1 //===- MachineLICM.cpp - Machine Loop Invariant Code Motion 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 performs loop invariant code motion on machine instructions. We
10 // attempt to remove as much code from the body of a loop as possible.
11 //
12 // This pass is not intended to be a replacement or a complete alternative
13 // for the LLVM-IR-level LICM pass. It is only designed to hoist simple
14 // constructs that are not exposed before lowering and instruction selection.
15 //
16 //===----------------------------------------------------------------------===//
17 
18 #include "llvm/ADT/BitVector.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
27 #include "llvm/CodeGen/MachineDominators.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineFunctionPass.h"
31 #include "llvm/CodeGen/MachineInstr.h"
32 #include "llvm/CodeGen/MachineLoopInfo.h"
33 #include "llvm/CodeGen/MachineMemOperand.h"
34 #include "llvm/CodeGen/MachineOperand.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/CodeGen/TargetInstrInfo.h"
38 #include "llvm/CodeGen/TargetLowering.h"
39 #include "llvm/CodeGen/TargetRegisterInfo.h"
40 #include "llvm/CodeGen/TargetSchedule.h"
41 #include "llvm/CodeGen/TargetSubtargetInfo.h"
42 #include "llvm/IR/DebugLoc.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/MC/MCInstrDesc.h"
45 #include "llvm/MC/MCRegister.h"
46 #include "llvm/MC/MCRegisterInfo.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/Casting.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include <algorithm>
53 #include <cassert>
54 #include <limits>
55 #include <vector>
56 
57 using namespace llvm;
58 
59 #define DEBUG_TYPE "machinelicm"
60 
61 static cl::opt<bool>
62 AvoidSpeculation("avoid-speculation",
63                  cl::desc("MachineLICM should avoid speculation"),
64                  cl::init(true), cl::Hidden);
65 
66 static cl::opt<bool>
67 HoistCheapInsts("hoist-cheap-insts",
68                 cl::desc("MachineLICM should hoist even cheap instructions"),
69                 cl::init(false), cl::Hidden);
70 
71 static cl::opt<bool>
72 HoistConstStores("hoist-const-stores",
73                  cl::desc("Hoist invariant stores"),
74                  cl::init(true), cl::Hidden);
75 
76 static cl::opt<bool> HoistConstLoads("hoist-const-loads",
77                                      cl::desc("Hoist invariant loads"),
78                                      cl::init(true), cl::Hidden);
79 
80 // The default threshold of 100 (i.e. if target block is 100 times hotter)
81 // is based on empirical data on a single target and is subject to tuning.
82 static cl::opt<unsigned>
83 BlockFrequencyRatioThreshold("block-freq-ratio-threshold",
84                              cl::desc("Do not hoist instructions if target"
85                              "block is N times hotter than the source."),
86                              cl::init(100), cl::Hidden);
87 
88 enum class UseBFI { None, PGO, All };
89 
90 static cl::opt<UseBFI>
91 DisableHoistingToHotterBlocks("disable-hoisting-to-hotter-blocks",
92                               cl::desc("Disable hoisting instructions to"
93                               " hotter blocks"),
94                               cl::init(UseBFI::PGO), cl::Hidden,
95                               cl::values(clEnumValN(UseBFI::None, "none",
96                               "disable the feature"),
97                               clEnumValN(UseBFI::PGO, "pgo",
98                               "enable the feature when using profile data"),
99                               clEnumValN(UseBFI::All, "all",
100                               "enable the feature with/wo profile data")));
101 
102 STATISTIC(NumHoisted,
103           "Number of machine instructions hoisted out of loops");
104 STATISTIC(NumLowRP,
105           "Number of instructions hoisted in low reg pressure situation");
106 STATISTIC(NumHighLatency,
107           "Number of high latency instructions hoisted");
108 STATISTIC(NumCSEed,
109           "Number of hoisted machine instructions CSEed");
110 STATISTIC(NumPostRAHoisted,
111           "Number of machine instructions hoisted out of loops post regalloc");
112 STATISTIC(NumStoreConst,
113           "Number of stores of const phys reg hoisted out of loops");
114 STATISTIC(NumNotHoistedDueToHotness,
115           "Number of instructions not hoisted due to block frequency");
116 
117 namespace {
118   enum HoistResult { NotHoisted = 1, Hoisted = 2, ErasedMI = 4 };
119 
120   class MachineLICMBase : public MachineFunctionPass {
121     const TargetInstrInfo *TII = nullptr;
122     const TargetLoweringBase *TLI = nullptr;
123     const TargetRegisterInfo *TRI = nullptr;
124     const MachineFrameInfo *MFI = nullptr;
125     MachineRegisterInfo *MRI = nullptr;
126     TargetSchedModel SchedModel;
127     bool PreRegAlloc = false;
128     bool HasProfileData = false;
129 
130     // Various analyses that we use...
131     AliasAnalysis *AA = nullptr;               // Alias analysis info.
132     MachineBlockFrequencyInfo *MBFI = nullptr; // Machine block frequncy info
133     MachineLoopInfo *MLI = nullptr;            // Current MachineLoopInfo
134     MachineDominatorTree *DT = nullptr; // Machine dominator tree for the cur loop
135 
136     // State that is updated as we process loops
137     bool Changed = false;           // True if a loop is changed.
138     bool FirstInLoop = false;       // True if it's the first LICM in the loop.
139 
140     // Holds information about whether it is allowed to move load instructions
141     // out of the loop
142     SmallDenseMap<MachineLoop *, bool> AllowedToHoistLoads;
143 
144     // Exit blocks of each Loop.
145     DenseMap<MachineLoop *, SmallVector<MachineBasicBlock *, 8>> ExitBlockMap;
146 
147     bool isExitBlock(MachineLoop *CurLoop, const MachineBasicBlock *MBB) {
148       if (ExitBlockMap.contains(CurLoop))
149         return is_contained(ExitBlockMap[CurLoop], MBB);
150 
151       SmallVector<MachineBasicBlock *, 8> ExitBlocks;
152       CurLoop->getExitBlocks(ExitBlocks);
153       ExitBlockMap[CurLoop] = ExitBlocks;
154       return is_contained(ExitBlocks, MBB);
155     }
156 
157     // Track 'estimated' register pressure.
158     SmallSet<Register, 32> RegSeen;
159     SmallVector<unsigned, 8> RegPressure;
160 
161     // Register pressure "limit" per register pressure set. If the pressure
162     // is higher than the limit, then it's considered high.
163     SmallVector<unsigned, 8> RegLimit;
164 
165     // Register pressure on path leading from loop preheader to current BB.
166     SmallVector<SmallVector<unsigned, 8>, 16> BackTrace;
167 
168     // For each opcode per preheader, keep a list of potential CSE instructions.
169     DenseMap<MachineBasicBlock *,
170              DenseMap<unsigned, std::vector<MachineInstr *>>>
171         CSEMap;
172 
173     enum {
174       SpeculateFalse   = 0,
175       SpeculateTrue    = 1,
176       SpeculateUnknown = 2
177     };
178 
179     // If a MBB does not dominate loop exiting blocks then it may not safe
180     // to hoist loads from this block.
181     // Tri-state: 0 - false, 1 - true, 2 - unknown
182     unsigned SpeculationState = SpeculateUnknown;
183 
184   public:
185     MachineLICMBase(char &PassID, bool PreRegAlloc)
186         : MachineFunctionPass(PassID), PreRegAlloc(PreRegAlloc) {}
187 
188     bool runOnMachineFunction(MachineFunction &MF) override;
189 
190     void getAnalysisUsage(AnalysisUsage &AU) const override {
191       AU.addRequired<MachineLoopInfo>();
192       if (DisableHoistingToHotterBlocks != UseBFI::None)
193         AU.addRequired<MachineBlockFrequencyInfo>();
194       AU.addRequired<MachineDominatorTree>();
195       AU.addRequired<AAResultsWrapperPass>();
196       AU.addPreserved<MachineLoopInfo>();
197       MachineFunctionPass::getAnalysisUsage(AU);
198     }
199 
200     void releaseMemory() override {
201       RegSeen.clear();
202       RegPressure.clear();
203       RegLimit.clear();
204       BackTrace.clear();
205       CSEMap.clear();
206       ExitBlockMap.clear();
207     }
208 
209   private:
210     /// Keep track of information about hoisting candidates.
