xref: /llvm-project/llvm/lib/Transforms/Scalar/LoopSink.cpp (revision e29292969b92aa15afba734d4f6863fc405f087c)
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/Statistic.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/AliasSetTracker.h"
37 #include "llvm/Analysis/BasicAliasAnalysis.h"
38 #include "llvm/Analysis/BlockFrequencyInfo.h"
39 #include "llvm/Analysis/Loads.h"
40 #include "llvm/Analysis/LoopInfo.h"
41 #include "llvm/Analysis/LoopPass.h"
42 #include "llvm/Analysis/MemorySSA.h"
43 #include "llvm/Analysis/MemorySSAUpdater.h"
44 #include "llvm/Analysis/ScalarEvolution.h"
45 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
46 #include "llvm/IR/Dominators.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/Metadata.h"
50 #include "llvm/InitializePasses.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Transforms/Scalar.h"
53 #include "llvm/Transforms/Scalar/LoopPassManager.h"
54 #include "llvm/Transforms/Utils/Local.h"
55 #include "llvm/Transforms/Utils/LoopUtils.h"
56 using namespace llvm;
57 
58 #define DEBUG_TYPE "loopsink"
59 
60 STATISTIC(NumLoopSunk, "Number of instructions sunk into loop");
61 STATISTIC(NumLoopSunkCloned, "Number of cloned instructions sunk into loop");
62 
63 static cl::opt<unsigned> SinkFrequencyPercentThreshold(
64     "sink-freq-percent-threshold", cl::Hidden, cl::init(90),
65     cl::desc("Do not sink instructions that require cloning unless they "
66              "execute less than this percent of the time."));
67 
68 static cl::opt<unsigned> MaxNumberOfUseBBsForSinking(
69     "max-uses-for-sinking", cl::Hidden, cl::init(30),
70     cl::desc("Do not sink instructions that have too many uses."));
71 
72 static cl::opt<bool> EnableMSSAInLoopSink(
73     "enable-mssa-in-loop-sink", cl::Hidden, cl::init(false),
74     cl::desc("Enable MemorySSA for LoopSink in new pass manager"));
75 
76 static cl::opt<bool> EnableMSSAInLegacyLoopSink(
77     "enable-mssa-in-legacy-loop-sink", cl::Hidden, cl::init(false),
78     cl::desc("Enable MemorySSA for LoopSink in legacy pass manager"));
79 
80 /// Return adjusted total frequency of \p BBs.
81 ///
82 /// * If there is only one BB, sinking instruction will not introduce code
83 ///   size increase. Thus there is no need to adjust the frequency.
84 /// * If there are more than one BB, sinking would lead to code size increase.
85 ///   In this case, we add some "tax" to the total frequency to make it harder
86 ///   to sink. E.g.
87 ///     Freq(Preheader) = 100
88 ///     Freq(BBs) = sum(50, 49) = 99
89 ///   Even if Freq(BBs) < Freq(Preheader), we will not sink from Preheade to
90 ///   BBs as the difference is too small to justify the code size increase.
91 ///   To model this, The adjusted Freq(BBs) will be:
92 ///     AdjustedFreq(BBs) = 99 / SinkFrequencyPercentThreshold%
93 static BlockFrequency adjustedSumFreq(SmallPtrSetImpl<BasicBlock *> &BBs,
94                                       BlockFrequencyInfo &BFI) {
95   BlockFrequency T = 0;
96   for (BasicBlock *B : BBs)
97     T += BFI.getBlockFreq(B);
98   if (BBs.size() > 1)
99     T /= BranchProbability(SinkFrequencyPercentThreshold, 100);
100   return T;
101 }
102 
103 /// Return a set of basic blocks to insert sinked instructions.
104 ///
105 /// The returned set of basic blocks (BBsToSinkInto) should satisfy:
106 ///
107 /// * Inside the loop \p L
108 /// * For each UseBB in \p UseBBs, there is at least one BB in BBsToSinkInto
109 ///   that domintates the UseBB
110 /// * Has minimum total frequency that is no greater than preheader frequency
111 ///
112 /// The purpose of the function is to find the optimal sinking points to
113 /// minimize execution cost, which is defined as "sum of frequency of
114 /// BBsToSinkInto".
