1 //===-- MachineLICM.cpp - Machine Loop Invariant Code Motion Pass ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs loop invariant code motion on machine instructions. We
11 // attempt to remove as much code from the body of a loop as possible.
13 // This pass is not intended to be a replacement or a complete alternative
14 // for the LLVM-IR-level LICM pass. It is only designed to hoist simple
15 // constructs that are not exposed before lowering and instruction selection.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineDominators.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineLoopInfo.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/PseudoSourceValue.h"
30 #include "llvm/MC/MCInstrItineraries.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetInstrInfo.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetRegisterInfo.h"
38 #include "llvm/Target/TargetSubtargetInfo.h"
41 #define DEBUG_TYPE "machine-licm"
44 AvoidSpeculation("avoid-speculation",
45 cl::desc("MachineLICM should avoid speculation"),
46 cl::init(true), cl::Hidden);
49 HoistCheapInsts("hoist-cheap-insts",
50 cl::desc("MachineLICM should hoist even cheap instructions"),
51 cl::init(false), cl::Hidden);
54 SinkInstsToAvoidSpills("sink-insts-to-avoid-spills",
55 cl::desc("MachineLICM should sink instructions into "
56 "loops to avoid register spills"),
57 cl::init(false), cl::Hidden);
60 "Number of machine instructions hoisted out of loops");
62 "Number of instructions hoisted in low reg pressure situation");
63 STATISTIC(NumHighLatency,
64 "Number of high latency instructions hoisted");
66 "Number of hoisted machine instructions CSEed");
67 STATISTIC(NumPostRAHoisted,
68 "Number of machine instructions hoisted out of loops post regalloc");
71 class MachineLICM : public MachineFunctionPass {
72 const TargetInstrInfo *TII;
73 const TargetLoweringBase *TLI;
74 const TargetRegisterInfo *TRI;
75 const MachineFrameInfo *MFI;
76 MachineRegisterInfo *MRI;
77 const InstrItineraryData *InstrItins;
80 // Various analyses that we use...
81 AliasAnalysis *AA; // Alias analysis info.
82 MachineLoopInfo *MLI; // Current MachineLoopInfo
83 MachineDominatorTree *DT; // Machine dominator tree for the cur loop
85 // State that is updated as we process loops
86 bool Changed; // True if a loop is changed.
87 bool FirstInLoop; // True if it's the first LICM in the loop.
88 MachineLoop *CurLoop; // The current loop we are working on.
89 MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
91 // Exit blocks for CurLoop.
92 SmallVector<MachineBasicBlock*, 8> ExitBlocks;
94 bool isExitBlock(const MachineBasicBlock *MBB) const {
95 return std::find(ExitBlocks.begin(), ExitBlocks.end(), MBB) !=
99 // Track 'estimated' register pressure.
100 SmallSet<unsigned, 32> RegSeen;
101 SmallVector<unsigned, 8> RegPressure;
103 // Register pressure "limit" per register pressure set. If the pressure
104 // is higher than the limit, then it's considered high.
105 SmallVector<unsigned, 8> RegLimit;
107 // Register pressure on path leading from loop preheader to current BB.
108 SmallVector<SmallVector<unsigned, 8>, 16> BackTrace;
110 // For each opcode, keep a list of potential CSE instructions.
111 DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
119 // If a MBB does not dominate loop exiting blocks then it may not safe
120 // to hoist loads from this block.
121 // Tri-state: 0 - false, 1 - true, 2 - unknown
122 unsigned SpeculationState;
125 static char ID; // Pass identification, replacement for typeid
127 MachineFunctionPass(ID), PreRegAlloc(true) {
128 initializeMachineLICMPass(*PassRegistry::getPassRegistry());
131 explicit MachineLICM(bool PreRA) :
132 MachineFunctionPass(ID), PreRegAlloc(PreRA) {
133 initializeMachineLICMPass(*PassRegistry::getPassRegistry());
136 bool runOnMachineFunction(MachineFunction &MF) override;
138 void getAnalysisUsage(AnalysisUsage &AU) const override {
139 AU.addRequired<MachineLoopInfo>();
140 AU.addRequired<MachineDominatorTree>();
141 AU.addRequired<AliasAnalysis>();
142 AU.addPreserved<MachineLoopInfo>();
143 AU.addPreserved<MachineDominatorTree>();
144 MachineFunctionPass::getAnalysisUsage(AU);
147 void releaseMemory() override {
156 /// CandidateInfo - Keep track of information about hoisting candidates.
157 struct CandidateInfo {
161 CandidateInfo(MachineInstr *mi, unsigned def, int fi)
162 : MI(mi), Def(def), FI(fi) {}
165 /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
166 /// invariants out to the preheader.
167 void HoistRegionPostRA();
169 /// HoistPostRA - When an instruction is found to only use loop invariant
170 /// operands that is safe to hoist, this instruction is called to do the
172 void HoistPostRA(MachineInstr *MI, unsigned Def);
174 /// ProcessMI - Examine the instruction for potentai LICM candidate. Also
175 /// gather register def and frame object update information.