211     struct CandidateInfo {
212       MachineInstr *MI;
213       unsigned      Def;
214       int           FI;
215 
216       CandidateInfo(MachineInstr *mi, unsigned def, int fi)
217         : MI(mi), Def(def), FI(fi) {}
218     };
219 
220     void HoistRegionPostRA(MachineLoop *CurLoop,
221                            MachineBasicBlock *CurPreheader);
222 
223     void HoistPostRA(MachineInstr *MI, unsigned Def, MachineLoop *CurLoop,
224                      MachineBasicBlock *CurPreheader);
225 
226     void ProcessMI(MachineInstr *MI, BitVector &PhysRegDefs,
227                    BitVector &PhysRegClobbers, SmallSet<int, 32> &StoredFIs,
228                    SmallVectorImpl<CandidateInfo> &Candidates,
229                    MachineLoop *CurLoop);
230 
231     void AddToLiveIns(MCRegister Reg, MachineLoop *CurLoop);
232 
233     bool IsLICMCandidate(MachineInstr &I, MachineLoop *CurLoop);
234 
235     bool IsLoopInvariantInst(MachineInstr &I, MachineLoop *CurLoop);
236 
237     bool HasLoopPHIUse(const MachineInstr *MI, MachineLoop *CurLoop);
238 
239     bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, Register Reg,
240                                MachineLoop *CurLoop) const;
241 
242     bool IsCheapInstruction(MachineInstr &MI) const;
243 
244     bool CanCauseHighRegPressure(const DenseMap<unsigned, int> &Cost,
245                                  bool Cheap);
246 
247     void UpdateBackTraceRegPressure(const MachineInstr *MI);
248 
249     bool IsProfitableToHoist(MachineInstr &MI, MachineLoop *CurLoop);
250 
251     bool IsGuaranteedToExecute(MachineBasicBlock *BB, MachineLoop *CurLoop);
252 
253     bool isTriviallyReMaterializable(const MachineInstr &MI) const;
254 
255     void EnterScope(MachineBasicBlock *MBB);
256 
257     void ExitScope(MachineBasicBlock *MBB);
258 
259     void ExitScopeIfDone(
260         MachineDomTreeNode *Node,
261         DenseMap<MachineDomTreeNode *, unsigned> &OpenChildren,
262         const DenseMap<MachineDomTreeNode *, MachineDomTreeNode *> &ParentMap);
263 
264     void HoistOutOfLoop(MachineDomTreeNode *HeaderN, MachineLoop *CurLoop,
265                         MachineBasicBlock *CurPreheader);
266 
267     void InitRegPressure(MachineBasicBlock *BB);
268 
269     DenseMap<unsigned, int> calcRegisterCost(const MachineInstr *MI,
270                                              bool ConsiderSeen,
271                                              bool ConsiderUnseenAsDef);
272 
273     void UpdateRegPressure(const MachineInstr *MI,
274                            bool ConsiderUnseenAsDef = false);
275 
276     MachineInstr *ExtractHoistableLoad(MachineInstr *MI, MachineLoop *CurLoop);
277 
278     MachineInstr *LookForDuplicate(const MachineInstr *MI,
279                                    std::vector<MachineInstr *> &PrevMIs);
280 
281     bool
282     EliminateCSE(MachineInstr *MI,
283                  DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI);
284 
285     bool MayCSE(MachineInstr *MI);
286 
287     unsigned Hoist(MachineInstr *MI, MachineBasicBlock *Preheader,
288                    MachineLoop *CurLoop);
289 
290     void InitCSEMap(MachineBasicBlock *BB);
291 
292     void InitializeLoadsHoistableLoops();
293 
294     bool isTgtHotterThanSrc(MachineBasicBlock *SrcBlock,
295                             MachineBasicBlock *TgtBlock);
296     MachineBasicBlock *getCurPreheader(MachineLoop *CurLoop,
297                                        MachineBasicBlock *CurPreheader);
298   };
299 
300   class MachineLICM : public MachineLICMBase {
301   public:
302     static char ID;
303     MachineLICM() : MachineLICMBase(ID, false) {
304       initializeMachineLICMPass(*PassRegistry::getPassRegistry());
305     }
306   };
307 
308   class EarlyMachineLICM : public MachineLICMBase {
309   public:
310     static char ID;
311     EarlyMachineLICM() : MachineLICMBase(ID, true) {
312       initializeEarlyMachineLICMPass(*PassRegistry::getPassRegistry());
313     }
314   };
315 
316 } // end anonymous namespace
317 
318 char MachineLICM::ID;
319 char EarlyMachineLICM::ID;
320 
321 char &llvm::MachineLICMID = MachineLICM::ID;
322 char &llvm::EarlyMachineLICMID = EarlyMachineLICM::ID;
323 
324 INITIALIZE_PASS_BEGIN(MachineLICM, DEBUG_TYPE,
325                       "Machine Loop Invariant Code Motion", false, false)
326 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
327 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
328 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
329 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
330 INITIALIZE_PASS_END(MachineLICM, DEBUG_TYPE,
331                     "Machine Loop Invariant Code Motion", false, false)
332 
333 INITIALIZE_PASS_BEGIN(EarlyMachineLICM, "early-machinelicm",
334                       "Early Machine Loop Invariant Code Motion", false, false)
335 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
336 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
337 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
338 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
339 INITIALIZE_PASS_END(EarlyMachineLICM, "early-machinelicm",
340                     "Early Machine Loop Invariant Code Motion", false, false)
341 
342 bool MachineLICMBase::runOnMachineFunction(MachineFunction &MF) {
343   if (skipFunction(MF.getFunction()))
344     return false;
345 
346   Changed = FirstInLoop = false;
347   const TargetSubtargetInfo &ST = MF.getSubtarget();
348   TII = ST.getInstrInfo();
349   TLI = ST.getTargetLowering();
350   TRI = ST.getRegisterInfo();
351   MFI = &MF.getFrameInfo();
352   MRI = &MF.getRegInfo();
353   SchedModel.init(&ST);
354 
355   PreRegAlloc = MRI->isSSA();
356   HasProfileData = MF.getFunction().hasProfileData();
357 
358   if (PreRegAlloc)
359     LLVM_DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
360   else
361     LLVM_DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
362   LLVM_DEBUG(dbgs() << MF.getName() << " ********\n");
363 
364   if (PreRegAlloc) {
365     // Estimate register pressure during pre-regalloc pass.
366     unsigned NumRPS = TRI->getNumRegPressureSets();
367     RegPressure.resize(NumRPS);
368     std::fill(RegPressure.begin(), RegPressure.end(), 0);
369     RegLimit.resize(NumRPS);
370     for (unsigned i = 0, e = NumRPS; i != e; ++i)
371       RegLimit[i] = TRI->getRegPressureSetLimit(MF, i);
372   }
373 
374   // Get our Loop information...
375   if (DisableHoistingToHotterBlocks != UseBFI::None)
376     MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
377   MLI = &getAnalysis<MachineLoopInfo>();
378   DT  = &getAnalysis<MachineDominatorTree>();
379   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
380 
381   if (HoistConstLoads)
382     InitializeLoadsHoistableLoops();
383 
384   SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
385   while (!Worklist.empty()) {
386     MachineLoop *CurLoop = Worklist.pop_back_val();
387     MachineBasicBlock *CurPreheader = nullptr;
388 
389     if (!PreRegAlloc)
390       HoistRegionPostRA(CurLoop, CurPreheader);
391     else {
392       // CSEMap is initialized for loop header when the first instruction is
393       // being hoisted.
394       MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
395       FirstInLoop = true;
396       HoistOutOfLoop(N, CurLoop, CurPreheader);
397       CSEMap.clear();
398     }
399   }
400 
401   return Changed;
402 }
403 
404 /// Return true if instruction stores to the specified frame.
405 static bool InstructionStoresToFI(const MachineInstr *MI, int FI) {
406   // Check mayStore before memory operands so that e.g. DBG_VALUEs will return
407   // true since they have no memory operands.
408   if (!MI->mayStore())
409      return false;
410   // If we lost memory operands, conservatively assume that the instruction
411   // writes to all slots.
412   if (MI->memoperands_empty())
413     return true;
414   for (const MachineMemOperand *MemOp : MI->memoperands()) {
415     if (!MemOp->isStore() || !MemOp->getPseudoValue())
416       continue;
417     if (const FixedStackPseudoSourceValue *Value =
418         dyn_cast<FixedStackPseudoSourceValue>(MemOp->getPseudoValue())) {
419       if (Value->getFrameIndex() == FI)
420         return true;
421     }
422   }
423   return false;
424 }
425 
426 /// Examine the instruction for potentai LICM candidate. Also
427 /// gather register def and frame object update information.