115 /// As a result, the returned BBsToSinkInto needs to have minimum total
116 /// frequency.
117 /// Additionally, if the total frequency of BBsToSinkInto exceeds preheader
118 /// frequency, the optimal solution is not sinking (return empty set).
119 ///
120 /// \p ColdLoopBBs is used to help find the optimal sinking locations.
121 /// It stores a list of BBs that is:
122 ///
123 /// * Inside the loop \p L
124 /// * Has a frequency no larger than the loop's preheader
125 /// * Sorted by BB frequency
126 ///
127 /// The complexity of the function is O(UseBBs.size() * ColdLoopBBs.size()).
128 /// To avoid expensive computation, we cap the maximum UseBBs.size() in its
129 /// caller.
130 static SmallPtrSet<BasicBlock *, 2>
131 findBBsToSinkInto(const Loop &L, const SmallPtrSetImpl<BasicBlock *> &UseBBs,
132                   const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
133                   DominatorTree &DT, BlockFrequencyInfo &BFI) {
134   SmallPtrSet<BasicBlock *, 2> BBsToSinkInto;
135   if (UseBBs.size() == 0)
136     return BBsToSinkInto;
137 
138   BBsToSinkInto.insert(UseBBs.begin(), UseBBs.end());
139   SmallPtrSet<BasicBlock *, 2> BBsDominatedByColdestBB;
140 
141   // For every iteration:
142   //   * Pick the ColdestBB from ColdLoopBBs
143   //   * Find the set BBsDominatedByColdestBB that satisfy:
144   //     - BBsDominatedByColdestBB is a subset of BBsToSinkInto
145   //     - Every BB in BBsDominatedByColdestBB is dominated by ColdestBB
146   //   * If Freq(ColdestBB) < Freq(BBsDominatedByColdestBB), remove
147   //     BBsDominatedByColdestBB from BBsToSinkInto, add ColdestBB to
148   //     BBsToSinkInto
149   for (BasicBlock *ColdestBB : ColdLoopBBs) {
150     BBsDominatedByColdestBB.clear();
151     for (BasicBlock *SinkedBB : BBsToSinkInto)
152       if (DT.dominates(ColdestBB, SinkedBB))
153         BBsDominatedByColdestBB.insert(SinkedBB);
154     if (BBsDominatedByColdestBB.size() == 0)
155       continue;
156     if (adjustedSumFreq(BBsDominatedByColdestBB, BFI) >
157         BFI.getBlockFreq(ColdestBB)) {
158       for (BasicBlock *DominatedBB : BBsDominatedByColdestBB) {
159         BBsToSinkInto.erase(DominatedBB);
160       }
161       BBsToSinkInto.insert(ColdestBB);
162     }
163   }
164 
165   // Can't sink into blocks that have no valid insertion point.
166   for (BasicBlock *BB : BBsToSinkInto) {
167     if (BB->getFirstInsertionPt() == BB->end()) {
168       BBsToSinkInto.clear();
169       break;
170     }
171   }
172 
173   // If the total frequency of BBsToSinkInto is larger than preheader frequency,
174   // do not sink.
175   if (adjustedSumFreq(BBsToSinkInto, BFI) >
176       BFI.getBlockFreq(L.getLoopPreheader()))
177     BBsToSinkInto.clear();
178   return BBsToSinkInto;
179 }
180 
181 // Sinks \p I from the loop \p L's preheader to its uses. Returns true if
182 // sinking is successful.
183 // \p LoopBlockNumber is used to sort the insertion blocks to ensure
184 // determinism.
185 static bool sinkInstruction(
186     Loop &L, Instruction &I, const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
187     const SmallDenseMap<BasicBlock *, int, 16> &LoopBlockNumber, LoopInfo &LI,
188     DominatorTree &DT, BlockFrequencyInfo &BFI, MemorySSAUpdater *MSSAU) {
189   // Compute the set of blocks in loop L which contain a use of I.