176 void ProcessMI(MachineInstr *MI,
177 BitVector &PhysRegDefs,
178 BitVector &PhysRegClobbers,
179 SmallSet<int, 32> &StoredFIs,
180 SmallVectorImpl<CandidateInfo> &Candidates);
182 /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the
184 void AddToLiveIns(unsigned Reg);
186 /// IsLICMCandidate - Returns true if the instruction may be a suitable
187 /// candidate for LICM. e.g. If the instruction is a call, then it's
188 /// obviously not safe to hoist it.
189 bool IsLICMCandidate(MachineInstr &I);
191 /// IsLoopInvariantInst - Returns true if the instruction is loop
192 /// invariant. I.e., all virtual register operands are defined outside of
193 /// the loop, physical registers aren't accessed (explicitly or implicitly),
194 /// and the instruction is hoistable.
196 bool IsLoopInvariantInst(MachineInstr &I);
198 /// HasLoopPHIUse - Return true if the specified instruction is used by any
199 /// phi node in the current loop.
200 bool HasLoopPHIUse(const MachineInstr *MI) const;
202 /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
203 /// and an use in the current loop, return true if the target considered
205 bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
208 bool IsCheapInstruction(MachineInstr &MI) const;
210 /// CanCauseHighRegPressure - Visit BBs from header to current BB,
211 /// check if hoisting an instruction of the given cost matrix can cause high
212 /// register pressure.
213 bool CanCauseHighRegPressure(const DenseMap<unsigned, int> &Cost,
216 /// UpdateBackTraceRegPressure - Traverse the back trace from header to
217 /// the current block and update their register pressures to reflect the
218 /// effect of hoisting MI from the current block to the preheader.
219 void UpdateBackTraceRegPressure(const MachineInstr *MI);
221 /// IsProfitableToHoist - Return true if it is potentially profitable to
222 /// hoist the given loop invariant.
223 bool IsProfitableToHoist(MachineInstr &MI);
225 /// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
226 /// If not then a load from this mbb may not be safe to hoist.
227 bool IsGuaranteedToExecute(MachineBasicBlock *BB);
229 void EnterScope(MachineBasicBlock *MBB);
231 void ExitScope(MachineBasicBlock *MBB);
233 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to given
234 /// dominator tree node if its a leaf or all of its children are done. Walk
235 /// up the dominator tree to destroy ancestors which are now done.
236 void ExitScopeIfDone(MachineDomTreeNode *Node,
237 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
238 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap);
240 /// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all
241 /// blocks dominated by the specified header block, and that are in the
242 /// current loop) in depth first order w.r.t the DominatorTree. This allows
243 /// us to visit definitions before uses, allowing us to hoist a loop body in
244 /// one pass without iteration.
246 void HoistOutOfLoop(MachineDomTreeNode *LoopHeaderNode);
247 void HoistRegion(MachineDomTreeNode *N, bool IsHeader);
249 /// SinkIntoLoop - Sink instructions into loops if profitable. This
250 /// especially tries to prevent register spills caused by register pressure
251 /// if there is little to no overhead moving instructions into loops.
254 /// InitRegPressure - Find all virtual register references that are liveout
255 /// of the preheader to initialize the starting "register pressure". Note
256 /// this does not count live through (livein but not used) registers.
257 void InitRegPressure(MachineBasicBlock *BB);
259 /// calcRegisterCost - Calculate the additional register pressure that the
260 /// registers used in MI cause.
262 /// If 'ConsiderSeen' is true, updates 'RegSeen' and uses the information to
263 /// figure out which usages are live-ins.
264 /// FIXME: Figure out a way to consider 'RegSeen' from all code paths.
265 DenseMap<unsigned, int> calcRegisterCost(const MachineInstr *MI,
267 bool ConsiderUnseenAsDef);
269 /// UpdateRegPressure - Update estimate of register pressure after the
270 /// specified instruction.
271 void UpdateRegPressure(const MachineInstr *MI,
272 bool ConsiderUnseenAsDef = false);
274 /// ExtractHoistableLoad - Unfold a load from the given machineinstr if
275 /// the load itself could be hoisted. Return the unfolded and hoistable
276 /// load, or null if the load couldn't be unfolded or if it wouldn't
278 MachineInstr *ExtractHoistableLoad(MachineInstr *MI);
280 /// LookForDuplicate - Find an instruction amount PrevMIs that is a
281 /// duplicate of MI. Return this instruction if it's found.
282 const MachineInstr *LookForDuplicate(const MachineInstr *MI,
283 std::vector<const MachineInstr*> &PrevMIs);
285 /// EliminateCSE - Given a LICM'ed instruction, look for an instruction on
286 /// the preheader that compute the same value. If it's found, do a RAU on
287 /// with the definition of the existing instruction rather than hoisting
288 /// the instruction to the preheader.