428 void MachineLICMBase::ProcessMI(MachineInstr *MI, BitVector &PhysRegDefs,
429                                 BitVector &PhysRegClobbers,
430                                 SmallSet<int, 32> &StoredFIs,
431                                 SmallVectorImpl<CandidateInfo> &Candidates,
432                                 MachineLoop *CurLoop) {
433   bool RuledOut = false;
434   bool HasNonInvariantUse = false;
435   unsigned Def = 0;
436   for (const MachineOperand &MO : MI->operands()) {
437     if (MO.isFI()) {
438       // Remember if the instruction stores to the frame index.
439       int FI = MO.getIndex();
440       if (!StoredFIs.count(FI) &&
441           MFI->isSpillSlotObjectIndex(FI) &&
442           InstructionStoresToFI(MI, FI))
443         StoredFIs.insert(FI);
444       HasNonInvariantUse = true;
445       continue;
446     }
447 
448     // We can't hoist an instruction defining a physreg that is clobbered in
449     // the loop.
450     if (MO.isRegMask()) {
451       PhysRegClobbers.setBitsNotInMask(MO.getRegMask());
452       continue;
453     }
454 
455     if (!MO.isReg())
456       continue;
457     Register Reg = MO.getReg();
458     if (!Reg)
459       continue;
460     assert(Reg.isPhysical() && "Not expecting virtual register!");
461 
462     if (!MO.isDef()) {
463       if (Reg && (PhysRegDefs.test(Reg) || PhysRegClobbers.test(Reg)))
464         // If it's using a non-loop-invariant register, then it's obviously not
465         // safe to hoist.
466         HasNonInvariantUse = true;
467       continue;
468     }
469 
470     if (MO.isImplicit()) {
471       for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
472         PhysRegClobbers.set(*AI);
473       if (!MO.isDead())
474         // Non-dead implicit def? This cannot be hoisted.
475         RuledOut = true;
476       // No need to check if a dead implicit def is also defined by
477       // another instruction.
478       continue;
479     }
480 
481     // FIXME: For now, avoid instructions with multiple defs, unless
482     // it's a dead implicit def.
483     if (Def)
484       RuledOut = true;
485     else
486       Def = Reg;
487 
488     // If we have already seen another instruction that defines the same
489     // register, then this is not safe.  Two defs is indicated by setting a
490     // PhysRegClobbers bit.
491     for (MCRegAliasIterator AS(Reg, TRI, true); AS.isValid(); ++AS) {
492       if (PhysRegDefs.test(*AS))
493         PhysRegClobbers.set(*AS);
494     }
495     // Need a second loop because MCRegAliasIterator can visit the same
496     // register twice.
497     for (MCRegAliasIterator AS(Reg, TRI, true); AS.isValid(); ++AS)
498       PhysRegDefs.set(*AS);
499 
500     if (PhysRegClobbers.test(Reg))
501       // MI defined register is seen defined by another instruction in
502       // the loop, it cannot be a LICM candidate.
503       RuledOut = true;
504   }
505 
506   // Only consider reloads for now and remats which do not have register
507   // operands. FIXME: Consider unfold load folding instructions.
508   if (Def && !RuledOut) {
509     int FI = std::numeric_limits<int>::min();
510     if ((!HasNonInvariantUse && IsLICMCandidate(*MI, CurLoop)) ||
511         (TII->isLoadFromStackSlot(*MI, FI) && MFI->isSpillSlotObjectIndex(FI)))
512       Candidates.push_back(CandidateInfo(MI, Def, FI));
513   }
514 }
515 
516 /// Walk the specified region of the CFG and hoist loop invariants out to the
517 /// preheader.
518 void MachineLICMBase::HoistRegionPostRA(MachineLoop *CurLoop,
519                                         MachineBasicBlock *CurPreheader) {
520   MachineBasicBlock *Preheader = getCurPreheader(CurLoop, CurPreheader);
521   if (!Preheader)
522     return;
523 
524   unsigned NumRegs = TRI->getNumRegs();
525   BitVector PhysRegDefs(NumRegs); // Regs defined once in the loop.
526   BitVector PhysRegClobbers(NumRegs); // Regs defined more than once.
527 
528   SmallVector<CandidateInfo, 32> Candidates;
529   SmallSet<int, 32> StoredFIs;
530 
531   // Walk the entire region, count number of defs for each register, and
532   // collect potential LICM candidates.
533   for (MachineBasicBlock *BB : CurLoop->getBlocks()) {
534     // If the header of the loop containing this basic block is a landing pad,
535     // then don't try to hoist instructions out of this loop.
536     const MachineLoop *ML = MLI->getLoopFor(BB);
537     if (ML && ML->getHeader()->isEHPad()) continue;
538 
539     // Conservatively treat live-in's as an external def.
540     // FIXME: That means a reload that're reused in successor block(s) will not
541     // be LICM'ed.
542     for (const auto &LI : BB->liveins()) {
543       for (MCRegAliasIterator AI(LI.PhysReg, TRI, true); AI.isValid(); ++AI)
544         PhysRegDefs.set(*AI);
545     }
546 
547     // Funclet entry blocks will clobber all registers
548     if (const uint32_t *Mask = BB->getBeginClobberMask(TRI))
549       PhysRegClobbers.setBitsNotInMask(Mask);
550 
551     SpeculationState = SpeculateUnknown;
552     for (MachineInstr &MI : *BB)
553       ProcessMI(&MI, PhysRegDefs, PhysRegClobbers, StoredFIs, Candidates,
554                 CurLoop);
555   }
556 
557   // Gather the registers read / clobbered by the terminator.
558   BitVector TermRegs(NumRegs);
559   MachineBasicBlock::iterator TI = Preheader->getFirstTerminator();
560   if (TI != Preheader->end()) {
561     for (const MachineOperand &MO : TI->operands()) {
562       if (!MO.isReg())
563         continue;
564       Register Reg = MO.getReg();
565       if (!Reg)
566         continue;
567       for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
568         TermRegs.set(*AI);
569     }
570   }
571 
572   // Now evaluate whether the potential candidates qualify.
573   // 1. Check if the candidate defined register is defined by another
574   //    instruction in the loop.
575   // 2. If the candidate is a load from stack slot (always true for now),
576   //    check if the slot is stored anywhere in the loop.
577   // 3. Make sure candidate def should not clobber
578   //    registers read by the terminator. Similarly its def should not be
579   //    clobbered by the terminator.
580   for (CandidateInfo &Candidate : Candidates) {
581     if (Candidate.FI != std::numeric_limits<int>::min() &&
582         StoredFIs.count(Candidate.FI))
583       continue;
584 
585     unsigned Def = Candidate.Def;
586     if (!PhysRegClobbers.test(Def) && !TermRegs.test(Def)) {
587       bool Safe = true;
588       MachineInstr *MI = Candidate.MI;
589       for (const MachineOperand &MO : MI->all_uses()) {
590         if (!MO.getReg())
591           continue;
592         Register Reg = MO.getReg();
593         if (PhysRegDefs.test(Reg) ||
594             PhysRegClobbers.test(Reg)) {
595           // If it's using a non-loop-invariant register, then it's obviously
596           // not safe to hoist.
597           Safe = false;
598           break;
599         }
600       }
601       if (Safe)
602         HoistPostRA(MI, Candidate.Def, CurLoop, CurPreheader);
603     }
604   }
605 }
606 
607 /// Add register 'Reg' to the livein sets of BBs in the current loop, and make
608 /// sure it is not killed by any instructions in the loop.
609 void MachineLICMBase::AddToLiveIns(MCRegister Reg, MachineLoop *CurLoop) {
610   for (MachineBasicBlock *BB : CurLoop->getBlocks()) {
611     if (!BB->isLiveIn(Reg))
612       BB->addLiveIn(Reg);
613     for (MachineInstr &MI : *BB) {
614       for (MachineOperand &MO : MI.all_uses()) {
615         if (!MO.getReg())
616           continue;
617         if (TRI->regsOverlap(Reg, MO.getReg()))
618           MO.setIsKill(false);
619       }
620     }
621   }
622 }
623 
624 /// When an instruction is found to only use loop invariant operands that is
625 /// safe to hoist, this instruction is called to do the dirty work.
