xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopSink.cpp (revision a0ca4af9455b844c5e094fc1b09b1390ffa979fc)
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/BlockFrequencyInfo.h"
38 #include "llvm/Analysis/LoopInfo.h"
39 #include "llvm/Analysis/MemorySSA.h"
40 #include "llvm/Analysis/MemorySSAUpdater.h"
41 #include "llvm/Analysis/ScalarEvolution.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/Support/BranchProbability.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Transforms/Scalar.h"
47 #include "llvm/Transforms/Utils/Local.h"
48 #include "llvm/Transforms/Utils/LoopUtils.h"
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "loopsink"
52 
53 STATISTIC(NumLoopSunk, "Number of instructions sunk into loop");
54 STATISTIC(NumLoopSunkCloned, "Number of cloned instructions sunk into loop");
55 
56 static cl::opt<unsigned> SinkFrequencyPercentThreshold(
57     "sink-freq-percent-threshold", cl::Hidden, cl::init(90),
58     cl::desc("Do not sink instructions that require cloning unless they "
59              "execute less than this percent of the time."));
60 
61 static cl::opt<unsigned> MaxNumberOfUseBBsForSinking(
62     "max-uses-for-sinking", cl::Hidden, cl::init(30),
63     cl::desc("Do not sink instructions that have too many uses."));
64 
65 /// Return adjusted total frequency of \p BBs.
66 ///
67 /// * If there is only one BB, sinking instruction will not introduce code
68 ///   size increase. Thus there is no need to adjust the frequency.
69 /// * If there are more than one BB, sinking would lead to code size increase.
70 ///   In this case, we add some "tax" to the total frequency to make it harder
71 ///   to sink. E.g.
72 ///     Freq(Preheader) = 100
73 ///     Freq(BBs) = sum(50, 49) = 99
74 ///   Even if Freq(BBs) < Freq(Preheader), we will not sink from Preheade to
75 ///   BBs as the difference is too small to justify the code size increase.
76 ///   To model this, The adjusted Freq(BBs) will be:
77 ///     AdjustedFreq(BBs) = 99 / SinkFrequencyPercentThreshold%
78 static BlockFrequency adjustedSumFreq(SmallPtrSetImpl<BasicBlock *> &BBs,
79                                       BlockFrequencyInfo &BFI) {
80   BlockFrequency T(0);
81   for (BasicBlock *B : BBs)
82     T += BFI.getBlockFreq(B);
83   if (BBs.size() > 1)
84     T /= BranchProbability(SinkFrequencyPercentThreshold, 100);
85   return T;
86 }
87 
88 /// Return a set of basic blocks to insert sinked instructions.
89 ///
90 /// The returned set of basic blocks (BBsToSinkInto) should satisfy:
91 ///
92 /// * Inside the loop \p L
93 /// * For each UseBB in \p UseBBs, there is at least one BB in BBsToSinkInto
94 ///   that domintates the UseBB
95 /// * Has minimum total frequency that is no greater than preheader frequency
96 ///
97 /// The purpose of the function is to find the optimal sinking points to
98 /// minimize execution cost, which is defined as "sum of frequency of
99 /// BBsToSinkInto".
100 /// As a result, the returned BBsToSinkInto needs to have minimum total
101 /// frequency.
102 /// Additionally, if the total frequency of BBsToSinkInto exceeds preheader
103 /// frequency, the optimal solution is not sinking (return empty set).
104 ///
105 /// \p ColdLoopBBs is used to help find the optimal sinking locations.
106 /// It stores a list of BBs that is:
107 ///
108 /// * Inside the loop \p L
109 /// * Has a frequency no larger than the loop's preheader
110 /// * Sorted by BB frequency
111 ///
112 /// The complexity of the function is O(UseBBs.size() * ColdLoopBBs.size()).
113 /// To avoid expensive computation, we cap the maximum UseBBs.size() in its
114 /// caller.
115 static SmallPtrSet<BasicBlock *, 2>
116 findBBsToSinkInto(const Loop &L, const SmallPtrSetImpl<BasicBlock *> &UseBBs,
117                   const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
118                   DominatorTree &DT, BlockFrequencyInfo &BFI) {
119   SmallPtrSet<BasicBlock *, 2> BBsToSinkInto;
120   if (UseBBs.size() == 0)
121     return BBsToSinkInto;
122 
123   BBsToSinkInto.insert(UseBBs.begin(), UseBBs.end());
124   SmallPtrSet<BasicBlock *, 2> BBsDominatedByColdestBB;
125 
126   // For every iteration:
127   //   * Pick the ColdestBB from ColdLoopBBs
128   //   * Find the set BBsDominatedByColdestBB that satisfy:
129   //     - BBsDominatedByColdestBB is a subset of BBsToSinkInto
130   //     - Every BB in BBsDominatedByColdestBB is dominated by ColdestBB
131   //   * If Freq(ColdestBB) < Freq(BBsDominatedByColdestBB), remove
132   //     BBsDominatedByColdestBB from BBsToSinkInto, add ColdestBB to
133   //     BBsToSinkInto
134   for (BasicBlock *ColdestBB : ColdLoopBBs) {
135     BBsDominatedByColdestBB.clear();
136     for (BasicBlock *SinkedBB : BBsToSinkInto)
137       if (DT.dominates(ColdestBB, SinkedBB))
138         BBsDominatedByColdestBB.insert(SinkedBB);
139     if (BBsDominatedByColdestBB.size() == 0)
140       continue;
141     if (adjustedSumFreq(BBsDominatedByColdestBB, BFI) >
142         BFI.getBlockFreq(ColdestBB)) {
143       for (BasicBlock *DominatedBB : BBsDominatedByColdestBB) {
144         BBsToSinkInto.erase(DominatedBB);
145       }
146       BBsToSinkInto.insert(ColdestBB);
147     }
148   }
149 
150   // Can't sink into blocks that have no valid insertion point.
