1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
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 pass performs loop invariant code motion, attempting to remove as much
10 // code from the body of a loop as possible. It does this by either hoisting
11 // code into the preheader block, or by sinking code to the exit blocks if it is
12 // safe. This pass also promotes must-aliased memory locations in the loop to
13 // live in registers, thus hoisting and sinking "invariant" loads and stores.
14 //
15 // Hoisting operations out of loops is a canonicalization transform. It
16 // enables and simplifies subsequent optimizations in the middle-end.
17 // Rematerialization of hoisted instructions to reduce register pressure is the
18 // responsibility of the back-end, which has more accurate information about
19 // register pressure and also handles other optimizations than LICM that
20 // increase live-ranges.
21 //
22 // This pass uses alias analysis for two purposes:
23 //
24 // 1. Moving loop invariant loads and calls out of loops. If we can determine
25 // that a load or call inside of a loop never aliases anything stored to,
26 // we can hoist it or sink it like any other instruction.
27 // 2. Scalar Promotion of Memory - If there is a store instruction inside of
28 // the loop, we try to move the store to happen AFTER the loop instead of
29 // inside of the loop. This can only happen if a few conditions are true:
30 // A. The pointer stored through is loop invariant
31 // B. There are no stores or loads in the loop which _may_ alias the
32 // pointer. There are no calls in the loop which mod/ref the pointer.
33 // If these conditions are true, we can promote the loads and stores in the
34 // loop of the pointer to use a temporary alloca'd variable. We then use
35 // the SSAUpdater to construct the appropriate SSA form for the value.
36 //
37 //===----------------------------------------------------------------------===//
38
39 #include "llvm/Transforms/Scalar/LICM.h"
40 #include "llvm/ADT/SetOperations.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/Analysis/AliasAnalysis.h"
43 #include "llvm/Analysis/AliasSetTracker.h"
44 #include "llvm/Analysis/BasicAliasAnalysis.h"
45 #include "llvm/Analysis/BlockFrequencyInfo.h"
46 #include "llvm/Analysis/CaptureTracking.h"
47 #include "llvm/Analysis/ConstantFolding.h"
48 #include "llvm/Analysis/GlobalsModRef.h"
49 #include "llvm/Analysis/GuardUtils.h"
50 #include "llvm/Analysis/LazyBlockFrequencyInfo.h"
51 #include "llvm/Analysis/Loads.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopIterator.h"
54 #include "llvm/Analysis/LoopPass.h"
55 #include "llvm/Analysis/MemoryBuiltins.h"
56 #include "llvm/Analysis/MemorySSA.h"
57 #include "llvm/Analysis/MemorySSAUpdater.h"
58 #include "llvm/Analysis/MustExecute.h"
59 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
60 #include "llvm/Analysis/ScalarEvolution.h"
61 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
62 #include "llvm/Analysis/TargetLibraryInfo.h"
63 #include "llvm/Analysis/ValueTracking.h"
64 #include "llvm/IR/CFG.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DebugInfoMetadata.h"
68 #include "llvm/IR/DerivedTypes.h"
69 #include "llvm/IR/Dominators.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/IntrinsicInst.h"
72 #include "llvm/IR/LLVMContext.h"
73 #include "llvm/IR/Metadata.h"
74 #include "llvm/IR/PatternMatch.h"
75 #include "llvm/IR/PredIteratorCache.h"
76 #include "llvm/InitializePasses.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/Debug.h"
79 #include "llvm/Support/raw_ostream.h"
80 #include "llvm/Transforms/Scalar.h"
81 #include "llvm/Transforms/Scalar/LoopPassManager.h"
82 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
83 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
84 #include "llvm/Transforms/Utils/Local.h"
85 #include "llvm/Transforms/Utils/LoopUtils.h"
86 #include "llvm/Transforms/Utils/SSAUpdater.h"
87 #include <algorithm>
88 #include <utility>
89 using namespace llvm;
90
91 #define DEBUG_TYPE "licm"
92
93 STATISTIC(NumCreatedBlocks, "Number of blocks created");
94 STATISTIC(NumClonedBranches, "Number of branches cloned");
95 STATISTIC(NumSunk, "Number of instructions sunk out of loop");
96 STATISTIC(NumHoisted, "Number of instructions hoisted out of loop");
97 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
98 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
99 STATISTIC(NumPromoted, "Number of memory locations promoted to registers");
100
101 /// Memory promotion is enabled by default.
102 static cl::opt<bool>
103 DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false),
104 cl::desc("Disable memory promotion in LICM pass"));
105
106 static cl::opt<bool> ControlFlowHoisting(
107 "licm-control-flow-hoisting", cl::Hidden, cl::init(false),
108 cl::desc("Enable control flow (and PHI) hoisting in LICM"));
109
110 static cl::opt<unsigned> HoistSinkColdnessThreshold(
111 "licm-coldness-threshold", cl::Hidden, cl::init(4),
112 cl::desc("Relative coldness Threshold of hoisting/sinking destination "
113 "block for LICM to be considered beneficial"));
114
115 static cl::opt<uint32_t> MaxNumUsesTraversed(
116 "licm-max-num-uses-traversed", cl::Hidden, cl::init(8),
117 cl::desc("Max num uses visited for identifying load "
118 "invariance in loop using invariant start (default = 8)"));
119
120 // Default value of zero implies we use the regular alias set tracker mechanism
121 // instead of the cross product using AA to identify aliasing of the memory
122 // location we are interested in.
123 static cl::opt<int>
124 LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0),
125 cl::desc("How many instruction to cross product using AA"));
126
127 // Experimental option to allow imprecision in LICM in pathological cases, in
128 // exchange for faster compile. This is to be removed if MemorySSA starts to
129 // address the same issue. This flag applies only when LICM uses MemorySSA
130 // instead on AliasSetTracker. LICM calls MemorySSAWalker's
131 // getClobberingMemoryAccess, up to the value of the Cap, getting perfect
132 // accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess,
133 // which may not be precise, since optimizeUses is capped. The result is
134 // correct, but we may not get as "far up" as possible to get which access is
135 // clobbering the one queried.
136 cl::opt<unsigned> llvm::SetLicmMssaOptCap(
137 "licm-mssa-optimization-cap", cl::init(100), cl::Hidden,
138 cl::desc("Enable imprecision in LICM in pathological cases, in exchange "
139 "for faster compile. Caps the MemorySSA clobbering calls."));
140
141 // Experimentally, memory promotion carries less importance than sinking and
142 // hoisting. Limit when we do promotion when using MemorySSA, in order to save
143 // compile time.
144 cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap(
145 "licm-mssa-max-acc-promotion", cl::init(250), cl::Hidden,
146 cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no "
147 "effect. When MSSA in LICM is enabled, then this is the maximum "
148 "number of accesses allowed to be present in a loop in order to "
149 "enable memory promotion."));
150
151 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
152 static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
153 const LoopSafetyInfo *SafetyInfo,
154 TargetTransformInfo *TTI, bool &FreeInLoop);
155 static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
156 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
157 MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
158 OptimizationRemarkEmitter *ORE);
159 static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
160 BlockFrequencyInfo *BFI, const Loop *CurLoop,
161 ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU,
162 OptimizationRemarkEmitter *ORE);
163 static bool isSafeToExecuteUnconditionally(Instruction &Inst,
164 const DominatorTree *DT,
165 const Loop *CurLoop,
166 const LoopSafetyInfo *SafetyInfo,
167 OptimizationRemarkEmitter *ORE,
168 const Instruction *CtxI = nullptr);
169 static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
170 AliasSetTracker *CurAST, Loop *CurLoop,
171 AAResults *AA);
172 static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
173 Loop *CurLoop, Instruction &I,
174 SinkAndHoistLICMFlags &Flags);
175 static bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA,
176 MemoryUse &MU);
177 static Instruction *cloneInstructionInExitBlock(
178 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
179 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU);
180
181 static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
182 AliasSetTracker *AST, MemorySSAUpdater *MSSAU);
183
184 static void moveInstructionBefore(Instruction &I, Instruction &Dest,
185 ICFLoopSafetyInfo &SafetyInfo,
186 MemorySSAUpdater *MSSAU, ScalarEvolution *SE);
187
188 namespace {
189 struct LoopInvariantCodeMotion {
190 bool runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT,
191 BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI,
192 TargetTransformInfo *TTI, ScalarEvolution *SE, MemorySSA *MSSA,
193 OptimizationRemarkEmitter *ORE);
194
LoopInvariantCodeMotion__anonfa121ca40111::LoopInvariantCodeMotion195 LoopInvariantCodeMotion(unsigned LicmMssaOptCap,
196 unsigned LicmMssaNoAccForPromotionCap)
197 : LicmMssaOptCap(LicmMssaOptCap),
198 LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {}
199
200 private:
201 unsigned LicmMssaOptCap;
202 unsigned LicmMssaNoAccForPromotionCap;
203
204 std::unique_ptr<AliasSetTracker>
205 collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AAResults *AA);
206 std::unique_ptr<AliasSetTracker>
207 collectAliasInfoForLoopWithMSSA(Loop *L, AAResults *AA,
208 MemorySSAUpdater *MSSAU);
209 };
210
211 struct LegacyLICMPass : public LoopPass {
212 static char ID; // Pass identification, replacement for typeid
LegacyLICMPass__anonfa121ca40111::LegacyLICMPass213 LegacyLICMPass(
214 unsigned LicmMssaOptCap = SetLicmMssaOptCap,
215 unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap)
216 : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) {
217 initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
218 }
219
runOnLoop__anonfa121ca40111::LegacyLICMPass220 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
221 if (skipLoop(L))
222 return false;
223
224 LLVM_DEBUG(dbgs() << "Perform LICM on Loop with header at block "
225 << L->getHeader()->getNameOrAsOperand() << "\n");
226
227 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
228 MemorySSA *MSSA = EnableMSSALoopDependency
229 ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA())
230 : nullptr;
231 bool hasProfileData = L->getHeader()->getParent()->hasProfileData();
232 BlockFrequencyInfo *BFI =
233 hasProfileData ? &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI()
234 : nullptr;
235 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
236 // pass. Function analyses need to be preserved across loop transformations
237 // but ORE cannot be preserved (see comment before the pass definition).
238 OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
239 return LICM.runOnLoop(
240 L, &getAnalysis<AAResultsWrapperPass>().getAAResults(),
241 &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
242 &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), BFI,
243 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
244 *L->getHeader()->getParent()),
245 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
246 *L->getHeader()->getParent()),
247 SE ? &SE->getSE() : nullptr, MSSA, &ORE);
248 }
249
250 /// This transformation requires natural loop information & requires that
251 /// loop preheaders be inserted into the CFG...
252 ///
getAnalysisUsage__anonfa121ca40111::LegacyLICMPass253 void getAnalysisUsage(AnalysisUsage &AU) const override {
254 AU.addPreserved<DominatorTreeWrapperPass>();
255 AU.addPreserved<LoopInfoWrapperPass>();
256 AU.addRequired<TargetLibraryInfoWrapperPass>();
257 if (EnableMSSALoopDependency) {
258 AU.addRequired<MemorySSAWrapperPass>();
259 AU.addPreserved<MemorySSAWrapperPass>();
260 }
261 AU.addRequired<TargetTransformInfoWrapperPass>();
262 getLoopAnalysisUsage(AU);
263 LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
264 AU.addPreserved<LazyBlockFrequencyInfoPass>();
265 AU.addPreserved<LazyBranchProbabilityInfoPass>();
266 }
267
268 private:
269 LoopInvariantCodeMotion LICM;
270 };
271 } // namespace
272
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater &)273 PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM,
274 LoopStandardAnalysisResults &AR, LPMUpdater &) {
275 // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
276 // pass. Function analyses need to be preserved across loop transformations
277 // but ORE cannot be preserved (see comment before the pass definition).