626 void MachineLICMBase::HoistPostRA(MachineInstr *MI, unsigned Def,
627                                   MachineLoop *CurLoop,
628                                   MachineBasicBlock *CurPreheader) {
629   MachineBasicBlock *Preheader = getCurPreheader(CurLoop, CurPreheader);
630 
631   // Now move the instructions to the predecessor, inserting it before any
632   // terminator instructions.
633   LLVM_DEBUG(dbgs() << "Hoisting to " << printMBBReference(*Preheader)
634                     << " from " << printMBBReference(*MI->getParent()) << ": "
635                     << *MI);
636 
637   // Splice the instruction to the preheader.
638   MachineBasicBlock *MBB = MI->getParent();
639   Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
640 
641   // Since we are moving the instruction out of its basic block, we do not
642   // retain its debug location. Doing so would degrade the debugging
643   // experience and adversely affect the accuracy of profiling information.
644   assert(!MI->isDebugInstr() && "Should not hoist debug inst");
645   MI->setDebugLoc(DebugLoc());
646 
647   // Add register to livein list to all the BBs in the current loop since a
648   // loop invariant must be kept live throughout the whole loop. This is
649   // important to ensure later passes do not scavenge the def register.
650   AddToLiveIns(Def, CurLoop);
651 
652   ++NumPostRAHoisted;
653   Changed = true;
654 }
655 
656 /// Check if this mbb is guaranteed to execute. If not then a load from this mbb
657 /// may not be safe to hoist.
658 bool MachineLICMBase::IsGuaranteedToExecute(MachineBasicBlock *BB,
659                                             MachineLoop *CurLoop) {
660   if (SpeculationState != SpeculateUnknown)
661     return SpeculationState == SpeculateFalse;
662 
663   if (BB != CurLoop->getHeader()) {
664     // Check loop exiting blocks.
665     SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks;
666     CurLoop->getExitingBlocks(CurrentLoopExitingBlocks);
667     for (MachineBasicBlock *CurrentLoopExitingBlock : CurrentLoopExitingBlocks)
668       if (!DT->dominates(BB, CurrentLoopExitingBlock)) {
669         SpeculationState = SpeculateTrue;
670         return false;
671       }
672   }
673 
674   SpeculationState = SpeculateFalse;
675   return true;
676 }
677 
678 /// Check if \p MI is trivially remateralizable and if it does not have any
679 /// virtual register uses. Even though rematerializable RA might not actually
680 /// rematerialize it in this scenario. In that case we do not want to hoist such
681 /// instruction out of the loop in a belief RA will sink it back if needed.
682 bool MachineLICMBase::isTriviallyReMaterializable(
683     const MachineInstr &MI) const {
684   if (!TII->isTriviallyReMaterializable(MI))
685     return false;
686 
687   for (const MachineOperand &MO : MI.all_uses()) {
688     if (MO.getReg().isVirtual())
689       return false;
690   }
691 
692   return true;
693 }
694 
695 void MachineLICMBase::EnterScope(MachineBasicBlock *MBB) {
696   LLVM_DEBUG(dbgs() << "Entering " << printMBBReference(*MBB) << '\n');
697 
698   // Remember livein register pressure.
699   BackTrace.push_back(RegPressure);
700 }
701 
702 void MachineLICMBase::ExitScope(MachineBasicBlock *MBB) {
703   LLVM_DEBUG(dbgs() << "Exiting " << printMBBReference(*MBB) << '\n');
704   BackTrace.pop_back();
705 }
706 
707 /// Destroy scope for the MBB that corresponds to the given dominator tree node
708 /// if its a leaf or all of its children are done. Walk up the dominator tree to
709 /// destroy ancestors which are now done.
710 void MachineLICMBase::ExitScopeIfDone(MachineDomTreeNode *Node,
711     DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
712     const DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
713   if (OpenChildren[Node])
714     return;
715 
716   for(;;) {
717     ExitScope(Node->getBlock());
718     // Now traverse upwards to pop ancestors whose offsprings are all done.
719     MachineDomTreeNode *Parent = ParentMap.lookup(Node);
720     if (!Parent || --OpenChildren[Parent] != 0)
721       break;
722     Node = Parent;
723   }
724 }
725 
726 /// Walk the specified loop in the CFG (defined by all blocks dominated by the
727 /// specified header block, and that are in the current loop) in depth first
728 /// order w.r.t the DominatorTree. This allows us to visit definitions before
729 /// uses, allowing us to hoist a loop body in one pass without iteration.
730 void MachineLICMBase::HoistOutOfLoop(MachineDomTreeNode *HeaderN,
731                                      MachineLoop *CurLoop,
732                                      MachineBasicBlock *CurPreheader) {
733   MachineBasicBlock *Preheader = getCurPreheader(CurLoop, CurPreheader);
734   if (!Preheader)
735     return;
736 
737   SmallVector<MachineDomTreeNode*, 32> Scopes;
738   SmallVector<MachineDomTreeNode*, 8> WorkList;
739   DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
740   DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
741 
742   // Perform a DFS walk to determine the order of visit.
743   WorkList.push_back(HeaderN);
744   while (!WorkList.empty()) {
745     MachineDomTreeNode *Node = WorkList.pop_back_val();
746     assert(Node && "Null dominator tree node?");
747     MachineBasicBlock *BB = Node->getBlock();
748 
749     // If the header of the loop containing this basic block is a landing pad,
750     // then don't try to hoist instructions out of this loop.
751     const MachineLoop *ML = MLI->getLoopFor(BB);
752     if (ML && ML->getHeader()->isEHPad())
753       continue;
754 
755     // If this subregion is not in the top level loop at all, exit.
756     if (!CurLoop->contains(BB))
757       continue;
758 
759     Scopes.push_back(Node);
760     unsigned NumChildren = Node->getNumChildren();
761 
762     // Don't hoist things out of a large switch statement.  This often causes
763     // code to be hoisted that wasn't going to be executed, and increases
764     // register pressure in a situation where it's likely to matter.
765     if (BB->succ_size() >= 25)
766       NumChildren = 0;
767 
768     OpenChildren[Node] = NumChildren;
769     if (NumChildren) {
770       // Add children in reverse order as then the next popped worklist node is
771       // the first child of this node.  This means we ultimately traverse the
772       // DOM tree in exactly the same order as if we'd recursed.
773       for (MachineDomTreeNode *Child : reverse(Node->children())) {
774         ParentMap[Child] = Node;
775         WorkList.push_back(Child);
776       }
777     }
778   }
779 
780   if (Scopes.size() == 0)
781     return;
782 
783   // Compute registers which are livein into the loop headers.
784   RegSeen.clear();
785   BackTrace.clear();
786   InitRegPressure(Preheader);
787 
788   // Now perform LICM.
789   for (MachineDomTreeNode *Node : Scopes) {
790     MachineBasicBlock *MBB = Node->getBlock();
791 
792     EnterScope(MBB);
793 
794     // Process the block
795     SpeculationState = SpeculateUnknown;
796     for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) {
797       unsigned HoistRes = HoistResult::NotHoisted;
798       HoistRes = Hoist(&MI, Preheader, CurLoop);
799       if (HoistRes & HoistResult::NotHoisted) {
800         // We have failed to hoist MI to outermost loop's preheader. If MI is in
801         // a subloop, try to hoist it to subloop's preheader.
802         SmallVector<MachineLoop *> InnerLoopWorkList;
803         for (MachineLoop *L = MLI->getLoopFor(MI.getParent()); L != CurLoop;
804              L = L->getParentLoop())
805           InnerLoopWorkList.push_back(L);
806 
807         while (!InnerLoopWorkList.empty()) {
808           MachineLoop *InnerLoop = InnerLoopWorkList.pop_back_val();
809           MachineBasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
810           if (InnerLoopPreheader) {
811             HoistRes = Hoist(&MI, InnerLoopPreheader, InnerLoop);
812             if (HoistRes & HoistResult::Hoisted)
813               break;
814           }
815         }
816       }
817 
818       if (HoistRes & HoistResult::ErasedMI)
819         continue;
820 
821       UpdateRegPressure(&MI);
822     }
823 
824     // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
825     ExitScopeIfDone(Node, OpenChildren, ParentMap);
826   }
827 }
828 
829 static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
830   return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
831 }
832 
833 /// Find all virtual register references that are liveout of the preheader to
834 /// initialize the starting "register pressure". Note this does not count live
835 /// through (livein but not used) registers.