151   for (BasicBlock *BB : BBsToSinkInto) {
152     if (BB->getFirstInsertionPt() == BB->end()) {
153       BBsToSinkInto.clear();
154       break;
155     }
156   }
157 
158   // If the total frequency of BBsToSinkInto is larger than preheader frequency,
159   // do not sink.
160   if (adjustedSumFreq(BBsToSinkInto, BFI) >
161       BFI.getBlockFreq(L.getLoopPreheader()))
162     BBsToSinkInto.clear();
163   return BBsToSinkInto;
164 }
165 
166 // Sinks \p I from the loop \p L's preheader to its uses. Returns true if
167 // sinking is successful.
168 // \p LoopBlockNumber is used to sort the insertion blocks to ensure
169 // determinism.
170 static bool sinkInstruction(
171     Loop &L, Instruction &I, const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
172     const SmallDenseMap<BasicBlock *, int, 16> &LoopBlockNumber, LoopInfo &LI,
173     DominatorTree &DT, BlockFrequencyInfo &BFI, MemorySSAUpdater *MSSAU) {
174   // Compute the set of blocks in loop L which contain a use of I.
175   SmallPtrSet<BasicBlock *, 2> BBs;
176   for (auto &U : I.uses()) {
177     Instruction *UI = cast<Instruction>(U.getUser());
178 
179     // We cannot sink I if it has uses outside of the loop.
180     if (!L.contains(LI.getLoopFor(UI->getParent())))
181       return false;
182 
183     if (!isa<PHINode>(UI)) {
184       BBs.insert(UI->getParent());
185       continue;
186     }
187 
188     // We cannot sink I to PHI-uses, try to look through PHI to find the incoming
189     // block of the value being used.
190     PHINode *PN = dyn_cast<PHINode>(UI);
191     BasicBlock *PhiBB = PN->getIncomingBlock(U);
192 
193     // If value's incoming block is from loop preheader directly, there's no
194     // place to sink to, bailout.
195     if (L.getLoopPreheader() == PhiBB)
196       return false;
197 
198     BBs.insert(PhiBB);
199   }
200 
201   // findBBsToSinkInto is O(BBs.size() * ColdLoopBBs.size()). We cap the max
202   // BBs.size() to avoid expensive computation.
203   // FIXME: Handle code size growth for min_size and opt_size.
204   if (BBs.size() > MaxNumberOfUseBBsForSinking)
205     return false;
206 
207   // Find the set of BBs that we should insert a copy of I.
208   SmallPtrSet<BasicBlock *, 2> BBsToSinkInto =
209       findBBsToSinkInto(L, BBs, ColdLoopBBs, DT, BFI);
210   if (BBsToSinkInto.empty())
211     return false;
212 
213   // Return if any of the candidate blocks to sink into is non-cold.
214   if (BBsToSinkInto.size() > 1 &&
215       !llvm::set_is_subset(BBsToSinkInto, LoopBlockNumber))
216     return false;
217 
218   // Copy the final BBs into a vector and sort them using the total ordering
219   // of the loop block numbers as iterating the set doesn't give a useful
220   // order. No need to stable sort as the block numbers are a total ordering.
221   SmallVector<BasicBlock *, 2> SortedBBsToSinkInto;
222   llvm::append_range(SortedBBsToSinkInto, BBsToSinkInto);
223   if (SortedBBsToSinkInto.size() > 1) {
224     llvm::sort(SortedBBsToSinkInto, [&](BasicBlock *A, BasicBlock *B) {
225       return LoopBlockNumber.find(A)->second < LoopBlockNumber.find(B)->second;
226     });
227   }
228 
229   BasicBlock *MoveBB = *SortedBBsToSinkInto.begin();
230   // FIXME: Optimize the efficiency for cloned value replacement. The current
231   //        implementation is O(SortedBBsToSinkInto.size() * I.num_uses()).
232   for (BasicBlock *N : ArrayRef(SortedBBsToSinkInto).drop_front(1)) {
233     assert(LoopBlockNumber.find(N)->second >
234                LoopBlockNumber.find(MoveBB)->second &&
235            "BBs not sorted!");
236     // Clone I and replace its uses.
237     Instruction *IC = I.clone();
238     IC->setName(I.getName());
239     IC->insertBefore(&*N->getFirstInsertionPt());
240 
241     if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
242       // Create a new MemoryAccess and let MemorySSA set its defining access.
