1 //===-- LoopSink.cpp - Loop Sink Pass -------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass does the inverse transformation of what LICM does.
11 // It traverses all of the instructions in the loop's preheader and sinks
12 // them to the loop body where frequency is lower than the loop's preheader.
13 // This pass is a reverse-transformation of LICM. It differs from the Sink
14 // pass in the following ways:
16 // * It only handles sinking of instructions from the loop's preheader to the
18 // * It uses alias set tracker to get more accurate alias info
19 // * It uses block frequency info to find the optimal sinking locations
23 // For I in Preheader:
24 // InsertBBs = BBs that uses I
25 // For BB in sorted(LoopBBs):
26 // DomBBs = BBs in InsertBBs that are dominated by BB
27 // if freq(DomBBs) > freq(BB)
28 // InsertBBs = UseBBs - DomBBs + BB
29 // For BB in InsertBBs:
30 // Insert I at BB's beginning
32 //===----------------------------------------------------------------------===//
34 #include "llvm/Transforms/Scalar/LoopSink.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/ScalarEvolution.h"
44 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
45 #include "llvm/IR/Dominators.h"
46 #include "llvm/IR/Instructions.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/Metadata.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Transforms/Scalar.h"
51 #include "llvm/Transforms/Scalar/LoopPassManager.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/LoopUtils.h"
56 #define DEBUG_TYPE "loopsink"
58 STATISTIC(NumLoopSunk, "Number of instructions sunk into loop");
59 STATISTIC(NumLoopSunkCloned, "Number of cloned instructions sunk into loop");
61 static cl::opt<unsigned> SinkFrequencyPercentThreshold(
62 "sink-freq-percent-threshold", cl::Hidden, cl::init(90),
63 cl::desc("Do not sink instructions that require cloning unless they "
64 "execute less than this percent of the time."));
66 static cl::opt<unsigned> MaxNumberOfUseBBsForSinking(
67 "max-uses-for-sinking", cl::Hidden, cl::init(30),
68 cl::desc("Do not sink instructions that have too many uses."));
70 /// Return adjusted total frequency of \p BBs.
72 /// * If there is only one BB, sinking instruction will not introduce code
73 /// size increase. Thus there is no need to adjust the frequency.
74 /// * If there are more than one BB, sinking would lead to code size increase.
75 /// In this case, we add some "tax" to the total frequency to make it harder
77 /// Freq(Preheader) = 100
78 /// Freq(BBs) = sum(50, 49) = 99
79 /// Even if Freq(BBs) < Freq(Preheader), we will not sink from Preheade to
80 /// BBs as the difference is too small to justify the code size increase.
81 /// To model this, The adjusted Freq(BBs) will be:
82 /// AdjustedFreq(BBs) = 99 / SinkFrequencyPercentThreshold%
83 static BlockFrequency adjustedSumFreq(SmallPtrSetImpl<BasicBlock *> &BBs,
84 BlockFrequencyInfo &BFI) {
86 for (BasicBlock *B : BBs)
87 T += BFI.getBlockFreq(B);
89 T /= BranchProbability(SinkFrequencyPercentThreshold, 100);
93 /// Return a set of basic blocks to insert sinked instructions.
95 /// The returned set of basic blocks (BBsToSinkInto) should satisfy:
97 /// * Inside the loop \p L
98 /// * For each UseBB in \p UseBBs, there is at least one BB in BBsToSinkInto
99 /// that domintates the UseBB
100 /// * Has minimum total frequency that is no greater than preheader frequency
102 /// The purpose of the function is to find the optimal sinking points to
103 /// minimize execution cost, which is defined as "sum of frequency of
105 /// As a result, the returned BBsToSinkInto needs to have minimum total
107 /// Additionally, if the total frequency of BBsToSinkInto exceeds preheader
108 /// frequency, the optimal solution is not sinking (return empty set).
110 /// \p ColdLoopBBs is used to help find the optimal sinking locations.
111 /// It stores a list of BBs that is:
113 /// * Inside the loop \p L
114 /// * Has a frequency no larger than the loop's preheader
115 /// * Sorted by BB frequency
117 /// The complexity of the function is O(UseBBs.size() * ColdLoopBBs.size()).
