1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass performs loop invariant code motion, attempting to remove as much
11 // code from the body of a loop as possible. It does this by either hoisting
12 // code into the preheader block, or by sinking code to the exit blocks if it is
13 // safe. This pass also promotes must-aliased memory locations in the loop to
14 // live in registers, thus hoisting and sinking "invariant" loads and stores.
15 //
16 // This pass uses alias analysis for two purposes:
17 //
18 // 1. Moving loop invariant loads and calls out of loops. If we can determine
19 // that a load or call inside of a loop never aliases anything stored to,
20 // we can hoist it or sink it like any other instruction.
21 // 2. Scalar Promotion of Memory - If there is a store instruction inside of
22 // the loop, we try to move the store to happen AFTER the loop instead of
23 // inside of the loop. This can only happen if a few conditions are true:
24 // A. The pointer stored through is loop invariant
25 // B. There are no stores or loads in the loop which _may_ alias the
26 // pointer. There are no calls in the loop which mod/ref the pointer.
27 // If these conditions are true, we can promote the loads and stores in the
28 // loop of the pointer to use a temporary alloca'd variable. We then use
29 // the SSAUpdater to construct the appropriate SSA form for the value.
30 //
31 //===----------------------------------------------------------------------===//
32
33 #include "llvm/Transforms/Scalar/LICM.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/AliasSetTracker.h"
37 #include "llvm/Analysis/BasicAliasAnalysis.h"
38 #include "llvm/Analysis/CaptureTracking.h"
39 #include "llvm/Analysis/ConstantFolding.h"
40 #include "llvm/Analysis/GlobalsModRef.h"
41 #include "llvm/Analysis/Loads.h"
42 #include "llvm/Analysis/LoopInfo.h"
43 #include "llvm/Analysis/LoopPass.h"
44 #include "llvm/Analysis/LoopPassManager.h"
45 #include "llvm/Analysis/MemoryBuiltins.h"
46 #include "llvm/Analysis/ScalarEvolution.h"
47 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Analysis/ValueTracking.h"
50 #include "llvm/IR/CFG.h"
51 #include "llvm/IR/Constants.h"
52 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/DerivedTypes.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/LLVMContext.h"
58 #include "llvm/IR/Metadata.h"
59 #include "llvm/IR/PredIteratorCache.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Transforms/Scalar.h"
64 #include "llvm/Transforms/Utils/Local.h"
65 #include "llvm/Transforms/Utils/LoopUtils.h"
66 #include "llvm/Transforms/Utils/SSAUpdater.h"
67 #include <algorithm>
68 #include <utility>
69 using namespace llvm;
70
71 #define DEBUG_TYPE "licm"
72
73 STATISTIC(NumSunk, "Number of instructions sunk out of loop");
74 STATISTIC(NumHoisted, "Number of instructions hoisted out of loop");
75 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
76 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
77 STATISTIC(NumPromoted, "Number of memory locations promoted to registers");
78
79 static cl::opt<bool>
80 DisablePromotion("disable-licm-promotion", cl::Hidden,
81 cl::desc("Disable memory promotion in LICM pass"));
82
83 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
84 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop,
85 const LoopSafetyInfo *SafetyInfo);
86 static bool hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
87 const LoopSafetyInfo *SafetyInfo);
88 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
89 const Loop *CurLoop, AliasSetTracker *CurAST,
90 const LoopSafetyInfo *SafetyInfo);
91 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
92 const DominatorTree *DT,
93 const Loop *CurLoop,
94 const LoopSafetyInfo *SafetyInfo,
95 const Instruction *CtxI = nullptr);
96 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
97 const AAMDNodes &AAInfo,
98 AliasSetTracker *CurAST);
99 static Instruction *
100 CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN,
101 const LoopInfo *LI,
102 const LoopSafetyInfo *SafetyInfo);
103 static bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA,
104 DominatorTree *DT, TargetLibraryInfo *TLI,
105 Loop *CurLoop, AliasSetTracker *CurAST,
106 LoopSafetyInfo *SafetyInfo);
107
108 namespace {
109 struct LoopInvariantCodeMotion {
110 bool runOnLoop(Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT,
111 TargetLibraryInfo *TLI, ScalarEvolution *SE, bool DeleteAST);
112
getLoopToAliasSetMap__anon8d593ef90111::LoopInvariantCodeMotion113 DenseMap<Loop *, AliasSetTracker *> &getLoopToAliasSetMap() {
114 return LoopToAliasSetMap;
115 }
116
117 private:
118 DenseMap<Loop *, AliasSetTracker *> LoopToAliasSetMap;
119
120 AliasSetTracker *collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
121 AliasAnalysis *AA);
122 };
123
124 struct LegacyLICMPass : public LoopPass {
125 static char ID; // Pass identification, replacement for typeid
LegacyLICMPass__anon8d593ef90111::LegacyLICMPass126 LegacyLICMPass() : LoopPass(ID) {
127 initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
128 }
129
runOnLoop__anon8d593ef90111::LegacyLICMPass130 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
131 if (skipLoop(L))
132 return false;
133
134 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
135 return LICM.runOnLoop(L,
136 &getAnalysis<AAResultsWrapperPass>().getAAResults(),
137 &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
138 &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
139 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
140 SE ? &SE->getSE() : nullptr, false);
141 }
142
143 /// This transformation requires natural loop information & requires that
144 /// loop preheaders be inserted into the CFG...
145 ///
getAnalysisUsage__anon8d593ef90111::LegacyLICMPass146 void getAnalysisUsage(AnalysisUsage &AU) const override {
147 AU.setPreservesCFG();
148 AU.addRequired<TargetLibraryInfoWrapperPass>();
149 getLoopAnalysisUsage(AU);
150 }
151
152 using llvm::Pass::doFinalization;
153
doFinalization__anon8d593ef90111::LegacyLICMPass154 bool doFinalization() override {
155 assert(LICM.getLoopToAliasSetMap().empty() &&
156 "Didn't free loop alias sets");
157 return false;
158 }
159
160 private:
161 LoopInvariantCodeMotion LICM;
162
163 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
164 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
165 Loop *L) override;
166
167 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
168 /// set.
