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(Reg.isPhysical() && "Not expecting virtual register!");
456 
457     if (!MO.isDef()) {
458       if (Reg && (PhysRegDefs.test(Reg) || PhysRegClobbers.test(Reg)))
459         // If it's using a non-loop-invariant register, then it's obviously not
460         // safe to hoist.
461         HasNonInvariantUse = true;
462       continue;
463     }
464 
465     if (MO.isImplicit()) {
466       for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
467         PhysRegClobbers.set(*AI);
468       if (!MO.isDead())
469         // Non-dead implicit def? This cannot be hoisted.
470         RuledOut = true;
471       // No need to check if a dead implicit def is also defined by
472       // another instruction.
473       continue;
474     }
475 
476     // FIXME: For now, avoid instructions with multiple defs, unless
477     // it's a dead implicit def.
478     if (Def)
479       RuledOut = true;
480     else
481       Def = Reg;
482 
483     // If we have already seen another instruction that defines the same
484     // register, then this is not safe.  Two defs is indicated by setting a
485     // PhysRegClobbers bit.
486     for (MCRegAliasIterator AS(Reg, TRI, true); AS.isValid(); ++AS) {
487       if (PhysRegDefs.test(*AS))
488         PhysRegClobbers.set(*AS);
489     }
490     // Need a second loop because MCRegAliasIterator can visit the same
491     // register twice.
492     for (MCRegAliasIterator AS(Reg, TRI, true); AS.isValid(); ++AS)
493       PhysRegDefs.set(*AS);
494 
495     if (PhysRegClobbers.test(Reg))
496       // MI defined register is seen defined by another instruction in
497       // the loop, it cannot be a LICM candidate.
498       RuledOut = true;
499   }
500 
501   // Only consider reloads for now and remats which do not have register
502   // operands. FIXME: Consider unfold load folding instructions.
503   if (Def && !RuledOut) {
504     int FI = std::numeric_limits<int>::min();
505     if ((!HasNonInvariantUse && IsLICMCandidate(*MI)) ||
506         (TII->isLoadFromStackSlot(*MI, FI) && MFI->isSpillSlotObjectIndex(FI)))
507       Candidates.push_back(CandidateInfo(MI, Def, FI));
508   }
509 }
510 
511 /// Walk the specified region of the CFG and hoist loop invariants out to the
512 /// preheader.
513 void MachineLICMBase::HoistRegionPostRA() {
514   MachineBasicBlock *Preheader = getCurPreheader();
515   if (!Preheader)
516     return;
517 
518   unsigned NumRegs = TRI->getNumRegs();
519   BitVector PhysRegDefs(NumRegs); // Regs defined once in the loop.
520   BitVector PhysRegClobbers(NumRegs); // Regs defined more than once.
521 
522   SmallVector<CandidateInfo, 32> Candidates;
523   SmallSet<int, 32> StoredFIs;
524 
525   // Walk the entire region, count number of defs for each register, and
526   // collect potential LICM candidates.
527   for (MachineBasicBlock *BB : CurLoop->getBlocks()) {
528     // If the header of the loop containing this basic block is a landing pad,
529     // then don't try to hoist instructions out of this loop.
530     const MachineLoop *ML = MLI->getLoopFor(BB);
531     if (ML && ML->getHeader()->isEHPad()) continue;
532 
533     // Conservatively treat live-in's as an external def.
534     // FIXME: That means a reload that're reused in successor block(s) will not
535     // be LICM'ed.
536     for (const auto &LI : BB->liveins()) {
537       for (MCRegAliasIterator AI(LI.PhysReg, TRI, true); AI.isValid(); ++AI)
538         PhysRegDefs.set(*AI);
539     }
540 
541     SpeculationState = SpeculateUnknown;
542     for (MachineInstr &MI : *BB)
543       ProcessMI(&MI, PhysRegDefs, PhysRegClobbers, StoredFIs, Candidates);
544   }
545 
546   // Gather the registers read / clobbered by the terminator.
547   BitVector TermRegs(NumRegs);
548   MachineBasicBlock::iterator TI = Preheader->getFirstTerminator();
549   if (TI != Preheader->end()) {
550     for (const MachineOperand &MO : TI->operands()) {
551       if (!MO.isReg())
552         continue;
553       Register Reg = MO.getReg();
554       if (!Reg)
555         continue;
556       for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
557         TermRegs.set(*AI);
558     }
559   }
560 
561   // Now evaluate whether the potential candidates qualify.
562   // 1. Check if the candidate defined register is defined by another
563   //    instruction in the loop.
564   // 2. If the candidate is a load from stack slot (always true for now),
565   //    check if the slot is stored anywhere in the loop.
566   // 3. Make sure candidate def should not clobber
567   //    registers read by the terminator. Similarly its def should not be
568   //    clobbered by the terminator.
569   for (CandidateInfo &Candidate : Candidates) {
570     if (Candidate.FI != std::numeric_limits<int>::min() &&
571         StoredFIs.count(Candidate.FI))
572       continue;
573 
574     unsigned Def = Candidate.Def;
575     if (!PhysRegClobbers.test(Def) && !TermRegs.test(Def)) {
576       bool Safe = true;
577       MachineInstr *MI = Candidate.MI;
578       for (const MachineOperand &MO : MI->operands()) {
579         if (!MO.isReg() || MO.isDef() || !MO.getReg())
580           continue;
581         Register Reg = MO.getReg();
582         if (PhysRegDefs.test(Reg) ||
583             PhysRegClobbers.test(Reg)) {
584           // If it's using a non-loop-invariant register, then it's obviously
585           // not safe to hoist.