278 OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
279
280 LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
281 if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, AR.BFI, &AR.TLI, &AR.TTI,
282 &AR.SE, AR.MSSA, &ORE))
283 return PreservedAnalyses::all();
284
285 auto PA = getLoopPassPreservedAnalyses();
286
287 PA.preserve<DominatorTreeAnalysis>();
288 PA.preserve<LoopAnalysis>();
289 if (AR.MSSA)
290 PA.preserve<MemorySSAAnalysis>();
291
292 return PA;
293 }
294
295 char LegacyLICMPass::ID = 0;
296 INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",
297 false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)298 INITIALIZE_PASS_DEPENDENCY(LoopPass)
299 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
300 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
301 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
302 INITIALIZE_PASS_DEPENDENCY(LazyBFIPass)
303 INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,
304 false)
305
306 Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
createLICMPass(unsigned LicmMssaOptCap,unsigned LicmMssaNoAccForPromotionCap)307 Pass *llvm::createLICMPass(unsigned LicmMssaOptCap,
308 unsigned LicmMssaNoAccForPromotionCap) {
309 return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
310 }
311
SinkAndHoistLICMFlags(bool IsSink,Loop * L,MemorySSA * MSSA)312 llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(bool IsSink, Loop *L,
313 MemorySSA *MSSA)
314 : SinkAndHoistLICMFlags(SetLicmMssaOptCap, SetLicmMssaNoAccForPromotionCap,
315 IsSink, L, MSSA) {}
316
SinkAndHoistLICMFlags(unsigned LicmMssaOptCap,unsigned LicmMssaNoAccForPromotionCap,bool IsSink,Loop * L,MemorySSA * MSSA)317 llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(
318 unsigned LicmMssaOptCap, unsigned LicmMssaNoAccForPromotionCap, bool IsSink,
319 Loop *L, MemorySSA *MSSA)
320 : LicmMssaOptCap(LicmMssaOptCap),
321 LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap),
322 IsSink(IsSink) {
323 assert(((L != nullptr) == (MSSA != nullptr)) &&
324 "Unexpected values for SinkAndHoistLICMFlags");
325 if (!MSSA)
326 return;
327
328 unsigned AccessCapCount = 0;
329 for (auto *BB : L->getBlocks())
330 if (const auto *Accesses = MSSA->getBlockAccesses(BB))
331 for (const auto &MA : *Accesses) {
332 (void)MA;
333 ++AccessCapCount;
334 if (AccessCapCount > LicmMssaNoAccForPromotionCap) {
335 NoOfMemAccTooLarge = true;
336 return;
337 }
338 }
339 }
340
341 /// Hoist expressions out of the specified loop. Note, alias info for inner
342 /// loop is not preserved so it is not a good idea to run LICM multiple
343 /// times on one loop.
runOnLoop(Loop * L,AAResults * AA,LoopInfo * LI,DominatorTree * DT,BlockFrequencyInfo * BFI,TargetLibraryInfo * TLI,TargetTransformInfo * TTI,ScalarEvolution * SE,MemorySSA * MSSA,OptimizationRemarkEmitter * ORE)344 bool LoopInvariantCodeMotion::runOnLoop(
345 Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT,
346 BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
347 ScalarEvolution *SE, MemorySSA *MSSA, OptimizationRemarkEmitter *ORE) {
348 bool Changed = false;
349
350 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
351
352 // If this loop has metadata indicating that LICM is not to be performed then
353 // just exit.
354 if (hasDisableLICMTransformsHint(L)) {
355 return false;
356 }
357
358 std::unique_ptr<AliasSetTracker> CurAST;
359 std::unique_ptr<MemorySSAUpdater> MSSAU;
360 std::unique_ptr<SinkAndHoistLICMFlags> Flags;
361
362 if (!MSSA) {
363 LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n");
364 CurAST = collectAliasInfoForLoop(L, LI, AA);
365 Flags = std::make_unique<SinkAndHoistLICMFlags>(
366 LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true);
367 } else {
368 LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n");
369 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
370 Flags = std::make_unique<SinkAndHoistLICMFlags>(
371 LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true, L, MSSA);
372 }
373
374 // Get the preheader block to move instructions into...
375 BasicBlock *Preheader = L->getLoopPreheader();
376
377 // Compute loop safety information.
378 ICFLoopSafetyInfo SafetyInfo;
379 SafetyInfo.computeLoopSafetyInfo(L);
380
381 // We want to visit all of the instructions in this loop... that are not parts
382 // of our subloops (they have already had their invariants hoisted out of
383 // their loop, into this loop, so there is no need to process the BODIES of
384 // the subloops).
385 //
386 // Traverse the body of the loop in depth first order on the dominator tree so
387 // that we are guaranteed to see definitions before we see uses. This allows
388 // us to sink instructions in one pass, without iteration. After sinking
389 // instructions, we perform another pass to hoist them out of the loop.
390 if (L->hasDedicatedExits())
391 Changed |=
392 sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, TTI, L,
393 CurAST.get(), MSSAU.get(), &SafetyInfo, *Flags.get(), ORE);
394 Flags->setIsSink(false);
395 if (Preheader)
396 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, L,
397 CurAST.get(), MSSAU.get(), SE, &SafetyInfo,
398 *Flags.get(), ORE);
399
400 // Now that all loop invariants have been removed from the loop, promote any
401 // memory references to scalars that we can.
402 // Don't sink stores from loops without dedicated block exits. Exits
403 // containing indirect branches are not transformed by loop simplify,
404 // make sure we catch that. An additional load may be generated in the
405 // preheader for SSA updater, so also avoid sinking when no preheader
406 // is available.
407 if (!DisablePromotion && Preheader && L->hasDedicatedExits() &&
408 !Flags->tooManyMemoryAccesses()) {
409 // Figure out the loop exits and their insertion points
410 SmallVector<BasicBlock *, 8> ExitBlocks;
411 L->getUniqueExitBlocks(ExitBlocks);
412
413 // We can't insert into a catchswitch.
414 bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) {
415 return isa<CatchSwitchInst>(Exit->getTerminator());
416 });
417
418 if (!HasCatchSwitch) {
419 SmallVector<Instruction *, 8> InsertPts;
420 SmallVector<MemoryAccess *, 8> MSSAInsertPts;
421 InsertPts.reserve(ExitBlocks.size());
422 if (MSSAU)
423 MSSAInsertPts.reserve(ExitBlocks.size());
424 for (BasicBlock *ExitBlock : ExitBlocks) {
425 InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
426 if (MSSAU)
427 MSSAInsertPts.push_back(nullptr);
428 }
429
430 PredIteratorCache PIC;
431
432 bool Promoted = false;
433
434 // Build an AST using MSSA.
435 if (!CurAST.get())
436 CurAST = collectAliasInfoForLoopWithMSSA(L, AA, MSSAU.get());
437
438 // Loop over all of the alias sets in the tracker object.
439 for (AliasSet &AS : *CurAST) {
440 // We can promote this alias set if it has a store, if it is a "Must"
441 // alias set, if the pointer is loop invariant, and if we are not
442 // eliminating any volatile loads or stores.
443 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
444 !L->isLoopInvariant(AS.begin()->getValue()))
445 continue;
446
447 assert(
448 !AS.empty() &&
449 "Must alias set should have at least one pointer element in it!");
450
451 SmallSetVector<Value *, 8> PointerMustAliases;
452 for (const auto &ASI : AS)
453 PointerMustAliases.insert(ASI.getValue());
454
455 Promoted |= promoteLoopAccessesToScalars(
456 PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI,
457 DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE);
458 }
459
460 // Once we have promoted values across the loop body we have to
461 // recursively reform LCSSA as any nested loop may now have values defined
462 // within the loop used in the outer loop.
463 // FIXME: This is really heavy handed. It would be a bit better to use an
464 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
465 // it as it went.
466 if (Promoted)
467 formLCSSARecursively(*L, *DT, LI, SE);
468
469 Changed |= Promoted;
470 }
471 }
472
473 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
474 // specifically moving instructions across the loop boundary and so it is
475 // especially in need of sanity checking here.
476 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
477 assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) &&
478 "Parent loop not left in LCSSA form after LICM!");
479
480 if (MSSAU.get() && VerifyMemorySSA)
481 MSSAU->getMemorySSA()->verifyMemorySSA();
482
483 if (Changed && SE)
484 SE->forgetLoopDispositions(L);
485 return Changed;
486 }
487
488 /// Walk the specified region of the CFG (defined by all blocks dominated by
489 /// the specified block, and that are in the current loop) in reverse depth
490 /// first order w.r.t the DominatorTree. This allows us to visit uses before
491 /// definitions, allowing us to sink a loop body in one pass without iteration.
492 ///
sinkRegion(DomTreeNode * N,AAResults * AA,LoopInfo * LI,DominatorTree * DT,BlockFrequencyInfo * BFI,TargetLibraryInfo * TLI,TargetTransformInfo * TTI,Loop * CurLoop,AliasSetTracker * CurAST,MemorySSAUpdater * MSSAU,ICFLoopSafetyInfo * SafetyInfo,SinkAndHoistLICMFlags & Flags,OptimizationRemarkEmitter * ORE)493 bool llvm::sinkRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI,
494 DominatorTree *DT, BlockFrequencyInfo *BFI,
495 TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
496 Loop *CurLoop, AliasSetTracker *CurAST,
497 MemorySSAUpdater *MSSAU, ICFLoopSafetyInfo *SafetyInfo,
498 SinkAndHoistLICMFlags &Flags,
499 OptimizationRemarkEmitter *ORE) {
500
501 // Verify inputs.
502 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
503 CurLoop != nullptr && SafetyInfo != nullptr &&
504 "Unexpected input to sinkRegion.");
505 assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
506 "Either AliasSetTracker or MemorySSA should be initialized.");
507
508 // We want to visit children before parents. We will enque all the parents
509 // before their children in the worklist and process the worklist in reverse
510 // order.
511 SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop);
512
513 bool Changed = false;
514 for (DomTreeNode *DTN : reverse(Worklist)) {
515 BasicBlock *BB = DTN->getBlock();
516 // Only need to process the contents of this block if it is not part of a
517 // subloop (which would already have been processed).
518 if (inSubLoop(BB, CurLoop, LI))
519 continue;
520
521 for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
522 Instruction &I = *--II;
523
524 // If the instruction is dead, we would try to sink it because it isn't
525 // used in the loop, instead, just delete it.
526 if (isInstructionTriviallyDead(&I, TLI)) {
527 LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
528 salvageKnowledge(&I);
529 salvageDebugInfo(I);
530 ++II;
531 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
532 Changed = true;
533 continue;
534 }
535
536 // Check to see if we can sink this instruction to the exit blocks
537 // of the loop. We can do this if the all users of the instruction are
538 // outside of the loop. In this case, it doesn't even matter if the
539 // operands of the instruction are loop invariant.
540 //
541 bool FreeInLoop = false;
542 if (!I.mayHaveSideEffects() &&
543 isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) &&
544 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
545 ORE)) {
546 if (sink(I, LI, DT, BFI, CurLoop, SafetyInfo, MSSAU, ORE)) {
547 if (!FreeInLoop) {
548 ++II;
549 salvageDebugInfo(I);
550 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
551 }
552 Changed = true;
553 }
554 }
555 }
556 }
557 if (MSSAU && VerifyMemorySSA)
558 MSSAU->getMemorySSA()->verifyMemorySSA();
559 return Changed;
560 }
561
562 namespace {
563 // This is a helper class for hoistRegion to make it able to hoist control flow
564 // in order to be able to hoist phis. The way this works is that we initially
565 // start hoisting to the loop preheader, and when we see a loop invariant branch
566 // we make note of this. When we then come to hoist an instruction that's
567 // conditional on such a branch we duplicate the branch and the relevant control
568 // flow, then hoist the instruction into the block corresponding to its original
569 // block in the duplicated control flow.
570 class ControlFlowHoister {
571 private:
572 // Information about the loop we are hoisting from
573 LoopInfo *LI;
574 DominatorTree *DT;
575 Loop *CurLoop;
576 MemorySSAUpdater *MSSAU;
577
578 // A map of blocks in the loop to the block their instructions will be hoisted
579 // to.
580 DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap;
581
582 // The branches that we can hoist, mapped to the block that marks a
583 // convergence point of their control flow.
584 DenseMap<BranchInst *, BasicBlock *> HoistableBranches;
585
586 public:
ControlFlowHoister(LoopInfo * LI,DominatorTree * DT,Loop * CurLoop,MemorySSAUpdater * MSSAU)587 ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop,
588 MemorySSAUpdater *MSSAU)
589 : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {}
590
registerPossiblyHoistableBranch(BranchInst * BI)591 void registerPossiblyHoistableBranch(BranchInst *BI) {
592 // We can only hoist conditional branches with loop invariant operands.
593 if (!ControlFlowHoisting || !BI->isConditional() ||
594 !CurLoop->hasLoopInvariantOperands(BI))
595 return;
596
597 // The branch destinations need to be in the loop, and we don't gain
598 // anything by duplicating conditional branches with duplicate successors,
599 // as it's essentially the same as an unconditional branch.
600 BasicBlock *TrueDest = BI->getSuccessor(0);
601 BasicBlock *FalseDest = BI->getSuccessor(1);
602 if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) ||
603 TrueDest == FalseDest)
604 return;
605
606 // We can hoist BI if one branch destination is the successor of the other,
607 // or both have common successor which we check by seeing if the
608 // intersection of their successors is non-empty.