289 bool EliminateCSE(MachineInstr *MI,
290 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI);
292 /// MayCSE - Return true if the given instruction will be CSE'd if it's
293 /// hoisted out of the loop.
294 bool MayCSE(MachineInstr *MI);
296 /// Hoist - When an instruction is found to only use loop invariant operands
297 /// that is safe to hoist, this instruction is called to do the dirty work.
298 /// It returns true if the instruction is hoisted.
299 bool Hoist(MachineInstr *MI, MachineBasicBlock *Preheader);
301 /// InitCSEMap - Initialize the CSE map with instructions that are in the
302 /// current loop preheader that may become duplicates of instructions that
303 /// are hoisted out of the loop.
304 void InitCSEMap(MachineBasicBlock *BB);
306 /// getCurPreheader - Get the preheader for the current loop, splitting
307 /// a critical edge if needed.
308 MachineBasicBlock *getCurPreheader();
310 } // end anonymous namespace
312 char MachineLICM::ID = 0;
313 char &llvm::MachineLICMID = MachineLICM::ID;
314 INITIALIZE_PASS_BEGIN(MachineLICM, "machinelicm",
315 "Machine Loop Invariant Code Motion", false, false)
316 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
317 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
318 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
319 INITIALIZE_PASS_END(MachineLICM, "machinelicm",
320 "Machine Loop Invariant Code Motion", false, false)
322 /// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most
323 /// loop that has a unique predecessor.
324 static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) {
325 // Check whether this loop even has a unique predecessor.
326 if (!CurLoop->getLoopPredecessor())
328 // Ok, now check to see if any of its outer loops do.
329 for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop())
330 if (L->getLoopPredecessor())
332 // None of them did, so this is the outermost with a unique predecessor.
336 bool MachineLICM::runOnMachineFunction(MachineFunction &MF) {
337 if (skipOptnoneFunction(*MF.getFunction()))
340 Changed = FirstInLoop = false;
341 TII = MF.getSubtarget().getInstrInfo();
342 TLI = MF.getSubtarget().getTargetLowering();
343 TRI = MF.getSubtarget().getRegisterInfo();
344 MFI = MF.getFrameInfo();
345 MRI = &MF.getRegInfo();
346 InstrItins = MF.getSubtarget().getInstrItineraryData();
348 PreRegAlloc = MRI->isSSA();
351 DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
353 DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
354 DEBUG(dbgs() << MF.getName() << " ********\n");
357 // Estimate register pressure during pre-regalloc pass.
358 unsigned NumRPS = TRI->getNumRegPressureSets();
359 RegPressure.resize(NumRPS);
360 std::fill(RegPressure.begin(), RegPressure.end(), 0);
361 RegLimit.resize(NumRPS);
362 for (unsigned i = 0, e = NumRPS; i != e; ++i)
363 RegLimit[i] = TRI->getRegPressureSetLimit(MF, i);
366 // Get our Loop information...
367 MLI = &getAnalysis<MachineLoopInfo>();
368 DT = &getAnalysis<MachineDominatorTree>();
369 AA = &getAnalysis<AliasAnalysis>();
371 SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
372 while (!Worklist.empty()) {
373 CurLoop = Worklist.pop_back_val();
374 CurPreheader = nullptr;
377 // If this is done before regalloc, only visit outer-most preheader-sporting
379 if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) {
380 Worklist.append(CurLoop->begin(), CurLoop->end());
384 CurLoop->getExitBlocks(ExitBlocks);
389 // CSEMap is initialized for loop header when the first instruction is
391 MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
396 if (SinkInstsToAvoidSpills)
404 /// InstructionStoresToFI - Return true if instruction stores to the
406 static bool InstructionStoresToFI(const MachineInstr *MI, int FI) {
407 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
408 oe = MI->memoperands_end(); o != oe; ++o) {
409 if (!(*o)->isStore() || !(*o)->getPseudoValue())
411 if (const FixedStackPseudoSourceValue *Value =
412 dyn_cast<FixedStackPseudoSourceValue>((*o)->getPseudoValue())) {
413 if (Value->getFrameIndex() == FI)
420 /// ProcessMI - Examine the instruction for potentai LICM candidate. Also
421 /// gather register def and frame object update information.
422 void MachineLICM::ProcessMI(MachineInstr *MI,
423 BitVector &PhysRegDefs,
424 BitVector &PhysRegClobbers,
425 SmallSet<int, 32> &StoredFIs,
426 SmallVectorImpl<CandidateInfo> &Candidates) {
427 bool RuledOut = false;
428 bool HasNonInvariantUse = false;
430 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
431 const MachineOperand &MO = MI->getOperand(i);
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;
443 // We can't hoist an instruction defining a physreg that is clobbered in
445 if (MO.isRegMask()) {
446 PhysRegClobbers.setBitsNotInMask(MO.getRegMask());
452 unsigned Reg = MO.getReg();
455 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
456 "Not expecting virtual register!");
459 if (Reg && (PhysRegDefs.test(Reg) || PhysRegClobbers.test(Reg)))
460 // If it's using a non-loop-invariant register, then it's obviously not
462 HasNonInvariantUse = true;
466 if (MO.isImplicit()) {
467 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
468 PhysRegClobbers.set(*AI);
470 // Non-dead implicit def? This cannot be hoisted.