586           Safe = false;
587           break;
588         }
589       }
590       if (Safe)
591         HoistPostRA(MI, Candidate.Def);
592     }
593   }
594 }
595 
596 /// Add register 'Reg' to the livein sets of BBs in the current loop, and make
597 /// sure it is not killed by any instructions in the loop.
598 void MachineLICMBase::AddToLiveIns(MCRegister Reg) {
599   for (MachineBasicBlock *BB : CurLoop->getBlocks()) {
600     if (!BB->isLiveIn(Reg))
601       BB->addLiveIn(Reg);
602     for (MachineInstr &MI : *BB) {
603       for (MachineOperand &MO : MI.operands()) {
604         if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue;
605         if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg()))
606           MO.setIsKill(false);
607       }
608     }
609   }
610 }
611 
612 /// When an instruction is found to only use loop invariant operands that is
613 /// safe to hoist, this instruction is called to do the dirty work.
614 void MachineLICMBase::HoistPostRA(MachineInstr *MI, unsigned Def) {
615   MachineBasicBlock *Preheader = getCurPreheader();
616 
617   // Now move the instructions to the predecessor, inserting it before any
618   // terminator instructions.
619   LLVM_DEBUG(dbgs() << "Hoisting to " << printMBBReference(*Preheader)
620                     << " from " << printMBBReference(*MI->getParent()) << ": "
621                     << *MI);
622 
623   // Splice the instruction to the preheader.
624   MachineBasicBlock *MBB = MI->getParent();
625   Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
626 
627   // Since we are moving the instruction out of its basic block, we do not
628   // retain its debug location. Doing so would degrade the debugging
629   // experience and adversely affect the accuracy of profiling information.
630   assert(!MI->isDebugInstr() && "Should not hoist debug inst");
631   MI->setDebugLoc(DebugLoc());
632 
633   // Add register to livein list to all the BBs in the current loop since a
634   // loop invariant must be kept live throughout the whole loop. This is
635   // important to ensure later passes do not scavenge the def register.
636   AddToLiveIns(Def);
637 
638   ++NumPostRAHoisted;
639   Changed = true;
640 }
641 
642 /// Check if this mbb is guaranteed to execute. If not then a load from this mbb
643 /// may not be safe to hoist.
644 bool MachineLICMBase::IsGuaranteedToExecute(MachineBasicBlock *BB) {
645   if (SpeculationState != SpeculateUnknown)
646     return SpeculationState == SpeculateFalse;
647 
648   if (BB != CurLoop->getHeader()) {
649     // Check loop exiting blocks.
650     SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks;
651     CurLoop->getExitingBlocks(CurrentLoopExitingBlocks);
652     for (MachineBasicBlock *CurrentLoopExitingBlock : CurrentLoopExitingBlocks)
653       if (!DT->dominates(BB, CurrentLoopExitingBlock)) {
654         SpeculationState = SpeculateTrue;
655         return false;
656       }
657   }
658 
659   SpeculationState = SpeculateFalse;
660   return true;
661 }
662 
663 /// Check if \p MI is trivially remateralizable and if it does not have any
664 /// virtual register uses. Even though rematerializable RA might not actually
665 /// rematerialize it in this scenario. In that case we do not want to hoist such
666 /// instruction out of the loop in a belief RA will sink it back if needed.
667 bool MachineLICMBase::isTriviallyReMaterializable(
668     const MachineInstr &MI) const {
669   if (!TII->isTriviallyReMaterializable(MI))
670     return false;
671 
672   for (const MachineOperand &MO : MI.operands()) {
673     if (MO.isReg() && MO.isUse() && MO.getReg().isVirtual())
674       return false;
675   }
676 
677   return true;
678 }
679 
680 void MachineLICMBase::EnterScope(MachineBasicBlock *MBB) {
681   LLVM_DEBUG(dbgs() << "Entering " << printMBBReference(*MBB) << '\n');
682 
683   // Remember livein register pressure.
684   BackTrace.push_back(RegPressure);
685 }
686 
687 void MachineLICMBase::ExitScope(MachineBasicBlock *MBB) {
688   LLVM_DEBUG(dbgs() << "Exiting " << printMBBReference(*MBB) << '\n');
689   BackTrace.pop_back();
690 }
691 
692 /// Destroy scope for the MBB that corresponds to the given dominator tree node
693 /// if its a leaf or all of its children are done. Walk up the dominator tree to
694 /// destroy ancestors which are now done.
695 void MachineLICMBase::ExitScopeIfDone(MachineDomTreeNode *Node,
696     DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
697     const DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
698   if (OpenChildren[Node])
699     return;
700 
701   for(;;) {
702     ExitScope(Node->getBlock());
703     // Now traverse upwards to pop ancestors whose offsprings are all done.
704     MachineDomTreeNode *Parent = ParentMap.lookup(Node);
705     if (!Parent || --OpenChildren[Parent] != 0)
706       break;
707     Node = Parent;
708   }
709 }
710 
711 /// Walk the specified loop in the CFG (defined by all blocks dominated by the
712 /// specified header block, and that are in the current loop) in depth first
713 /// order w.r.t the DominatorTree. This allows us to visit definitions before
714 /// uses, allowing us to hoist a loop body in one pass without iteration.