169 void deleteAnalysisValue(Value *V, Loop *L) override;
170
171 /// Simple Analysis hook. Delete loop L from alias set map.
172 void deleteAnalysisLoop(Loop *L) override;
173 };
174 }
175
run(Loop & L,AnalysisManager<Loop> & AM)176 PreservedAnalyses LICMPass::run(Loop &L, AnalysisManager<Loop> &AM) {
177 const auto &FAM =
178 AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager();
179 Function *F = L.getHeader()->getParent();
180
181 auto *AA = FAM.getCachedResult<AAManager>(*F);
182 auto *LI = FAM.getCachedResult<LoopAnalysis>(*F);
183 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(*F);
184 auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(*F);
185 auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(*F);
186 assert((AA && LI && DT && TLI && SE) && "Analyses for LICM not available");
187
188 LoopInvariantCodeMotion LICM;
189
190 if (!LICM.runOnLoop(&L, AA, LI, DT, TLI, SE, true))
191 return PreservedAnalyses::all();
192
193 // FIXME: There is no setPreservesCFG in the new PM. When that becomes
194 // available, it should be used here.
195 return getLoopPassPreservedAnalyses();
196 }
197
198 char LegacyLICMPass::ID = 0;
199 INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",
200 false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)201 INITIALIZE_PASS_DEPENDENCY(LoopPass)
202 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
203 INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,
204 false)
205
206 Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
207
208 /// Hoist expressions out of the specified loop. Note, alias info for inner
209 /// loop is not preserved so it is not a good idea to run LICM multiple
210 /// times on one loop.
211 /// We should delete AST for inner loops in the new pass manager to avoid
212 /// memory leak.
213 ///
runOnLoop(Loop * L,AliasAnalysis * AA,LoopInfo * LI,DominatorTree * DT,TargetLibraryInfo * TLI,ScalarEvolution * SE,bool DeleteAST)214 bool LoopInvariantCodeMotion::runOnLoop(Loop *L, AliasAnalysis *AA,
215 LoopInfo *LI, DominatorTree *DT,
216 TargetLibraryInfo *TLI,
217 ScalarEvolution *SE, bool DeleteAST) {
218 bool Changed = false;
219
220 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
221
222 AliasSetTracker *CurAST = collectAliasInfoForLoop(L, LI, AA);
223
224 // Get the preheader block to move instructions into...
225 BasicBlock *Preheader = L->getLoopPreheader();
226
227 // Compute loop safety information.
228 LoopSafetyInfo SafetyInfo;
229 computeLoopSafetyInfo(&SafetyInfo, L);
230
231 // We want to visit all of the instructions in this loop... that are not parts
232 // of our subloops (they have already had their invariants hoisted out of
233 // their loop, into this loop, so there is no need to process the BODIES of
234 // the subloops).
235 //
236 // Traverse the body of the loop in depth first order on the dominator tree so
237 // that we are guaranteed to see definitions before we see uses. This allows
238 // us to sink instructions in one pass, without iteration. After sinking
239 // instructions, we perform another pass to hoist them out of the loop.
240 //
241 if (L->hasDedicatedExits())
242 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L,
243 CurAST, &SafetyInfo);
244 if (Preheader)
245 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L,
246 CurAST, &SafetyInfo);
247
248 // Now that all loop invariants have been removed from the loop, promote any
249 // memory references to scalars that we can.
250 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) {
251 SmallVector<BasicBlock *, 8> ExitBlocks;
252 SmallVector<Instruction *, 8> InsertPts;
253 PredIteratorCache PIC;
254
255 // Loop over all of the alias sets in the tracker object.
256 for (AliasSet &AS : *CurAST)
257 Changed |= promoteLoopAccessesToScalars(
258 AS, ExitBlocks, InsertPts, PIC, LI, DT, TLI, L, CurAST, &SafetyInfo);
259
260 // Once we have promoted values across the loop body we have to recursively
261 // reform LCSSA as any nested loop may now have values defined within the
262 // loop used in the outer loop.
263 // FIXME: This is really heavy handed. It would be a bit better to use an
264 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
265 // it as it went.
266 if (Changed) {
267 formLCSSARecursively(*L, *DT, LI, SE);
268 }
269 }
270
271 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
272 // specifically moving instructions across the loop boundary and so it is
273 // especially in need of sanity checking here.
274 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
275 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
276 "Parent loop not left in LCSSA form after LICM!");
277
278 // If this loop is nested inside of another one, save the alias information
279 // for when we process the outer loop.
280 if (L->getParentLoop() && !DeleteAST)
281 LoopToAliasSetMap[L] = CurAST;
282 else
283 delete CurAST;
284
285 if (Changed && SE)
286 SE->forgetLoopDispositions(L);
287 return Changed;
288 }
289
290 /// Walk the specified region of the CFG (defined by all blocks dominated by
291 /// the specified block, and that are in the current loop) in reverse depth
292 /// first order w.r.t the DominatorTree. This allows us to visit uses before
293 /// definitions, allowing us to sink a loop body in one pass without iteration.
294 ///
sinkRegion(DomTreeNode * N,AliasAnalysis * AA,LoopInfo * LI,DominatorTree * DT,TargetLibraryInfo * TLI,Loop * CurLoop,AliasSetTracker * CurAST,LoopSafetyInfo * SafetyInfo)295 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
296 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
297 AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo) {
298
299 // Verify inputs.
300 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
301 CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr &&
302 "Unexpected input to sinkRegion");
303
304 BasicBlock *BB = N->getBlock();
305 // If this subregion is not in the top level loop at all, exit.
306 if (!CurLoop->contains(BB))
307 return false;
308
309 // We are processing blocks in reverse dfo, so process children first.
