xref: /llvm-project/llvm/lib/Transforms/Scalar/LoopSink.cpp (revision a494ae43bef09c8d0f6a6a98e92f3a89758247d5)
1 //===-- LoopSink.cpp - Loop Sink 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 does the inverse transformation of what LICM does.
10 // It traverses all of the instructions in the loop's preheader and sinks
11 // them to the loop body where frequency is lower than the loop's preheader.
12 // This pass is a reverse-transformation of LICM. It differs from the Sink
13 // pass in the following ways:
14 //
15 // * It only handles sinking of instructions from the loop's preheader to the
16 //   loop's body
17 // * It uses alias set tracker to get more accurate alias info
18 // * It uses block frequency info to find the optimal sinking locations
19 //
20 // Overall algorithm:
21 //
22 // For I in Preheader:
23 //   InsertBBs = BBs that uses I
24 //   For BB in sorted(LoopBBs):
25 //     DomBBs = BBs in InsertBBs that are dominated by BB
26 //     if freq(DomBBs) > freq(BB)
27 //       InsertBBs = UseBBs - DomBBs + BB
28 //   For BB in InsertBBs:
29 //     Insert I at BB's beginning
30 //
31 //===----------------------------------------------------------------------===//
32 
33 #include "llvm/Transforms/Scalar/LoopSink.h"
34 #include "llvm/ADT/SetOperations.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/Analysis/AliasSetTracker.h"
38 #include "llvm/Analysis/BasicAliasAnalysis.h"
39 #include "llvm/Analysis/BlockFrequencyInfo.h"
40 #include "llvm/Analysis/Loads.h"
41 #include "llvm/Analysis/LoopInfo.h"
42 #include "llvm/Analysis/LoopPass.h"
43 #include "llvm/Analysis/MemorySSA.h"
44 #include "llvm/Analysis/MemorySSAUpdater.h"
45 #include "llvm/Analysis/ScalarEvolution.h"
46 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
47 #include "llvm/IR/Dominators.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/LLVMContext.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/InitializePasses.h"
52 #include "llvm/Support/BranchProbability.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/Transforms/Scalar/LoopPassManager.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/LoopUtils.h"
58 using namespace llvm;
59 
60 #define DEBUG_TYPE "loopsink"
61 
62 STATISTIC(NumLoopSunk, "Number of instructions sunk into loop");
63 STATISTIC(NumLoopSunkCloned, "Number of cloned instructions sunk into loop");
64 
65 static cl::opt<unsigned> SinkFrequencyPercentThreshold(
66     "sink-freq-percent-threshold", cl::Hidden, cl::init(90),
67     cl::desc("Do not sink instructions that require cloning unless they "
68              "execute less than this percent of the time."));
69 
70 static cl::opt<unsigned> MaxNumberOfUseBBsForSinking(
71     "max-uses-for-sinking", cl::Hidden, cl::init(30),
72     cl::desc("Do not sink instructions that have too many uses."));
73 
74 static cl::opt<bool> EnableMSSAInLoopSink(
75     "enable-mssa-in-loop-sink", cl::Hidden, cl::init(true),
76     cl::desc("Enable MemorySSA for LoopSink in new pass manager"));
77 
78 static cl::opt<bool> EnableMSSAInLegacyLoopSink(
79     "enable-mssa-in-legacy-loop-sink", cl::Hidden, cl::init(false),
80     cl::desc("Enable MemorySSA for LoopSink in legacy pass manager"));
81 
82 /// Return adjusted total frequency of \p BBs.
83 ///
84 /// * If there is only one BB, sinking instruction will not introduce code
85 ///   size increase. Thus there is no need to adjust the frequency.
86 /// * If there are more than one BB, sinking would lead to code size increase.
87 ///   In this case, we add some "tax" to the total frequency to make it harder
88 ///   to sink. E.g.
89 ///     Freq(Preheader) = 100
90 ///     Freq(BBs) = sum(50, 49) = 99
91 ///   Even if Freq(BBs) < Freq(Preheader), we will not sink from Preheade to
92 ///   BBs as the difference is too small to justify the code size increase.