715 void MachineLICMBase::HoistOutOfLoop(MachineDomTreeNode *HeaderN) {
716   MachineBasicBlock *Preheader = getCurPreheader();
717   if (!Preheader)
718     return;
719 
720   SmallVector<MachineDomTreeNode*, 32> Scopes;
721   SmallVector<MachineDomTreeNode*, 8> WorkList;
722   DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
723   DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
724 
725   // Perform a DFS walk to determine the order of visit.
726   WorkList.push_back(HeaderN);
727   while (!WorkList.empty()) {
728     MachineDomTreeNode *Node = WorkList.pop_back_val();
729     assert(Node && "Null dominator tree node?");
730     MachineBasicBlock *BB = Node->getBlock();
731 
732     // If the header of the loop containing this basic block is a landing pad,
733     // then don't try to hoist instructions out of this loop.
734     const MachineLoop *ML = MLI->getLoopFor(BB);
735     if (ML && ML->getHeader()->isEHPad())
736       continue;
737 
738     // If this subregion is not in the top level loop at all, exit.
739     if (!CurLoop->contains(BB))
740       continue;
741 
742     Scopes.push_back(Node);
743     unsigned NumChildren = Node->getNumChildren();
744 
745     // Don't hoist things out of a large switch statement.  This often causes
746     // code to be hoisted that wasn't going to be executed, and increases
747     // register pressure in a situation where it's likely to matter.
748     if (BB->succ_size() >= 25)
749       NumChildren = 0;
750 
751     OpenChildren[Node] = NumChildren;
752     if (NumChildren) {
753       // Add children in reverse order as then the next popped worklist node is
754       // the first child of this node.  This means we ultimately traverse the
755       // DOM tree in exactly the same order as if we'd recursed.
756       for (MachineDomTreeNode *Child : reverse(Node->children())) {
757         ParentMap[Child] = Node;
758         WorkList.push_back(Child);
759       }
760     }
761   }
762 
763   if (Scopes.size() == 0)
764     return;
765 
766   // Compute registers which are livein into the loop headers.
767   RegSeen.clear();
768   BackTrace.clear();
769   InitRegPressure(Preheader);
770 
771   // Now perform LICM.
772   for (MachineDomTreeNode *Node : Scopes) {
773     MachineBasicBlock *MBB = Node->getBlock();
774 
775     EnterScope(MBB);
776 
777     // Process the block
778     SpeculationState = SpeculateUnknown;
779     for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) {
780       if (!Hoist(&MI, Preheader))
781         UpdateRegPressure(&MI);
782       // If we have hoisted an instruction that may store, it can only be a
783       // constant store.
784     }
785 
786     // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
787     ExitScopeIfDone(Node, OpenChildren, ParentMap);
788   }
789 }
790 
791 static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
792   return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
793 }
794 
795 /// Find all virtual register references that are liveout of the preheader to
796 /// initialize the starting "register pressure". Note this does not count live
797 /// through (livein but not used) registers.
798 void MachineLICMBase::InitRegPressure(MachineBasicBlock *BB) {
799   std::fill(RegPressure.begin(), RegPressure.end(), 0);
800 
801   // If the preheader has only a single predecessor and it ends with a
802   // fallthrough or an unconditional branch, then scan its predecessor for live
803   // defs as well. This happens whenever the preheader is created by splitting
804   // the critical edge from the loop predecessor to the loop header.
805   if (BB->pred_size() == 1) {
806     MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
807     SmallVector<MachineOperand, 4> Cond;
808     if (!TII->analyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty())
809       InitRegPressure(*BB->pred_begin());
810   }
811 
812   for (const MachineInstr &MI : *BB)
813     UpdateRegPressure(&MI, /*ConsiderUnseenAsDef=*/true);
814 }
815 
816 /// Update estimate of register pressure after the specified instruction.
817 void MachineLICMBase::UpdateRegPressure(const MachineInstr *MI,
818                                         bool ConsiderUnseenAsDef) {
819   auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/true, ConsiderUnseenAsDef);
820   for (const auto &RPIdAndCost : Cost) {
821     unsigned Class = RPIdAndCost.first;
822     if (static_cast<int>(RegPressure[Class]) < -RPIdAndCost.second)
823       RegPressure[Class] = 0;
824     else
825       RegPressure[Class] += RPIdAndCost.second;
826   }
827 }
828 
829 /// Calculate the additional register pressure that the registers used in MI
830 /// cause.
831 ///
832 /// If 'ConsiderSeen' is true, updates 'RegSeen' and uses the information to
833 /// figure out which usages are live-ins.
834 /// FIXME: Figure out a way to consider 'RegSeen' from all code paths.
835 DenseMap<unsigned, int>
836 MachineLICMBase::calcRegisterCost(const MachineInstr *MI, bool ConsiderSeen,
837                                   bool ConsiderUnseenAsDef) {
838   DenseMap<unsigned, int> Cost;
839   if (MI->isImplicitDef())
840     return Cost;
841   for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
842     const MachineOperand &MO = MI->getOperand(i);
843     if (!MO.isReg() || MO.isImplicit())
844       continue;
845     Register Reg = MO.getReg();
846     if (!Reg.isVirtual())
847       continue;
848 
849     // FIXME: It seems bad to use RegSeen only for some of these calculations.