609 // TODO: This could be expanded to allowing branches where both ends
610 // eventually converge to a single block.
611 SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc;
612 TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest));
613 FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest));
614 BasicBlock *CommonSucc = nullptr;
615 if (TrueDestSucc.count(FalseDest)) {
616 CommonSucc = FalseDest;
617 } else if (FalseDestSucc.count(TrueDest)) {
618 CommonSucc = TrueDest;
619 } else {
620 set_intersect(TrueDestSucc, FalseDestSucc);
621 // If there's one common successor use that.
622 if (TrueDestSucc.size() == 1)
623 CommonSucc = *TrueDestSucc.begin();
624 // If there's more than one pick whichever appears first in the block list
625 // (we can't use the value returned by TrueDestSucc.begin() as it's
626 // unpredicatable which element gets returned).
627 else if (!TrueDestSucc.empty()) {
628 Function *F = TrueDest->getParent();
629 auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); };
630 auto It = llvm::find_if(*F, IsSucc);
631 assert(It != F->end() && "Could not find successor in function");
632 CommonSucc = &*It;
633 }
634 }
635 // The common successor has to be dominated by the branch, as otherwise
636 // there will be some other path to the successor that will not be
637 // controlled by this branch so any phi we hoist would be controlled by the
638 // wrong condition. This also takes care of avoiding hoisting of loop back
639 // edges.
640 // TODO: In some cases this could be relaxed if the successor is dominated
641 // by another block that's been hoisted and we can guarantee that the
642 // control flow has been replicated exactly.
643 if (CommonSucc && DT->dominates(BI, CommonSucc))
644 HoistableBranches[BI] = CommonSucc;
645 }
646
canHoistPHI(PHINode * PN)647 bool canHoistPHI(PHINode *PN) {
648 // The phi must have loop invariant operands.
649 if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN))
650 return false;
651 // We can hoist phis if the block they are in is the target of hoistable
652 // branches which cover all of the predecessors of the block.
653 SmallPtrSet<BasicBlock *, 8> PredecessorBlocks;
654 BasicBlock *BB = PN->getParent();
655 for (BasicBlock *PredBB : predecessors(BB))
656 PredecessorBlocks.insert(PredBB);
657 // If we have less predecessor blocks than predecessors then the phi will
658 // have more than one incoming value for the same block which we can't
659 // handle.
660 // TODO: This could be handled be erasing some of the duplicate incoming
661 // values.
662 if (PredecessorBlocks.size() != pred_size(BB))
663 return false;
664 for (auto &Pair : HoistableBranches) {
665 if (Pair.second == BB) {
666 // Which blocks are predecessors via this branch depends on if the
667 // branch is triangle-like or diamond-like.
668 if (Pair.first->getSuccessor(0) == BB) {
669 PredecessorBlocks.erase(Pair.first->getParent());
670 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
671 } else if (Pair.first->getSuccessor(1) == BB) {
672 PredecessorBlocks.erase(Pair.first->getParent());
673 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
674 } else {
675 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
676 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
677 }
678 }
679 }
680 // PredecessorBlocks will now be empty if for every predecessor of BB we
681 // found a hoistable branch source.
682 return PredecessorBlocks.empty();
683 }
684
getOrCreateHoistedBlock(BasicBlock * BB)685 BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) {
686 if (!ControlFlowHoisting)
687 return CurLoop->getLoopPreheader();
688 // If BB has already been hoisted, return that
689 if (HoistDestinationMap.count(BB))
690 return HoistDestinationMap[BB];
691
692 // Check if this block is conditional based on a pending branch
693 auto HasBBAsSuccessor =
694 [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) {
695 return BB != Pair.second && (Pair.first->getSuccessor(0) == BB ||
696 Pair.first->getSuccessor(1) == BB);
697 };
698 auto It = llvm::find_if(HoistableBranches, HasBBAsSuccessor);
699
700 // If not involved in a pending branch, hoist to preheader
701 BasicBlock *InitialPreheader = CurLoop->getLoopPreheader();
702 if (It == HoistableBranches.end()) {
703 LLVM_DEBUG(dbgs() << "LICM using "
704 << InitialPreheader->getNameOrAsOperand()
705 << " as hoist destination for "
706 << BB->getNameOrAsOperand() << "\n");
707 HoistDestinationMap[BB] = InitialPreheader;
708 return InitialPreheader;
709 }
710 BranchInst *BI = It->first;
711 assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==
712 HoistableBranches.end() &&
713 "BB is expected to be the target of at most one branch");
714
715 LLVMContext &C = BB->getContext();
716 BasicBlock *TrueDest = BI->getSuccessor(0);
717 BasicBlock *FalseDest = BI->getSuccessor(1);
718 BasicBlock *CommonSucc = HoistableBranches[BI];
719 BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent());
720
721 // Create hoisted versions of blocks that currently don't have them
722 auto CreateHoistedBlock = [&](BasicBlock *Orig) {
723 if (HoistDestinationMap.count(Orig))
724 return HoistDestinationMap[Orig];
725 BasicBlock *New =
726 BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent());
727 HoistDestinationMap[Orig] = New;
728 DT->addNewBlock(New, HoistTarget);
729 if (CurLoop->getParentLoop())
730 CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI);
731 ++NumCreatedBlocks;
732 LLVM_DEBUG(dbgs() << "LICM created " << New->getName()
733 << " as hoist destination for " << Orig->getName()
734 << "\n");
735 return New;
736 };
737 BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest);
738 BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest);
739 BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc);
740
741 // Link up these blocks with branches.
742 if (!HoistCommonSucc->getTerminator()) {
743 // The new common successor we've generated will branch to whatever that
744 // hoist target branched to.
745 BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor();
746 assert(TargetSucc && "Expected hoist target to have a single successor");
747 HoistCommonSucc->moveBefore(TargetSucc);
748 BranchInst::Create(TargetSucc, HoistCommonSucc);
749 }
750 if (!HoistTrueDest->getTerminator()) {
751 HoistTrueDest->moveBefore(HoistCommonSucc);
752 BranchInst::Create(HoistCommonSucc, HoistTrueDest);
753 }
754 if (!HoistFalseDest->getTerminator()) {
755 HoistFalseDest->moveBefore(HoistCommonSucc);
756 BranchInst::Create(HoistCommonSucc, HoistFalseDest);
757 }
758
759 // If BI is being cloned to what was originally the preheader then
760 // HoistCommonSucc will now be the new preheader.
761 if (HoistTarget == InitialPreheader) {
762 // Phis in the loop header now need to use the new preheader.
763 InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc);
764 if (MSSAU)
765 MSSAU->wireOldPredecessorsToNewImmediatePredecessor(
766 HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget});
767 // The new preheader dominates the loop header.
768 DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc);
769 DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader());
770 DT->changeImmediateDominator(HeaderNode, PreheaderNode);
771 // The preheader hoist destination is now the new preheader, with the
772 // exception of the hoist destination of this branch.
773 for (auto &Pair : HoistDestinationMap)
774 if (Pair.second == InitialPreheader && Pair.first != BI->getParent())
775 Pair.second = HoistCommonSucc;
776 }
777
778 // Now finally clone BI.
779 ReplaceInstWithInst(
780 HoistTarget->getTerminator(),
781 BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition()));
782 ++NumClonedBranches;
783
784 assert(CurLoop->getLoopPreheader() &&
785 "Hoisting blocks should not have destroyed preheader");
786 return HoistDestinationMap[BB];
787 }
788 };
789 } // namespace
790
791 // Hoisting/sinking instruction out of a loop isn't always beneficial. It's only
792 // only worthwhile if the destination block is actually colder than current
793 // block.
worthSinkOrHoistInst(Instruction & I,BasicBlock * DstBlock,OptimizationRemarkEmitter * ORE,BlockFrequencyInfo * BFI)794 static bool worthSinkOrHoistInst(Instruction &I, BasicBlock *DstBlock,
795 OptimizationRemarkEmitter *ORE,
796 BlockFrequencyInfo *BFI) {
797 // Check block frequency only when runtime profile is available
798 // to avoid pathological cases. With static profile, lean towards
799 // hosting because it helps canonicalize the loop for vectorizer.
800 if (!DstBlock->getParent()->hasProfileData())
801 return true;
802
803 if (!HoistSinkColdnessThreshold || !BFI)
804 return true;
805
806 BasicBlock *SrcBlock = I.getParent();
807 if (BFI->getBlockFreq(DstBlock).getFrequency() / HoistSinkColdnessThreshold >
808 BFI->getBlockFreq(SrcBlock).getFrequency()) {
809 ORE->emit([&]() {
810 return OptimizationRemarkMissed(DEBUG_TYPE, "SinkHoistInst", &I)
811 << "failed to sink or hoist instruction because containing block "
812 "has lower frequency than destination block";
813 });
814 return false;
815 }
816
817 return true;
818 }
819
820 /// Walk the specified region of the CFG (defined by all blocks dominated by
821 /// the specified block, and that are in the current loop) in depth first
822 /// order w.r.t the DominatorTree. This allows us to visit definitions before
823 /// uses, allowing us to hoist a loop body in one pass without iteration.
824 ///
hoistRegion(DomTreeNode * N,AAResults * AA,LoopInfo * LI,DominatorTree * DT,BlockFrequencyInfo * BFI,TargetLibraryInfo * TLI,Loop * CurLoop,AliasSetTracker * CurAST,MemorySSAUpdater * MSSAU,ScalarEvolution * SE,ICFLoopSafetyInfo * SafetyInfo,SinkAndHoistLICMFlags & Flags,OptimizationRemarkEmitter * ORE)825 bool llvm::hoistRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI,
826 DominatorTree *DT, BlockFrequencyInfo *BFI,
827 TargetLibraryInfo *TLI, Loop *CurLoop,
828 AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
829 ScalarEvolution *SE, ICFLoopSafetyInfo *SafetyInfo,
830 SinkAndHoistLICMFlags &Flags,
831 OptimizationRemarkEmitter *ORE) {
832 // Verify inputs.
833 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
834 CurLoop != nullptr && SafetyInfo != nullptr &&
835 "Unexpected input to hoistRegion.");
836 assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
837 "Either AliasSetTracker or MemorySSA should be initialized.");
838
839 ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU);
840
841 // Keep track of instructions that have been hoisted, as they may need to be
842 // re-hoisted if they end up not dominating all of their uses.
843 SmallVector<Instruction *, 16> HoistedInstructions;
844
845 // For PHI hoisting to work we need to hoist blocks before their successors.
846 // We can do this by iterating through the blocks in the loop in reverse
847 // post-order.
848 LoopBlocksRPO Worklist(CurLoop);
849 Worklist.perform(LI);
850 bool Changed = false;
851 for (BasicBlock *BB : Worklist) {
852 // Only need to process the contents of this block if it is not part of a
853 // subloop (which would already have been processed).
854 if (inSubLoop(BB, CurLoop, LI))
855 continue;
856
857 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
858 Instruction &I = *II++;
859 // Try constant folding this instruction. If all the operands are
860 // constants, it is technically hoistable, but it would be better to
861 // just fold it.
862 if (Constant *C = ConstantFoldInstruction(
863 &I, I.getModule()->getDataLayout(), TLI)) {
864 LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C
865 << '\n');
866 if (CurAST)
867 CurAST->copyValue(&I, C);
868 // FIXME MSSA: Such replacements may make accesses unoptimized (D51960).
869 I.replaceAllUsesWith(C);
870 if (isInstructionTriviallyDead(&I, TLI))
871 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
872 Changed = true;
873 continue;
874 }
875
876 // Try hoisting the instruction out to the preheader. We can only do
877 // this if all of the operands of the instruction are loop invariant and
878 // if it is safe to hoist the instruction. We also check block frequency
879 // to make sure instruction only gets hoisted into colder blocks.
880 // TODO: It may be safe to hoist if we are hoisting to a conditional block
881 // and we have accurately duplicated the control flow from the loop header
882 // to that block.
883 if (CurLoop->hasLoopInvariantOperands(&I) &&
884 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
885 ORE) &&
886 worthSinkOrHoistInst(I, CurLoop->getLoopPreheader(), ORE, BFI) &&
887 isSafeToExecuteUnconditionally(
888 I, DT, CurLoop, SafetyInfo, ORE,
889 CurLoop->getLoopPreheader()->getTerminator())) {
890 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
891 MSSAU, SE, ORE);
892 HoistedInstructions.push_back(&I);
893 Changed = true;
894 continue;
895 }
896
897 // Attempt to remove floating point division out of the loop by
898 // converting it to a reciprocal multiplication.