836 void MachineLICMBase::InitRegPressure(MachineBasicBlock *BB) {
837   std::fill(RegPressure.begin(), RegPressure.end(), 0);
838 
839   // If the preheader has only a single predecessor and it ends with a
840   // fallthrough or an unconditional branch, then scan its predecessor for live
841   // defs as well. This happens whenever the preheader is created by splitting
842   // the critical edge from the loop predecessor to the loop header.
843   if (BB->pred_size() == 1) {
844     MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
845     SmallVector<MachineOperand, 4> Cond;
846     if (!TII->analyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty())
847       InitRegPressure(*BB->pred_begin());
848   }
849 
850   for (const MachineInstr &MI : *BB)
851     UpdateRegPressure(&MI, /*ConsiderUnseenAsDef=*/true);
852 }
853 
854 /// Update estimate of register pressure after the specified instruction.
855 void MachineLICMBase::UpdateRegPressure(const MachineInstr *MI,
856                                         bool ConsiderUnseenAsDef) {
857   auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/true, ConsiderUnseenAsDef);
858   for (const auto &RPIdAndCost : Cost) {
859     unsigned Class = RPIdAndCost.first;
860     if (static_cast<int>(RegPressure[Class]) < -RPIdAndCost.second)
861       RegPressure[Class] = 0;
862     else
863       RegPressure[Class] += RPIdAndCost.second;
864   }
865 }
866 
867 /// Calculate the additional register pressure that the registers used in MI
868 /// cause.
869 ///
870 /// If 'ConsiderSeen' is true, updates 'RegSeen' and uses the information to
871 /// figure out which usages are live-ins.
872 /// FIXME: Figure out a way to consider 'RegSeen' from all code paths.
873 DenseMap<unsigned, int>
874 MachineLICMBase::calcRegisterCost(const MachineInstr *MI, bool ConsiderSeen,
875                                   bool ConsiderUnseenAsDef) {
876   DenseMap<unsigned, int> Cost;
877   if (MI->isImplicitDef())
878     return Cost;
879   for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
880     const MachineOperand &MO = MI->getOperand(i);
881     if (!MO.isReg() || MO.isImplicit())
882       continue;
883     Register Reg = MO.getReg();
884     if (!Reg.isVirtual())
885       continue;
886 
887     // FIXME: It seems bad to use RegSeen only for some of these calculations.
888     bool isNew = ConsiderSeen ? RegSeen.insert(Reg).second : false;
889     const TargetRegisterClass *RC = MRI->getRegClass(Reg);
890 
891     RegClassWeight W = TRI->getRegClassWeight(RC);
892     int RCCost = 0;
893     if (MO.isDef())
894       RCCost = W.RegWeight;
895     else {
896       bool isKill = isOperandKill(MO, MRI);
897       if (isNew && !isKill && ConsiderUnseenAsDef)
898         // Haven't seen this, it must be a livein.
899         RCCost = W.RegWeight;
900       else if (!isNew && isKill)
901         RCCost = -W.RegWeight;
902     }
903     if (RCCost == 0)
904       continue;
905     const int *PS = TRI->getRegClassPressureSets(RC);
906     for (; *PS != -1; ++PS) {
907       if (!Cost.contains(*PS))
908         Cost[*PS] = RCCost;
909       else
910         Cost[*PS] += RCCost;
911     }
912   }
913   return Cost;
914 }
915 
916 /// Return true if this machine instruction loads from global offset table or
917 /// constant pool.
918 static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) {
919   assert(MI.mayLoad() && "Expected MI that loads!");
920 
921   // If we lost memory operands, conservatively assume that the instruction
922   // reads from everything..
923   if (MI.memoperands_empty())
924     return true;
925 
926   for (MachineMemOperand *MemOp : MI.memoperands())
927     if (const PseudoSourceValue *PSV = MemOp->getPseudoValue())
928       if (PSV->isGOT() || PSV->isConstantPool())
929         return true;
930 
931   return false;
932 }
933 
934 // This function iterates through all the operands of the input store MI and
935 // checks that each register operand statisfies isCallerPreservedPhysReg.
936 // This means, the value being stored and the address where it is being stored
937 // is constant throughout the body of the function (not including prologue and
938 // epilogue). When called with an MI that isn't a store, it returns false.
939 // A future improvement can be to check if the store registers are constant
940 // throughout the loop rather than throughout the funtion.
941 static bool isInvariantStore(const MachineInstr &MI,
942                              const TargetRegisterInfo *TRI,
943                              const MachineRegisterInfo *MRI) {
944 
945   bool FoundCallerPresReg = false;
946   if (!MI.mayStore() || MI.hasUnmodeledSideEffects() ||
947       (MI.getNumOperands() == 0))
948     return false;
949 
950   // Check that all register operands are caller-preserved physical registers.
951   for (const MachineOperand &MO : MI.operands()) {
952     if (MO.isReg()) {
953       Register Reg = MO.getReg();
954       // If operand is a virtual register, check if it comes from a copy of a
955       // physical register.
956       if (Reg.isVirtual())
957         Reg = TRI->lookThruCopyLike(MO.getReg(), MRI);
958       if (Reg.isVirtual())
959         return false;
960       if (!TRI->isCallerPreservedPhysReg(Reg.asMCReg(), *MI.getMF()))
961         return false;
962       else
963         FoundCallerPresReg = true;
964     } else if (!MO.isImm()) {
965         return false;
966     }
967   }
968   return FoundCallerPresReg;
969 }
970 
971 // Return true if the input MI is a copy instruction that feeds an invariant
972 // store instruction. This means that the src of the copy has to satisfy
973 // isCallerPreservedPhysReg and atleast one of it's users should satisfy
974 // isInvariantStore.
975 static bool isCopyFeedingInvariantStore(const MachineInstr &MI,
976                                         const MachineRegisterInfo *MRI,
977                                         const TargetRegisterInfo *TRI) {
978 
979   // FIXME: If targets would like to look through instructions that aren't
980   // pure copies, this can be updated to a query.
981   if (!MI.isCopy())
982     return false;
983 
984   const MachineFunction *MF = MI.getMF();
985   // Check that we are copying a constant physical register.
986   Register CopySrcReg = MI.getOperand(1).getReg();
987   if (CopySrcReg.isVirtual())
988     return false;
989 
990   if (!TRI->isCallerPreservedPhysReg(CopySrcReg.asMCReg(), *MF))
991     return false;
992 
993   Register CopyDstReg = MI.getOperand(0).getReg();
994   // Check if any of the uses of the copy are invariant stores.
995   assert(CopyDstReg.isVirtual() && "copy dst is not a virtual reg");
996 
997   for (MachineInstr &UseMI : MRI->use_instructions(CopyDstReg)) {
998     if (UseMI.mayStore() && isInvariantStore(UseMI, TRI, MRI))
999       return true;
1000   }
1001   return false;
1002 }
1003 
1004 /// Returns true if the instruction may be a suitable candidate for LICM.
1005 /// e.g. If the instruction is a call, then it's obviously not safe to hoist it.
1006 bool MachineLICMBase::IsLICMCandidate(MachineInstr &I, MachineLoop *CurLoop) {
1007   // Check if it's safe to move the instruction.
1008   bool DontMoveAcrossStore = !HoistConstLoads || !AllowedToHoistLoads[CurLoop];
1009   if ((!I.isSafeToMove(AA, DontMoveAcrossStore)) &&
1010       !(HoistConstStores && isInvariantStore(I, TRI, MRI))) {
1011     LLVM_DEBUG(dbgs() << "LICM: Instruction not safe to move.\n");
1012     return false;
1013   }
1014 
1015   // If it is a load then check if it is guaranteed to execute by making sure
1016   // that it dominates all exiting blocks. If it doesn't, then there is a path
1017   // out of the loop which does not execute this load, so we can't hoist it.
1018   // Loads from constant memory are safe to speculate, for example indexed load
1019   // from a jump table.
1020   // Stores and side effects are already checked by isSafeToMove.
1021   if (I.mayLoad() && !mayLoadFromGOTOrConstantPool(I) &&
1022       !IsGuaranteedToExecute(I.getParent(), CurLoop)) {
1023     LLVM_DEBUG(dbgs() << "LICM: Load not guaranteed to execute.\n");
1024     return false;
1025   }
1026 
1027   // Convergent attribute has been used on operations that involve inter-thread
1028   // communication which results are implicitly affected by the enclosing
1029   // control flows. It is not safe to hoist or sink such operations across
1030   // control flow.