850     bool isNew = ConsiderSeen ? RegSeen.insert(Reg).second : false;
851     const TargetRegisterClass *RC = MRI->getRegClass(Reg);
852 
853     RegClassWeight W = TRI->getRegClassWeight(RC);
854     int RCCost = 0;
855     if (MO.isDef())
856       RCCost = W.RegWeight;
857     else {
858       bool isKill = isOperandKill(MO, MRI);
859       if (isNew && !isKill && ConsiderUnseenAsDef)
860         // Haven't seen this, it must be a livein.
861         RCCost = W.RegWeight;
862       else if (!isNew && isKill)
863         RCCost = -W.RegWeight;
864     }
865     if (RCCost == 0)
866       continue;
867     const int *PS = TRI->getRegClassPressureSets(RC);
868     for (; *PS != -1; ++PS) {
869       if (Cost.find(*PS) == Cost.end())
870         Cost[*PS] = RCCost;
871       else
872         Cost[*PS] += RCCost;
873     }
874   }
875   return Cost;
876 }
877 
878 /// Return true if this machine instruction loads from global offset table or
879 /// constant pool.
880 static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) {
881   assert(MI.mayLoad() && "Expected MI that loads!");
882 
883   // If we lost memory operands, conservatively assume that the instruction
884   // reads from everything..
885   if (MI.memoperands_empty())
886     return true;
887 
888   for (MachineMemOperand *MemOp : MI.memoperands())
889     if (const PseudoSourceValue *PSV = MemOp->getPseudoValue())
890       if (PSV->isGOT() || PSV->isConstantPool())
891         return true;
892 
893   return false;
894 }
895 
896 // This function iterates through all the operands of the input store MI and
897 // checks that each register operand statisfies isCallerPreservedPhysReg.
898 // This means, the value being stored and the address where it is being stored
899 // is constant throughout the body of the function (not including prologue and
900 // epilogue). When called with an MI that isn't a store, it returns false.
901 // A future improvement can be to check if the store registers are constant
902 // throughout the loop rather than throughout the funtion.
903 static bool isInvariantStore(const MachineInstr &MI,
904                              const TargetRegisterInfo *TRI,
905                              const MachineRegisterInfo *MRI) {
906 
907   bool FoundCallerPresReg = false;
908   if (!MI.mayStore() || MI.hasUnmodeledSideEffects() ||
909       (MI.getNumOperands() == 0))
910     return false;
911 
912   // Check that all register operands are caller-preserved physical registers.
913   for (const MachineOperand &MO : MI.operands()) {
914     if (MO.isReg()) {
915       Register Reg = MO.getReg();
916       // If operand is a virtual register, check if it comes from a copy of a
917       // physical register.
918       if (Reg.isVirtual())
919         Reg = TRI->lookThruCopyLike(MO.getReg(), MRI);
920       if (Reg.isVirtual())
921         return false;
922       if (!TRI->isCallerPreservedPhysReg(Reg.asMCReg(), *MI.getMF()))
923         return false;
924       else
925         FoundCallerPresReg = true;
926     } else if (!MO.isImm()) {
927         return false;
928     }
929   }
930   return FoundCallerPresReg;
931 }
932 
933 // Return true if the input MI is a copy instruction that feeds an invariant
934 // store instruction. This means that the src of the copy has to satisfy
935 // isCallerPreservedPhysReg and atleast one of it's users should satisfy
936 // isInvariantStore.
937 static bool isCopyFeedingInvariantStore(const MachineInstr &MI,
938                                         const MachineRegisterInfo *MRI,
939                                         const TargetRegisterInfo *TRI) {
940 
941   // FIXME: If targets would like to look through instructions that aren't
942   // pure copies, this can be updated to a query.
943   if (!MI.isCopy())
944     return false;
945 
946   const MachineFunction *MF = MI.getMF();
947   // Check that we are copying a constant physical register.
948   Register CopySrcReg = MI.getOperand(1).getReg();
949   if (CopySrcReg.isVirtual())
950     return false;
951 
952   if (!TRI->isCallerPreservedPhysReg(CopySrcReg.asMCReg(), *MF))
953     return false;
954 
955   Register CopyDstReg = MI.getOperand(0).getReg();
956   // Check if any of the uses of the copy are invariant stores.
957   assert(CopyDstReg.isVirtual() && "copy dst is not a virtual reg");
958 
959   for (MachineInstr &UseMI : MRI->use_instructions(CopyDstReg)) {
960     if (UseMI.mayStore() && isInvariantStore(UseMI, TRI, MRI))
961       return true;
962   }
963   return false;
964 }
965 
966 /// Returns true if the instruction may be a suitable candidate for LICM.
967 /// e.g. If the instruction is a call, then it's obviously not safe to hoist it.
968 bool MachineLICMBase::IsLICMCandidate(MachineInstr &I) {
969   // Check if it's safe to move the instruction.
970   bool DontMoveAcrossStore = true;
971   if ((!I.isSafeToMove(AA, DontMoveAcrossStore)) &&
972       !(HoistConstStores && isInvariantStore(I, TRI, MRI))) {
973     LLVM_DEBUG(dbgs() << "LICM: Instruction not safe to move.\n");
974     return false;
975   }
976 
977   // If it is a load then check if it is guaranteed to execute by making sure
978   // that it dominates all exiting blocks. If it doesn't, then there is a path
979   // out of the loop which does not execute this load, so we can't hoist it.
980   // Loads from constant memory are safe to speculate, for example indexed load
981   // from a jump table.
982   // Stores and side effects are already checked by isSafeToMove.