118 /// To avoid expensive computation, we cap the maximum UseBBs.size() in its
120 static SmallPtrSet<BasicBlock *, 2>
121 findBBsToSinkInto(const Loop &L, const SmallPtrSetImpl<BasicBlock *> &UseBBs,
122 const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
123 DominatorTree &DT, BlockFrequencyInfo &BFI) {
124 SmallPtrSet<BasicBlock *, 2> BBsToSinkInto;
125 if (UseBBs.size() == 0)
126 return BBsToSinkInto;
128 BBsToSinkInto.insert(UseBBs.begin(), UseBBs.end());
129 SmallPtrSet<BasicBlock *, 2> BBsDominatedByColdestBB;
131 // For every iteration:
132 // * Pick the ColdestBB from ColdLoopBBs
133 // * Find the set BBsDominatedByColdestBB that satisfy:
134 // - BBsDominatedByColdestBB is a subset of BBsToSinkInto
135 // - Every BB in BBsDominatedByColdestBB is dominated by ColdestBB
136 // * If Freq(ColdestBB) < Freq(BBsDominatedByColdestBB), remove
137 // BBsDominatedByColdestBB from BBsToSinkInto, add ColdestBB to
139 for (BasicBlock *ColdestBB : ColdLoopBBs) {
140 BBsDominatedByColdestBB.clear();
141 for (BasicBlock *SinkedBB : BBsToSinkInto)
142 if (DT.dominates(ColdestBB, SinkedBB))
143 BBsDominatedByColdestBB.insert(SinkedBB);
144 if (BBsDominatedByColdestBB.size() == 0)
146 if (adjustedSumFreq(BBsDominatedByColdestBB, BFI) >
147 BFI.getBlockFreq(ColdestBB)) {
148 for (BasicBlock *DominatedBB : BBsDominatedByColdestBB) {
149 BBsToSinkInto.erase(DominatedBB);
151 BBsToSinkInto.insert(ColdestBB);
155 // If the total frequency of BBsToSinkInto is larger than preheader frequency,
157 if (adjustedSumFreq(BBsToSinkInto, BFI) >
158 BFI.getBlockFreq(L.getLoopPreheader()))
159 BBsToSinkInto.clear();
160 return BBsToSinkInto;
163 // Sinks \p I from the loop \p L's preheader to its uses. Returns true if
164 // sinking is successful.
165 // \p LoopBlockNumber is used to sort the insertion blocks to ensure
167 static bool sinkInstruction(Loop &L, Instruction &I,
168 const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
169 const SmallDenseMap<BasicBlock *, int, 16> &LoopBlockNumber,
170 LoopInfo &LI, DominatorTree &DT,
171 BlockFrequencyInfo &BFI) {
172 // Compute the set of blocks in loop L which contain a use of I.
173 SmallPtrSet<BasicBlock *, 2> BBs;
174 for (auto &U : I.uses()) {
175 Instruction *UI = cast<Instruction>(U.getUser());
176 // We cannot sink I to PHI-uses.
177 if (dyn_cast<PHINode>(UI))
179 // We cannot sink I if it has uses outside of the loop.
180 if (!L.contains(LI.getLoopFor(UI->getParent())))
182 BBs.insert(UI->getParent());
185 // findBBsToSinkInto is O(BBs.size() * ColdLoopBBs.size()). We cap the max
186 // BBs.size() to avoid expensive computation.
187 // FIXME: Handle code size growth for min_size and opt_size.
188 if (BBs.size() > MaxNumberOfUseBBsForSinking)
191 // Find the set of BBs that we should insert a copy of I.
192 SmallPtrSet<BasicBlock *, 2> BBsToSinkInto =
193 findBBsToSinkInto(L, BBs, ColdLoopBBs, DT, BFI);
194 if (BBsToSinkInto.empty())
197 // Copy the final BBs into a vector and sort them using the total ordering
198 // of the loop block numbers as iterating the set doesn't give a useful
199 // order. No need to stable sort as the block numbers are a total ordering.
200 SmallVector<BasicBlock *, 2> SortedBBsToSinkInto;
201 SortedBBsToSinkInto.insert(SortedBBsToSinkInto.begin(), BBsToSinkInto.begin(),
202 BBsToSinkInto.end());
203 std::sort(SortedBBsToSinkInto.begin(), SortedBBsToSinkInto.end(),
204 [&](BasicBlock *A, BasicBlock *B) {
205 return *LoopBlockNumber.find(A) < *LoopBlockNumber.find(B);
208 BasicBlock *MoveBB = *SortedBBsToSinkInto.begin();
209 // FIXME: Optimize the efficiency for cloned value replacement. The current
210 // implementation is O(SortedBBsToSinkInto.size() * I.num_uses()).
211 for (BasicBlock *N : SortedBBsToSinkInto) {
214 // Clone I and replace its uses.
215 Instruction *IC = I.clone();
216 IC->setName(I.getName());
217 IC->insertBefore(&*N->getFirstInsertionPt());
218 // Replaces uses of I with IC in N
219 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;) {
221 auto *I = cast<Instruction>(U.getUser());
222 if (I->getParent() == N)
225 // Replaces uses of I with IC in blocks dominated by N
226 replaceDominatedUsesWith(&I, IC, DT, N);
227 DEBUG(dbgs() << "Sinking a clone of " << I << " To: " << N->getName()
231 DEBUG(dbgs() << "Sinking " << I << " To: " << MoveBB->getName() << '\n');
233 I.moveBefore(&*MoveBB->getFirstInsertionPt());
238 /// Sinks instructions from loop's preheader to the loop body if the
239 /// sum frequency of inserted copy is smaller than preheader's frequency.