472 // No need to check if a dead implicit def is also defined by
473 // another instruction.
477 // FIXME: For now, avoid instructions with multiple defs, unless
478 // it's a dead implicit def.
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 PhysRegDefs.set(*AS);
492 if (PhysRegClobbers.test(Reg))
493 // MI defined register is seen defined by another instruction in
494 // the loop, it cannot be a LICM candidate.
498 // Only consider reloads for now and remats which do not have register
499 // operands. FIXME: Consider unfold load folding instructions.
500 if (Def && !RuledOut) {
502 if ((!HasNonInvariantUse && IsLICMCandidate(*MI)) ||
503 (TII->isLoadFromStackSlot(MI, FI) && MFI->isSpillSlotObjectIndex(FI)))
504 Candidates.push_back(CandidateInfo(MI, Def, FI));
508 /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
509 /// invariants out to the preheader.
510 void MachineLICM::HoistRegionPostRA() {
511 MachineBasicBlock *Preheader = getCurPreheader();
515 unsigned NumRegs = TRI->getNumRegs();
516 BitVector PhysRegDefs(NumRegs); // Regs defined once in the loop.
517 BitVector PhysRegClobbers(NumRegs); // Regs defined more than once.
519 SmallVector<CandidateInfo, 32> Candidates;
520 SmallSet<int, 32> StoredFIs;
522 // Walk the entire region, count number of defs for each register, and
523 // collect potential LICM candidates.
524 const std::vector<MachineBasicBlock *> &Blocks = CurLoop->getBlocks();
525 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
526 MachineBasicBlock *BB = Blocks[i];
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()->isLandingPad()) continue;
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
536 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
537 E = BB->livein_end(); I != E; ++I) {
539 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
540 PhysRegDefs.set(*AI);
543 SpeculationState = SpeculateUnknown;
544 for (MachineBasicBlock::iterator
545 MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
546 MachineInstr *MI = &*MII;
547 ProcessMI(MI, PhysRegDefs, PhysRegClobbers, StoredFIs, Candidates);
551 // Gather the registers read / clobbered by the terminator.
552 BitVector TermRegs(NumRegs);
553 MachineBasicBlock::iterator TI = Preheader->getFirstTerminator();
554 if (TI != Preheader->end()) {
555 for (unsigned i = 0, e = TI->getNumOperands(); i != e; ++i) {
556 const MachineOperand &MO = TI->getOperand(i);
559 unsigned Reg = MO.getReg();
562 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
567 // Now evaluate whether the potential candidates qualify.
568 // 1. Check if the candidate defined register is defined by another
569 // instruction in the loop.
570 // 2. If the candidate is a load from stack slot (always true for now),
571 // check if the slot is stored anywhere in the loop.
572 // 3. Make sure candidate def should not clobber
573 // registers read by the terminator. Similarly its def should not be
574 // clobbered by the terminator.
575 for (unsigned i = 0, e = Candidates.size(); i != e; ++i) {
576 if (Candidates[i].FI != INT_MIN &&
577 StoredFIs.count(Candidates[i].FI))
580 unsigned Def = Candidates[i].Def;
581 if (!PhysRegClobbers.test(Def) && !TermRegs.test(Def)) {
583 MachineInstr *MI = Candidates[i].MI;
584 for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) {
585 const MachineOperand &MO = MI->getOperand(j);
586 if (!MO.isReg() || MO.isDef() || !MO.getReg())
588 unsigned Reg = MO.getReg();
589 if (PhysRegDefs.test(Reg) ||
590 PhysRegClobbers.test(Reg)) {
591 // If it's using a non-loop-invariant register, then it's obviously
592 // not safe to hoist.
598 HoistPostRA(MI, Candidates[i].Def);
603 /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the current
604 /// loop, and make sure it is not killed by any instructions in the loop.
605 void MachineLICM::AddToLiveIns(unsigned Reg) {
606 const std::vector<MachineBasicBlock *> &Blocks = CurLoop->getBlocks();
607 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
608 MachineBasicBlock *BB = Blocks[i];
609 if (!BB->isLiveIn(Reg))
611 for (MachineBasicBlock::iterator
612 MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
613 MachineInstr *MI = &*MII;
614 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
615 MachineOperand &MO = MI->getOperand(i);
616 if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue;
617 if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg()))
624 /// HoistPostRA - When an instruction is found to only use loop invariant
625 /// operands that is safe to hoist, this instruction is called to do the
627 void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) {
628 MachineBasicBlock *Preheader = getCurPreheader();
630 // Now move the instructions to the predecessor, inserting it before any
631 // terminator instructions.