243       MemoryAccess *NewMemAcc =
244           MSSAU->createMemoryAccessInBB(IC, nullptr, N, MemorySSA::Beginning);
245       if (NewMemAcc) {
246         if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
247           MSSAU->insertDef(MemDef, /*RenameUses=*/true);
248         else {
249           auto *MemUse = cast<MemoryUse>(NewMemAcc);
250           MSSAU->insertUse(MemUse, /*RenameUses=*/true);
251         }
252       }
253     }
254 
255     // Replaces uses of I with IC in N, except PHI-use which is being taken
256     // care of by defs in PHI's incoming blocks.
257     I.replaceUsesWithIf(IC, [N](Use &U) {
258       Instruction *UIToReplace = cast<Instruction>(U.getUser());
259       return UIToReplace->getParent() == N && !isa<PHINode>(UIToReplace);
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                                           MemorySSA &MSSA,
285                                           ScalarEvolution *SE) {
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   MemorySSAUpdater MSSAU(&MSSA);
302   SinkAndHoistLICMFlags LICMFlags(/*IsSink=*/true, L, MSSA);
303 
304   bool Changed = false;
305 
306   // Sort loop's basic blocks by frequency
307   SmallVector<BasicBlock *, 10> ColdLoopBBs;
308   SmallDenseMap<BasicBlock *, int, 16> LoopBlockNumber;
309   int i = 0;
310   for (BasicBlock *B : L.blocks())
311     if (BFI.getBlockFreq(B) < BFI.getBlockFreq(L.getLoopPreheader())) {
312       ColdLoopBBs.push_back(B);
313       LoopBlockNumber[B] = ++i;
314     }
315   llvm::stable_sort(ColdLoopBBs, [&](BasicBlock *A, BasicBlock *B) {
316     return BFI.getBlockFreq(A) < BFI.getBlockFreq(B);
317   });
318 
319   // Traverse preheader's instructions in reverse order because if A depends
320   // on B (A appears after B), A needs to be sunk first before B can be
321   // sinked.
322   for (Instruction &I : llvm::make_early_inc_range(llvm::reverse(*Preheader))) {
323     if (isa<PHINode>(&I))
324       continue;
325     // No need to check for instruction's operands are loop invariant.
326     assert(L.hasLoopInvariantOperands(&I) &&
327            "Insts in a loop's preheader should have loop invariant operands!");
328     if (!canSinkOrHoistInst(I, &AA, &DT, &L, MSSAU, false, LICMFlags))
329       continue;
330     if (sinkInstruction(L, I, ColdLoopBBs, LoopBlockNumber, LI, DT, BFI,
331                         &MSSAU)) {
332       Changed = true;
333       if (SE)
334         SE->forgetBlockAndLoopDispositions(&I);
335     }
336   }
337 
338   return Changed;
339 }
340 
341 PreservedAnalyses LoopSinkPass::run(Function &F, FunctionAnalysisManager &FAM) {
342   // Enable LoopSink only when runtime profile is available.
343   // With static profile, the sinking decision may be sub-optimal.
344   if (!F.hasProfileData())
345     return PreservedAnalyses::all();
346 
347   LoopInfo &LI = FAM.getResult<LoopAnalysis>(F);
348   // Nothing to do if there are no loops.
349   if (LI.empty())
350     return PreservedAnalyses::all();
351 
352   AAResults &AA = FAM.getResult<AAManager>(F);
353   DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F);
354   BlockFrequencyInfo &BFI = FAM.getResult<BlockFrequencyAnalysis>(F);
355   MemorySSA &MSSA = FAM.getResult<MemorySSAAnalysis>(F).getMSSA();
356 
357   // We want to do a postorder walk over the loops. Since loops are a tree this
358   // is equivalent to a reversed preorder walk and preorder is easy to compute
359   // without recursion. Since we reverse the preorder, we will visit siblings
360   // in reverse program order. This isn't expected to matter at all but is more
361   // consistent with sinking algorithms which generally work bottom-up.
362   SmallVector<Loop *, 4> PreorderLoops = LI.getLoopsInPreorder();
363 
364   bool Changed = false;
365   do {
366     Loop &L = *PreorderLoops.pop_back_val();
367 
368     BasicBlock *Preheader = L.getLoopPreheader();
369     if (!Preheader)
370       continue;
371 
372     // Note that we don't pass SCEV here because it is only used to invalidate
373     // loops in SCEV and we don't preserve (or request) SCEV at all making that
374     // unnecessary.
375     Changed |= sinkLoopInvariantInstructions(L, AA, LI, DT, BFI, MSSA,
376                                              /*ScalarEvolution*/ nullptr);
377   } while (!PreorderLoops.empty());
378 
379   if (!Changed)
380     return PreservedAnalyses::all();
381 
382   PreservedAnalyses PA;
383   PA.preserveSet<CFGAnalyses>();
384   PA.preserve<MemorySSAAnalysis>();
385 
386   if (VerifyMemorySSA)
387     MSSA.verifyMemorySSA();
388 
389   return PA;
390 }
391