310 bool Changed = false;
311 const std::vector<DomTreeNode *> &Children = N->getChildren();
312 for (DomTreeNode *Child : Children)
313 Changed |= sinkRegion(Child, AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
314
315 // Only need to process the contents of this block if it is not part of a
316 // subloop (which would already have been processed).
317 if (inSubLoop(BB, CurLoop, LI))
318 return Changed;
319
320 for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
321 Instruction &I = *--II;
322
323 // If the instruction is dead, we would try to sink it because it isn't used
324 // in the loop, instead, just delete it.
325 if (isInstructionTriviallyDead(&I, TLI)) {
326 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
327 ++II;
328 CurAST->deleteValue(&I);
329 I.eraseFromParent();
330 Changed = true;
331 continue;
332 }
333
334 // Check to see if we can sink this instruction to the exit blocks
335 // of the loop. We can do this if the all users of the instruction are
336 // outside of the loop. In this case, it doesn't even matter if the
337 // operands of the instruction are loop invariant.
338 //
339 if (isNotUsedInLoop(I, CurLoop, SafetyInfo) &&
340 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo)) {
341 ++II;
342 Changed |= sink(I, LI, DT, CurLoop, CurAST, SafetyInfo);
343 }
344 }
345 return Changed;
346 }
347
348 /// Walk the specified region of the CFG (defined by all blocks dominated by
349 /// the specified block, and that are in the current loop) in depth first
350 /// order w.r.t the DominatorTree. This allows us to visit definitions before
351 /// uses, allowing us to hoist a loop body in one pass without iteration.
352 ///
hoistRegion(DomTreeNode * N,AliasAnalysis * AA,LoopInfo * LI,DominatorTree * DT,TargetLibraryInfo * TLI,Loop * CurLoop,AliasSetTracker * CurAST,LoopSafetyInfo * SafetyInfo)353 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
354 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
355 AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo) {
356 // Verify inputs.
357 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
358 CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr &&
359 "Unexpected input to hoistRegion");
360
361 BasicBlock *BB = N->getBlock();
362
363 // If this subregion is not in the top level loop at all, exit.
364 if (!CurLoop->contains(BB))
365 return false;
366
367 // Only need to process the contents of this block if it is not part of a
368 // subloop (which would already have been processed).
369 bool Changed = false;
370 if (!inSubLoop(BB, CurLoop, LI))
371 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
372 Instruction &I = *II++;
373 // Try constant folding this instruction. If all the operands are
374 // constants, it is technically hoistable, but it would be better to just
375 // fold it.
376 if (Constant *C = ConstantFoldInstruction(
377 &I, I.getModule()->getDataLayout(), TLI)) {
378 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
379 CurAST->copyValue(&I, C);
380 CurAST->deleteValue(&I);
381 I.replaceAllUsesWith(C);
382 I.eraseFromParent();
383 continue;
384 }
385
386 // Try hoisting the instruction out to the preheader. We can only do this
387 // if all of the operands of the instruction are loop invariant and if it
388 // is safe to hoist the instruction.
389 //
390 if (CurLoop->hasLoopInvariantOperands(&I) &&
391 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo) &&
392 isSafeToExecuteUnconditionally(
393 I, DT, CurLoop, SafetyInfo,
394 CurLoop->getLoopPreheader()->getTerminator()))
395 Changed |= hoist(I, DT, CurLoop, SafetyInfo);
396 }
397
398 const std::vector<DomTreeNode *> &Children = N->getChildren();
399 for (DomTreeNode *Child : Children)
400 Changed |= hoistRegion(Child, AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
401 return Changed;
402 }
403
404 /// Computes loop safety information, checks loop body & header
405 /// for the possibility of may throw exception.
406 ///
computeLoopSafetyInfo(LoopSafetyInfo * SafetyInfo,Loop * CurLoop)407 void llvm::computeLoopSafetyInfo(LoopSafetyInfo *SafetyInfo, Loop *CurLoop) {
408 assert(CurLoop != nullptr && "CurLoop cant be null");
409 BasicBlock *Header = CurLoop->getHeader();
410 // Setting default safety values.
411 SafetyInfo->MayThrow = false;
412 SafetyInfo->HeaderMayThrow = false;
413 // Iterate over header and compute safety info.
414 for (BasicBlock::iterator I = Header->begin(), E = Header->end();
415 (I != E) && !SafetyInfo->HeaderMayThrow; ++I)
416 SafetyInfo->HeaderMayThrow |=
417 !isGuaranteedToTransferExecutionToSuccessor(&*I);
418
419 SafetyInfo->MayThrow = SafetyInfo->HeaderMayThrow;
420 // Iterate over loop instructions and compute safety info.
421 for (Loop::block_iterator BB = CurLoop->block_begin(),
422 BBE = CurLoop->block_end();
423 (BB != BBE) && !SafetyInfo->MayThrow; ++BB)
424 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
425 (I != E) && !SafetyInfo->MayThrow; ++I)
426 SafetyInfo->MayThrow |= !isGuaranteedToTransferExecutionToSuccessor(&*I);
427
428 // Compute funclet colors if we might sink/hoist in a function with a funclet
429 // personality routine.
430 Function *Fn = CurLoop->getHeader()->getParent();
431 if (Fn->hasPersonalityFn())
432 if (Constant *PersonalityFn = Fn->getPersonalityFn())
433 if (isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)))
434 SafetyInfo->BlockColors = colorEHFunclets(*Fn);
435 }
436
437 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
438 /// instruction.
439 ///
canSinkOrHoistInst(Instruction & I,AliasAnalysis * AA,DominatorTree * DT,TargetLibraryInfo * TLI,Loop * CurLoop,AliasSetTracker * CurAST,LoopSafetyInfo * SafetyInfo)440 bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA, DominatorTree *DT,
441 TargetLibraryInfo *TLI, Loop *CurLoop,
442 AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo) {
443 // Loads have extra constraints we have to verify before we can hoist them.