93 ///   To model this, The adjusted Freq(BBs) will be:
94 ///     AdjustedFreq(BBs) = 99 / SinkFrequencyPercentThreshold%
95 static BlockFrequency adjustedSumFreq(SmallPtrSetImpl<BasicBlock *> &BBs,
96                                       BlockFrequencyInfo &BFI) {
97   BlockFrequency T = 0;
98   for (BasicBlock *B : BBs)
99     T += BFI.getBlockFreq(B);
100   if (BBs.size() > 1)
101     T /= BranchProbability(SinkFrequencyPercentThreshold, 100);
102   return T;
103 }
104 
105 /// Return a set of basic blocks to insert sinked instructions.
106 ///
107 /// The returned set of basic blocks (BBsToSinkInto) should satisfy:
108 ///
109 /// * Inside the loop \p L
110 /// * For each UseBB in \p UseBBs, there is at least one BB in BBsToSinkInto
111 ///   that domintates the UseBB
112 /// * Has minimum total frequency that is no greater than preheader frequency
113 ///
114 /// The purpose of the function is to find the optimal sinking points to
115 /// minimize execution cost, which is defined as "sum of frequency of
116 /// BBsToSinkInto".
117 /// As a result, the returned BBsToSinkInto needs to have minimum total
118 /// frequency.
119 /// Additionally, if the total frequency of BBsToSinkInto exceeds preheader
120 /// frequency, the optimal solution is not sinking (return empty set).
121 ///
122 /// \p ColdLoopBBs is used to help find the optimal sinking locations.
123 /// It stores a list of BBs that is:
124 ///
125 /// * Inside the loop \p L
126 /// * Has a frequency no larger than the loop's preheader
127 /// * Sorted by BB frequency
128 ///
129 /// The complexity of the function is O(UseBBs.size() * ColdLoopBBs.size()).
130 /// To avoid expensive computation, we cap the maximum UseBBs.size() in its
131 /// caller.
132 static SmallPtrSet<BasicBlock *, 2>
133 findBBsToSinkInto(const Loop &L, const SmallPtrSetImpl<BasicBlock *> &UseBBs,
134                   const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
135                   DominatorTree &DT, BlockFrequencyInfo &BFI) {
136   SmallPtrSet<BasicBlock *, 2> BBsToSinkInto;
137   if (UseBBs.size() == 0)
138     return BBsToSinkInto;
139 
140   BBsToSinkInto.insert(UseBBs.begin(), UseBBs.end());
141   SmallPtrSet<BasicBlock *, 2> BBsDominatedByColdestBB;
142 
143   // For every iteration:
144   //   * Pick the ColdestBB from ColdLoopBBs
145   //   * Find the set BBsDominatedByColdestBB that satisfy:
146   //     - BBsDominatedByColdestBB is a subset of BBsToSinkInto
147   //     - Every BB in BBsDominatedByColdestBB is dominated by ColdestBB
148   //   * If Freq(ColdestBB) < Freq(BBsDominatedByColdestBB), remove
149   //     BBsDominatedByColdestBB from BBsToSinkInto, add ColdestBB to
150   //     BBsToSinkInto
151   for (BasicBlock *ColdestBB : ColdLoopBBs) {
152     BBsDominatedByColdestBB.clear();
153     for (BasicBlock *SinkedBB : BBsToSinkInto)
154       if (DT.dominates(ColdestBB, SinkedBB))
155         BBsDominatedByColdestBB.insert(SinkedBB);
156     if (BBsDominatedByColdestBB.size() == 0)
157       continue;
158     if (adjustedSumFreq(BBsDominatedByColdestBB, BFI) >
159         BFI.getBlockFreq(ColdestBB)) {
160       for (BasicBlock *DominatedBB : BBsDominatedByColdestBB) {
161         BBsToSinkInto.erase(DominatedBB);
162       }
163       BBsToSinkInto.insert(ColdestBB);
164     }
165   }
166 
167   // Can't sink into blocks that have no valid insertion point.
168   for (BasicBlock *BB : BBsToSinkInto) {
169     if (BB->getFirstInsertionPt() == BB->end()) {
170       BBsToSinkInto.clear();
171       break;
172     }
173   }
174 
175   // If the total frequency of BBsToSinkInto is larger than preheader frequency,
176   // do not sink.
177   if (adjustedSumFreq(BBsToSinkInto, BFI) >
178       BFI.getBlockFreq(L.getLoopPreheader()))
179     BBsToSinkInto.clear();
180   return BBsToSinkInto;
181 }
182 
183 // Sinks \p I from the loop \p L's preheader to its uses. Returns true if
184 // sinking is successful.
