInfo(CurrentDef->getMemoryInst(), Loc));
2115 }))
2116 return {};
2117 }
2118
2119 for (Use &U : Current->uses())
2120 AccessesToCheck.push_back(cast<MemoryAccess>(U.getUser()));
2121 }
2122
2123 // We could also allow loops with known trip counts without mustprogress,
2124 // but ScalarEvolution may not be available.
2125 Info.PathIsNoop &= isMustProgress(&L);
2126
2127 // If the path is considered a no-op so far, check if it reaches a
2128 // single exit block without any phis. This ensures no values from the
2129 // loop are used outside of the loop.
2130 if (Info.PathIsNoop) {
2131 for (auto *Exiting : ExitingBlocks) {
2132 if (!Seen.contains(Exiting))
2133 continue;
2134 for (auto *Succ : successors(Exiting)) {
2135 if (L.contains(Succ))
2136 continue;
2137
2138 Info.PathIsNoop &= Succ->phis().empty() &&
2139 (!Info.ExitForPath || Info.ExitForPath == Succ);
2140 if (!Info.PathIsNoop)
2141 break;
2142 assert((!Info.ExitForPath || Info.ExitForPath == Succ) &&
2143 "cannot have multiple exit blocks");
2144 Info.ExitForPath = Succ;
2145 }
2146 }
2147 }
2148 if (!Info.ExitForPath)
2149 Info.PathIsNoop = false;
2150
2151 Info.InstToDuplicate = InstToDuplicate;
2152 return Info;
2153 };
2154
2155 // If we branch to the same successor, partial unswitching will not be
2156 // beneficial.
2157 if (TI->getSuccessor(0) == TI->getSuccessor(1))
2158 return {};
2159
2160 if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(0), L.getHeader(),
2161 AccessesToCheck)) {
2162 Info->KnownValue = ConstantInt::getTrue(TI->getContext());
2163 return Info;
2164 }
2165 if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(1), L.getHeader(),
2166 AccessesToCheck)) {
2167 Info->KnownValue = ConstantInt::getFalse(TI->getContext());
2168 return Info;
2169 }
2170
2171 return {};
2172 }
2173