444 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
445 if (!LI->isUnordered())
446 return false; // Don't hoist volatile/atomic loads!
447
448 // Loads from constant memory are always safe to move, even if they end up
449 // in the same alias set as something that ends up being modified.
450 if (AA->pointsToConstantMemory(LI->getOperand(0)))
451 return true;
452 if (LI->getMetadata(LLVMContext::MD_invariant_load))
453 return true;
454
455 // Don't hoist loads which have may-aliased stores in loop.
456 uint64_t Size = 0;
457 if (LI->getType()->isSized())
458 Size = I.getModule()->getDataLayout().getTypeStoreSize(LI->getType());
459
460 AAMDNodes AAInfo;
461 LI->getAAMetadata(AAInfo);
462
463 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo, CurAST);
464 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
465 // Don't sink or hoist dbg info; it's legal, but not useful.
466 if (isa<DbgInfoIntrinsic>(I))
467 return false;
468
469 // Don't sink calls which can throw.
470 if (CI->mayThrow())
471 return false;
472
473 // Handle simple cases by querying alias analysis.
474 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
475 if (Behavior == FMRB_DoesNotAccessMemory)
476 return true;
477 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
478 // A readonly argmemonly function only reads from memory pointed to by
479 // it's arguments with arbitrary offsets. If we can prove there are no
480 // writes to this memory in the loop, we can hoist or sink.
481 if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) {
482 for (Value *Op : CI->arg_operands())
483 if (Op->getType()->isPointerTy() &&
484 pointerInvalidatedByLoop(Op, MemoryLocation::UnknownSize,
485 AAMDNodes(), CurAST))
486 return false;
487 return true;
488 }
489 // If this call only reads from memory and there are no writes to memory
490 // in the loop, we can hoist or sink the call as appropriate.
491 bool FoundMod = false;
492 for (AliasSet &AS : *CurAST) {
493 if (!AS.isForwardingAliasSet() && AS.isMod()) {
494 FoundMod = true;
495 break;
496 }
497 }
498 if (!FoundMod)
499 return true;
500 }
501
502 // FIXME: This should use mod/ref information to see if we can hoist or
503 // sink the call.
504
505 return false;
506 }
507
508 // Only these instructions are hoistable/sinkable.
509 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
510 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
511 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
512 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
513 !isa<InsertValueInst>(I))
514 return false;
515
516 // TODO: Plumb the context instruction through to make hoisting and sinking
517 // more powerful. Hoisting of loads already works due to the special casing
518 // above.
519 return isSafeToExecuteUnconditionally(I, DT, CurLoop, SafetyInfo, nullptr);
520 }
521
522 /// Returns true if a PHINode is a trivially replaceable with an
523 /// Instruction.
524 /// This is true when all incoming values are that instruction.
525 /// This pattern occurs most often with LCSSA PHI nodes.
526 ///
isTriviallyReplacablePHI(const PHINode & PN,const Instruction & I)527 static bool isTriviallyReplacablePHI(const PHINode &PN, const Instruction &I) {
528 for (const Value *IncValue : PN.incoming_values())
529 if (IncValue != &I)
530 return false;
531
532 return true;
533 }
534
535 /// Return true if the only users of this instruction are outside of
536 /// the loop. If this is true, we can sink the instruction to the exit
537 /// blocks of the loop.
538 ///
isNotUsedInLoop(const Instruction & I,const Loop * CurLoop,const LoopSafetyInfo * SafetyInfo)539 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop,
540 const LoopSafetyInfo *SafetyInfo) {
541 const auto &BlockColors = SafetyInfo->BlockColors;
542 for (const User *U : I.users()) {
543 const Instruction *UI = cast<Instruction>(U);
544 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
545 const BasicBlock *BB = PN->getParent();
546 // We cannot sink uses in catchswitches.
547 if (isa<CatchSwitchInst>(BB->getTerminator()))
548 return false;
549
550 // We need to sink a callsite to a unique funclet. Avoid sinking if the
551 // phi use is too muddled.
552 if (isa<CallInst>(I))
553 if (!BlockColors.empty() &&
554 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
555 return false;
556
557 // A PHI node where all of the incoming values are this instruction are
558 // special -- they can just be RAUW'ed with the instruction and thus
559 // don't require a use in the predecessor. This is a particular important
560 // special case because it is the pattern found in LCSSA form.
561 if (isTriviallyReplacablePHI(*PN, I)) {
562 if (CurLoop->contains(PN))
563 return false;
564 else
565 continue;
566 }
567
568 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
569 // values. Check for such a use being inside the loop.
570 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
571 if (PN->getIncomingValue(i) == &I)
572 if (CurLoop->contains(PN->getIncomingBlock(i)))
573 return false;
574
575 continue;
576 }
577
578 if (CurLoop->contains(UI))
579 return false;
580 }
581 return true;
582 }
583
584 static Instruction *
CloneInstructionInExitBlock(Instruction & I,BasicBlock & ExitBlock,PHINode & PN,const LoopInfo * LI,const LoopSafetyInfo * SafetyInfo)585 CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN,
586 const LoopInfo *LI,
587 const LoopSafetyInfo *SafetyInfo) {
588 Instruction *New;
589 if (auto *CI = dyn_cast<CallInst>(&I)) {
590 const auto &BlockColors = SafetyInfo->BlockColors;
591
592 // Sinking call-sites need to be handled differently from other
593 // instructions. The cloned call-site needs a funclet bundle operand
594 // appropriate for it's location in the CFG.