1031   if (I.isConvergent())
1032     return false;
1033 
1034   if (!TII->shouldHoist(I, CurLoop))
1035     return false;
1036 
1037   return true;
1038 }
1039 
1040 /// Returns true if the instruction is loop invariant.
1041 bool MachineLICMBase::IsLoopInvariantInst(MachineInstr &I,
1042                                           MachineLoop *CurLoop) {
1043   if (!IsLICMCandidate(I, CurLoop)) {
1044     LLVM_DEBUG(dbgs() << "LICM: Instruction not a LICM candidate\n");
1045     return false;
1046   }
1047   return CurLoop->isLoopInvariant(I);
1048 }
1049 
1050 /// Return true if the specified instruction is used by a phi node and hoisting
1051 /// it could cause a copy to be inserted.
1052 bool MachineLICMBase::HasLoopPHIUse(const MachineInstr *MI,
1053                                     MachineLoop *CurLoop) {
1054   SmallVector<const MachineInstr *, 8> Work(1, MI);
1055   do {
1056     MI = Work.pop_back_val();
1057     for (const MachineOperand &MO : MI->all_defs()) {
1058       Register Reg = MO.getReg();
1059       if (!Reg.isVirtual())
1060         continue;
1061       for (MachineInstr &UseMI : MRI->use_instructions(Reg)) {
1062         // A PHI may cause a copy to be inserted.
1063         if (UseMI.isPHI()) {
1064           // A PHI inside the loop causes a copy because the live range of Reg is
1065           // extended across the PHI.
1066           if (CurLoop->contains(&UseMI))
1067             return true;
1068           // A PHI in an exit block can cause a copy to be inserted if the PHI
1069           // has multiple predecessors in the loop with different values.
1070           // For now, approximate by rejecting all exit blocks.
1071           if (isExitBlock(CurLoop, UseMI.getParent()))
1072             return true;
1073           continue;
1074         }
1075         // Look past copies as well.
1076         if (UseMI.isCopy() && CurLoop->contains(&UseMI))
1077           Work.push_back(&UseMI);
1078       }
1079     }
1080   } while (!Work.empty());
1081   return false;
1082 }
1083 
1084 /// Compute operand latency between a def of 'Reg' and an use in the current
1085 /// loop, return true if the target considered it high.
1086 bool MachineLICMBase::HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
1087                                             Register Reg,
1088                                             MachineLoop *CurLoop) const {
1089   if (MRI->use_nodbg_empty(Reg))
1090     return false;
1091 
1092   for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) {
1093     if (UseMI.isCopyLike())
1094       continue;
1095     if (!CurLoop->contains(UseMI.getParent()))
1096       continue;
1097     for (unsigned i = 0, e = UseMI.getNumOperands(); i != e; ++i) {
1098       const MachineOperand &MO = UseMI.getOperand(i);
1099       if (!MO.isReg() || !MO.isUse())
1100         continue;
1101       Register MOReg = MO.getReg();
1102       if (MOReg != Reg)
1103         continue;
1104 
1105       if (TII->hasHighOperandLatency(SchedModel, MRI, MI, DefIdx, UseMI, i))
1106         return true;
1107     }
1108 
1109     // Only look at the first in loop use.
1110     break;
1111   }
1112 
1113   return false;
1114 }
1115 
1116 /// Return true if the instruction is marked "cheap" or the operand latency
1117 /// between its def and a use is one or less.
1118 bool MachineLICMBase::IsCheapInstruction(MachineInstr &MI) const {
1119   if (TII->isAsCheapAsAMove(MI) || MI.isCopyLike())
1120     return true;
1121 
1122   bool isCheap = false;
1123   unsigned NumDefs = MI.getDesc().getNumDefs();
1124   for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
1125     MachineOperand &DefMO = MI.getOperand(i);
1126     if (!DefMO.isReg() || !DefMO.isDef())
1127       continue;
1128     --NumDefs;
1129     Register Reg = DefMO.getReg();
1130     if (Reg.isPhysical())
1131       continue;
1132 
1133     if (!TII->hasLowDefLatency(SchedModel, MI, i))
1134       return false;
1135     isCheap = true;
1136   }
1137 
1138   return isCheap;
1139 }
1140 
1141 /// Visit BBs from header to current BB, check if hoisting an instruction of the
1142 /// given cost matrix can cause high register pressure.
1143 bool
1144 MachineLICMBase::CanCauseHighRegPressure(const DenseMap<unsigned, int>& Cost,
1145                                          bool CheapInstr) {
1146   for (const auto &RPIdAndCost : Cost) {
1147     if (RPIdAndCost.second <= 0)
1148       continue;
1149 
1150     unsigned Class = RPIdAndCost.first;
1151     int Limit = RegLimit[Class];
1152 
1153     // Don't hoist cheap instructions if they would increase register pressure,
1154     // even if we're under the limit.
1155     if (CheapInstr && !HoistCheapInsts)
1156       return true;
1157 
1158     for (const auto &RP : BackTrace)
1159       if (static_cast<int>(RP[Class]) + RPIdAndCost.second >= Limit)
1160         return true;
1161   }
1162 
1163   return false;
1164 }
1165 
1166 /// Traverse the back trace from header to the current block and update their
1167 /// register pressures to reflect the effect of hoisting MI from the current
1168 /// block to the preheader.
1169 void MachineLICMBase::UpdateBackTraceRegPressure(const MachineInstr *MI) {
1170   // First compute the 'cost' of the instruction, i.e. its contribution
1171   // to register pressure.
1172   auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/false,
1173                                /*ConsiderUnseenAsDef=*/false);
1174 
1175   // Update register pressure of blocks from loop header to current block.
1176   for (auto &RP : BackTrace)
1177     for (const auto &RPIdAndCost : Cost)
1178       RP[RPIdAndCost.first] += RPIdAndCost.second;
1179 }
1180 
1181 /// Return true if it is potentially profitable to hoist the given loop
1182 /// invariant.
1183 bool MachineLICMBase::IsProfitableToHoist(MachineInstr &MI,
1184                                           MachineLoop *CurLoop) {
1185   if (MI.isImplicitDef())
1186     return true;
1187 
1188   // Besides removing computation from the loop, hoisting an instruction has
1189   // these effects:
1190   //
1191   // - The value defined by the instruction becomes live across the entire
1192   //   loop. This increases register pressure in the loop.
1193   //
1194   // - If the value is used by a PHI in the loop, a copy will be required for
1195   //   lowering the PHI after extending the live range.
1196   //
1197   // - When hoisting the last use of a value in the loop, that value no longer
1198   //   needs to be live in the loop. This lowers register pressure in the loop.
1199 
1200   if (HoistConstStores &&  isCopyFeedingInvariantStore(MI, MRI, TRI))
1201     return true;
1202 
1203   bool CheapInstr = IsCheapInstruction(MI);
1204   bool CreatesCopy = HasLoopPHIUse(&MI, CurLoop);
1205 
1206   // Don't hoist a cheap instruction if it would create a copy in the loop.
1207   if (CheapInstr && CreatesCopy) {
1208     LLVM_DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI);
1209     return false;
1210   }
1211 
1212   // Rematerializable instructions should always be hoisted providing the
1213   // register allocator can just pull them down again when needed.
1214   if (isTriviallyReMaterializable(MI))
1215     return true;
1216 
1217   // FIXME: If there are long latency loop-invariant instructions inside the
1218   // loop at this point, why didn't the optimizer's LICM hoist them?
1219   for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
1220     const MachineOperand &MO = MI.getOperand(i);
1221     if (!MO.isReg() || MO.isImplicit())
1222       continue;
1223     Register Reg = MO.getReg();
1224     if (!Reg.isVirtual())
1225       continue;
1226     if (MO.isDef() && HasHighOperandLatency(MI, i, Reg, CurLoop)) {
1227       LLVM_DEBUG(dbgs() << "Hoist High Latency: " << MI);
1228       ++NumHighLatency;
1229       return true;
1230     }
1231   }
1232 
1233   // Estimate register pressure to determine whether to LICM the instruction.
1234   // In low register pressure situation, we can be more aggressive about
1235   // hoisting. Also, favors hoisting long latency instructions even in
1236   // moderately high pressure situation.
1237   // Cheap instructions will only be hoisted if they don't increase register
1238   // pressure at all.
1239   auto Cost = calcRegisterCost(&MI, /*ConsiderSeen=*/false,
1240                                /*ConsiderUnseenAsDef=*/false);
1241 
1242   // Visit BBs from header to current BB, if hoisting this doesn't cause
1243   // high register pressure, then it's safe to proceed.