632 DEBUG(dbgs() << "Hoisting to BB#" << Preheader->getNumber() << " from BB#"
633 << MI->getParent()->getNumber() << ": " << *MI);
635 // Splice the instruction to the preheader.
636 MachineBasicBlock *MBB = MI->getParent();
637 Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
639 // Add register to livein list to all the BBs in the current loop since a
640 // loop invariant must be kept live throughout the whole loop. This is
641 // important to ensure later passes do not scavenge the def register.
648 // IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
649 // If not then a load from this mbb may not be safe to hoist.
650 bool MachineLICM::IsGuaranteedToExecute(MachineBasicBlock *BB) {
651 if (SpeculationState != SpeculateUnknown)
652 return SpeculationState == SpeculateFalse;
654 if (BB != CurLoop->getHeader()) {
655 // Check loop exiting blocks.
656 SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks;
657 CurLoop->getExitingBlocks(CurrentLoopExitingBlocks);
658 for (unsigned i = 0, e = CurrentLoopExitingBlocks.size(); i != e; ++i)
659 if (!DT->dominates(BB, CurrentLoopExitingBlocks[i])) {
660 SpeculationState = SpeculateTrue;
665 SpeculationState = SpeculateFalse;
669 void MachineLICM::EnterScope(MachineBasicBlock *MBB) {
670 DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
672 // Remember livein register pressure.
673 BackTrace.push_back(RegPressure);
676 void MachineLICM::ExitScope(MachineBasicBlock *MBB) {
677 DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
678 BackTrace.pop_back();
681 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
682 /// dominator tree node if its a leaf or all of its children are done. Walk
683 /// up the dominator tree to destroy ancestors which are now done.
684 void MachineLICM::ExitScopeIfDone(MachineDomTreeNode *Node,
685 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
686 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
687 if (OpenChildren[Node])
691 ExitScope(Node->getBlock());
693 // Now traverse upwards to pop ancestors whose offsprings are all done.
694 while (MachineDomTreeNode *Parent = ParentMap[Node]) {
695 unsigned Left = --OpenChildren[Parent];
698 ExitScope(Parent->getBlock());
703 /// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all
704 /// blocks dominated by the specified header block, and that are in the
705 /// current loop) in depth first order w.r.t the DominatorTree. This allows
706 /// us to visit definitions before uses, allowing us to hoist a loop body in
707 /// one pass without iteration.
709 void MachineLICM::HoistOutOfLoop(MachineDomTreeNode *HeaderN) {
710 MachineBasicBlock *Preheader = getCurPreheader();
714 SmallVector<MachineDomTreeNode*, 32> Scopes;
715 SmallVector<MachineDomTreeNode*, 8> WorkList;
716 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
717 DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
719 // Perform a DFS walk to determine the order of visit.
720 WorkList.push_back(HeaderN);
721 while (!WorkList.empty()) {
722 MachineDomTreeNode *Node = WorkList.pop_back_val();
723 assert(Node && "Null dominator tree node?");
724 MachineBasicBlock *BB = Node->getBlock();
726 // If the header of the loop containing this basic block is a landing pad,
727 // then don't try to hoist instructions out of this loop.
728 const MachineLoop *ML = MLI->getLoopFor(BB);
729 if (ML && ML->getHeader()->isLandingPad())
732 // If this subregion is not in the top level loop at all, exit.
733 if (!CurLoop->contains(BB))
736 Scopes.push_back(Node);
737 const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
738 unsigned NumChildren = Children.size();
740 // Don't hoist things out of a large switch statement. This often causes
741 // code to be hoisted that wasn't going to be executed, and increases
742 // register pressure in a situation where it's likely to matter.
743 if (BB->succ_size() >= 25)
746 OpenChildren[Node] = NumChildren;
747 // Add children in reverse order as then the next popped worklist node is
748 // the first child of this node. This means we ultimately traverse the
749 // DOM tree in exactly the same order as if we'd recursed.
750 for (int i = (int)NumChildren-1; i >= 0; --i) {
751 MachineDomTreeNode *Child = Children[i];
752 ParentMap[Child] = Node;
753 WorkList.push_back(Child);
757 if (Scopes.size() == 0)
760 // Compute registers which are livein into the loop headers.
763 InitRegPressure(Preheader);
766 for (unsigned i = 0, e = Scopes.size(); i != e; ++i) {
767 MachineDomTreeNode *Node = Scopes[i];
768 MachineBasicBlock *MBB = Node->getBlock();
773 SpeculationState = SpeculateUnknown;
774 for (MachineBasicBlock::iterator
775 MII = MBB->begin(), E = MBB->end(); MII != E; ) {
776 MachineBasicBlock::iterator NextMII = MII; ++NextMII;
777 MachineInstr *MI = &*MII;
778 if (!Hoist(MI, Preheader))
779 UpdateRegPressure(MI);
783 // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
784 ExitScopeIfDone(Node, OpenChildren, ParentMap);
788 void MachineLICM::SinkIntoLoop() {
789 MachineBasicBlock *Preheader = getCurPreheader();
793 SmallVector<MachineInstr *, 8> Candidates;
794 for (MachineBasicBlock::instr_iterator I = Preheader->instr_begin();
795 I != Preheader->instr_end(); ++I) {
796 // We need to ensure that we can safely move this instruction into the loop.