595 SmallVector<OperandBundleDef, 1> OpBundles;
596 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
597 BundleIdx != BundleEnd; ++BundleIdx) {
598 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
599 if (Bundle.getTagID() == LLVMContext::OB_funclet)
600 continue;
601
602 OpBundles.emplace_back(Bundle);
603 }
604
605 if (!BlockColors.empty()) {
606 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
607 assert(CV.size() == 1 && "non-unique color for exit block!");
608 BasicBlock *BBColor = CV.front();
609 Instruction *EHPad = BBColor->getFirstNonPHI();
610 if (EHPad->isEHPad())
611 OpBundles.emplace_back("funclet", EHPad);
612 }
613
614 New = CallInst::Create(CI, OpBundles);
615 } else {
616 New = I.clone();
617 }
618
619 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
620 if (!I.getName().empty())
621 New->setName(I.getName() + ".le");
622
623 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
624 // particularly cheap because we can rip off the PHI node that we're
625 // replacing for the number and blocks of the predecessors.
626 // OPT: If this shows up in a profile, we can instead finish sinking all
627 // invariant instructions, and then walk their operands to re-establish
628 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
629 // sinking bottom-up.
630 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
631 ++OI)
632 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
633 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
634 if (!OLoop->contains(&PN)) {
635 PHINode *OpPN =
636 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
637 OInst->getName() + ".lcssa", &ExitBlock.front());
638 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
639 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
640 *OI = OpPN;
641 }
642 return New;
643 }
644
645 /// When an instruction is found to only be used outside of the loop, this
646 /// function moves it to the exit blocks and patches up SSA form as needed.
647 /// This method is guaranteed to remove the original instruction from its
648 /// position, and may either delete it or move it to outside of the loop.
649 ///
sink(Instruction & I,const LoopInfo * LI,const DominatorTree * DT,const Loop * CurLoop,AliasSetTracker * CurAST,const LoopSafetyInfo * SafetyInfo)650 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
651 const Loop *CurLoop, AliasSetTracker *CurAST,
652 const LoopSafetyInfo *SafetyInfo) {
653 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
654 bool Changed = false;
655 if (isa<LoadInst>(I))
656 ++NumMovedLoads;
657 else if (isa<CallInst>(I))
658 ++NumMovedCalls;
659 ++NumSunk;
660 Changed = true;
661
662 #ifndef NDEBUG
663 SmallVector<BasicBlock *, 32> ExitBlocks;
664 CurLoop->getUniqueExitBlocks(ExitBlocks);
665 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
666 ExitBlocks.end());
667 #endif
668
669 // Clones of this instruction. Don't create more than one per exit block!
670 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
671
672 // If this instruction is only used outside of the loop, then all users are
673 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
674 // the instruction.
675 while (!I.use_empty()) {
676 Value::user_iterator UI = I.user_begin();
677 auto *User = cast<Instruction>(*UI);
678 if (!DT->isReachableFromEntry(User->getParent())) {
679 User->replaceUsesOfWith(&I, UndefValue::get(I.getType()));
680 continue;
681 }
682 // The user must be a PHI node.
683 PHINode *PN = cast<PHINode>(User);
684
685 // Surprisingly, instructions can be used outside of loops without any
686 // exits. This can only happen in PHI nodes if the incoming block is
687 // unreachable.
688 Use &U = UI.getUse();
689 BasicBlock *BB = PN->getIncomingBlock(U);
690 if (!DT->isReachableFromEntry(BB)) {
691 U = UndefValue::get(I.getType());
692 continue;
693 }
694
695 BasicBlock *ExitBlock = PN->getParent();
696 assert(ExitBlockSet.count(ExitBlock) &&
697 "The LCSSA PHI is not in an exit block!");
698
699 Instruction *New;
700 auto It = SunkCopies.find(ExitBlock);
701 if (It != SunkCopies.end())
702 New = It->second;
703 else
704 New = SunkCopies[ExitBlock] =
705 CloneInstructionInExitBlock(I, *ExitBlock, *PN, LI, SafetyInfo);
706
707 PN->replaceAllUsesWith(New);
708 PN->eraseFromParent();
709 }
710
711 CurAST->deleteValue(&I);
712 I.eraseFromParent();
713 return Changed;
714 }
715
716 /// When an instruction is found to only use loop invariant operands that
717 /// is safe to hoist, this instruction is called to do the dirty work.
718 ///
hoist(Instruction & I,const DominatorTree * DT,const Loop * CurLoop,const LoopSafetyInfo * SafetyInfo)719 static bool hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
720 const LoopSafetyInfo *SafetyInfo) {
721 auto *Preheader = CurLoop->getLoopPreheader();
722 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " << I
723 << "\n");
724
725 // Metadata can be dependent on conditions we are hoisting above.
726 // Conservatively strip all metadata on the instruction unless we were
727 // guaranteed to execute I if we entered the loop, in which case the metadata
728 // is valid in the loop preheader.
729 if (I.hasMetadataOtherThanDebugLoc() &&
730 // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
731 // time in isGuaranteedToExecute if we don't actually have anything to
732 // drop. It is a compile time optimization, not required for correctness.
733 !isGuaranteedToExecute(I, DT, CurLoop, SafetyInfo))
734 I.dropUnknownNonDebugMetadata();
735
736 // Move the new node to the Preheader, before its terminator.
737 I.moveBefore(Preheader->getTerminator());
738
739 if (isa<LoadInst>(I))
740 ++NumMovedLoads;
741 else if (isa<CallInst>(I))
742 ++NumMovedCalls;
743 ++NumHoisted;
744 return true;
745 }
746
747 /// Only sink or hoist an instruction if it is not a trapping instruction,
748 /// or if the instruction is known not to trap when moved to the preheader.