240 static bool sinkLoopInvariantInstructions(Loop &L, AAResults &AA, LoopInfo &LI,
242 BlockFrequencyInfo &BFI,
243 ScalarEvolution *SE) {
244 BasicBlock *Preheader = L.getLoopPreheader();
248 // Enable LoopSink only when runtime profile is available.
249 // With static profile, the sinking decision may be sub-optimal.
250 if (!Preheader->getParent()->getEntryCount())
253 const BlockFrequency PreheaderFreq = BFI.getBlockFreq(Preheader);
254 // If there are no basic blocks with lower frequency than the preheader then
255 // we can avoid the detailed analysis as we will never find profitable sinking
257 if (all_of(L.blocks(), [&](const BasicBlock *BB) {
258 return BFI.getBlockFreq(BB) > PreheaderFreq;
262 bool Changed = false;
263 AliasSetTracker CurAST(AA);
265 // Compute alias set.
266 for (BasicBlock *BB : L.blocks())
269 // Sort loop's basic blocks by frequency
270 SmallVector<BasicBlock *, 10> ColdLoopBBs;
271 SmallDenseMap<BasicBlock *, int, 16> LoopBlockNumber;
273 for (BasicBlock *B : L.blocks())
274 if (BFI.getBlockFreq(B) < BFI.getBlockFreq(L.getLoopPreheader())) {
275 ColdLoopBBs.push_back(B);
276 LoopBlockNumber[B] = ++i;
278 std::stable_sort(ColdLoopBBs.begin(), ColdLoopBBs.end(),
279 [&](BasicBlock *A, BasicBlock *B) {
280 return BFI.getBlockFreq(A) < BFI.getBlockFreq(B);
283 // Traverse preheader's instructions in reverse order becaue if A depends
284 // on B (A appears after B), A needs to be sinked first before B can be
286 for (auto II = Preheader->rbegin(), E = Preheader->rend(); II != E;) {
287 Instruction *I = &*II++;
288 // No need to check for instruction's operands are loop invariant.
289 assert(L.hasLoopInvariantOperands(I) &&
290 "Insts in a loop's preheader should have loop invariant operands!");
291 if (!canSinkOrHoistInst(*I, &AA, &DT, &L, &CurAST, nullptr))
293 if (sinkInstruction(L, *I, ColdLoopBBs, LoopBlockNumber, LI, DT, BFI))
298 SE->forgetLoopDispositions(&L);
302 PreservedAnalyses LoopSinkPass::run(Function &F, FunctionAnalysisManager &FAM) {
303 LoopInfo &LI = FAM.getResult<LoopAnalysis>(F);
304 // Nothing to do if there are no loops.
306 return PreservedAnalyses::all();
308 AAResults &AA = FAM.getResult<AAManager>(F);
309 DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F);
310 BlockFrequencyInfo &BFI = FAM.getResult<BlockFrequencyAnalysis>(F);
312 // We want to do a postorder walk over the loops. Since loops are a tree this
313 // is equivalent to a reversed preorder walk and preorder is easy to compute
314 // without recursion. Since we reverse the preorder, we will visit siblings
315 // in reverse program order. This isn't expected to matter at all but is more
316 // consistent with sinking algorithms which generally work bottom-up.
317 SmallVector<Loop *, 4> PreorderLoops = LI.getLoopsInPreorder();
319 bool Changed = false;
321 Loop &L = *PreorderLoops.pop_back_val();
323 // Note that we don't pass SCEV here because it is only used to invalidate
324 // loops in SCEV and we don't preserve (or request) SCEV at all making that
326 Changed |= sinkLoopInvariantInstructions(L, AA, LI, DT, BFI,
327 /*ScalarEvolution*/ nullptr);
328 } while (!PreorderLoops.empty());
331 return PreservedAnalyses::all();
333 PreservedAnalyses PA;
334 PA.preserveSet<CFGAnalyses>();
339 struct LegacyLoopSinkPass : public LoopPass {
341 LegacyLoopSinkPass() : LoopPass(ID) {
342 initializeLegacyLoopSinkPassPass(*PassRegistry::getPassRegistry());
345 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
349 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
350 return sinkLoopInvariantInstructions(
351 *L, getAnalysis<AAResultsWrapperPass>().getAAResults(),
352 getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
353 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
354 getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(),
355 SE ? &SE->getSE() : nullptr);
358 void getAnalysisUsage(AnalysisUsage &AU) const override {
359 AU.setPreservesCFG();
360 AU.addRequired<BlockFrequencyInfoWrapperPass>();
361 getLoopAnalysisUsage(AU);
366 char LegacyLoopSinkPass::ID = 0;
367 INITIALIZE_PASS_BEGIN(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false,
369 INITIALIZE_PASS_DEPENDENCY(LoopPass)
370 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
371 INITIALIZE_PASS_END(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false, false)
373 Pass *llvm::createLoopSinkPass() { return new LegacyLoopSinkPass(); }