190   SmallPtrSet<BasicBlock *, 2> BBs;
191   for (auto &U : I.uses()) {
192     Instruction *UI = cast<Instruction>(U.getUser());
193     // We cannot sink I to PHI-uses.
194     if (dyn_cast<PHINode>(UI))
195       return false;
196     // We cannot sink I if it has uses outside of the loop.
197     if (!L.contains(LI.getLoopFor(UI->getParent())))
198       return false;
199     BBs.insert(UI->getParent());
200   }
201 
202   // findBBsToSinkInto is O(BBs.size() * ColdLoopBBs.size()). We cap the max
203   // BBs.size() to avoid expensive computation.
204   // FIXME: Handle code size growth for min_size and opt_size.
205   if (BBs.size() > MaxNumberOfUseBBsForSinking)
206     return false;
207 
208   // Find the set of BBs that we should insert a copy of I.
209   SmallPtrSet<BasicBlock *, 2> BBsToSinkInto =
210       findBBsToSinkInto(L, BBs, ColdLoopBBs, DT, BFI);
211   if (BBsToSinkInto.empty())
212     return false;
213 
214   // Return if any of the candidate blocks to sink into is non-cold.
215   if (BBsToSinkInto.size() > 1) {
216     for (auto *BB : BBsToSinkInto)
217       if (!LoopBlockNumber.count(BB))
218         return false;
219   }
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   SortedBBsToSinkInto.insert(SortedBBsToSinkInto.begin(), BBsToSinkInto.begin(),
226                              BBsToSinkInto.end());
227   llvm::sort(SortedBBsToSinkInto, [&](BasicBlock *A, BasicBlock *B) {
228     return LoopBlockNumber.find(A)->second < LoopBlockNumber.find(B)->second;
229   });
230 
231   BasicBlock *MoveBB = *SortedBBsToSinkInto.begin();
232   // FIXME: Optimize the efficiency for cloned value replacement. The current
233   //        implementation is O(SortedBBsToSinkInto.size() * I.num_uses()).
234   for (BasicBlock *N : makeArrayRef(SortedBBsToSinkInto).drop_front(1)) {
235     assert(LoopBlockNumber.find(N)->second >
236                LoopBlockNumber.find(MoveBB)->second &&
237            "BBs not sorted!");
238     // Clone I and replace its uses.
239     Instruction *IC = I.clone();
240     IC->setName(I.getName());
241     IC->insertBefore(&*N->getFirstInsertionPt());
242 
243     if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
244       // Create a new MemoryAccess and let MemorySSA set its defining access.
245       MemoryAccess *NewMemAcc =
246           MSSAU->createMemoryAccessInBB(IC, nullptr, N, MemorySSA::Beginning);
247       if (NewMemAcc) {
248         if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
249           MSSAU->insertDef(MemDef, /*RenameUses=*/true);
250         else {
251           auto *MemUse = cast<MemoryUse>(NewMemAcc);
252           MSSAU->insertUse(MemUse, /*RenameUses=*/true);
253         }
254       }
255     }
256 
257     // Replaces uses of I with IC in N
258     I.replaceUsesWithIf(IC, [N](Use &U) {
259       return cast<Instruction>(U.getUser())->getParent() == N;
260     });
261     // Replaces uses of I with IC in blocks dominated by N
262     replaceDominatedUsesWith(&I, IC, DT, N);
263     LLVM_DEBUG(dbgs() << "Sinking a clone of " << I << " To: " << N->getName()
264                       << '\n');
265     NumLoopSunkCloned++;
266   }
267   LLVM_DEBUG(dbgs() << "Sinking " << I << " To: " << MoveBB->getName() << '\n');
268   NumLoopSunk++;
269   I.moveBefore(&*MoveBB->getFirstInsertionPt());
270 
271   if (MSSAU)
272     if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
273             MSSAU->getMemorySSA()->getMemoryAccess(&I)))
274       MSSAU->moveToPlace(OldMemAcc, MoveBB, MemorySSA::Beginning);
275 
276   return true;
277 }
278 
279 /// Sinks instructions from loop's preheader to the loop body if the
280 /// sum frequency of inserted copy is smaller than preheader's frequency.