749 /// or if it is a trapping instruction and is guaranteed to execute.
isSafeToExecuteUnconditionally(const Instruction & Inst,const DominatorTree * DT,const Loop * CurLoop,const LoopSafetyInfo * SafetyInfo,const Instruction * CtxI)750 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
751 const DominatorTree *DT,
752 const Loop *CurLoop,
753 const LoopSafetyInfo *SafetyInfo,
754 const Instruction *CtxI) {
755 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT))
756 return true;
757
758 return isGuaranteedToExecute(Inst, DT, CurLoop, SafetyInfo);
759 }
760
761 namespace {
762 class LoopPromoter : public LoadAndStorePromoter {
763 Value *SomePtr; // Designated pointer to store to.
764 SmallPtrSetImpl<Value *> &PointerMustAliases;
765 SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
766 SmallVectorImpl<Instruction *> &LoopInsertPts;
767 PredIteratorCache &PredCache;
768 AliasSetTracker &AST;
769 LoopInfo &LI;
770 DebugLoc DL;
771 int Alignment;
772 AAMDNodes AATags;
773
maybeInsertLCSSAPHI(Value * V,BasicBlock * BB) const774 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
775 if (Instruction *I = dyn_cast<Instruction>(V))
776 if (Loop *L = LI.getLoopFor(I->getParent()))
777 if (!L->contains(BB)) {
778 // We need to create an LCSSA PHI node for the incoming value and
779 // store that.
780 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
781 I->getName() + ".lcssa", &BB->front());
782 for (BasicBlock *Pred : PredCache.get(BB))
783 PN->addIncoming(I, Pred);
784 return PN;
785 }
786 return V;
787 }
788
789 public:
LoopPromoter(Value * SP,ArrayRef<const Instruction * > Insts,SSAUpdater & S,SmallPtrSetImpl<Value * > & PMA,SmallVectorImpl<BasicBlock * > & LEB,SmallVectorImpl<Instruction * > & LIP,PredIteratorCache & PIC,AliasSetTracker & ast,LoopInfo & li,DebugLoc dl,int alignment,const AAMDNodes & AATags)790 LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
791 SmallPtrSetImpl<Value *> &PMA,
792 SmallVectorImpl<BasicBlock *> &LEB,
793 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
794 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
795 const AAMDNodes &AATags)
796 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
797 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
798 LI(li), DL(std::move(dl)), Alignment(alignment), AATags(AATags) {}
799
isInstInList(Instruction * I,const SmallVectorImpl<Instruction * > &) const800 bool isInstInList(Instruction *I,
801 const SmallVectorImpl<Instruction *> &) const override {
802 Value *Ptr;
803 if (LoadInst *LI = dyn_cast<LoadInst>(I))
804 Ptr = LI->getOperand(0);
805 else
806 Ptr = cast<StoreInst>(I)->getPointerOperand();
807 return PointerMustAliases.count(Ptr);
808 }
809
doExtraRewritesBeforeFinalDeletion() const810 void doExtraRewritesBeforeFinalDeletion() const override {
811 // Insert stores after in the loop exit blocks. Each exit block gets a
812 // store of the live-out values that feed them. Since we've already told
813 // the SSA updater about the defs in the loop and the preheader
814 // definition, it is all set and we can start using it.
815 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
816 BasicBlock *ExitBlock = LoopExitBlocks[i];
817 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
818 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
819 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
820 Instruction *InsertPos = LoopInsertPts[i];
821 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
822 NewSI->setAlignment(Alignment);
823 NewSI->setDebugLoc(DL);
824 if (AATags)
825 NewSI->setAAMetadata(AATags);
826 }
827 }
828
replaceLoadWithValue(LoadInst * LI,Value * V) const829 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
830 // Update alias analysis.
831 AST.copyValue(LI, V);
832 }
instructionDeleted(Instruction * I) const833 void instructionDeleted(Instruction *I) const override { AST.deleteValue(I); }
834 };
835 } // end anon namespace
836
837 /// Try to promote memory values to scalars by sinking stores out of the
838 /// loop and moving loads to before the loop. We do this by looping over
839 /// the stores in the loop, looking for stores to Must pointers which are
840 /// loop invariant.
841 ///
promoteLoopAccessesToScalars(AliasSet & AS,SmallVectorImpl<BasicBlock * > & ExitBlocks,SmallVectorImpl<Instruction * > & InsertPts,PredIteratorCache & PIC,LoopInfo * LI,DominatorTree * DT,const TargetLibraryInfo * TLI,Loop * CurLoop,AliasSetTracker * CurAST,LoopSafetyInfo * SafetyInfo)842 bool llvm::promoteLoopAccessesToScalars(
843 AliasSet &AS, SmallVectorImpl<BasicBlock *> &ExitBlocks,
844 SmallVectorImpl<Instruction *> &InsertPts, PredIteratorCache &PIC,
845 LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
846 Loop *CurLoop, AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo) {
847 // Verify inputs.
848 assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&
849 CurAST != nullptr && SafetyInfo != nullptr &&
850 "Unexpected Input to promoteLoopAccessesToScalars");
851
852 // We can promote this alias set if it has a store, if it is a "Must" alias
853 // set, if the pointer is loop invariant, and if we are not eliminating any
854 // volatile loads or stores.
855 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
856 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
857 return false;
858
859 assert(!AS.empty() &&
860 "Must alias set should have at least one pointer element in it!");
861
862 Value *SomePtr = AS.begin()->getValue();
863 BasicBlock *Preheader = CurLoop->getLoopPreheader();
864
865 // It isn't safe to promote a load/store from the loop if the load/store is
866 // conditional. For example, turning:
867 //
868 // for () { if (c) *P += 1; }
869 //
870 // into:
871 //
872 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
873 //
874 // is not safe, because *P may only be valid to access if 'c' is true.
875 //
876 // The safety property divides into two parts:
877 // 1) The memory may not be dereferenceable on entry to the loop. In this
878 // case, we can't insert the required load in the preheader.
879 // 2) The memory model does not allow us to insert a store along any dynamic
880 // path which did not originally have one.
881 //
882 // It is safe to promote P if all uses are direct load/stores and if at
883 // least one is guaranteed to be executed.
884 bool GuaranteedToExecute = false;
885
886 // It is also safe to promote P if we can prove that speculating a load into
887 // the preheader is safe (i.e. proving dereferenceability on all
888 // paths through the loop), and that the memory can be proven thread local
889 // (so that the memory model requirement doesn't apply.) We first establish
890 // the former, and then run a capture analysis below to establish the later.