797 // As such, it must not have side-effects, e.g. such as a call has.
798 if (IsLoopInvariantInst(*I) && !HasLoopPHIUse(I))
799 Candidates.push_back(I);
802 for (MachineInstr *I : Candidates) {
803 const MachineOperand &MO = I->getOperand(0);
804 if (!MO.isDef() || !MO.isReg() || !MO.getReg())
806 if (!MRI->hasOneDef(MO.getReg()))
809 MachineBasicBlock *B = nullptr;
810 for (MachineInstr &MI : MRI->use_instructions(MO.getReg())) {
811 // FIXME: Come up with a proper cost model that estimates whether sinking
812 // the instruction (and thus possibly executing it on every loop
813 // iteration) is more expensive than a register.
814 // For now assumes that copies are cheap and thus almost always worth it.
823 B = DT->findNearestCommonDominator(B, MI.getParent());
829 if (!CanSink || !B || B == Preheader)
831 B->splice(B->getFirstNonPHI(), Preheader, I);
835 static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
836 return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
839 /// InitRegPressure - Find all virtual register references that are liveout of
840 /// the preheader to initialize the starting "register pressure". Note this
841 /// does not count live through (livein but not used) registers.
842 void MachineLICM::InitRegPressure(MachineBasicBlock *BB) {
843 std::fill(RegPressure.begin(), RegPressure.end(), 0);
845 // If the preheader has only a single predecessor and it ends with a
846 // fallthrough or an unconditional branch, then scan its predecessor for live
847 // defs as well. This happens whenever the preheader is created by splitting
848 // the critical edge from the loop predecessor to the loop header.
849 if (BB->pred_size() == 1) {
850 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
851 SmallVector<MachineOperand, 4> Cond;
852 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty())
853 InitRegPressure(*BB->pred_begin());
856 for (const MachineInstr &MI : *BB)
857 UpdateRegPressure(&MI, /*ConsiderUnseenAsDef=*/true);
860 /// UpdateRegPressure - Update estimate of register pressure after the
861 /// specified instruction.
862 void MachineLICM::UpdateRegPressure(const MachineInstr *MI,
863 bool ConsiderUnseenAsDef) {
864 auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/true, ConsiderUnseenAsDef);
865 for (const auto &RPIdAndCost : Cost) {
866 unsigned Class = RPIdAndCost.first;
867 if (static_cast<int>(RegPressure[Class]) < -RPIdAndCost.second)
868 RegPressure[Class] = 0;
870 RegPressure[Class] += RPIdAndCost.second;
874 DenseMap<unsigned, int>
875 MachineLICM::calcRegisterCost(const MachineInstr *MI, bool ConsiderSeen,
876 bool ConsiderUnseenAsDef) {
877 DenseMap<unsigned, int> Cost;
878 if (MI->isImplicitDef())
880 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
881 const MachineOperand &MO = MI->getOperand(i);
882 if (!MO.isReg() || MO.isImplicit())
884 unsigned Reg = MO.getReg();
885 if (!TargetRegisterInfo::isVirtualRegister(Reg))
888 // FIXME: It seems bad to use RegSeen only for some of these calculations.
889 bool isNew = ConsiderSeen ? RegSeen.insert(Reg).second : false;
890 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
892 RegClassWeight W = TRI->getRegClassWeight(RC);
895 RCCost = W.RegWeight;
897 bool isKill = isOperandKill(MO, MRI);
898 if (isNew && !isKill && ConsiderUnseenAsDef)
899 // Haven't seen this, it must be a livein.
900 RCCost = W.RegWeight;
901 else if (!isNew && isKill)
902 RCCost = -W.RegWeight;
906 const int *PS = TRI->getRegClassPressureSets(RC);
907 for (; *PS != -1; ++PS) {
908 if (Cost.find(*PS) == Cost.end())
917 /// isLoadFromGOTOrConstantPool - Return true if this machine instruction
918 /// loads from global offset table or constant pool.
919 static bool isLoadFromGOTOrConstantPool(MachineInstr &MI) {
920 assert (MI.mayLoad() && "Expected MI that loads!");
921 for (MachineInstr::mmo_iterator I = MI.memoperands_begin(),
922 E = MI.memoperands_end(); I != E; ++I) {
923 if (const PseudoSourceValue *PSV = (*I)->getPseudoValue()) {
924 if (PSV == PSV->getGOT() || PSV == PSV->getConstantPool())
931 /// IsLICMCandidate - Returns true if the instruction may be a suitable
932 /// candidate for LICM. e.g. If the instruction is a call, then it's obviously
933 /// not safe to hoist it.