899 if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() &&
900 CurLoop->isLoopInvariant(I.getOperand(1))) {
901 auto Divisor = I.getOperand(1);
902 auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0);
903 auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor);
904 ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags());
905 SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent());
906 ReciprocalDivisor->insertBefore(&I);
907
908 auto Product =
909 BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor);
910 Product->setFastMathFlags(I.getFastMathFlags());
911 SafetyInfo->insertInstructionTo(Product, I.getParent());
912 Product->insertAfter(&I);
913 I.replaceAllUsesWith(Product);
914 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
915
916 hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB),
917 SafetyInfo, MSSAU, SE, ORE);
918 HoistedInstructions.push_back(ReciprocalDivisor);
919 Changed = true;
920 continue;
921 }
922
923 auto IsInvariantStart = [&](Instruction &I) {
924 using namespace PatternMatch;
925 return I.use_empty() &&
926 match(&I, m_Intrinsic<Intrinsic::invariant_start>());
927 };
928 auto MustExecuteWithoutWritesBefore = [&](Instruction &I) {
929 return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) &&
930 SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop);
931 };
932 if ((IsInvariantStart(I) || isGuard(&I)) &&
933 CurLoop->hasLoopInvariantOperands(&I) &&
934 MustExecuteWithoutWritesBefore(I)) {
935 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
936 MSSAU, SE, ORE);
937 HoistedInstructions.push_back(&I);
938 Changed = true;
939 continue;
940 }
941
942 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
943 if (CFH.canHoistPHI(PN)) {
944 // Redirect incoming blocks first to ensure that we create hoisted
945 // versions of those blocks before we hoist the phi.
946 for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i)
947 PN->setIncomingBlock(
948 i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i)));
949 hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
950 MSSAU, SE, ORE);
951 assert(DT->dominates(PN, BB) && "Conditional PHIs not expected");
952 Changed = true;
953 continue;
954 }
955 }
956
957 // Remember possibly hoistable branches so we can actually hoist them
958 // later if needed.
959 if (BranchInst *BI = dyn_cast<BranchInst>(&I))
960 CFH.registerPossiblyHoistableBranch(BI);
961 }
962 }
963
964 // If we hoisted instructions to a conditional block they may not dominate
965 // their uses that weren't hoisted (such as phis where some operands are not
966 // loop invariant). If so make them unconditional by moving them to their
967 // immediate dominator. We iterate through the instructions in reverse order
968 // which ensures that when we rehoist an instruction we rehoist its operands,
969 // and also keep track of where in the block we are rehoisting to to make sure
970 // that we rehoist instructions before the instructions that use them.
971 Instruction *HoistPoint = nullptr;
972 if (ControlFlowHoisting) {
973 for (Instruction *I : reverse(HoistedInstructions)) {
974 if (!llvm::all_of(I->uses(),
975 [&](Use &U) { return DT->dominates(I, U); })) {
976 BasicBlock *Dominator =
977 DT->getNode(I->getParent())->getIDom()->getBlock();
978 if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) {
979 if (HoistPoint)
980 assert(DT->dominates(Dominator, HoistPoint->getParent()) &&
981 "New hoist point expected to dominate old hoist point");
982 HoistPoint = Dominator->getTerminator();
983 }
984 LLVM_DEBUG(dbgs() << "LICM rehoisting to "
985 << HoistPoint->getParent()->getNameOrAsOperand()
986 << ": " << *I << "\n");
987 moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU, SE);
988 HoistPoint = I;
989 Changed = true;
990 }
991 }
992 }
993 if (MSSAU && VerifyMemorySSA)
994 MSSAU->getMemorySSA()->verifyMemorySSA();
995
996 // Now that we've finished hoisting make sure that LI and DT are still
997 // valid.
998 #ifdef EXPENSIVE_CHECKS
999 if (Changed) {
1000 assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&
1001 "Dominator tree verification failed");
1002 LI->verify(*DT);
1003 }
1004 #endif
1005
1006 return Changed;
1007 }
1008
1009 // Return true if LI is invariant within scope of the loop. LI is invariant if
1010 // CurLoop is dominated by an invariant.start representing the same memory
1011 // location and size as the memory location LI loads from, and also the
1012 // invariant.start has no uses.
isLoadInvariantInLoop(LoadInst * LI,DominatorTree * DT,Loop * CurLoop)1013 static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT,
1014 Loop *CurLoop) {
1015 Value *Addr = LI->getOperand(0);
1016 const DataLayout &DL = LI->getModule()->getDataLayout();
1017 const TypeSize LocSizeInBits = DL.getTypeSizeInBits(LI->getType());
1018
1019 // It is not currently possible for clang to generate an invariant.start
1020 // intrinsic with scalable vector types because we don't support thread local
1021 // sizeless types and we don't permit sizeless types in structs or classes.
1022 // Furthermore, even if support is added for this in future the intrinsic
1023 // itself is defined to have a size of -1 for variable sized objects. This
1024 // makes it impossible to verify if the intrinsic envelops our region of
1025 // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8>
1026 // types would have a -1 parameter, but the former is clearly double the size
1027 // of the latter.
1028 if (LocSizeInBits.isScalable())
1029 return false;
1030
1031 // if the type is i8 addrspace(x)*, we know this is the type of
1032 // llvm.invariant.start operand
1033 auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()),
1034 LI->getPointerAddressSpace());
1035 unsigned BitcastsVisited = 0;
1036 // Look through bitcasts until we reach the i8* type (this is invariant.start
1037 // operand type).
1038 while (Addr->getType() != PtrInt8Ty) {
1039 auto *BC = dyn_cast<BitCastInst>(Addr);
1040 // Avoid traversing high number of bitcast uses.
1041 if (++BitcastsVisited > MaxNumUsesTraversed || !BC)
1042 return false;
1043 Addr = BC->getOperand(0);
1044 }
1045
1046 unsigned UsesVisited = 0;
1047 // Traverse all uses of the load operand value, to see if invariant.start is
1048 // one of the uses, and whether it dominates the load instruction.
1049 for (auto *U : Addr->users()) {
1050 // Avoid traversing for Load operand with high number of users.
1051 if (++UsesVisited > MaxNumUsesTraversed)
1052 return false;
1053 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1054 // If there are escaping uses of invariant.start instruction, the load maybe
1055 // non-invariant.
1056 if (!II || II->getIntrinsicID() != Intrinsic::invariant_start ||
1057 !II->use_empty())
1058 continue;
1059 ConstantInt *InvariantSize = cast<ConstantInt>(II->getArgOperand(0));
1060 // The intrinsic supports having a -1 argument for variable sized objects
1061 // so we should check for that here.
1062 if (InvariantSize->isNegative())
1063 continue;
1064 uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8;
1065 // Confirm the invariant.start location size contains the load operand size
1066 // in bits. Also, the invariant.start should dominate the load, and we
1067 // should not hoist the load out of a loop that contains this dominating
1068 // invariant.start.
1069 if (LocSizeInBits.getFixedSize() <= InvariantSizeInBits &&
1070 DT->properlyDominates(II->getParent(), CurLoop->getHeader()))
1071 return true;
1072 }
1073
1074 return false;
1075 }
1076
1077 namespace {
1078 /// Return true if-and-only-if we know how to (mechanically) both hoist and
1079 /// sink a given instruction out of a loop. Does not address legality
1080 /// concerns such as aliasing or speculation safety.
isHoistableAndSinkableInst(Instruction & I)1081 bool isHoistableAndSinkableInst(Instruction &I) {
1082 // Only these instructions are hoistable/sinkable.
1083 return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) ||
1084 isa<FenceInst>(I) || isa<CastInst>(I) || isa<UnaryOperator>(I) ||
1085 isa<BinaryOperator>(I) || isa<SelectInst>(I) ||
1086 isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
1087 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
1088 isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) ||
1089 isa<InsertValueInst>(I) || isa<FreezeInst>(I));
1090 }
1091 /// Return true if all of the alias sets within this AST are known not to
1092 /// contain a Mod, or if MSSA knows thare are no MemoryDefs in the loop.
isReadOnly(AliasSetTracker * CurAST,const MemorySSAUpdater * MSSAU,const Loop * L)1093 bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU,
1094 const Loop *L) {
1095 if (CurAST) {
1096 for (AliasSet &AS : *CurAST) {
1097 if (!AS.isForwardingAliasSet() && AS.isMod()) {
1098 return false;
1099 }
1100 }
1101 return true;
1102 } else { /*MSSAU*/
1103 for (auto *BB : L->getBlocks())
1104 if (MSSAU->getMemorySSA()->getBlockDefs(BB))
1105 return false;
1106 return true;
1107 }
1108 }
1109
1110 /// Return true if I is the only Instruction with a MemoryAccess in L.
isOnlyMemoryAccess(const Instruction * I,const Loop * L,const MemorySSAUpdater * MSSAU)1111 bool isOnlyMemoryAccess(const Instruction *I, const Loop *L,
1112 const MemorySSAUpdater *MSSAU) {
1113 for (auto *BB : L->getBlocks())
1114 if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) {
1115 int NotAPhi = 0;
1116 for (const auto &Acc : *Accs) {
1117 if (isa<MemoryPhi>(&Acc))
1118 continue;
1119 const auto *MUD = cast<MemoryUseOrDef>(&Acc);
1120 if (MUD->getMemoryInst() != I || NotAPhi++ == 1)
1121 return false;
1122 }
1123 }
1124 return true;
1125 }
1126 }
1127
canSinkOrHoistInst(Instruction & I,AAResults * AA,DominatorTree * DT,Loop * CurLoop,AliasSetTracker * CurAST,MemorySSAUpdater * MSSAU,bool TargetExecutesOncePerLoop,SinkAndHoistLICMFlags * Flags,OptimizationRemarkEmitter * ORE)1128 bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT,
1129 Loop *CurLoop, AliasSetTracker *CurAST,
1130 MemorySSAUpdater *MSSAU,
1131 bool TargetExecutesOncePerLoop,
1132 SinkAndHoistLICMFlags *Flags,
1133 OptimizationRemarkEmitter *ORE) {
1134 assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
1135 "Either AliasSetTracker or MemorySSA should be initialized.");
1136
1137 // If we don't understand the instruction, bail early.
1138 if (!isHoistableAndSinkableInst(I))
1139 return false;
1140
1141 MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr;
1142 if (MSSA)
1143 assert(Flags != nullptr && "Flags cannot be null.");
1144
1145 // Loads have extra constraints we have to verify before we can hoist them.
1146 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
1147 if (!LI->isUnordered())
1148 return false; // Don't sink/hoist volatile or ordered atomic loads!
1149
1150 // Loads from constant memory are always safe to move, even if they end up
1151 // in the same alias set as something that ends up being modified.
1152 if (AA->pointsToConstantMemory(LI->getOperand(0)))
1153 return true;
1154 if (LI->hasMetadata(LLVMContext::MD_invariant_load))
1155 return true;
1156
1157 if (LI->isAtomic() && !TargetExecutesOncePerLoop)
1158 return false; // Don't risk duplicating unordered loads
1159
1160 // This checks for an invariant.start dominating the load.
1161 if (isLoadInvariantInLoop(LI, DT, CurLoop))
1162 return true;
1163
1164 bool Invalidated;
1165 if (CurAST)
1166 Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST,
1167 CurLoop, AA);
1168 else
1169 Invalidated = pointerInvalidatedByLoopWithMSSA(
1170 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, I, *Flags);
1171 // Check loop-invariant address because this may also be a sinkable load
1172 // whose address is not necessarily loop-invariant.
1173 if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1174 ORE->emit([&]() {
1175 return OptimizationRemarkMissed(
1176 DEBUG_TYPE, "LoadWithLoopInvariantAddressInvalidated", LI)
1177 << "failed to move load with loop-invariant address "
1178 "because the loop may invalidate its value";
1179 });
1180
1181 return !Invalidated;
1182 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1183 // Don't sink or hoist dbg info; it's legal, but not useful.
1184 if (isa<DbgInfoIntrinsic>(I))
1185 return false;
1186
1187 // Don't sink calls which can throw.
1188 if (CI->mayThrow())
1189 return false;
1190
1191 // Convergent attribute has been used on operations that involve
1192 // inter-thread communication which results are implicitly affected by the
1193 // enclosing control flows. It is not safe to hoist or sink such operations
1194 // across control flow.
1195 if (CI->isConvergent())
1196 return false;
1197
1198 using namespace PatternMatch;
1199 if (match(CI, m_Intrinsic<Intrinsic::assume>()))
1200 // Assumes don't actually alias anything or throw
1201 return true;
1202
1203 if (match(CI, m_Intrinsic<Intrinsic::experimental_widenable_condition>()))
1204 // Widenable conditions don't actually alias anything or throw
1205 return true;
1206
1207 // Handle simple cases by querying alias analysis.