891 // We can use any access within the alias set to prove dereferenceability
892 // since they're all must alias.
893 bool CanSpeculateLoad = false;
894
895 SmallVector<Instruction *, 64> LoopUses;
896 SmallPtrSet<Value *, 4> PointerMustAliases;
897
898 // We start with an alignment of one and try to find instructions that allow
899 // us to prove better alignment.
900 unsigned Alignment = 1;
901 AAMDNodes AATags;
902 bool HasDedicatedExits = CurLoop->hasDedicatedExits();
903
904 // Don't sink stores from loops without dedicated block exits. Exits
905 // containing indirect branches are not transformed by loop simplify,
906 // make sure we catch that. An additional load may be generated in the
907 // preheader for SSA updater, so also avoid sinking when no preheader
908 // is available.
909 if (!HasDedicatedExits || !Preheader)
910 return false;
911
912 const DataLayout &MDL = Preheader->getModule()->getDataLayout();
913
914 if (SafetyInfo->MayThrow) {
915 // If a loop can throw, we have to insert a store along each unwind edge.
916 // That said, we can't actually make the unwind edge explicit. Therefore,
917 // we have to prove that the store is dead along the unwind edge.
918 //
919 // Currently, this code just special-cases alloca instructions.
920 if (!isa<AllocaInst>(GetUnderlyingObject(SomePtr, MDL)))
921 return false;
922 }
923
924 // Check that all of the pointers in the alias set have the same type. We
925 // cannot (yet) promote a memory location that is loaded and stored in
926 // different sizes. While we are at it, collect alignment and AA info.
927 bool Changed = false;
928 for (const auto &ASI : AS) {
929 Value *ASIV = ASI.getValue();
930 PointerMustAliases.insert(ASIV);
931
932 // Check that all of the pointers in the alias set have the same type. We
933 // cannot (yet) promote a memory location that is loaded and stored in
934 // different sizes.
935 if (SomePtr->getType() != ASIV->getType())
936 return Changed;
937
938 for (User *U : ASIV->users()) {
939 // Ignore instructions that are outside the loop.
940 Instruction *UI = dyn_cast<Instruction>(U);
941 if (!UI || !CurLoop->contains(UI))
942 continue;
943
944 // If there is an non-load/store instruction in the loop, we can't promote
945 // it.
946 if (const LoadInst *Load = dyn_cast<LoadInst>(UI)) {
947 assert(!Load->isVolatile() && "AST broken");
948 if (!Load->isSimple())
949 return Changed;
950
951 if (!GuaranteedToExecute && !CanSpeculateLoad)
952 CanSpeculateLoad = isSafeToExecuteUnconditionally(
953 *Load, DT, CurLoop, SafetyInfo, Preheader->getTerminator());
954 } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
955 // Stores *of* the pointer are not interesting, only stores *to* the
956 // pointer.
957 if (UI->getOperand(1) != ASIV)
958 continue;
959 assert(!Store->isVolatile() && "AST broken");
960 if (!Store->isSimple())
961 return Changed;
962
963 // Note that we only check GuaranteedToExecute inside the store case
964 // so that we do not introduce stores where they did not exist before
965 // (which would break the LLVM concurrency model).
966
967 // If the alignment of this instruction allows us to specify a more
968 // restrictive (and performant) alignment and if we are sure this
969 // instruction will be executed, update the alignment.
970 // Larger is better, with the exception of 0 being the best alignment.
971 unsigned InstAlignment = Store->getAlignment();
972 if ((InstAlignment > Alignment || InstAlignment == 0) &&
973 Alignment != 0) {
974 if (isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo)) {
975 GuaranteedToExecute = true;
976 Alignment = InstAlignment;
977 }
978 } else if (!GuaranteedToExecute) {
979 GuaranteedToExecute =
980 isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo);
981 }
982
983 if (!GuaranteedToExecute && !CanSpeculateLoad) {
984 CanSpeculateLoad = isDereferenceableAndAlignedPointer(
985 Store->getPointerOperand(), Store->getAlignment(), MDL,
986 Preheader->getTerminator(), DT);
987 }
988 } else
989 return Changed; // Not a load or store.
990
991 // Merge the AA tags.
992 if (LoopUses.empty()) {
993 // On the first load/store, just take its AA tags.
994 UI->getAAMetadata(AATags);
995 } else if (AATags) {
996 UI->getAAMetadata(AATags, /* Merge = */ true);
997 }
998
999 LoopUses.push_back(UI);
1000 }
1001 }
1002
1003 // Check legality per comment above. Otherwise, we can't promote.
1004 bool PromotionIsLegal = GuaranteedToExecute;
1005 if (!PromotionIsLegal && CanSpeculateLoad) {
1006 // If this is a thread local location, then we can insert stores along
1007 // paths which originally didn't have them without violating the memory
1008 // model.
1009 Value *Object = GetUnderlyingObject(SomePtr, MDL);
1010 PromotionIsLegal =
1011 isAllocLikeFn(Object, TLI) && !PointerMayBeCaptured(Object, true, true);
1012 }
1013 if (!PromotionIsLegal)
1014 return Changed;
1015
1016 // Figure out the loop exits and their insertion points, if this is the
1017 // first promotion.
1018 if (ExitBlocks.empty()) {
1019 CurLoop->getUniqueExitBlocks(ExitBlocks);
1020 InsertPts.clear();
1021 InsertPts.reserve(ExitBlocks.size());
1022 for (BasicBlock *ExitBlock : ExitBlocks)
1023 InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
1024 }
1025
1026 // Can't insert into a catchswitch.
1027 for (BasicBlock *ExitBlock : ExitBlocks)
1028 if (isa<CatchSwitchInst>(ExitBlock->getTerminator()))
1029 return Changed;
1030
1031 // Otherwise, this is safe to promote, lets do it!