934 bool MachineLICM::IsLICMCandidate(MachineInstr &I) {
935 // Check if it's safe to move the instruction.
936 bool DontMoveAcrossStore = true;
937 if (!I.isSafeToMove(AA, DontMoveAcrossStore))
940 // If it is load then check if it is guaranteed to execute by making sure that
941 // it dominates all exiting blocks. If it doesn't, then there is a path out of
942 // the loop which does not execute this load, so we can't hoist it. Loads
943 // from constant memory are not safe to speculate all the time, for example
944 // indexed load from a jump table.
945 // Stores and side effects are already checked by isSafeToMove.
946 if (I.mayLoad() && !isLoadFromGOTOrConstantPool(I) &&
947 !IsGuaranteedToExecute(I.getParent()))
953 /// IsLoopInvariantInst - Returns true if the instruction is loop
954 /// invariant. I.e., all virtual register operands are defined outside of the
955 /// loop, physical registers aren't accessed explicitly, and there are no side
956 /// effects that aren't captured by the operands or other flags.
958 bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) {
959 if (!IsLICMCandidate(I))
962 // The instruction is loop invariant if all of its operands are.
963 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
964 const MachineOperand &MO = I.getOperand(i);
969 unsigned Reg = MO.getReg();
970 if (Reg == 0) continue;
972 // Don't hoist an instruction that uses or defines a physical register.
973 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
975 // If the physreg has no defs anywhere, it's just an ambient register
976 // and we can freely move its uses. Alternatively, if it's allocatable,
977 // it could get allocated to something with a def during allocation.
978 if (!MRI->isConstantPhysReg(Reg, *I.getParent()->getParent()))
980 // Otherwise it's safe to move.
982 } else if (!MO.isDead()) {
983 // A def that isn't dead. We can't move it.
985 } else if (CurLoop->getHeader()->isLiveIn(Reg)) {
986 // If the reg is live into the loop, we can't hoist an instruction
987 // which would clobber it.
995 assert(MRI->getVRegDef(Reg) &&
996 "Machine instr not mapped for this vreg?!");
998 // If the loop contains the definition of an operand, then the instruction
999 // isn't loop invariant.
1000 if (CurLoop->contains(MRI->getVRegDef(Reg)))
1004 // If we got this far, the instruction is loop invariant!
1009 /// HasLoopPHIUse - Return true if the specified instruction is used by a
1010 /// phi node and hoisting it could cause a copy to be inserted.
1011 bool MachineLICM::HasLoopPHIUse(const MachineInstr *MI) const {
1012 SmallVector<const MachineInstr*, 8> Work(1, MI);
1014 MI = Work.pop_back_val();
1015 for (ConstMIOperands MO(MI); MO.isValid(); ++MO) {
1016 if (!MO->isReg() || !MO->isDef())
1018 unsigned Reg = MO->getReg();
1019 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1021 for (MachineInstr &UseMI : MRI->use_instructions(Reg)) {
1022 // A PHI may cause a copy to be inserted.
1023 if (UseMI.isPHI()) {
1024 // A PHI inside the loop causes a copy because the live range of Reg is
1025 // extended across the PHI.
1026 if (CurLoop->contains(&UseMI))
1028 // A PHI in an exit block can cause a copy to be inserted if the PHI
1029 // has multiple predecessors in the loop with different values.
1030 // For now, approximate by rejecting all exit blocks.
1031 if (isExitBlock(UseMI.getParent()))
1035 // Look past copies as well.
1036 if (UseMI.isCopy() && CurLoop->contains(&UseMI))
1037 Work.push_back(&UseMI);
1040 } while (!Work.empty());
1044 /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
1045 /// and an use in the current loop, return true if the target considered
1047 bool MachineLICM::HasHighOperandLatency(MachineInstr &MI,
1048 unsigned DefIdx, unsigned Reg) const {
1049 if (!InstrItins || InstrItins->isEmpty() || MRI->use_nodbg_empty(Reg))
1052 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) {
1053 if (UseMI.isCopyLike())
1055 if (!CurLoop->contains(UseMI.getParent()))
1057 for (unsigned i = 0, e = UseMI.getNumOperands(); i != e; ++i) {
1058 const MachineOperand &MO = UseMI.getOperand(i);
1059 if (!MO.isReg() || !MO.isUse())
1061 unsigned MOReg = MO.getReg();
1065 if (TII->hasHighOperandLatency(InstrItins, MRI, &MI, DefIdx, &UseMI, i))
1069 // Only look at the first in loop use.
1076 /// IsCheapInstruction - Return true if the instruction is marked "cheap" or
1077 /// the operand latency between its def and a use is one or less.