1208 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
1209 if (Behavior == FMRB_DoesNotAccessMemory)
1210 return true;
1211 if (AAResults::onlyReadsMemory(Behavior)) {
1212 // A readonly argmemonly function only reads from memory pointed to by
1213 // it's arguments with arbitrary offsets. If we can prove there are no
1214 // writes to this memory in the loop, we can hoist or sink.
1215 if (AAResults::onlyAccessesArgPointees(Behavior)) {
1216 // TODO: expand to writeable arguments
1217 for (Value *Op : CI->arg_operands())
1218 if (Op->getType()->isPointerTy()) {
1219 bool Invalidated;
1220 if (CurAST)
1221 Invalidated = pointerInvalidatedByLoop(
1222 MemoryLocation::getBeforeOrAfter(Op), CurAST, CurLoop, AA);
1223 else
1224 Invalidated = pointerInvalidatedByLoopWithMSSA(
1225 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop, I,
1226 *Flags);
1227 if (Invalidated)
1228 return false;
1229 }
1230 return true;
1231 }
1232
1233 // If this call only reads from memory and there are no writes to memory
1234 // in the loop, we can hoist or sink the call as appropriate.
1235 if (isReadOnly(CurAST, MSSAU, CurLoop))
1236 return true;
1237 }
1238
1239 // FIXME: This should use mod/ref information to see if we can hoist or
1240 // sink the call.
1241
1242 return false;
1243 } else if (auto *FI = dyn_cast<FenceInst>(&I)) {
1244 // Fences alias (most) everything to provide ordering. For the moment,
1245 // just give up if there are any other memory operations in the loop.
1246 if (CurAST) {
1247 auto Begin = CurAST->begin();
1248 assert(Begin != CurAST->end() && "must contain FI");
1249 if (std::next(Begin) != CurAST->end())
1250 // constant memory for instance, TODO: handle better
1251 return false;
1252 auto *UniqueI = Begin->getUniqueInstruction();
1253 if (!UniqueI)
1254 // other memory op, give up
1255 return false;
1256 (void)FI; // suppress unused variable warning
1257 assert(UniqueI == FI && "AS must contain FI");
1258 return true;
1259 } else // MSSAU
1260 return isOnlyMemoryAccess(FI, CurLoop, MSSAU);
1261 } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
1262 if (!SI->isUnordered())
1263 return false; // Don't sink/hoist volatile or ordered atomic store!
1264
1265 // We can only hoist a store that we can prove writes a value which is not
1266 // read or overwritten within the loop. For those cases, we fallback to
1267 // load store promotion instead. TODO: We can extend this to cases where
1268 // there is exactly one write to the location and that write dominates an
1269 // arbitrary number of reads in the loop.
1270 if (CurAST) {
1271 auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI));
1272
1273 if (AS.isRef() || !AS.isMustAlias())
1274 // Quick exit test, handled by the full path below as well.
1275 return false;
1276 auto *UniqueI = AS.getUniqueInstruction();
1277 if (!UniqueI)
1278 // other memory op, give up
1279 return false;
1280 assert(UniqueI == SI && "AS must contain SI");
1281 return true;
1282 } else { // MSSAU
1283 if (isOnlyMemoryAccess(SI, CurLoop, MSSAU))
1284 return true;
1285 // If there are more accesses than the Promotion cap or no "quota" to
1286 // check clobber, then give up as we're not walking a list that long.
1287 if (Flags->tooManyMemoryAccesses() || Flags->tooManyClobberingCalls())
1288 return false;
1289 // If there are interfering Uses (i.e. their defining access is in the
1290 // loop), or ordered loads (stored as Defs!), don't move this store.
1291 // Could do better here, but this is conservatively correct.
1292 // TODO: Cache set of Uses on the first walk in runOnLoop, update when
1293 // moving accesses. Can also extend to dominating uses.
1294 auto *SIMD = MSSA->getMemoryAccess(SI);
1295 for (auto *BB : CurLoop->getBlocks())
1296 if (auto *Accesses = MSSA->getBlockAccesses(BB)) {
1297 for (const auto &MA : *Accesses)
1298 if (const auto *MU = dyn_cast<MemoryUse>(&MA)) {
1299 auto *MD = MU->getDefiningAccess();
1300 if (!MSSA->isLiveOnEntryDef(MD) &&
1301 CurLoop->contains(MD->getBlock()))
1302 return false;
1303 // Disable hoisting past potentially interfering loads. Optimized
1304 // Uses may point to an access outside the loop, as getClobbering
1305 // checks the previous iteration when walking the backedge.
1306 // FIXME: More precise: no Uses that alias SI.
1307 if (!Flags->getIsSink() && !MSSA->dominates(SIMD, MU))
1308 return false;
1309 } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) {
1310 if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) {
1311 (void)LI; // Silence warning.
1312 assert(!LI->isUnordered() && "Expected unordered load");
1313 return false;
1314 }
1315 // Any call, while it may not be clobbering SI, it may be a use.
1316 if (auto *CI = dyn_cast<CallInst>(MD->getMemoryInst())) {
1317 // Check if the call may read from the memory locattion written
1318 // to by SI. Check CI's attributes and arguments; the number of
1319 // such checks performed is limited above by NoOfMemAccTooLarge.
1320 ModRefInfo MRI = AA->getModRefInfo(CI, MemoryLocation::get(SI));
1321 if (isModOrRefSet(MRI))
1322 return false;
1323 }
1324 }
1325 }
1326 auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI);
1327 Flags->incrementClobberingCalls();
1328 // If there are no clobbering Defs in the loop, store is safe to hoist.
1329 return MSSA->isLiveOnEntryDef(Source) ||
1330 !CurLoop->contains(Source->getBlock());
1331 }
1332 }
1333
1334 assert(!I.mayReadOrWriteMemory() && "unhandled aliasing");
1335
1336 // We've established mechanical ability and aliasing, it's up to the caller
1337 // to check fault safety
1338 return true;
1339 }
1340
1341 /// Returns true if a PHINode is a trivially replaceable with an
1342 /// Instruction.
1343 /// This is true when all incoming values are that instruction.
1344 /// This pattern occurs most often with LCSSA PHI nodes.
1345 ///
isTriviallyReplaceablePHI(const PHINode & PN,const Instruction & I)1346 static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) {
1347 for (const Value *IncValue : PN.incoming_values())
1348 if (IncValue != &I)
1349 return false;
1350
1351 return true;
1352 }
1353
1354 /// Return true if the instruction is free in the loop.
isFreeInLoop(const Instruction & I,const Loop * CurLoop,const TargetTransformInfo * TTI)1355 static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop,
1356 const TargetTransformInfo *TTI) {
1357
1358 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1359 if (TTI->getUserCost(GEP, TargetTransformInfo::TCK_SizeAndLatency) !=
1360 TargetTransformInfo::TCC_Free)
1361 return false;
1362 // For a GEP, we cannot simply use getUserCost because currently it
1363 // optimistically assume that a GEP will fold into addressing mode
1364 // regardless of its users.
1365 const BasicBlock *BB = GEP->getParent();
1366 for (const User *U : GEP->users()) {
1367 const Instruction *UI = cast<Instruction>(U);
1368 if (CurLoop->contains(UI) &&
1369 (BB != UI->getParent() ||
1370 (!isa<StoreInst>(UI) && !isa<LoadInst>(UI))))
1371 return false;
1372 }
1373 return true;
1374 } else
1375 return TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency) ==
1376 TargetTransformInfo::TCC_Free;
1377 }
1378
1379 /// Return true if the only users of this instruction are outside of
1380 /// the loop. If this is true, we can sink the instruction to the exit
1381 /// blocks of the loop.
1382 ///
1383 /// We also return true if the instruction could be folded away in lowering.
1384 /// (e.g., a GEP can be folded into a load as an addressing mode in the loop).
isNotUsedOrFreeInLoop(const Instruction & I,const Loop * CurLoop,const LoopSafetyInfo * SafetyInfo,TargetTransformInfo * TTI,bool & FreeInLoop)1385 static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
1386 const LoopSafetyInfo *SafetyInfo,
1387 TargetTransformInfo *TTI, bool &FreeInLoop) {
1388 const auto &BlockColors = SafetyInfo->getBlockColors();
1389 bool IsFree = isFreeInLoop(I, CurLoop, TTI);
1390 for (const User *U : I.users()) {
1391 const Instruction *UI = cast<Instruction>(U);
1392 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1393 const BasicBlock *BB = PN->getParent();
1394 // We cannot sink uses in catchswitches.
1395 if (isa<CatchSwitchInst>(BB->getTerminator()))
1396 return false;
1397
1398 // We need to sink a callsite to a unique funclet. Avoid sinking if the
1399 // phi use is too muddled.
1400 if (isa<CallInst>(I))
1401 if (!BlockColors.empty() &&
1402 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
1403 return false;
1404 }
1405
1406 if (CurLoop->contains(UI)) {
1407 if (IsFree) {
1408 FreeInLoop = true;
1409 continue;
1410 }
1411 return false;
1412 }
1413 }
1414 return true;
1415 }
1416
cloneInstructionInExitBlock(Instruction & I,BasicBlock & ExitBlock,PHINode & PN,const LoopInfo * LI,const LoopSafetyInfo * SafetyInfo,MemorySSAUpdater * MSSAU)1417 static Instruction *cloneInstructionInExitBlock(
1418 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
1419 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) {
1420 Instruction *New;
1421 if (auto *CI = dyn_cast<CallInst>(&I)) {
1422 const auto &BlockColors = SafetyInfo->getBlockColors();
1423
1424 // Sinking call-sites need to be handled differently from other
1425 // instructions. The cloned call-site needs a funclet bundle operand
1426 // appropriate for its location in the CFG.
1427 SmallVector<OperandBundleDef, 1> OpBundles;
1428 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
1429 BundleIdx != BundleEnd; ++BundleIdx) {
1430 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
1431 if (Bundle.getTagID() == LLVMContext::OB_funclet)
1432 continue;
1433
1434 OpBundles.emplace_back(Bundle);
1435 }
1436
1437 if (!BlockColors.empty()) {
1438 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
1439 assert(CV.size() == 1 && "non-unique color for exit block!");
1440 BasicBlock *BBColor = CV.front();
1441 Instruction *EHPad = BBColor->getFirstNonPHI();
1442 if (EHPad->isEHPad())
1443 OpBundles.emplace_back("funclet", EHPad);
1444 }
1445
1446 New = CallInst::Create(CI, OpBundles);
1447 } else {
1448 New = I.clone();
1449 }
1450
1451 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
1452 if (!I.getName().empty())
1453 New->setName(I.getName() + ".le");
1454
1455 if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
1456 // Create a new MemoryAccess and let MemorySSA set its defining access.
1457 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1458 New, nullptr, New->getParent(), MemorySSA::Beginning);
1459 if (NewMemAcc) {
1460 if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
1461 MSSAU->insertDef(MemDef, /*RenameUses=*/true);
1462 else {
1463 auto *MemUse = cast<MemoryUse>(NewMemAcc);
1464 MSSAU->insertUse(MemUse, /*RenameUses=*/true);
1465 }
1466 }
1467 }
1468
1469 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
1470 // particularly cheap because we can rip off the PHI node that we're
1471 // replacing for the number and blocks of the predecessors.
1472 // OPT: If this shows up in a profile, we can instead finish sinking all
1473 // invariant instructions, and then walk their operands to re-establish
1474 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
1475 // sinking bottom-up.