1032 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr
1033 << '\n');
1034 Changed = true;
1035 ++NumPromoted;
1036
1037 // Grab a debug location for the inserted loads/stores; given that the
1038 // inserted loads/stores have little relation to the original loads/stores,
1039 // this code just arbitrarily picks a location from one, since any debug
1040 // location is better than none.
1041 DebugLoc DL = LoopUses[0]->getDebugLoc();
1042
1043 // We use the SSAUpdater interface to insert phi nodes as required.
1044 SmallVector<PHINode *, 16> NewPHIs;
1045 SSAUpdater SSA(&NewPHIs);
1046 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
1047 InsertPts, PIC, *CurAST, *LI, DL, Alignment, AATags);
1048
1049 // Set up the preheader to have a definition of the value. It is the live-out
1050 // value from the preheader that uses in the loop will use.
1051 LoadInst *PreheaderLoad = new LoadInst(
1052 SomePtr, SomePtr->getName() + ".promoted", Preheader->getTerminator());
1053 PreheaderLoad->setAlignment(Alignment);
1054 PreheaderLoad->setDebugLoc(DL);
1055 if (AATags)
1056 PreheaderLoad->setAAMetadata(AATags);
1057 SSA.AddAvailableValue(Preheader, PreheaderLoad);
1058
1059 // Rewrite all the loads in the loop and remember all the definitions from
1060 // stores in the loop.
1061 Promoter.run(LoopUses);
1062
1063 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
1064 if (PreheaderLoad->use_empty())
1065 PreheaderLoad->eraseFromParent();
1066
1067 return Changed;
1068 }
1069
1070 /// Returns an owning pointer to an alias set which incorporates aliasing info
1071 /// from L and all subloops of L.
1072 /// FIXME: In new pass manager, there is no helper functions to handle loop
1073 /// analysis such as cloneBasicBlockAnalysis. So the AST needs to be recompute
1074 /// from scratch for every loop. Hook up with the helper functions when
1075 /// available in the new pass manager to avoid redundant computation.
1076 AliasSetTracker *
collectAliasInfoForLoop(Loop * L,LoopInfo * LI,AliasAnalysis * AA)1077 LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
1078 AliasAnalysis *AA) {
1079 AliasSetTracker *CurAST = nullptr;
1080 SmallVector<Loop *, 4> RecomputeLoops;
1081 for (Loop *InnerL : L->getSubLoops()) {
1082 auto MapI = LoopToAliasSetMap.find(InnerL);
1083 // If the AST for this inner loop is missing it may have been merged into
1084 // some other loop's AST and then that loop unrolled, and so we need to
1085 // recompute it.
1086 if (MapI == LoopToAliasSetMap.end()) {
1087 RecomputeLoops.push_back(InnerL);
1088 continue;
1089 }
1090 AliasSetTracker *InnerAST = MapI->second;
1091
1092 if (CurAST != nullptr) {
1093 // What if InnerLoop was modified by other passes ?
1094 CurAST->add(*InnerAST);
1095
1096 // Once we've incorporated the inner loop's AST into ours, we don't need
1097 // the subloop's anymore.
1098 delete InnerAST;
1099 } else {
1100 CurAST = InnerAST;
1101 }
1102 LoopToAliasSetMap.erase(MapI);
1103 }
1104 if (CurAST == nullptr)
1105 CurAST = new AliasSetTracker(*AA);
1106
1107 auto mergeLoop = [&](Loop *L) {
1108 // Loop over the body of this loop, looking for calls, invokes, and stores.
1109 // Because subloops have already been incorporated into AST, we skip blocks
1110 // in subloops.
1111 for (BasicBlock *BB : L->blocks())
1112 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
1113 CurAST->add(*BB); // Incorporate the specified basic block
1114 };
1115
1116 // Add everything from the sub loops that are no longer directly available.
1117 for (Loop *InnerL : RecomputeLoops)
1118 mergeLoop(InnerL);
1119
1120 // And merge in this loop.
1121 mergeLoop(L);
1122
1123 return CurAST;
1124 }
1125
1126 /// Simple analysis hook. Clone alias set info.
1127 ///
cloneBasicBlockAnalysis(BasicBlock * From,BasicBlock * To,Loop * L)1128 void LegacyLICMPass::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
1129 Loop *L) {
1130 AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L);
1131 if (!AST)
1132 return;
1133
1134 AST->copyValue(From, To);
1135 }
1136
1137 /// Simple Analysis hook. Delete value V from alias set
1138 ///
deleteAnalysisValue(Value * V,Loop * L)1139 void LegacyLICMPass::deleteAnalysisValue(Value *V, Loop *L) {
1140 AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L);
1141 if (!AST)
1142 return;
1143
1144 AST->deleteValue(V);
1145 }
1146
1147 /// Simple Analysis hook. Delete value L from alias set map.
1148 ///
deleteAnalysisLoop(Loop * L)1149 void LegacyLICMPass::deleteAnalysisLoop(Loop *L) {
1150 AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L);
1151 if (!AST)
1152 return;
1153
1154 delete AST;
1155 LICM.getLoopToAliasSetMap().erase(L);
1156 }
1157
1158 /// Return true if the body of this loop may store into the memory
1159 /// location pointed to by V.
1160 ///
pointerInvalidatedByLoop(Value * V,uint64_t Size,const AAMDNodes & AAInfo,AliasSetTracker * CurAST)1161 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
1162 const AAMDNodes &AAInfo,
1163 AliasSetTracker *CurAST) {
1164 // Check to see if any of the basic blocks in CurLoop invalidate *V.
1165 return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod();
1166 }
1167
1168 /// Little predicate that returns true if the specified basic block is in
1169 /// a subloop of the current one, not the current one itself.
1170 ///
inSubLoop(BasicBlock * BB,Loop * CurLoop,LoopInfo * LI)1171 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
1172 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
1173 return LI->getLoopFor(BB) != CurLoop;
1174 }
1175