983   if (I.mayLoad() && !mayLoadFromGOTOrConstantPool(I) &&
984       !IsGuaranteedToExecute(I.getParent())) {
985     LLVM_DEBUG(dbgs() << "LICM: Load not guaranteed to execute.\n");
986     return false;
987   }
988 
989   // Convergent attribute has been used on operations that involve inter-thread
990   // communication which results are implicitly affected by the enclosing
991   // control flows. It is not safe to hoist or sink such operations across
992   // control flow.
993   if (I.isConvergent())
994     return false;
995 
996   if (!TII->shouldHoist(I, CurLoop))
997     return false;
998 
999   return true;
1000 }
1001 
1002 /// Returns true if the instruction is loop invariant.
1003 bool MachineLICMBase::IsLoopInvariantInst(MachineInstr &I) {
1004   if (!IsLICMCandidate(I)) {
1005     LLVM_DEBUG(dbgs() << "LICM: Instruction not a LICM candidate\n");
1006     return false;
1007   }
1008   return CurLoop->isLoopInvariant(I);
1009 }
1010 
1011 /// Return true if the specified instruction is used by a phi node and hoisting
1012 /// it could cause a copy to be inserted.
1013 bool MachineLICMBase::HasLoopPHIUse(const MachineInstr *MI) const {
1014   SmallVector<const MachineInstr*, 8> Work(1, MI);
1015   do {
1016     MI = Work.pop_back_val();
1017     for (const MachineOperand &MO : MI->operands()) {
1018       if (!MO.isReg() || !MO.isDef())
1019         continue;
1020       Register Reg = MO.getReg();
1021       if (!Reg.isVirtual())
1022         continue;
1023       for (MachineInstr &UseMI : MRI->use_instructions(Reg)) {
1024         // A PHI may cause a copy to be inserted.
1025         if (UseMI.isPHI()) {
1026           // A PHI inside the loop causes a copy because the live range of Reg is
1027           // extended across the PHI.
1028           if (CurLoop->contains(&UseMI))
1029             return true;
1030           // A PHI in an exit block can cause a copy to be inserted if the PHI
1031           // has multiple predecessors in the loop with different values.
1032           // For now, approximate by rejecting all exit blocks.
1033           if (isExitBlock(UseMI.getParent()))
1034             return true;
1035           continue;
1036         }
1037         // Look past copies as well.
1038         if (UseMI.isCopy() && CurLoop->contains(&UseMI))
1039           Work.push_back(&UseMI);
1040       }
1041     }
1042   } while (!Work.empty());
1043   return false;
1044 }
1045 
1046 /// Compute operand latency between a def of 'Reg' and an use in the current
1047 /// loop, return true if the target considered it high.
1048 bool MachineLICMBase::HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
1049                                             Register Reg) const {
1050   if (MRI->use_nodbg_empty(Reg))
1051     return false;
1052 
1053   for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) {
1054     if (UseMI.isCopyLike())
1055       continue;
1056     if (!CurLoop->contains(UseMI.getParent()))
1057       continue;
1058     for (unsigned i = 0, e = UseMI.getNumOperands(); i != e; ++i) {
1059       const MachineOperand &MO = UseMI.getOperand(i);
1060       if (!MO.isReg() || !MO.isUse())
1061         continue;
1062       Register MOReg = MO.getReg();
1063       if (MOReg != Reg)
1064         continue;
1065 
1066       if (TII->hasHighOperandLatency(SchedModel, MRI, MI, DefIdx, UseMI, i))
1067         return true;
1068     }
1069 
1070     // Only look at the first in loop use.
1071     break;
1072   }
1073 
1074   return false;
1075 }
1076 
1077 /// Return true if the instruction is marked "cheap" or the operand latency
1078 /// between its def and a use is one or less.
1079 bool MachineLICMBase::IsCheapInstruction(MachineInstr &MI) const {
1080   if (TII->isAsCheapAsAMove(MI) || MI.isCopyLike())
1081     return true;
1082 
1083   bool isCheap = false;
1084   unsigned NumDefs = MI.getDesc().getNumDefs();
1085   for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
1086     MachineOperand &DefMO = MI.getOperand(i);
1087     if (!DefMO.isReg() || !DefMO.isDef())
1088       continue;
1089     --NumDefs;
1090     Register Reg = DefMO.getReg();
1091     if (Reg.isPhysical())
1092       continue;
1093 
1094     if (!TII->hasLowDefLatency(SchedModel, MI, i))
1095       return false;
1096     isCheap = true;
1097   }
1098 
1099   return isCheap;
1100 }
1101 
1102 /// Visit BBs from header to current BB, check if hoisting an instruction of the
1103 /// given cost matrix can cause high register pressure.
1104 bool
1105 MachineLICMBase::CanCauseHighRegPressure(const DenseMap<unsigned, int>& Cost,
1106                                          bool CheapInstr) {
1107   for (const auto &RPIdAndCost : Cost) {
1108     if (RPIdAndCost.second <= 0)
1109       continue;
1110 
1111     unsigned Class = RPIdAndCost.first;
1112     int Limit = RegLimit[Class];
1113 
1114     // Don't hoist cheap instructions if they would increase register pressure,
1115     // even if we're under the limit.