1476 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
1477 ++OI)
1478 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
1479 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
1480 if (!OLoop->contains(&PN)) {
1481 PHINode *OpPN =
1482 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
1483 OInst->getName() + ".lcssa", &ExitBlock.front());
1484 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1485 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
1486 *OI = OpPN;
1487 }
1488 return New;
1489 }
1490
eraseInstruction(Instruction & I,ICFLoopSafetyInfo & SafetyInfo,AliasSetTracker * AST,MemorySSAUpdater * MSSAU)1491 static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
1492 AliasSetTracker *AST, MemorySSAUpdater *MSSAU) {
1493 if (AST)
1494 AST->deleteValue(&I);
1495 if (MSSAU)
1496 MSSAU->removeMemoryAccess(&I);
1497 SafetyInfo.removeInstruction(&I);
1498 I.eraseFromParent();
1499 }
1500
moveInstructionBefore(Instruction & I,Instruction & Dest,ICFLoopSafetyInfo & SafetyInfo,MemorySSAUpdater * MSSAU,ScalarEvolution * SE)1501 static void moveInstructionBefore(Instruction &I, Instruction &Dest,
1502 ICFLoopSafetyInfo &SafetyInfo,
1503 MemorySSAUpdater *MSSAU,
1504 ScalarEvolution *SE) {
1505 SafetyInfo.removeInstruction(&I);
1506 SafetyInfo.insertInstructionTo(&I, Dest.getParent());
1507 I.moveBefore(&Dest);
1508 if (MSSAU)
1509 if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
1510 MSSAU->getMemorySSA()->getMemoryAccess(&I)))
1511 MSSAU->moveToPlace(OldMemAcc, Dest.getParent(),
1512 MemorySSA::BeforeTerminator);
1513 if (SE)
1514 SE->forgetValue(&I);
1515 }
1516
sinkThroughTriviallyReplaceablePHI(PHINode * TPN,Instruction * I,LoopInfo * LI,SmallDenseMap<BasicBlock *,Instruction *,32> & SunkCopies,const LoopSafetyInfo * SafetyInfo,const Loop * CurLoop,MemorySSAUpdater * MSSAU)1517 static Instruction *sinkThroughTriviallyReplaceablePHI(
1518 PHINode *TPN, Instruction *I, LoopInfo *LI,
1519 SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies,
1520 const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop,
1521 MemorySSAUpdater *MSSAU) {
1522 assert(isTriviallyReplaceablePHI(*TPN, *I) &&
1523 "Expect only trivially replaceable PHI");
1524 BasicBlock *ExitBlock = TPN->getParent();
1525 Instruction *New;
1526 auto It = SunkCopies.find(ExitBlock);
1527 if (It != SunkCopies.end())
1528 New = It->second;
1529 else
1530 New = SunkCopies[ExitBlock] = cloneInstructionInExitBlock(
1531 *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU);
1532 return New;
1533 }
1534
canSplitPredecessors(PHINode * PN,LoopSafetyInfo * SafetyInfo)1535 static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) {
1536 BasicBlock *BB = PN->getParent();
1537 if (!BB->canSplitPredecessors())
1538 return false;
1539 // It's not impossible to split EHPad blocks, but if BlockColors already exist
1540 // it require updating BlockColors for all offspring blocks accordingly. By
1541 // skipping such corner case, we can make updating BlockColors after splitting
1542 // predecessor fairly simple.
1543 if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad())
1544 return false;
1545 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1546 BasicBlock *BBPred = *PI;
1547 if (isa<IndirectBrInst>(BBPred->getTerminator()) ||
1548 isa<CallBrInst>(BBPred->getTerminator()))
1549 return false;
1550 }
1551 return true;
1552 }
1553
splitPredecessorsOfLoopExit(PHINode * PN,DominatorTree * DT,LoopInfo * LI,const Loop * CurLoop,LoopSafetyInfo * SafetyInfo,MemorySSAUpdater * MSSAU)1554 static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT,
1555 LoopInfo *LI, const Loop *CurLoop,
1556 LoopSafetyInfo *SafetyInfo,
1557 MemorySSAUpdater *MSSAU) {
1558 #ifndef NDEBUG
1559 SmallVector<BasicBlock *, 32> ExitBlocks;
1560 CurLoop->getUniqueExitBlocks(ExitBlocks);
1561 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1562 ExitBlocks.end());
1563 #endif
1564 BasicBlock *ExitBB = PN->getParent();
1565 assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.");
1566
1567 // Split predecessors of the loop exit to make instructions in the loop are
1568 // exposed to exit blocks through trivially replaceable PHIs while keeping the
1569 // loop in the canonical form where each predecessor of each exit block should
1570 // be contained within the loop. For example, this will convert the loop below
1571 // from
1572 //
1573 // LB1:
1574 // %v1 =
1575 // br %LE, %LB2
1576 // LB2:
1577 // %v2 =
1578 // br %LE, %LB1
1579 // LE:
1580 // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable
1581 //
1582 // to
1583 //
1584 // LB1:
1585 // %v1 =
1586 // br %LE.split, %LB2
1587 // LB2:
1588 // %v2 =
1589 // br %LE.split2, %LB1
1590 // LE.split:
1591 // %p1 = phi [%v1, %LB1] <-- trivially replaceable
1592 // br %LE
1593 // LE.split2:
1594 // %p2 = phi [%v2, %LB2] <-- trivially replaceable
1595 // br %LE
1596 // LE:
1597 // %p = phi [%p1, %LE.split], [%p2, %LE.split2]
1598 //
1599 const auto &BlockColors = SafetyInfo->getBlockColors();
1600 SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB));
1601 while (!PredBBs.empty()) {
1602 BasicBlock *PredBB = *PredBBs.begin();
1603 assert(CurLoop->contains(PredBB) &&
1604 "Expect all predecessors are in the loop");
1605 if (PN->getBasicBlockIndex(PredBB) >= 0) {
1606 BasicBlock *NewPred = SplitBlockPredecessors(
1607 ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true);
1608 // Since we do not allow splitting EH-block with BlockColors in
1609 // canSplitPredecessors(), we can simply assign predecessor's color to
1610 // the new block.
1611 if (!BlockColors.empty())
1612 // Grab a reference to the ColorVector to be inserted before getting the
1613 // reference to the vector we are copying because inserting the new
1614 // element in BlockColors might cause the map to be reallocated.
1615 SafetyInfo->copyColors(NewPred, PredBB);
1616 }
1617 PredBBs.remove(PredBB);
1618 }
1619 }
1620
1621 /// When an instruction is found to only be used outside of the loop, this
1622 /// function moves it to the exit blocks and patches up SSA form as needed.
1623 /// This method is guaranteed to remove the original instruction from its
1624 /// position, and may either delete it or move it to outside of the loop.
1625 ///
sink(Instruction & I,LoopInfo * LI,DominatorTree * DT,BlockFrequencyInfo * BFI,const Loop * CurLoop,ICFLoopSafetyInfo * SafetyInfo,MemorySSAUpdater * MSSAU,OptimizationRemarkEmitter * ORE)1626 static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
1627 BlockFrequencyInfo *BFI, const Loop *CurLoop,
1628 ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU,
1629 OptimizationRemarkEmitter *ORE) {
1630 LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
1631 ORE->emit([&]() {
1632 return OptimizationRemark(DEBUG_TYPE, "InstSunk", &I)
1633 << "sinking " << ore::NV("Inst", &I);
1634 });
1635 bool Changed = false;
1636 if (isa<LoadInst>(I))
1637 ++NumMovedLoads;
1638 else if (isa<CallInst>(I))
1639 ++NumMovedCalls;
1640 ++NumSunk;
1641
1642 // Iterate over users to be ready for actual sinking. Replace users via
1643 // unreachable blocks with undef and make all user PHIs trivially replaceable.
1644 SmallPtrSet<Instruction *, 8> VisitedUsers;
1645 for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) {
1646 auto *User = cast<Instruction>(*UI);
1647 Use &U = UI.getUse();
1648 ++UI;
1649
1650 if (VisitedUsers.count(User) || CurLoop->contains(User))
1651 continue;
1652
1653 if (!DT->isReachableFromEntry(User->getParent())) {
1654 U = UndefValue::get(I.getType());
1655 Changed = true;
1656 continue;
1657 }
1658
1659 // The user must be a PHI node.
1660 PHINode *PN = cast<PHINode>(User);
1661
1662 // Surprisingly, instructions can be used outside of loops without any
1663 // exits. This can only happen in PHI nodes if the incoming block is
1664 // unreachable.
1665 BasicBlock *BB = PN->getIncomingBlock(U);
1666 if (!DT->isReachableFromEntry(BB)) {
1667 U = UndefValue::get(I.getType());
1668 Changed = true;
1669 continue;
1670 }
1671
1672 VisitedUsers.insert(PN);
1673 if (isTriviallyReplaceablePHI(*PN, I))
1674 continue;
1675
1676 if (!canSplitPredecessors(PN, SafetyInfo))
1677 return Changed;
1678
1679 // Split predecessors of the PHI so that we can make users trivially
1680 // replaceable.
1681 splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU);
1682
1683 // Should rebuild the iterators, as they may be invalidated by
1684 // splitPredecessorsOfLoopExit().
1685 UI = I.user_begin();
1686 UE = I.user_end();
1687 }
1688
1689 if (VisitedUsers.empty())
1690 return Changed;
1691
1692 #ifndef NDEBUG
1693 SmallVector<BasicBlock *, 32> ExitBlocks;
1694 CurLoop->getUniqueExitBlocks(ExitBlocks);
1695 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1696 ExitBlocks.end());
1697 #endif
1698
1699 // Clones of this instruction. Don't create more than one per exit block!
1700 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
1701
1702 // If this instruction is only used outside of the loop, then all users are
1703 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
1704 // the instruction.
1705 // First check if I is worth sinking for all uses. Sink only when it is worth
1706 // across all uses.
1707 SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end());
1708 SmallVector<PHINode *, 8> ExitPNs;
1709 for (auto *UI : Users) {
1710 auto *User = cast<Instruction>(UI);
1711
1712 if (CurLoop->contains(User))
1713 continue;
1714
1715 PHINode *PN = cast<PHINode>(User);
1716 assert(ExitBlockSet.count(PN->getParent()) &&
1717 "The LCSSA PHI is not in an exit block!");
1718 if (!worthSinkOrHoistInst(I, PN->getParent(), ORE, BFI)) {
1719 return Changed;
1720 }
1721
1722 ExitPNs.push_back(PN);
1723 }
1724
1725 for (auto *PN : ExitPNs) {
1726
1727 // The PHI must be trivially replaceable.
1728 Instruction *New = sinkThroughTriviallyReplaceablePHI(
1729 PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU);
1730 PN->replaceAllUsesWith(New);
1731 eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr);
1732 Changed = true;
1733 }
1734 return Changed;
1735 }
1736
1737 /// When an instruction is found to only use loop invariant operands that
1738 /// is safe to hoist, this instruction is called to do the dirty work.
1739 ///
hoist(Instruction & I,const DominatorTree * DT,const Loop * CurLoop,BasicBlock * Dest,ICFLoopSafetyInfo * SafetyInfo,MemorySSAUpdater * MSSAU,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)1740 static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
1741 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
1742 MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
1743 OptimizationRemarkEmitter *ORE) {
1744 LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getNameOrAsOperand() << ": "
1745 << I << "\n");
1746 ORE->emit([&]() {
1747 return OptimizationRemark(DEBUG_TYPE, "Hoisted", &I) << "hoisting "
1748 << ore::NV("Inst", &I);
1749 });
1750
1751 // Metadata can be dependent on conditions we are hoisting above.
1752 // Conservatively strip all metadata on the instruction unless we were
1753 // guaranteed to execute I if we entered the loop, in which case the metadata
1754 // is valid in the loop preheader.
1755 if (I.hasMetadataOtherThanDebugLoc() &&
1756 // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
1757 // time in isGuaranteedToExecute if we don't actually have anything to
1758 // drop. It is a compile time optimization, not required for correctness.
1759 !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop))
1760 I.dropUnknownNonDebugMetadata();
1761
1762 if (isa<PHINode>(I))
1763 // Move the new node to the end of the phi list in the destination block.
1764 moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU, SE);
1765 else
1766 // Move the new node to the destination block, before its terminator.
1767 moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU, SE);
1768
1769 I.updateLocationAfterHoist();
1770
1771 if (isa<LoadInst>(I))
1772 ++NumMovedLoads;
1773 else if (isa<CallInst>(I))
1774 ++NumMovedCalls;
1775 ++NumHoisted;
1776 }
1777
1778 /// Only sink or hoist an instruction if it is not a trapping instruction,
1779 /// or if the instruction is known not to trap when moved to the preheader.