281 static bool sinkLoopInvariantInstructions(Loop &L, AAResults &AA, LoopInfo &LI,
282                                           DominatorTree &DT,
283                                           BlockFrequencyInfo &BFI,
284                                           ScalarEvolution *SE,
285                                           AliasSetTracker *CurAST,
286                                           MemorySSA *MSSA) {
287   BasicBlock *Preheader = L.getLoopPreheader();
288   assert(Preheader && "Expected loop to have preheader");
289 
290   // Enable LoopSink only when runtime profile is available.
291   // With static profile, the sinking decision may be sub-optimal.
292   if (!Preheader->getParent()->hasProfileData())
293     return false;
294 
295   const BlockFrequency PreheaderFreq = BFI.getBlockFreq(Preheader);
296   // If there are no basic blocks with lower frequency than the preheader then
297   // we can avoid the detailed analysis as we will never find profitable sinking
298   // opportunities.
299   if (all_of(L.blocks(), [&](const BasicBlock *BB) {
300         return BFI.getBlockFreq(BB) > PreheaderFreq;
301       }))
302     return false;
303 
304   std::unique_ptr<MemorySSAUpdater> MSSAU;
305   std::unique_ptr<SinkAndHoistLICMFlags> LICMFlags;
306   if (MSSA) {
307     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
308     LICMFlags =
309         std::make_unique<SinkAndHoistLICMFlags>(/*IsSink=*/true, &L, MSSA);
310   }
311 
312   bool Changed = false;
313 
314   // Sort loop's basic blocks by frequency
315   SmallVector<BasicBlock *, 10> ColdLoopBBs;
316   SmallDenseMap<BasicBlock *, int, 16> LoopBlockNumber;
317   int i = 0;
318   for (BasicBlock *B : L.blocks())
319     if (BFI.getBlockFreq(B) < BFI.getBlockFreq(L.getLoopPreheader())) {
320       ColdLoopBBs.push_back(B);
321       LoopBlockNumber[B] = ++i;
322     }
323   llvm::stable_sort(ColdLoopBBs, [&](BasicBlock *A, BasicBlock *B) {
324     return BFI.getBlockFreq(A) < BFI.getBlockFreq(B);
325   });
326 
327   // Traverse preheader's instructions in reverse order becaue if A depends
328   // on B (A appears after B), A needs to be sinked first before B can be
329   // sinked.
330   for (auto II = Preheader->rbegin(), E = Preheader->rend(); II != E;) {
331     Instruction *I = &*II++;
332     // No need to check for instruction's operands are loop invariant.
333     assert(L.hasLoopInvariantOperands(I) &&
334            "Insts in a loop's preheader should have loop invariant operands!");
335     if (!canSinkOrHoistInst(*I, &AA, &DT, &L, CurAST, MSSAU.get(), false,
336                             LICMFlags.get()))
337       continue;
338     if (sinkInstruction(L, *I, ColdLoopBBs, LoopBlockNumber, LI, DT, BFI,
339                         MSSAU.get()))
340       Changed = true;
341   }
342 
343   if (Changed && SE)
344     SE->forgetLoopDispositions(&L);
345   return Changed;
346 }
347 
348 static void computeAliasSet(Loop &L, BasicBlock &Preheader,
349                             AliasSetTracker &CurAST) {
350   for (BasicBlock *BB : L.blocks())
351     CurAST.add(*BB);
352   CurAST.add(Preheader);
353 }
354 
355 PreservedAnalyses LoopSinkPass::run(Function &F, FunctionAnalysisManager &FAM) {
356   LoopInfo &LI = FAM.getResult<LoopAnalysis>(F);
357   // Nothing to do if there are no loops.