185 // \p LoopBlockNumber is used to sort the insertion blocks to ensure
186 // determinism.
187 static bool sinkInstruction(
188     Loop &L, Instruction &I, const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
189     const SmallDenseMap<BasicBlock *, int, 16> &LoopBlockNumber, LoopInfo &LI,
190     DominatorTree &DT, BlockFrequencyInfo &BFI, MemorySSAUpdater *MSSAU) {
191   // Compute the set of blocks in loop L which contain a use of I.
192   SmallPtrSet<BasicBlock *, 2> BBs;
193   for (auto &U : I.uses()) {
194     Instruction *UI = cast<Instruction>(U.getUser());
195     // We cannot sink I to PHI-uses.
196     if (isa<PHINode>(UI))
197       return false;
198     // We cannot sink I if it has uses outside of the loop.
199     if (!L.contains(LI.getLoopFor(UI->getParent())))
200       return false;
201     BBs.insert(UI->getParent());
202   }
203 
204   // findBBsToSinkInto is O(BBs.size() * ColdLoopBBs.size()). We cap the max
205   // BBs.size() to avoid expensive computation.
206   // FIXME: Handle code size growth for min_size and opt_size.
207   if (BBs.size() > MaxNumberOfUseBBsForSinking)
208     return false;
209 
210   // Find the set of BBs that we should insert a copy of I.
211   SmallPtrSet<BasicBlock *, 2> BBsToSinkInto =
212       findBBsToSinkInto(L, BBs, ColdLoopBBs, DT, BFI);
213   if (BBsToSinkInto.empty())
214     return false;
215 
216   // Return if any of the candidate blocks to sink into is non-cold.
217   if (BBsToSinkInto.size() > 1 &&
218       !llvm::set_is_subset(BBsToSinkInto, LoopBlockNumber))
219     return false;
220 
221   // Copy the final BBs into a vector and sort them using the total ordering
222   // of the loop block numbers as iterating the set doesn't give a useful
223   // order. No need to stable sort as the block numbers are a total ordering.
224   SmallVector<BasicBlock *, 2> SortedBBsToSinkInto;
225   llvm::append_range(SortedBBsToSinkInto, BBsToSinkInto);
226   llvm::sort(SortedBBsToSinkInto, [&](BasicBlock *A, BasicBlock *B) {
227     return LoopBlockNumber.find(A)->second < LoopBlockNumber.find(B)->second;
228   });
229 
230   BasicBlock *MoveBB = *SortedBBsToSinkInto.begin();
231   // FIXME: Optimize the efficiency for cloned value replacement. The current
232   //        implementation is O(SortedBBsToSinkInto.size() * I.num_uses()).
233   for (BasicBlock *N : makeArrayRef(SortedBBsToSinkInto).drop_front(1)) {
234     assert(LoopBlockNumber.find(N)->second >
235                LoopBlockNumber.find(MoveBB)->second &&
236            "BBs not sorted!");
237     // Clone I and replace its uses.
238     Instruction *IC = I.clone();
239     IC->setName(I.getName());
240     IC->insertBefore(&*N->getFirstInsertionPt());
241 
242     if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
243       // Create a new MemoryAccess and let MemorySSA set its defining access.
244       MemoryAccess *NewMemAcc =
245           MSSAU->createMemoryAccessInBB(IC, nullptr, N, MemorySSA::Beginning);
246       if (NewMemAcc) {
247         if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
248           MSSAU->insertDef(MemDef, /*RenameUses=*/true);
249         else {
250           auto *MemUse = cast<MemoryUse>(NewMemAcc);
251           MSSAU->insertUse(MemUse, /*RenameUses=*/true);
252         }
253       }
254     }
255 
256     // Replaces uses of I with IC in N
257     I.replaceUsesWithIf(IC, [N](Use &U) {
258       return cast<Instruction>(U.getUser())->getParent() == N;
259     });
260     // Replaces uses of I with IC in blocks dominated by N
261     replaceDominatedUsesWith(&I, IC, DT, N);
262     LLVM_DEBUG(dbgs() << "Sinking a clone of " << I << " To: " << N->getName()
263                       << '\n');
264     NumLoopSunkCloned++;
265   }
266   LLVM_DEBUG(dbgs() << "Sinking " << I << " To: " << MoveBB->getName() << '\n');
267   NumLoopSunk++;
268   I.moveBefore(&*MoveBB->getFirstInsertionPt());
269 
270   if (MSSAU)
271     if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
272             MSSAU->getMemorySSA()->getMemoryAccess(&I)))
273       MSSAU->moveToPlace(OldMemAcc, MoveBB, MemorySSA::Beginning);
274 
275   return true;
276 }
277 
278 /// Sinks instructions from loop's preheader to the loop body if the
279 /// sum frequency of inserted copy is smaller than preheader's frequency.
