1 //===-- LoopSink.cpp - Loop Sink Pass -------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
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:
15 // * It only handles sinking of instructions from the loop's preheader to the
17 // * It uses alias set tracker to get more accurate alias info
18 // * It uses block frequency info to find the optimal sinking locations
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
31 //===----------------------------------------------------------------------===//
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/ScalarEvolution.h"
43 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/LLVMContext.h"
47 #include "llvm/IR/Metadata.h"
48 #include "llvm/InitializePasses.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 // Can't sink into blocks that have no valid insertion point.
156 for (BasicBlock *BB : BBsToSinkInto) {
157 if (BB->getFirstInsertionPt() == BB->end()) {
158 BBsToSinkInto.clear();
163 // If the total frequency of BBsToSinkInto is larger than preheader frequency,
165 if (adjustedSumFreq(BBsToSinkInto, BFI) >
166 BFI.getBlockFreq(L.getLoopPreheader()))
167 BBsToSinkInto.clear();
168 return BBsToSinkInto;
171 // Sinks \p I from the loop \p L's preheader to its uses. Returns true if
172 // sinking is successful.
173 // \p LoopBlockNumber is used to sort the insertion blocks to ensure
175 static bool sinkInstruction(Loop &L, Instruction &I,
176 const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
177 const SmallDenseMap<BasicBlock *, int, 16> &LoopBlockNumber,
178 LoopInfo &LI, DominatorTree &DT,
179 BlockFrequencyInfo &BFI) {
180 // Compute the set of blocks in loop L which contain a use of I.
181 SmallPtrSet<BasicBlock *, 2> BBs;
182 for (auto &U : I.uses()) {
183 Instruction *UI = cast<Instruction>(U.getUser());
184 // We cannot sink I to PHI-uses.
185 if (dyn_cast<PHINode>(UI))
187 // We cannot sink I if it has uses outside of the loop.
188 if (!L.contains(LI.getLoopFor(UI->getParent())))
190 BBs.insert(UI->getParent());
193 // findBBsToSinkInto is O(BBs.size() * ColdLoopBBs.size()). We cap the max
194 // BBs.size() to avoid expensive computation.
195 // FIXME: Handle code size growth for min_size and opt_size.
196 if (BBs.size() > MaxNumberOfUseBBsForSinking)
199 // Find the set of BBs that we should insert a copy of I.
200 SmallPtrSet<BasicBlock *, 2> BBsToSinkInto =
201 findBBsToSinkInto(L, BBs, ColdLoopBBs, DT, BFI);
202 if (BBsToSinkInto.empty())
205 // Return if any of the candidate blocks to sink into is non-cold.
206 if (BBsToSinkInto.size() > 1) {
207 for (auto *BB : BBsToSinkInto)
208 if (!LoopBlockNumber.count(BB))
212 // Copy the final BBs into a vector and sort them using the total ordering
213 // of the loop block numbers as iterating the set doesn't give a useful
214 // order. No need to stable sort as the block numbers are a total ordering.
215 SmallVector<BasicBlock *, 2> SortedBBsToSinkInto;
216 SortedBBsToSinkInto.insert(SortedBBsToSinkInto.begin(), BBsToSinkInto.begin(),
217 BBsToSinkInto.end());
218 llvm::sort(SortedBBsToSinkInto, [&](BasicBlock *A, BasicBlock *B) {
219 return LoopBlockNumber.find(A)->second < LoopBlockNumber.find(B)->second;
222 BasicBlock *MoveBB = *SortedBBsToSinkInto.begin();
223 // FIXME: Optimize the efficiency for cloned value replacement. The current
224 // implementation is O(SortedBBsToSinkInto.size() * I.num_uses()).
225 for (BasicBlock *N : makeArrayRef(SortedBBsToSinkInto).drop_front(1)) {
226 assert(LoopBlockNumber.find(N)->second >
227 LoopBlockNumber.find(MoveBB)->second &&
229 // Clone I and replace its uses.
230 Instruction *IC = I.clone();
231 IC->setName(I.getName());
232 IC->insertBefore(&*N->getFirstInsertionPt());
233 // Replaces uses of I with IC in N
234 I.replaceUsesWithIf(IC, [N](Use &U) {
235 return cast<Instruction>(U.getUser())->getParent() == N;
237 // Replaces uses of I with IC in blocks dominated by N
238 replaceDominatedUsesWith(&I, IC, DT, N);
239 LLVM_DEBUG(dbgs() << "Sinking a clone of " << I << " To: " << N->getName()
243 LLVM_DEBUG(dbgs() << "Sinking " << I << " To: " << MoveBB->getName() << '\n');
245 I.moveBefore(&*MoveBB->getFirstInsertionPt());
250 /// Sinks instructions from loop's preheader to the loop body if the
251 /// sum frequency of inserted copy is smaller than preheader's frequency.