1780 /// or if it is a trapping instruction and is guaranteed to execute.
isSafeToExecuteUnconditionally(Instruction & Inst,const DominatorTree * DT,const Loop * CurLoop,const LoopSafetyInfo * SafetyInfo,OptimizationRemarkEmitter * ORE,const Instruction * CtxI)1781 static bool isSafeToExecuteUnconditionally(Instruction &Inst,
1782 const DominatorTree *DT,
1783 const Loop *CurLoop,
1784 const LoopSafetyInfo *SafetyInfo,
1785 OptimizationRemarkEmitter *ORE,
1786 const Instruction *CtxI) {
1787 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT))
1788 return true;
1789
1790 bool GuaranteedToExecute =
1791 SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop);
1792
1793 if (!GuaranteedToExecute) {
1794 auto *LI = dyn_cast<LoadInst>(&Inst);
1795 if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1796 ORE->emit([&]() {
1797 return OptimizationRemarkMissed(
1798 DEBUG_TYPE, "LoadWithLoopInvariantAddressCondExecuted", LI)
1799 << "failed to hoist load with loop-invariant address "
1800 "because load is conditionally executed";
1801 });
1802 }
1803
1804 return GuaranteedToExecute;
1805 }
1806
1807 namespace {
1808 class LoopPromoter : public LoadAndStorePromoter {
1809 Value *SomePtr; // Designated pointer to store to.
1810 const SmallSetVector<Value *, 8> &PointerMustAliases;
1811 SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
1812 SmallVectorImpl<Instruction *> &LoopInsertPts;
1813 SmallVectorImpl<MemoryAccess *> &MSSAInsertPts;
1814 PredIteratorCache &PredCache;
1815 AliasSetTracker *AST;
1816 MemorySSAUpdater *MSSAU;
1817 LoopInfo &LI;
1818 DebugLoc DL;
1819 int Alignment;
1820 bool UnorderedAtomic;
1821 AAMDNodes AATags;
1822 ICFLoopSafetyInfo &SafetyInfo;
1823
maybeInsertLCSSAPHI(Value * V,BasicBlock * BB) const1824 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
1825 if (Instruction *I = dyn_cast<Instruction>(V))
1826 if (Loop *L = LI.getLoopFor(I->getParent()))
1827 if (!L->contains(BB)) {
1828 // We need to create an LCSSA PHI node for the incoming value and
1829 // store that.
1830 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
1831 I->getName() + ".lcssa", &BB->front());
1832 for (BasicBlock *Pred : PredCache.get(BB))
1833 PN->addIncoming(I, Pred);
1834 return PN;
1835 }
1836 return V;
1837 }
1838
1839 public:
LoopPromoter(Value * SP,ArrayRef<const Instruction * > Insts,SSAUpdater & S,const SmallSetVector<Value *,8> & PMA,SmallVectorImpl<BasicBlock * > & LEB,SmallVectorImpl<Instruction * > & LIP,SmallVectorImpl<MemoryAccess * > & MSSAIP,PredIteratorCache & PIC,AliasSetTracker * ast,MemorySSAUpdater * MSSAU,LoopInfo & li,DebugLoc dl,int alignment,bool UnorderedAtomic,const AAMDNodes & AATags,ICFLoopSafetyInfo & SafetyInfo)1840 LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
1841 const SmallSetVector<Value *, 8> &PMA,
1842 SmallVectorImpl<BasicBlock *> &LEB,
1843 SmallVectorImpl<Instruction *> &LIP,
1844 SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC,
1845 AliasSetTracker *ast, MemorySSAUpdater *MSSAU, LoopInfo &li,
1846 DebugLoc dl, int alignment, bool UnorderedAtomic,
1847 const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo)
1848 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
1849 LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP),
1850 PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)),
1851 Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags),
1852 SafetyInfo(SafetyInfo) {}
1853
isInstInList(Instruction * I,const SmallVectorImpl<Instruction * > &) const1854 bool isInstInList(Instruction *I,
1855 const SmallVectorImpl<Instruction *> &) const override {
1856 Value *Ptr;
1857 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1858 Ptr = LI->getOperand(0);
1859 else
1860 Ptr = cast<StoreInst>(I)->getPointerOperand();
1861 return PointerMustAliases.count(Ptr);
1862 }
1863
doExtraRewritesBeforeFinalDeletion()1864 void doExtraRewritesBeforeFinalDeletion() override {
1865 // Insert stores after in the loop exit blocks. Each exit block gets a
1866 // store of the live-out values that feed them. Since we've already told
1867 // the SSA updater about the defs in the loop and the preheader
1868 // definition, it is all set and we can start using it.
1869 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
1870 BasicBlock *ExitBlock = LoopExitBlocks[i];
1871 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
1872 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
1873 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
1874 Instruction *InsertPos = LoopInsertPts[i];
1875 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
1876 if (UnorderedAtomic)
1877 NewSI->setOrdering(AtomicOrdering::Unordered);
1878 NewSI->setAlignment(Align(Alignment));
1879 NewSI->setDebugLoc(DL);
1880 if (AATags)
1881 NewSI->setAAMetadata(AATags);
1882
1883 if (MSSAU) {
1884 MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i];
1885 MemoryAccess *NewMemAcc;
1886 if (!MSSAInsertPoint) {
1887 NewMemAcc = MSSAU->createMemoryAccessInBB(
1888 NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning);
1889 } else {
1890 NewMemAcc =
1891 MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint);
1892 }
1893 MSSAInsertPts[i] = NewMemAcc;
1894 MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1895 // FIXME: true for safety, false may still be correct.
1896 }
1897 }
1898 }
1899
replaceLoadWithValue(LoadInst * LI,Value * V) const1900 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
1901 // Update alias analysis.
1902 if (AST)
1903 AST->copyValue(LI, V);
1904 }
instructionDeleted(Instruction * I) const1905 void instructionDeleted(Instruction *I) const override {
1906 SafetyInfo.removeInstruction(I);
1907 if (AST)
1908 AST->deleteValue(I);
1909 if (MSSAU)
1910 MSSAU->removeMemoryAccess(I);
1911 }
1912 };
1913
1914
1915 /// Return true iff we can prove that a caller of this function can not inspect
1916 /// the contents of the provided object in a well defined program.
isKnownNonEscaping(Value * Object,const TargetLibraryInfo * TLI)1917 bool isKnownNonEscaping(Value *Object, const TargetLibraryInfo *TLI) {
1918 if (isa<AllocaInst>(Object))
1919 // Since the alloca goes out of scope, we know the caller can't retain a
1920 // reference to it and be well defined. Thus, we don't need to check for
1921 // capture.
1922 return true;
1923
1924 // For all other objects we need to know that the caller can't possibly
1925 // have gotten a reference to the object. There are two components of
1926 // that:
1927 // 1) Object can't be escaped by this function. This is what
1928 // PointerMayBeCaptured checks.
1929 // 2) Object can't have been captured at definition site. For this, we
1930 // need to know the return value is noalias. At the moment, we use a
1931 // weaker condition and handle only AllocLikeFunctions (which are
1932 // known to be noalias). TODO
1933 return isAllocLikeFn(Object, TLI) &&
1934 !PointerMayBeCaptured(Object, true, true);
1935 }
1936
1937 } // namespace
1938
1939 /// Try to promote memory values to scalars by sinking stores out of the
1940 /// loop and moving loads to before the loop. We do this by looping over
1941 /// the stores in the loop, looking for stores to Must pointers which are
1942 /// loop invariant.
1943 ///
promoteLoopAccessesToScalars(const SmallSetVector<Value *,8> & PointerMustAliases,SmallVectorImpl<BasicBlock * > & ExitBlocks,SmallVectorImpl<Instruction * > & InsertPts,SmallVectorImpl<MemoryAccess * > & MSSAInsertPts,PredIteratorCache & PIC,LoopInfo * LI,DominatorTree * DT,const TargetLibraryInfo * TLI,Loop * CurLoop,AliasSetTracker * CurAST,MemorySSAUpdater * MSSAU,ICFLoopSafetyInfo * SafetyInfo,OptimizationRemarkEmitter * ORE)1944 bool llvm::promoteLoopAccessesToScalars(
1945 const SmallSetVector<Value *, 8> &PointerMustAliases,
1946 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1947 SmallVectorImpl<Instruction *> &InsertPts,
1948 SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC,
1949 LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
1950 Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
1951 ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) {
1952 // Verify inputs.
1953 assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&
1954 SafetyInfo != nullptr &&
1955 "Unexpected Input to promoteLoopAccessesToScalars");
1956
1957 Value *SomePtr = *PointerMustAliases.begin();
1958 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1959
1960 // It is not safe to promote a load/store from the loop if the load/store is
1961 // conditional. For example, turning:
1962 //
1963 // for () { if (c) *P += 1; }
1964 //
1965 // into:
1966 //
1967 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
1968 //
1969 // is not safe, because *P may only be valid to access if 'c' is true.
1970 //
1971 // The safety property divides into two parts:
1972 // p1) The memory may not be dereferenceable on entry to the loop. In this
1973 // case, we can't insert the required load in the preheader.
1974 // p2) The memory model does not allow us to insert a store along any dynamic
1975 // path which did not originally have one.
1976 //
1977 // If at least one store is guaranteed to execute, both properties are
1978 // satisfied, and promotion is legal.
1979 //
1980 // This, however, is not a necessary condition. Even if no store/load is
1981 // guaranteed to execute, we can still establish these properties.
1982 // We can establish (p1) by proving that hoisting the load into the preheader
1983 // is safe (i.e. proving dereferenceability on all paths through the loop). We
1984 // can use any access within the alias set to prove dereferenceability,
1985 // since they're all must alias.
1986 //
1987 // There are two ways establish (p2):
1988 // a) Prove the location is thread-local. In this case the memory model
1989 // requirement does not apply, and stores are safe to insert.
1990 // b) Prove a store dominates every exit block. In this case, if an exit
1991 // blocks is reached, the original dynamic path would have taken us through
1992 // the store, so inserting a store into the exit block is safe. Note that this
1993 // is different from the store being guaranteed to execute. For instance,
1994 // if an exception is thrown on the first iteration of the loop, the original
1995 // store is never executed, but the exit blocks are not executed either.
1996
1997 bool DereferenceableInPH = false;
1998 bool SafeToInsertStore = false;
1999
2000 SmallVector<Instruction *, 64> LoopUses;
2001
2002 // We start with an alignment of one and try to find instructions that allow
2003 // us to prove better alignment.
2004 Align Alignment;
2005 // Keep track of which types of access we see
2006 bool SawUnorderedAtomic = false;
2007 bool SawNotAtomic = false;
2008 AAMDNodes AATags;
2009
2010 const DataLayout &MDL = Preheader->getModule()->getDataLayout();
2011
2012 bool IsKnownThreadLocalObject = false;
2013 if (SafetyInfo->anyBlockMayThrow()) {
2014 // If a loop can throw, we have to insert a store along each unwind edge.
2015 // That said, we can't actually make the unwind edge explicit. Therefore,
2016 // we have to prove that the store is dead along the unwind edge. We do
2017 // this by proving that the caller can't have a reference to the object
2018 // after return and thus can't possibly load from the object.
2019 Value *Object = getUnderlyingObject(SomePtr);
2020 if (!isKnownNonEscaping(Object, TLI))
2021 return false;
2022 // Subtlety: Alloca's aren't visible to callers, but *are* potentially
2023 // visible to other threads if captured and used during their lifetimes.
2024 IsKnownThreadLocalObject = !isa<AllocaInst>(Object);
2025 }
2026
2027 // Check that all of the pointers in the alias set have the same type. We
2028 // cannot (yet) promote a memory location that is loaded and stored in
2029 // different sizes. While we are at it, collect alignment and AA info.
2030 for (Value *ASIV : PointerMustAliases) {
2031 // Check that all of the pointers in the alias set have the same type. We
2032 // cannot (yet) promote a memory location that is loaded and stored in
2033 // different sizes.
2034 if (SomePtr->getType() != ASIV->getType())
2035 return false;
2036
2037 for (User *U : ASIV->users()) {
2038 // Ignore instructions that are outside the loop.
2039 Instruction *UI = dyn_cast<Instruction>(U);
2040 if (!UI || !CurLoop->contains(UI))
2041 continue;
2042
2043 // If there is an non-load/store instruction in the loop, we can't promote
2044 // it.
2045 if (LoadInst *Load = dyn_cast<LoadInst>(UI)) {
2046 if (!Load->isUnordered())
2047 return false;
2048
2049 SawUnorderedAtomic |= Load->isAtomic();
2050 SawNotAtomic |= !Load->isAtomic();
2051
2052 Align InstAlignment = Load->getAlign();
2053
2054 // Note that proving a load safe to speculate requires proving
2055 // sufficient alignment at the target location. Proving it guaranteed
2056 // to execute does as well. Thus we can increase our guaranteed
2057 // alignment as well.
2058 if (!DereferenceableInPH || (InstAlignment > Alignment))
2059 if (isSafeToExecuteUnconditionally(*Load, DT, CurLoop, SafetyInfo,
2060 ORE, Preheader->getTerminator())) {
2061 DereferenceableInPH = true;
2062 Alignment = std::max(Alignment, InstAlignment);
2063 }
2064 } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
2065 // Stores *of* the pointer are not interesting, only stores *to* the
2066 // pointer.
2067 if (UI->getOperand(1) != ASIV)
2068 continue;
2069 if (!Store->isUnordered())
2070 return false;
2071
2072 SawUnorderedAtomic |= Store->isAtomic();
2073 SawNotAtomic |= !Store->isAtomic();
2074
2075 // If the store is guaranteed to execute, both properties are satisfied.