1116     if (CheapInstr && !HoistCheapInsts)
1117       return true;
1118 
1119     for (const auto &RP : BackTrace)
1120       if (static_cast<int>(RP[Class]) + RPIdAndCost.second >= Limit)
1121         return true;
1122   }
1123 
1124   return false;
1125 }
1126 
1127 /// Traverse the back trace from header to the current block and update their
1128 /// register pressures to reflect the effect of hoisting MI from the current
1129 /// block to the preheader.
1130 void MachineLICMBase::UpdateBackTraceRegPressure(const MachineInstr *MI) {
1131   // First compute the 'cost' of the instruction, i.e. its contribution
1132   // to register pressure.
1133   auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/false,
1134                                /*ConsiderUnseenAsDef=*/false);
1135 
1136   // Update register pressure of blocks from loop header to current block.
1137   for (auto &RP : BackTrace)
1138     for (const auto &RPIdAndCost : Cost)
1139       RP[RPIdAndCost.first] += RPIdAndCost.second;
1140 }
1141 
1142 /// Return true if it is potentially profitable to hoist the given loop
1143 /// invariant.
1144 bool MachineLICMBase::IsProfitableToHoist(MachineInstr &MI) {
1145   if (MI.isImplicitDef())
1146     return true;
1147 
1148   // Besides removing computation from the loop, hoisting an instruction has
1149   // these effects:
1150   //
1151   // - The value defined by the instruction becomes live across the entire
1152   //   loop. This increases register pressure in the loop.
1153   //
1154   // - If the value is used by a PHI in the loop, a copy will be required for
1155   //   lowering the PHI after extending the live range.
1156   //
1157   // - When hoisting the last use of a value in the loop, that value no longer
1158   //   needs to be live in the loop. This lowers register pressure in the loop.
1159 
1160   if (HoistConstStores &&  isCopyFeedingInvariantStore(MI, MRI, TRI))
1161     return true;
1162 
1163   bool CheapInstr = IsCheapInstruction(MI);
1164   bool CreatesCopy = HasLoopPHIUse(&MI);
1165 
1166   // Don't hoist a cheap instruction if it would create a copy in the loop.
1167   if (CheapInstr && CreatesCopy) {
1168     LLVM_DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI);
1169     return false;
1170   }
1171 
1172   // Rematerializable instructions should always be hoisted providing the
1173   // register allocator can just pull them down again when needed.
1174   if (isTriviallyReMaterializable(MI))
1175     return true;
1176 
1177   // FIXME: If there are long latency loop-invariant instructions inside the
1178   // loop at this point, why didn't the optimizer's LICM hoist them?
1179   for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
1180     const MachineOperand &MO = MI.getOperand(i);
1181     if (!MO.isReg() || MO.isImplicit())
1182       continue;
1183     Register Reg = MO.getReg();
1184     if (!Reg.isVirtual())
1185       continue;
1186     if (MO.isDef() && HasHighOperandLatency(MI, i, Reg)) {
1187       LLVM_DEBUG(dbgs() << "Hoist High Latency: " << MI);
1188       ++NumHighLatency;
1189       return true;
1190     }
1191   }
1192 
1193   // Estimate register pressure to determine whether to LICM the instruction.
1194   // In low register pressure situation, we can be more aggressive about
1195   // hoisting. Also, favors hoisting long latency instructions even in
1196   // moderately high pressure situation.
1197   // Cheap instructions will only be hoisted if they don't increase register
1198   // pressure at all.
1199   auto Cost = calcRegisterCost(&MI, /*ConsiderSeen=*/false,
1200                                /*ConsiderUnseenAsDef=*/false);
1201 
1202   // Visit BBs from header to current BB, if hoisting this doesn't cause
1203   // high register pressure, then it's safe to proceed.
1204   if (!CanCauseHighRegPressure(Cost, CheapInstr)) {
1205     LLVM_DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI);
1206     ++NumLowRP;
1207     return true;
1208   }
1209 
1210   // Don't risk increasing register pressure if it would create copies.
1211   if (CreatesCopy) {
1212     LLVM_DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI);
1213     return false;
1214   }
1215 
1216   // Do not "speculate" in high register pressure situation. If an
1217   // instruction is not guaranteed to be executed in the loop, it's best to be
1218   // conservative.
1219   if (AvoidSpeculation &&
1220       (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI))) {
1221     LLVM_DEBUG(dbgs() << "Won't speculate: " << MI);
1222     return false;
1223   }
1224 
1225   // High register pressure situation, only hoist if the instruction is going
1226   // to be remat'ed.
1227   if (!isTriviallyReMaterializable(MI) &&
1228       !MI.isDereferenceableInvariantLoad()) {
1229     LLVM_DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI);
1230     return false;
1231   }
1232 
1233   return true;
1234 }
1235 
1236 /// Unfold a load from the given machineinstr if the load itself could be
1237 /// hoisted. Return the unfolded and hoistable load, or null if the load
1238 /// couldn't be unfolded or if it wouldn't be hoistable.
1239 MachineInstr *MachineLICMBase::ExtractHoistableLoad(MachineInstr *MI) {
1240   // Don't unfold simple loads.
1241   if (MI->canFoldAsLoad())
1242     return nullptr;
1243 
1244   // If not, we may be able to unfold a load and hoist that.
1245   // First test whether the instruction is loading from an amenable
1246   // memory location.
1247   if (!MI->isDereferenceableInvariantLoad())
1248     return nullptr;
1249 
1250   // Next determine the register class for a temporary register.