252 static bool sinkLoopInvariantInstructions(Loop &L, AAResults &AA, LoopInfo &LI,
254 BlockFrequencyInfo &BFI,
255 ScalarEvolution *SE) {
256 BasicBlock *Preheader = L.getLoopPreheader();
260 // Enable LoopSink only when runtime profile is available.
261 // With static profile, the sinking decision may be sub-optimal.
262 if (!Preheader->getParent()->hasProfileData())
265 const BlockFrequency PreheaderFreq = BFI.getBlockFreq(Preheader);
266 // If there are no basic blocks with lower frequency than the preheader then
267 // we can avoid the detailed analysis as we will never find profitable sinking
269 if (all_of(L.blocks(), [&](const BasicBlock *BB) {
270 return BFI.getBlockFreq(BB) > PreheaderFreq;
274 bool Changed = false;
275 AliasSetTracker CurAST(AA);
277 // Compute alias set.
278 for (BasicBlock *BB : L.blocks())
280 CurAST.add(*Preheader);
282 // Sort loop's basic blocks by frequency
283 SmallVector<BasicBlock *, 10> ColdLoopBBs;
284 SmallDenseMap<BasicBlock *, int, 16> LoopBlockNumber;
286 for (BasicBlock *B : L.blocks())
287 if (BFI.getBlockFreq(B) < BFI.getBlockFreq(L.getLoopPreheader())) {
288 ColdLoopBBs.push_back(B);
289 LoopBlockNumber[B] = ++i;
291 llvm::stable_sort(ColdLoopBBs, [&](BasicBlock *A, BasicBlock *B) {
292 return BFI.getBlockFreq(A) < BFI.getBlockFreq(B);
295 // Traverse preheader's instructions in reverse order becaue if A depends
296 // on B (A appears after B), A needs to be sinked first before B can be
298 for (auto II = Preheader->rbegin(), E = Preheader->rend(); II != E;) {
299 Instruction *I = &*II++;
300 // No need to check for instruction's operands are loop invariant.
301 assert(L.hasLoopInvariantOperands(I) &&
302 "Insts in a loop's preheader should have loop invariant operands!");
303 if (!canSinkOrHoistInst(*I, &AA, &DT, &L, &CurAST, nullptr, false))
305 if (sinkInstruction(L, *I, ColdLoopBBs, LoopBlockNumber, LI, DT, BFI))
310 SE->forgetLoopDispositions(&L);
314 PreservedAnalyses LoopSinkPass::run(Function &F, FunctionAnalysisManager &FAM) {
315 LoopInfo &LI = FAM.getResult<LoopAnalysis>(F);
316 // Nothing to do if there are no loops.
318 return PreservedAnalyses::all();
320 AAResults &AA = FAM.getResult<AAManager>(F);
321 DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F);
322 BlockFrequencyInfo &BFI = FAM.getResult<BlockFrequencyAnalysis>(F);
324 // We want to do a postorder walk over the loops. Since loops are a tree this
325 // is equivalent to a reversed preorder walk and preorder is easy to compute
326 // without recursion. Since we reverse the preorder, we will visit siblings
327 // in reverse program order. This isn't expected to matter at all but is more
328 // consistent with sinking algorithms which generally work bottom-up.
329 SmallVector<Loop *, 4> PreorderLoops = LI.getLoopsInPreorder();
331 bool Changed = false;
333 Loop &L = *PreorderLoops.pop_back_val();
335 // Note that we don't pass SCEV here because it is only used to invalidate
336 // loops in SCEV and we don't preserve (or request) SCEV at all making that
338 Changed |= sinkLoopInvariantInstructions(L, AA, LI, DT, BFI,
339 /*ScalarEvolution*/ nullptr);
340 } while (!PreorderLoops.empty());
343 return PreservedAnalyses::all();
345 PreservedAnalyses PA;
346 PA.preserveSet<CFGAnalyses>();
351 struct LegacyLoopSinkPass : public LoopPass {
353 LegacyLoopSinkPass() : LoopPass(ID) {
354 initializeLegacyLoopSinkPassPass(*PassRegistry::getPassRegistry());
357 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
361 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
362 return sinkLoopInvariantInstructions(
363 *L, getAnalysis<AAResultsWrapperPass>().getAAResults(),
364 getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
365 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
366 getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(),
367 SE ? &SE->getSE() : nullptr);
370 void getAnalysisUsage(AnalysisUsage &AU) const override {
371 AU.setPreservesCFG();
372 AU.addRequired<BlockFrequencyInfoWrapperPass>();
373 getLoopAnalysisUsage(AU);
378 char LegacyLoopSinkPass::ID = 0;
379 INITIALIZE_PASS_BEGIN(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false,
381 INITIALIZE_PASS_DEPENDENCY(LoopPass)
382 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
383 INITIALIZE_PASS_END(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false, false)
385 Pass *llvm::createLoopSinkPass() { return new LegacyLoopSinkPass(); }