280 static bool sinkLoopInvariantInstructions(Loop &L, AAResults &AA, LoopInfo &LI,
281                                           DominatorTree &DT,
282                                           BlockFrequencyInfo &BFI,
283                                           ScalarEvolution *SE,
284                                           AliasSetTracker *CurAST,
285                                           MemorySSA *MSSA) {
286   BasicBlock *Preheader = L.getLoopPreheader();
287   assert(Preheader && "Expected loop to have preheader");
288 
289   assert(Preheader->getParent()->hasProfileData() &&
290          "Unexpected call when profile data unavailable.");
291 
292   const BlockFrequency PreheaderFreq = BFI.getBlockFreq(Preheader);
293   // If there are no basic blocks with lower frequency than the preheader then
294   // we can avoid the detailed analysis as we will never find profitable sinking
295   // opportunities.
296   if (all_of(L.blocks(), [&](const BasicBlock *BB) {
297         return BFI.getBlockFreq(BB) > PreheaderFreq;
298       }))
299     return false;
300 
301   std::unique_ptr<MemorySSAUpdater> MSSAU;
302   std::unique_ptr<SinkAndHoistLICMFlags> LICMFlags;
303   if (MSSA) {
304     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
305     LICMFlags =
306         std::make_unique<SinkAndHoistLICMFlags>(/*IsSink=*/true, &L, MSSA);
307   }
308 
309   bool Changed = false;
310 
311   // Sort loop's basic blocks by frequency
312   SmallVector<BasicBlock *, 10> ColdLoopBBs;
313   SmallDenseMap<BasicBlock *, int, 16> LoopBlockNumber;
314   int i = 0;
315   for (BasicBlock *B : L.blocks())
316     if (BFI.getBlockFreq(B) < BFI.getBlockFreq(L.getLoopPreheader())) {
317       ColdLoopBBs.push_back(B);
318       LoopBlockNumber[B] = ++i;
319     }
320   llvm::stable_sort(ColdLoopBBs, [&](BasicBlock *A, BasicBlock *B) {
321     return BFI.getBlockFreq(A) < BFI.getBlockFreq(B);
322   });
323 
324   // Traverse preheader's instructions in reverse order becaue if A depends
325   // on B (A appears after B), A needs to be sinked first before B can be
326   // sinked.
327   for (Instruction &I : llvm::make_early_inc_range(llvm::reverse(*Preheader))) {
328     // No need to check for instruction's operands are loop invariant.
329     assert(L.hasLoopInvariantOperands(&I) &&
330            "Insts in a loop's preheader should have loop invariant operands!");
331     if (!canSinkOrHoistInst(I, &AA, &DT, &L, CurAST, MSSAU.get(), false,
332                             LICMFlags.get()))
333       continue;
334     if (sinkInstruction(L, I, ColdLoopBBs, LoopBlockNumber, LI, DT, BFI,
335                         MSSAU.get()))
336       Changed = true;
337   }
338 
339   if (Changed && SE)
340     SE->forgetLoopDispositions(&L);
341   return Changed;
342 }
343 
344 static void computeAliasSet(Loop &L, BasicBlock &Preheader,
345                             AliasSetTracker &CurAST) {
346   for (BasicBlock *BB : L.blocks())
347     CurAST.add(*BB);
348   CurAST.add(Preheader);
349 }
350 
351 PreservedAnalyses LoopSinkPass::run(Function &F, FunctionAnalysisManager &FAM) {
352   LoopInfo &LI = FAM.getResult<LoopAnalysis>(F);
353   // Nothing to do if there are no loops.