1251   unsigned LoadRegIndex;
1252   unsigned NewOpc =
1253     TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(),
1254                                     /*UnfoldLoad=*/true,
1255                                     /*UnfoldStore=*/false,
1256                                     &LoadRegIndex);
1257   if (NewOpc == 0) return nullptr;
1258   const MCInstrDesc &MID = TII->get(NewOpc);
1259   MachineFunction &MF = *MI->getMF();
1260   const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI, MF);
1261   // Ok, we're unfolding. Create a temporary register and do the unfold.
1262   Register Reg = MRI->createVirtualRegister(RC);
1263 
1264   SmallVector<MachineInstr *, 2> NewMIs;
1265   bool Success = TII->unfoldMemoryOperand(MF, *MI, Reg,
1266                                           /*UnfoldLoad=*/true,
1267                                           /*UnfoldStore=*/false, NewMIs);
1268   (void)Success;
1269   assert(Success &&
1270          "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold "
1271          "succeeded!");
1272   assert(NewMIs.size() == 2 &&
1273          "Unfolded a load into multiple instructions!");
1274   MachineBasicBlock *MBB = MI->getParent();
1275   MachineBasicBlock::iterator Pos = MI;
1276   MBB->insert(Pos, NewMIs[0]);
1277   MBB->insert(Pos, NewMIs[1]);
1278   // If unfolding produced a load that wasn't loop-invariant or profitable to
1279   // hoist, discard the new instructions and bail.
1280   if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) {
1281     NewMIs[0]->eraseFromParent();
1282     NewMIs[1]->eraseFromParent();
1283     return nullptr;
1284   }
1285 
1286   // Update register pressure for the unfolded instruction.
1287   UpdateRegPressure(NewMIs[1]);
1288 
1289   // Otherwise we successfully unfolded a load that we can hoist.
1290 
1291   // Update the call site info.
1292   if (MI->shouldUpdateCallSiteInfo())
1293     MF.eraseCallSiteInfo(MI);
1294 
1295   MI->eraseFromParent();
1296   return NewMIs[0];
1297 }
1298 
1299 /// Initialize the CSE map with instructions that are in the current loop
1300 /// preheader that may become duplicates of instructions that are hoisted
1301 /// out of the loop.
1302 void MachineLICMBase::InitCSEMap(MachineBasicBlock *BB) {
1303   for (MachineInstr &MI : *BB)
1304     CSEMap[MI.getOpcode()].push_back(&MI);
1305 }
1306 
1307 /// Find an instruction amount PrevMIs that is a duplicate of MI.
1308 /// Return this instruction if it's found.
1309 MachineInstr *
1310 MachineLICMBase::LookForDuplicate(const MachineInstr *MI,
1311                                   std::vector<MachineInstr *> &PrevMIs) {
1312   for (MachineInstr *PrevMI : PrevMIs)
1313     if (TII->produceSameValue(*MI, *PrevMI, (PreRegAlloc ? MRI : nullptr)))
1314       return PrevMI;
1315 
1316   return nullptr;
1317 }
1318 
1319 /// Given a LICM'ed instruction, look for an instruction on the preheader that
1320 /// computes the same value. If it's found, do a RAU on with the definition of
1321 /// the existing instruction rather than hoisting the instruction to the
1322 /// preheader.
1323 bool MachineLICMBase::EliminateCSE(
1324     MachineInstr *MI,
1325     DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI) {
1326   // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1327   // the undef property onto uses.
1328   if (CI == CSEMap.end() || MI->isImplicitDef())
1329     return false;
1330 
1331   if (MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
1332     LLVM_DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup);
1333 
1334     // Replace virtual registers defined by MI by their counterparts defined
1335     // by Dup.
1336     SmallVector<unsigned, 2> Defs;
1337     for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1338       const MachineOperand &MO = MI->getOperand(i);
1339 
1340       // Physical registers may not differ here.
1341       assert((!MO.isReg() || MO.getReg() == 0 || !MO.getReg().isPhysical() ||
1342               MO.getReg() == Dup->getOperand(i).getReg()) &&
1343              "Instructions with different phys regs are not identical!");
1344 
1345       if (MO.isReg() && MO.isDef() && !MO.getReg().isPhysical())
1346         Defs.push_back(i);
1347     }
1348 
1349     SmallVector<const TargetRegisterClass*, 2> OrigRCs;
1350     for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1351       unsigned Idx = Defs[i];
1352       Register Reg = MI->getOperand(Idx).getReg();
1353       Register DupReg = Dup->getOperand(Idx).getReg();
1354       OrigRCs.push_back(MRI->getRegClass(DupReg));
1355 
1356       if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) {
1357         // Restore old RCs if more than one defs.
1358         for (unsigned j = 0; j != i; ++j)
1359           MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]);
1360         return false;
1361       }
1362     }
1363 
1364     for (unsigned Idx : Defs) {
1365       Register Reg = MI->getOperand(Idx).getReg();
1366       Register DupReg = Dup->getOperand(Idx).getReg();
1367       MRI->replaceRegWith(Reg, DupReg);
1368       MRI->clearKillFlags(DupReg);
1369       // Clear Dup dead flag if any, we reuse it for Reg.
1370       if (!MRI->use_nodbg_empty(DupReg))
1371         Dup->getOperand(Idx).setIsDead(false);
1372     }
1373 
1374     MI->eraseFromParent();
1375     ++NumCSEed;
1376     return true;
1377   }
1378   return false;
1379 }
1380 
1381 /// Return true if the given instruction will be CSE'd if it's hoisted out of
1382 /// the loop.