358   if (LI.empty())
359     return PreservedAnalyses::all();
360 
361   AAResults &AA = FAM.getResult<AAManager>(F);
362   DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F);
363   BlockFrequencyInfo &BFI = FAM.getResult<BlockFrequencyAnalysis>(F);
364 
365   MemorySSA *MSSA = EnableMSSAInLoopSink
366                         ? &FAM.getResult<MemorySSAAnalysis>(F).getMSSA()
367                         : nullptr;
368 
369   // We want to do a postorder walk over the loops. Since loops are a tree this
370   // is equivalent to a reversed preorder walk and preorder is easy to compute
371   // without recursion. Since we reverse the preorder, we will visit siblings
372   // in reverse program order. This isn't expected to matter at all but is more
373   // consistent with sinking algorithms which generally work bottom-up.
374   SmallVector<Loop *, 4> PreorderLoops = LI.getLoopsInPreorder();
375 
376   bool Changed = false;
377   do {
378     Loop &L = *PreorderLoops.pop_back_val();
379 
380     BasicBlock *Preheader = L.getLoopPreheader();
381     if (!Preheader)
382       continue;
383 
384     std::unique_ptr<AliasSetTracker> CurAST;
385     if (!EnableMSSAInLoopSink) {
386       CurAST = std::make_unique<AliasSetTracker>(AA);
387       computeAliasSet(L, *Preheader, *CurAST.get());
388     }
389 
390     // Note that we don't pass SCEV here because it is only used to invalidate
391     // loops in SCEV and we don't preserve (or request) SCEV at all making that
392     // unnecessary.
393     Changed |= sinkLoopInvariantInstructions(L, AA, LI, DT, BFI,
394                                              /*ScalarEvolution*/ nullptr,
395                                              CurAST.get(), MSSA);
396   } while (!PreorderLoops.empty());
397 
398   if (!Changed)
399     return PreservedAnalyses::all();
400 
401   PreservedAnalyses PA;
402   PA.preserveSet<CFGAnalyses>();
403 
404   if (MSSA) {
405     PA.preserve<MemorySSAAnalysis>();
406 
407     if (VerifyMemorySSA)
408       MSSA->verifyMemorySSA();
409   }
410 
411   return PA;
412 }
413 
414 namespace {
415 struct LegacyLoopSinkPass : public LoopPass {
416   static char ID;
417   LegacyLoopSinkPass() : LoopPass(ID) {
418     initializeLegacyLoopSinkPassPass(*PassRegistry::getPassRegistry());
419   }
420 
421   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
422     if (skipLoop(L))
423       return false;
424 
425     BasicBlock *Preheader = L->getLoopPreheader();
426     if (!Preheader)
427       return false;
428 
429     AAResults &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
430     auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
431     std::unique_ptr<AliasSetTracker> CurAST;
432     MemorySSA *MSSA = nullptr;
433     if (EnableMSSAInLegacyLoopSink)
434       MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
435     else {
436       CurAST = std::make_unique<AliasSetTracker>(AA);
437       computeAliasSet(*L, *Preheader, *CurAST.get());
438     }
439 
440     bool Changed = sinkLoopInvariantInstructions(
441         *L, AA, getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
442         getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
443         getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(),
444         SE ? &SE->getSE() : nullptr, CurAST.get(), MSSA);
445 
446     if (MSSA && VerifyMemorySSA)
447       MSSA->verifyMemorySSA();
448 
449     return Changed;
450   }
451 
452   void getAnalysisUsage(AnalysisUsage &AU) const override {
453     AU.setPreservesCFG();
454     AU.addRequired<BlockFrequencyInfoWrapperPass>();
455     getLoopAnalysisUsage(AU);
456     if (EnableMSSAInLegacyLoopSink) {
457       AU.addRequired<MemorySSAWrapperPass>();
458       AU.addPreserved<MemorySSAWrapperPass>();
459     }
460   }
461 };
462 }
463 
464 char LegacyLoopSinkPass::ID = 0;
465 INITIALIZE_PASS_BEGIN(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false,
466                       false)
467 INITIALIZE_PASS_DEPENDENCY(LoopPass)
468 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
469 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
470 INITIALIZE_PASS_END(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false, false)
471 
472 Pass *llvm::createLoopSinkPass() { return new LegacyLoopSinkPass(); }
473