354   if (LI.empty())
355     return PreservedAnalyses::all();
356 
357   AAResults &AA = FAM.getResult<AAManager>(F);
358   DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F);
359   BlockFrequencyInfo &BFI = FAM.getResult<BlockFrequencyAnalysis>(F);
360 
361   MemorySSA *MSSA = EnableMSSAInLoopSink
362                         ? &FAM.getResult<MemorySSAAnalysis>(F).getMSSA()
363                         : nullptr;
364 
365   // We want to do a postorder walk over the loops. Since loops are a tree this
366   // is equivalent to a reversed preorder walk and preorder is easy to compute
367   // without recursion. Since we reverse the preorder, we will visit siblings
368   // in reverse program order. This isn't expected to matter at all but is more
369   // consistent with sinking algorithms which generally work bottom-up.
370   SmallVector<Loop *, 4> PreorderLoops = LI.getLoopsInPreorder();
371 
372   bool Changed = false;
373   do {
374     Loop &L = *PreorderLoops.pop_back_val();
375 
376     BasicBlock *Preheader = L.getLoopPreheader();
377     if (!Preheader)
378       continue;
379 
380     // Enable LoopSink only when runtime profile is available.
381     // With static profile, the sinking decision may be sub-optimal.
382     if (!Preheader->getParent()->hasProfileData())
383       continue;
384 
385     std::unique_ptr<AliasSetTracker> CurAST;
386     if (!EnableMSSAInLoopSink) {
387       CurAST = std::make_unique<AliasSetTracker>(AA);
388       computeAliasSet(L, *Preheader, *CurAST.get());
389     }
390 
391     // Note that we don't pass SCEV here because it is only used to invalidate
392     // loops in SCEV and we don't preserve (or request) SCEV at all making that
393     // unnecessary.
394     Changed |= sinkLoopInvariantInstructions(L, AA, LI, DT, BFI,
395                                              /*ScalarEvolution*/ nullptr,
396                                              CurAST.get(), MSSA);
397   } while (!PreorderLoops.empty());
398 
399   if (!Changed)
400     return PreservedAnalyses::all();
401 
402   PreservedAnalyses PA;
403   PA.preserveSet<CFGAnalyses>();
404 
405   if (MSSA) {
406     PA.preserve<MemorySSAAnalysis>();
407 
408     if (VerifyMemorySSA)
409       MSSA->verifyMemorySSA();
410   }
411 
412   return PA;
413 }
414 
415 namespace {
416 struct LegacyLoopSinkPass : public LoopPass {
417   static char ID;
418   LegacyLoopSinkPass() : LoopPass(ID) {
419     initializeLegacyLoopSinkPassPass(*PassRegistry::getPassRegistry());
420   }
421 
422   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
423     if (skipLoop(L))
424       return false;
425 
426     BasicBlock *Preheader = L->getLoopPreheader();
427     if (!Preheader)
428       return false;
429 
430     // Enable LoopSink only when runtime profile is available.
431     // With static profile, the sinking decision may be sub-optimal.
432     if (!Preheader->getParent()->hasProfileData())
433       return false;
434 
435     AAResults &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
436     auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
437     std::unique_ptr<AliasSetTracker> CurAST;
438     MemorySSA *MSSA = nullptr;
439     if (EnableMSSAInLegacyLoopSink)
440       MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
441     else {
442       CurAST = std::make_unique<AliasSetTracker>(AA);
443       computeAliasSet(*L, *Preheader, *CurAST.get());
444     }
445 
446     bool Changed = sinkLoopInvariantInstructions(
447         *L, AA, getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
448         getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
449         getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(),
450         SE ? &SE->getSE() : nullptr, CurAST.get(), MSSA);
451 
452     if (MSSA && VerifyMemorySSA)
453       MSSA->verifyMemorySSA();
454 
455     return Changed;
456   }
457 
458   void getAnalysisUsage(AnalysisUsage &AU) const override {
459     AU.setPreservesCFG();
460     AU.addRequired<BlockFrequencyInfoWrapperPass>();
461     getLoopAnalysisUsage(AU);
462     if (EnableMSSAInLegacyLoopSink) {
463       AU.addRequired<MemorySSAWrapperPass>();
464       AU.addPreserved<MemorySSAWrapperPass>();
465     }
466   }
467 };
468 }
469 
470 char LegacyLoopSinkPass::ID = 0;
471 INITIALIZE_PASS_BEGIN(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false,
472                       false)
473 INITIALIZE_PASS_DEPENDENCY(LoopPass)
474 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
475 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
476 INITIALIZE_PASS_END(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false, false)
477 
478 Pass *llvm::createLoopSinkPass() { return new LegacyLoopSinkPass(); }
479