1244   if (!CanCauseHighRegPressure(Cost, CheapInstr)) {
1245     LLVM_DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI);
1246     ++NumLowRP;
1247     return true;
1248   }
1249 
1250   // Don't risk increasing register pressure if it would create copies.
1251   if (CreatesCopy) {
1252     LLVM_DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI);
1253     return false;
1254   }
1255 
1256   // Do not "speculate" in high register pressure situation. If an
1257   // instruction is not guaranteed to be executed in the loop, it's best to be
1258   // conservative.
1259   if (AvoidSpeculation &&
1260       (!IsGuaranteedToExecute(MI.getParent(), CurLoop) && !MayCSE(&MI))) {
1261     LLVM_DEBUG(dbgs() << "Won't speculate: " << MI);
1262     return false;
1263   }
1264 
1265   // If we have a COPY with other uses in the loop, hoist to allow the users to
1266   // also be hoisted.
1267   if (MI.isCopy() && MI.getOperand(0).isReg() &&
1268       MI.getOperand(0).getReg().isVirtual() && MI.getOperand(1).isReg() &&
1269       MI.getOperand(1).getReg().isVirtual() &&
1270       IsLoopInvariantInst(MI, CurLoop) &&
1271       any_of(MRI->use_nodbg_instructions(MI.getOperand(0).getReg()),
1272              [&CurLoop](MachineInstr &UseMI) {
1273                return CurLoop->contains(&UseMI);
1274              }))
1275     return true;
1276 
1277   // High register pressure situation, only hoist if the instruction is going
1278   // to be remat'ed.
1279   if (!isTriviallyReMaterializable(MI) &&
1280       !MI.isDereferenceableInvariantLoad()) {
1281     LLVM_DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI);
1282     return false;
1283   }
1284 
1285   return true;
1286 }
1287 
1288 /// Unfold a load from the given machineinstr if the load itself could be
1289 /// hoisted. Return the unfolded and hoistable load, or null if the load
1290 /// couldn't be unfolded or if it wouldn't be hoistable.
1291 MachineInstr *MachineLICMBase::ExtractHoistableLoad(MachineInstr *MI,
1292                                                     MachineLoop *CurLoop) {
1293   // Don't unfold simple loads.
1294   if (MI->canFoldAsLoad())
1295     return nullptr;
1296 
1297   // If not, we may be able to unfold a load and hoist that.
1298   // First test whether the instruction is loading from an amenable
1299   // memory location.
1300   if (!MI->isDereferenceableInvariantLoad())
1301     return nullptr;
1302 
1303   // Next determine the register class for a temporary register.
1304   unsigned LoadRegIndex;
1305   unsigned NewOpc =
1306     TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(),
1307                                     /*UnfoldLoad=*/true,
1308                                     /*UnfoldStore=*/false,
1309                                     &LoadRegIndex);
1310   if (NewOpc == 0) return nullptr;
1311   const MCInstrDesc &MID = TII->get(NewOpc);
1312   MachineFunction &MF = *MI->getMF();
1313   const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI, MF);
1314   // Ok, we're unfolding. Create a temporary register and do the unfold.
1315   Register Reg = MRI->createVirtualRegister(RC);
1316 
1317   SmallVector<MachineInstr *, 2> NewMIs;
1318   bool Success = TII->unfoldMemoryOperand(MF, *MI, Reg,
1319                                           /*UnfoldLoad=*/true,
1320                                           /*UnfoldStore=*/false, NewMIs);
1321   (void)Success;
1322   assert(Success &&
1323          "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold "
1324          "succeeded!");
1325   assert(NewMIs.size() == 2 &&
1326          "Unfolded a load into multiple instructions!");
1327   MachineBasicBlock *MBB = MI->getParent();
1328   MachineBasicBlock::iterator Pos = MI;
1329   MBB->insert(Pos, NewMIs[0]);
1330   MBB->insert(Pos, NewMIs[1]);
1331   // If unfolding produced a load that wasn't loop-invariant or profitable to
1332   // hoist, discard the new instructions and bail.
1333   if (!IsLoopInvariantInst(*NewMIs[0], CurLoop) ||
1334       !IsProfitableToHoist(*NewMIs[0], CurLoop)) {
1335     NewMIs[0]->eraseFromParent();
1336     NewMIs[1]->eraseFromParent();
1337     return nullptr;
1338   }
1339 
1340   // Update register pressure for the unfolded instruction.
1341   UpdateRegPressure(NewMIs[1]);
1342 
1343   // Otherwise we successfully unfolded a load that we can hoist.
1344 
1345   // Update the call site info.
1346   if (MI->shouldUpdateCallSiteInfo())
1347     MF.eraseCallSiteInfo(MI);
1348 
1349   MI->eraseFromParent();
1350   return NewMIs[0];
1351 }
1352 
1353 /// Initialize the CSE map with instructions that are in the current loop
1354 /// preheader that may become duplicates of instructions that are hoisted
1355 /// out of the loop.
1356 void MachineLICMBase::InitCSEMap(MachineBasicBlock *BB) {
1357   for (MachineInstr &MI : *BB)
1358     CSEMap[BB][MI.getOpcode()].push_back(&MI);
1359 }
1360 
1361 /// Initialize AllowedToHoistLoads with information about whether invariant
1362 /// loads can be moved outside a given loop
1363 void MachineLICMBase::InitializeLoadsHoistableLoops() {
1364   SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
1365   SmallVector<MachineLoop *, 8> LoopsInPreOrder;
1366 
1367   // Mark all loops as hoistable initially and prepare a list of loops in
1368   // pre-order DFS.
1369   while (!Worklist.empty()) {
1370     auto *L = Worklist.pop_back_val();
1371     AllowedToHoistLoads[L] = true;
1372     LoopsInPreOrder.push_back(L);
1373     Worklist.insert(Worklist.end(), L->getSubLoops().begin(),
1374                     L->getSubLoops().end());
1375   }
1376 
1377   // Going from the innermost to outermost loops, check if a loop has
1378   // instructions preventing invariant load hoisting. If such instruction is
1379   // found, mark this loop and its parent as non-hoistable and continue
1380   // investigating the next loop.
1381   // Visiting in a reversed pre-ordered DFS manner
1382   // allows us to not process all the instructions of the outer loop if the
1383   // inner loop is proved to be non-load-hoistable.
1384   for (auto *Loop : reverse(LoopsInPreOrder)) {
1385     for (auto *MBB : Loop->blocks()) {
1386       // If this loop has already been marked as non-hoistable, skip it.
1387       if (!AllowedToHoistLoads[Loop])
1388         continue;
1389       for (auto &MI : *MBB) {
1390         if (!MI.mayStore() && !MI.isCall() &&
1391             !(MI.mayLoad() && MI.hasOrderedMemoryRef()))
1392           continue;
1393         for (MachineLoop *L = Loop; L != nullptr; L = L->getParentLoop())
1394           AllowedToHoistLoads[L] = false;
1395         break;
1396       }
1397     }
1398   }
1399 }
1400 
1401 /// Find an instruction amount PrevMIs that is a duplicate of MI.
1402 /// Return this instruction if it's found.
1403 MachineInstr *
1404 MachineLICMBase::LookForDuplicate(const MachineInstr *MI,
1405                                   std::vector<MachineInstr *> &PrevMIs) {
1406   for (MachineInstr *PrevMI : PrevMIs)
1407     if (TII->produceSameValue(*MI, *PrevMI, (PreRegAlloc ? MRI : nullptr)))
1408       return PrevMI;
1409 
1410   return nullptr;
1411 }
1412 
1413 /// Given a LICM'ed instruction, look for an instruction on the preheader that
1414 /// computes the same value. If it's found, do a RAU on with the definition of
1415 /// the existing instruction rather than hoisting the instruction to the
1416 /// preheader.
1417 bool MachineLICMBase::EliminateCSE(
1418     MachineInstr *MI,
1419     DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI) {
1420   // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1421   // the undef property onto uses.
1422   if (MI->isImplicitDef())
1423     return false;
1424 
1425   // Do not CSE normal loads because between them could be store instructions
1426   // that change the loaded value
1427   if (MI->mayLoad() && !MI->isDereferenceableInvariantLoad())
1428     return false;
1429 
1430   if (MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
1431     LLVM_DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup);
1432 
1433     // Replace virtual registers defined by MI by their counterparts defined
1434     // by Dup.