2076 // We may want to check if a store is guaranteed to execute even if we
2077 // already know that promotion is safe, since it may have higher
2078 // alignment than any other guaranteed stores, in which case we can
2079 // raise the alignment on the promoted store.
2080 Align InstAlignment = Store->getAlign();
2081
2082 if (!DereferenceableInPH || !SafeToInsertStore ||
2083 (InstAlignment > Alignment)) {
2084 if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) {
2085 DereferenceableInPH = true;
2086 SafeToInsertStore = true;
2087 Alignment = std::max(Alignment, InstAlignment);
2088 }
2089 }
2090
2091 // If a store dominates all exit blocks, it is safe to sink.
2092 // As explained above, if an exit block was executed, a dominating
2093 // store must have been executed at least once, so we are not
2094 // introducing stores on paths that did not have them.
2095 // Note that this only looks at explicit exit blocks. If we ever
2096 // start sinking stores into unwind edges (see above), this will break.
2097 if (!SafeToInsertStore)
2098 SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) {
2099 return DT->dominates(Store->getParent(), Exit);
2100 });
2101
2102 // If the store is not guaranteed to execute, we may still get
2103 // deref info through it.
2104 if (!DereferenceableInPH) {
2105 DereferenceableInPH = isDereferenceableAndAlignedPointer(
2106 Store->getPointerOperand(), Store->getValueOperand()->getType(),
2107 Store->getAlign(), MDL, Preheader->getTerminator(), DT);
2108 }
2109 } else
2110 return false; // Not a load or store.
2111
2112 // Merge the AA tags.
2113 if (LoopUses.empty()) {
2114 // On the first load/store, just take its AA tags.
2115 UI->getAAMetadata(AATags);
2116 } else if (AATags) {
2117 UI->getAAMetadata(AATags, /* Merge = */ true);
2118 }
2119
2120 LoopUses.push_back(UI);
2121 }
2122 }
2123
2124 // If we found both an unordered atomic instruction and a non-atomic memory
2125 // access, bail. We can't blindly promote non-atomic to atomic since we
2126 // might not be able to lower the result. We can't downgrade since that
2127 // would violate memory model. Also, align 0 is an error for atomics.
2128 if (SawUnorderedAtomic && SawNotAtomic)
2129 return false;
2130
2131 // If we're inserting an atomic load in the preheader, we must be able to
2132 // lower it. We're only guaranteed to be able to lower naturally aligned
2133 // atomics.
2134 auto *SomePtrElemType = SomePtr->getType()->getPointerElementType();
2135 if (SawUnorderedAtomic &&
2136 Alignment < MDL.getTypeStoreSize(SomePtrElemType))
2137 return false;
2138
2139 // If we couldn't prove we can hoist the load, bail.
2140 if (!DereferenceableInPH)
2141 return false;
2142
2143 // We know we can hoist the load, but don't have a guaranteed store.
2144 // Check whether the location is thread-local. If it is, then we can insert
2145 // stores along paths which originally didn't have them without violating the
2146 // memory model.
2147 if (!SafeToInsertStore) {
2148 if (IsKnownThreadLocalObject)
2149 SafeToInsertStore = true;
2150 else {
2151 Value *Object = getUnderlyingObject(SomePtr);
2152 SafeToInsertStore =
2153 (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) &&
2154 !PointerMayBeCaptured(Object, true, true);
2155 }
2156 }
2157
2158 // If we've still failed to prove we can sink the store, give up.
2159 if (!SafeToInsertStore)
2160 return false;
2161
2162 // Otherwise, this is safe to promote, lets do it!
2163 LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr
2164 << '\n');
2165 ORE->emit([&]() {
2166 return OptimizationRemark(DEBUG_TYPE, "PromoteLoopAccessesToScalar",
2167 LoopUses[0])
2168 << "Moving accesses to memory location out of the loop";
2169 });
2170 ++NumPromoted;
2171
2172 // Look at all the loop uses, and try to merge their locations.
2173 std::vector<const DILocation *> LoopUsesLocs;
2174 for (auto U : LoopUses)
2175 LoopUsesLocs.push_back(U->getDebugLoc().get());
2176 auto DL = DebugLoc(DILocation::getMergedLocations(LoopUsesLocs));
2177
2178 // We use the SSAUpdater interface to insert phi nodes as required.
2179 SmallVector<PHINode *, 16> NewPHIs;
2180 SSAUpdater SSA(&NewPHIs);
2181 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
2182 InsertPts, MSSAInsertPts, PIC, CurAST, MSSAU, *LI, DL,
2183 Alignment.value(), SawUnorderedAtomic, AATags,
2184 *SafetyInfo);
2185
2186 // Set up the preheader to have a definition of the value. It is the live-out
2187 // value from the preheader that uses in the loop will use.
2188 LoadInst *PreheaderLoad = new LoadInst(
2189 SomePtr->getType()->getPointerElementType(), SomePtr,
2190 SomePtr->getName() + ".promoted", Preheader->getTerminator());
2191 if (SawUnorderedAtomic)
2192 PreheaderLoad->setOrdering(AtomicOrdering::Unordered);
2193 PreheaderLoad->setAlignment(Alignment);
2194 PreheaderLoad->setDebugLoc(DebugLoc());
2195 if (AATags)
2196 PreheaderLoad->setAAMetadata(AATags);
2197 SSA.AddAvailableValue(Preheader, PreheaderLoad);
2198
2199 if (MSSAU) {
2200 MemoryAccess *PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB(
2201 PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End);
2202 MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess);
2203 MSSAU->insertUse(NewMemUse, /*RenameUses=*/true);
2204 }
2205
2206 if (MSSAU && VerifyMemorySSA)
2207 MSSAU->getMemorySSA()->verifyMemorySSA();
2208 // Rewrite all the loads in the loop and remember all the definitions from
2209 // stores in the loop.
2210 Promoter.run(LoopUses);
2211
2212 if (MSSAU && VerifyMemorySSA)
2213 MSSAU->getMemorySSA()->verifyMemorySSA();
2214 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
2215 if (PreheaderLoad->use_empty())
2216 eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU);
2217
2218 return true;
2219 }
2220
2221 /// Returns an owning pointer to an alias set which incorporates aliasing info
2222 /// from L and all subloops of L.
2223 std::unique_ptr<AliasSetTracker>
collectAliasInfoForLoop(Loop * L,LoopInfo * LI,AAResults * AA)2224 LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
2225 AAResults *AA) {
2226 auto CurAST = std::make_unique<AliasSetTracker>(*AA);
2227
2228 // Add everything from all the sub loops.
2229 for (Loop *InnerL : L->getSubLoops())
2230 for (BasicBlock *BB : InnerL->blocks())
2231 CurAST->add(*BB);
2232
2233 // And merge in this loop (without anything from inner loops).
2234 for (BasicBlock *BB : L->blocks())
2235 if (LI->getLoopFor(BB) == L)
2236 CurAST->add(*BB);
2237
2238 return CurAST;
2239 }
2240
2241 std::unique_ptr<AliasSetTracker>
collectAliasInfoForLoopWithMSSA(Loop * L,AAResults * AA,MemorySSAUpdater * MSSAU)2242 LoopInvariantCodeMotion::collectAliasInfoForLoopWithMSSA(
2243 Loop *L, AAResults *AA, MemorySSAUpdater *MSSAU) {
2244 auto *MSSA = MSSAU->getMemorySSA();
2245 auto CurAST = std::make_unique<AliasSetTracker>(*AA, MSSA, L);
2246 CurAST->addAllInstructionsInLoopUsingMSSA();
2247 return CurAST;
2248 }
2249
pointerInvalidatedByLoop(MemoryLocation MemLoc,AliasSetTracker * CurAST,Loop * CurLoop,AAResults * AA)2250 static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
2251 AliasSetTracker *CurAST, Loop *CurLoop,
2252 AAResults *AA) {
2253 // First check to see if any of the basic blocks in CurLoop invalidate *V.
2254 bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod();
2255
2256 if (!isInvalidatedAccordingToAST || !LICMN2Theshold)
2257 return isInvalidatedAccordingToAST;
2258
2259 // Check with a diagnostic analysis if we can refine the information above.
2260 // This is to identify the limitations of using the AST.
2261 // The alias set mechanism used by LICM has a major weakness in that it
2262 // combines all things which may alias into a single set *before* asking
2263 // modref questions. As a result, a single readonly call within a loop will
2264 // collapse all loads and stores into a single alias set and report
2265 // invalidation if the loop contains any store. For example, readonly calls
2266 // with deopt states have this form and create a general alias set with all
2267 // loads and stores. In order to get any LICM in loops containing possible
2268 // deopt states we need a more precise invalidation of checking the mod ref
2269 // info of each instruction within the loop and LI. This has a complexity of
2270 // O(N^2), so currently, it is used only as a diagnostic tool since the
2271 // default value of LICMN2Threshold is zero.
2272
2273 // Don't look at nested loops.
2274 if (CurLoop->begin() != CurLoop->end())
2275 return true;
2276
2277 int N = 0;
2278 for (BasicBlock *BB : CurLoop->getBlocks())
2279 for (Instruction &I : *BB) {
2280 if (N >= LICMN2Theshold) {
2281 LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "
2282 << *(MemLoc.Ptr) << "\n");
2283 return true;
2284 }
2285 N++;
2286 auto Res = AA->getModRefInfo(&I, MemLoc);
2287 if (isModSet(Res)) {
2288 LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "
2289 << *(MemLoc.Ptr) << "\n");
2290 return true;
2291 }
2292 }
2293 LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n");
2294 return false;
2295 }
2296
pointerInvalidatedByLoopWithMSSA(MemorySSA * MSSA,MemoryUse * MU,Loop * CurLoop,Instruction & I,SinkAndHoistLICMFlags & Flags)2297 bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
2298 Loop *CurLoop, Instruction &I,
2299 SinkAndHoistLICMFlags &Flags) {
2300 // For hoisting, use the walker to determine safety
2301 if (!Flags.getIsSink()) {
2302 MemoryAccess *Source;
2303 // See declaration of SetLicmMssaOptCap for usage details.
2304 if (Flags.tooManyClobberingCalls())
2305 Source = MU->getDefiningAccess();
2306 else {
2307 Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU);
2308 Flags.incrementClobberingCalls();
2309 }
2310 return !MSSA->isLiveOnEntryDef(Source) &&
2311 CurLoop->contains(Source->getBlock());
2312 }
2313
2314 // For sinking, we'd need to check all Defs below this use. The getClobbering
2315 // call will look on the backedge of the loop, but will check aliasing with
2316 // the instructions on the previous iteration.
2317 // For example:
2318 // for (i ... )
2319 // load a[i] ( Use (LoE)
2320 // store a[i] ( 1 = Def (2), with 2 = Phi for the loop.
2321 // i++;
2322 // The load sees no clobbering inside the loop, as the backedge alias check
2323 // does phi translation, and will check aliasing against store a[i-1].
2324 // However sinking the load outside the loop, below the store is incorrect.
2325
2326 // For now, only sink if there are no Defs in the loop, and the existing ones
2327 // precede the use and are in the same block.
2328 // FIXME: Increase precision: Safe to sink if Use post dominates the Def;
2329 // needs PostDominatorTreeAnalysis.
2330 // FIXME: More precise: no Defs that alias this Use.
2331 if (Flags.tooManyMemoryAccesses())
2332 return true;
2333 for (auto *BB : CurLoop->getBlocks())
2334 if (pointerInvalidatedByBlockWithMSSA(*BB, *MSSA, *MU))
2335 return true;
2336 // When sinking, the source block may not be part of the loop so check it.
2337 if (!CurLoop->contains(&I))
2338 return pointerInvalidatedByBlockWithMSSA(*I.getParent(), *MSSA, *MU);
2339
2340 return false;
2341 }
2342
pointerInvalidatedByBlockWithMSSA(BasicBlock & BB,MemorySSA & MSSA,MemoryUse & MU)2343 bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA,
2344 MemoryUse &MU) {
2345 if (const auto *Accesses = MSSA.getBlockDefs(&BB))
2346 for (const auto &MA : *Accesses)
2347 if (const auto *MD = dyn_cast<MemoryDef>(&MA))
2348 if (MU.getBlock() != MD->getBlock() || !MSSA.locallyDominates(MD, &MU))
2349 return true;
2350 return false;
2351 }
2352
2353 /// Little predicate that returns true if the specified basic block is in
2354 /// a subloop of the current one, not the current one itself.
2355 ///
inSubLoop(BasicBlock * BB,Loop * CurLoop,LoopInfo * LI)2356 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
2357 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
2358 return LI->getLoopFor(BB) != CurLoop;
2359 }
2360