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