1435     SmallVector<unsigned, 2> Defs;
1436     for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1437       const MachineOperand &MO = MI->getOperand(i);
1438 
1439       // Physical registers may not differ here.
1440       assert((!MO.isReg() || MO.getReg() == 0 || !MO.getReg().isPhysical() ||
1441               MO.getReg() == Dup->getOperand(i).getReg()) &&
1442              "Instructions with different phys regs are not identical!");
1443 
1444       if (MO.isReg() && MO.isDef() && !MO.getReg().isPhysical())
1445         Defs.push_back(i);
1446     }
1447 
1448     SmallVector<const TargetRegisterClass*, 2> OrigRCs;
1449     for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1450       unsigned Idx = Defs[i];
1451       Register Reg = MI->getOperand(Idx).getReg();
1452       Register DupReg = Dup->getOperand(Idx).getReg();
1453       OrigRCs.push_back(MRI->getRegClass(DupReg));
1454 
1455       if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) {
1456         // Restore old RCs if more than one defs.
1457         for (unsigned j = 0; j != i; ++j)
1458           MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]);
1459         return false;
1460       }
1461     }
1462 
1463     for (unsigned Idx : Defs) {
1464       Register Reg = MI->getOperand(Idx).getReg();
1465       Register DupReg = Dup->getOperand(Idx).getReg();
1466       MRI->replaceRegWith(Reg, DupReg);
1467       MRI->clearKillFlags(DupReg);
1468       // Clear Dup dead flag if any, we reuse it for Reg.
1469       if (!MRI->use_nodbg_empty(DupReg))
1470         Dup->getOperand(Idx).setIsDead(false);
1471     }
1472 
1473     MI->eraseFromParent();
1474     ++NumCSEed;
1475     return true;
1476   }
1477   return false;
1478 }
1479 
1480 /// Return true if the given instruction will be CSE'd if it's hoisted out of
1481 /// the loop.
1482 bool MachineLICMBase::MayCSE(MachineInstr *MI) {
1483   if (MI->mayLoad() && !MI->isDereferenceableInvariantLoad())
1484     return false;
1485 
1486   unsigned Opcode = MI->getOpcode();
1487   for (auto &Map : CSEMap) {
1488     // Check this CSEMap's preheader dominates MI's basic block.
1489     if (DT->dominates(Map.first, MI->getParent())) {
1490       DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI =
1491           Map.second.find(Opcode);
1492       // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1493       // the undef property onto uses.
1494       if (CI == Map.second.end() || MI->isImplicitDef())
1495         continue;
1496       if (LookForDuplicate(MI, CI->second) != nullptr)
1497         return true;
1498     }
1499   }
1500 
1501   return false;
1502 }
1503 
1504 /// When an instruction is found to use only loop invariant operands
1505 /// that are safe to hoist, this instruction is called to do the dirty work.
1506 /// It returns true if the instruction is hoisted.
1507 unsigned MachineLICMBase::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader,
1508                                 MachineLoop *CurLoop) {
1509   MachineBasicBlock *SrcBlock = MI->getParent();
1510 
1511   // Disable the instruction hoisting due to block hotness
1512   if ((DisableHoistingToHotterBlocks == UseBFI::All ||
1513       (DisableHoistingToHotterBlocks == UseBFI::PGO && HasProfileData)) &&
1514       isTgtHotterThanSrc(SrcBlock, Preheader)) {
1515     ++NumNotHoistedDueToHotness;
1516     return HoistResult::NotHoisted;
1517   }
1518   // First check whether we should hoist this instruction.
1519   bool HasExtractHoistableLoad = false;
1520   if (!IsLoopInvariantInst(*MI, CurLoop) ||
1521       !IsProfitableToHoist(*MI, CurLoop)) {
1522     // If not, try unfolding a hoistable load.
1523     MI = ExtractHoistableLoad(MI, CurLoop);
1524     if (!MI)
1525       return HoistResult::NotHoisted;
1526     HasExtractHoistableLoad = true;
1527   }
1528 
1529   // If we have hoisted an instruction that may store, it can only be a constant
1530   // store.
1531   if (MI->mayStore())
1532     NumStoreConst++;
1533 
1534   // Now move the instructions to the predecessor, inserting it before any
1535   // terminator instructions.
1536   LLVM_DEBUG({
1537     dbgs() << "Hoisting " << *MI;
1538     if (MI->getParent()->getBasicBlock())
1539       dbgs() << " from " << printMBBReference(*MI->getParent());
1540     if (Preheader->getBasicBlock())
1541       dbgs() << " to " << printMBBReference(*Preheader);
1542     dbgs() << "\n";
1543   });
1544 
1545   // If this is the first instruction being hoisted to the preheader,
1546   // initialize the CSE map with potential common expressions.
1547   if (FirstInLoop) {
1548     InitCSEMap(Preheader);
1549     FirstInLoop = false;
1550   }
1551 
1552   // Look for opportunity to CSE the hoisted instruction.
1553   unsigned Opcode = MI->getOpcode();
1554   bool HasCSEDone = false;
1555   for (auto &Map : CSEMap) {
1556     // Check this CSEMap's preheader dominates MI's basic block.
1557     if (DT->dominates(Map.first, MI->getParent())) {
1558       DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI =
1559           Map.second.find(Opcode);
1560       if (CI != Map.second.end()) {
1561         if (EliminateCSE(MI, CI)) {
1562           HasCSEDone = true;
1563           break;
1564         }
1565       }
1566     }
1567   }
1568 
1569   if (!HasCSEDone) {
1570     // Otherwise, splice the instruction to the preheader.
1571     Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
1572 
1573     // Since we are moving the instruction out of its basic block, we do not
1574     // retain its debug location. Doing so would degrade the debugging
1575     // experience and adversely affect the accuracy of profiling information.
1576     assert(!MI->isDebugInstr() && "Should not hoist debug inst");
1577     MI->setDebugLoc(DebugLoc());
1578 
1579     // Update register pressure for BBs from header to this block.
1580     UpdateBackTraceRegPressure(MI);
1581 
1582     // Clear the kill flags of any register this instruction defines,
1583     // since they may need to be live throughout the entire loop
1584     // rather than just live for part of it.
1585     for (MachineOperand &MO : MI->all_defs())
1586       if (!MO.isDead())
1587         MRI->clearKillFlags(MO.getReg());
1588 
1589     CSEMap[Preheader][Opcode].push_back(MI);
1590   }
1591 
1592   ++NumHoisted;
1593   Changed = true;
1594 
1595   if (HasCSEDone || HasExtractHoistableLoad)
1596     return HoistResult::Hoisted | HoistResult::ErasedMI;
1597   return HoistResult::Hoisted;
1598 }
1599 
1600 /// Get the preheader for the current loop, splitting a critical edge if needed.
1601 MachineBasicBlock *
1602 MachineLICMBase::getCurPreheader(MachineLoop *CurLoop,
1603                                  MachineBasicBlock *CurPreheader) {
1604   // Determine the block to which to hoist instructions. If we can't find a
1605   // suitable loop predecessor, we can't do any hoisting.
1606 
1607   // If we've tried to get a preheader and failed, don't try again.
1608   if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
1609     return nullptr;
1610 
1611   if (!CurPreheader) {
1612     CurPreheader = CurLoop->getLoopPreheader();
1613     if (!CurPreheader) {
1614       MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
1615       if (!Pred) {
1616         CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1617         return nullptr;
1618       }
1619 
1620       CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), *this);
1621       if (!CurPreheader) {
1622         CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1623         return nullptr;
1624       }
1625     }
1626   }
1627   return CurPreheader;
1628 }
1629 
1630 /// Is the target basic block at least "BlockFrequencyRatioThreshold"
1631 /// times hotter than the source basic block.
1632 bool MachineLICMBase::isTgtHotterThanSrc(MachineBasicBlock *SrcBlock,
1633                                          MachineBasicBlock *TgtBlock) {
1634   // Parse source and target basic block frequency from MBFI
1635   uint64_t SrcBF = MBFI->getBlockFreq(SrcBlock).getFrequency();
1636   uint64_t DstBF = MBFI->getBlockFreq(TgtBlock).getFrequency();
1637 
1638   // Disable the hoisting if source block frequency is zero
1639   if (!SrcBF)
1640     return true;
1641 
1642   double Ratio = (double)DstBF / SrcBF;
1643 
1644   // Compare the block frequency ratio with the threshold
1645   return Ratio > BlockFrequencyRatioThreshold;
1646 }
1647