1383 bool MachineLICMBase::MayCSE(MachineInstr *MI) {
1384   unsigned Opcode = MI->getOpcode();
1385   DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI =
1386       CSEMap.find(Opcode);
1387   // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1388   // the undef property onto uses.
1389   if (CI == CSEMap.end() || MI->isImplicitDef())
1390     return false;
1391 
1392   return LookForDuplicate(MI, CI->second) != nullptr;
1393 }
1394 
1395 /// When an instruction is found to use only loop invariant operands
1396 /// that are safe to hoist, this instruction is called to do the dirty work.
1397 /// It returns true if the instruction is hoisted.
1398 bool MachineLICMBase::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader) {
1399   MachineBasicBlock *SrcBlock = MI->getParent();
1400 
1401   // Disable the instruction hoisting due to block hotness
1402   if ((DisableHoistingToHotterBlocks == UseBFI::All ||
1403       (DisableHoistingToHotterBlocks == UseBFI::PGO && HasProfileData)) &&
1404       isTgtHotterThanSrc(SrcBlock, Preheader)) {
1405     ++NumNotHoistedDueToHotness;
1406     return false;
1407   }
1408   // First check whether we should hoist this instruction.
1409   if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) {
1410     // If not, try unfolding a hoistable load.
1411     MI = ExtractHoistableLoad(MI);
1412     if (!MI) return false;
1413   }
1414 
1415   // If we have hoisted an instruction that may store, it can only be a constant
1416   // store.
1417   if (MI->mayStore())
1418     NumStoreConst++;
1419 
1420   // Now move the instructions to the predecessor, inserting it before any
1421   // terminator instructions.
1422   LLVM_DEBUG({
1423     dbgs() << "Hoisting " << *MI;
1424     if (MI->getParent()->getBasicBlock())
1425       dbgs() << " from " << printMBBReference(*MI->getParent());
1426     if (Preheader->getBasicBlock())
1427       dbgs() << " to " << printMBBReference(*Preheader);
1428     dbgs() << "\n";
1429   });
1430 
1431   // If this is the first instruction being hoisted to the preheader,
1432   // initialize the CSE map with potential common expressions.
1433   if (FirstInLoop) {
1434     InitCSEMap(Preheader);
1435     FirstInLoop = false;
1436   }
1437 
1438   // Look for opportunity to CSE the hoisted instruction.
1439   unsigned Opcode = MI->getOpcode();
1440   DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI =
1441       CSEMap.find(Opcode);
1442   if (!EliminateCSE(MI, CI)) {
1443     // Otherwise, splice the instruction to the preheader.
1444     Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
1445 
1446     // Since we are moving the instruction out of its basic block, we do not
1447     // retain its debug location. Doing so would degrade the debugging
1448     // experience and adversely affect the accuracy of profiling information.
1449     assert(!MI->isDebugInstr() && "Should not hoist debug inst");
1450     MI->setDebugLoc(DebugLoc());
1451 
1452     // Update register pressure for BBs from header to this block.
1453     UpdateBackTraceRegPressure(MI);
1454 
1455     // Clear the kill flags of any register this instruction defines,
1456     // since they may need to be live throughout the entire loop
1457     // rather than just live for part of it.
1458     for (MachineOperand &MO : MI->operands())
1459       if (MO.isReg() && MO.isDef() && !MO.isDead())
1460         MRI->clearKillFlags(MO.getReg());
1461 
1462     // Add to the CSE map.
1463     if (CI != CSEMap.end())
1464       CI->second.push_back(MI);
1465     else
1466       CSEMap[Opcode].push_back(MI);
1467   }
1468 
1469   ++NumHoisted;
1470   Changed = true;
1471 
1472   return true;
1473 }
1474 
1475 /// Get the preheader for the current loop, splitting a critical edge if needed.
1476 MachineBasicBlock *MachineLICMBase::getCurPreheader() {
1477   // Determine the block to which to hoist instructions. If we can't find a
1478   // suitable loop predecessor, we can't do any hoisting.
1479 
1480   // If we've tried to get a preheader and failed, don't try again.
1481   if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
1482     return nullptr;
1483 
1484   if (!CurPreheader) {
1485     CurPreheader = CurLoop->getLoopPreheader();
1486     if (!CurPreheader) {
1487       MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
1488       if (!Pred) {
1489         CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1490         return nullptr;
1491       }
1492 
1493       CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), *this);
1494       if (!CurPreheader) {
1495         CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1496         return nullptr;
1497       }
1498     }
1499   }
1500   return CurPreheader;
1501 }
1502 
1503 /// Is the target basic block at least "BlockFrequencyRatioThreshold"
1504 /// times hotter than the source basic block.
1505 bool MachineLICMBase::isTgtHotterThanSrc(MachineBasicBlock *SrcBlock,
1506                                          MachineBasicBlock *TgtBlock) {
1507   // Parse source and target basic block frequency from MBFI
1508   uint64_t SrcBF = MBFI->getBlockFreq(SrcBlock).getFrequency();
1509   uint64_t DstBF = MBFI->getBlockFreq(TgtBlock).getFrequency();
1510 
1511   // Disable the hoisting if source block frequency is zero
1512   if (!SrcBF)
1513     return true;
1514 
1515   double Ratio = (double)DstBF / SrcBF;
1516 
1517   // Compare the block frequency ratio with the threshold
1518   return Ratio > BlockFrequencyRatioThreshold;
1519 }
1520