1078 bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const {
1079 if (TII->isAsCheapAsAMove(&MI) || MI.isCopyLike())
1081 if (!InstrItins || InstrItins->isEmpty())
1084 bool isCheap = false;
1085 unsigned NumDefs = MI.getDesc().getNumDefs();
1086 for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
1087 MachineOperand &DefMO = MI.getOperand(i);
1088 if (!DefMO.isReg() || !DefMO.isDef())
1091 unsigned Reg = DefMO.getReg();
1092 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1095 if (!TII->hasLowDefLatency(InstrItins, &MI, i))
1103 /// CanCauseHighRegPressure - Visit BBs from header to current BB, check
1104 /// if hoisting an instruction of the given cost matrix can cause high
1105 /// register pressure.
1106 bool MachineLICM::CanCauseHighRegPressure(const DenseMap<unsigned, int>& Cost,
1108 for (const auto &RPIdAndCost : Cost) {
1109 if (RPIdAndCost.second <= 0)
1112 unsigned Class = RPIdAndCost.first;
1113 int Limit = RegLimit[Class];
1115 // Don't hoist cheap instructions if they would increase register pressure,
1116 // even if we're under the limit.
1117 if (CheapInstr && !HoistCheapInsts)
1120 for (const auto &RP : BackTrace)
1121 if (static_cast<int>(RP[Class]) + RPIdAndCost.second >= Limit)
1128 /// UpdateBackTraceRegPressure - Traverse the back trace from header to the
1129 /// current block and update their register pressures to reflect the effect
1130 /// of hoisting MI from the current block to the preheader.
1131 void MachineLICM::UpdateBackTraceRegPressure(const MachineInstr *MI) {
1132 // First compute the 'cost' of the instruction, i.e. its contribution
1133 // to register pressure.
1134 auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/false,
1135 /*ConsiderUnseenAsDef=*/false);
1137 // Update register pressure of blocks from loop header to current block.
1138 for (auto &RP : BackTrace)
1139 for (const auto &RPIdAndCost : Cost)
1140 RP[RPIdAndCost.first] += RPIdAndCost.second;
1143 /// IsProfitableToHoist - Return true if it is potentially profitable to hoist
1144 /// the given loop invariant.
1145 bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
1146 if (MI.isImplicitDef())
1149 // Besides removing computation from the loop, hoisting an instruction has
1152 // - The value defined by the instruction becomes live across the entire
1153 // loop. This increases register pressure in the loop.
1155 // - If the value is used by a PHI in the loop, a copy will be required for
1156 // lowering the PHI after extending the live range.
1158 // - When hoisting the last use of a value in the loop, that value no longer
1159 // needs to be live in the loop. This lowers register pressure in the loop.
1161 bool CheapInstr = IsCheapInstruction(MI);
1162 bool CreatesCopy = HasLoopPHIUse(&MI);
1164 // Don't hoist a cheap instruction if it would create a copy in the loop.
1165 if (CheapInstr && CreatesCopy) {
1166 DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI);
1170 // Rematerializable instructions should always be hoisted since the register
1171 // allocator can just pull them down again when needed.
1172 if (TII->isTriviallyReMaterializable(&MI, AA))
1175 // FIXME: If there are long latency loop-invariant instructions inside the
1176 // loop at this point, why didn't the optimizer's LICM hoist them?
1177 for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
1178 const MachineOperand &MO = MI.getOperand(i);
1179 if (!MO.isReg() || MO.isImplicit())
1181 unsigned Reg = MO.getReg();
1182 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1184 if (MO.isDef() && HasHighOperandLatency(MI, i, Reg)) {
1185 DEBUG(dbgs() << "Hoist High Latency: " << MI);
1191 // Estimate register pressure to determine whether to LICM the instruction.
1192 // In low register pressure situation, we can be more aggressive about
1193 // hoisting. Also, favors hoisting long latency instructions even in
1194 // moderately high pressure situation.
1195 // Cheap instructions will only be hoisted if they don't increase register
1197 auto Cost = calcRegisterCost(&MI, /*ConsiderSeen=*/false,
1198 /*ConsiderUnseenAsDef=*/false);
1200 // Visit BBs from header to current BB, if hoisting this doesn't cause
1201 // high register pressure, then it's safe to proceed.
1202 if (!CanCauseHighRegPressure(Cost, CheapInstr)) {
1203 DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI);
1208 // Don't risk increasing register pressure if it would create copies.
1210 DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI);
1214 // Do not "speculate" in high register pressure situation. If an
1215 // instruction is not guaranteed to be executed in the loop, it's best to be
1217 if (AvoidSpeculation &&
1218 (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI))) {
1219 DEBUG(dbgs() << "Won't speculate: " << MI);
1223 // High register pressure situation, only hoist if the instruction is going
1225 if (!TII->isTriviallyReMaterializable(&MI, AA) &&
1226 !MI.isInvariantLoad(AA)) {
1227 DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI);
1234 MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) {
1235 // Don't unfold simple loads.
1236 if (MI->canFoldAsLoad())
1239 // If not, we may be able to unfold a load and hoist that.
1240 // First test whether the instruction is loading from an amenable
1242 if (!MI->isInvariantLoad(AA))
1245 // Next determine the register class for a temporary register.
1246 unsigned LoadRegIndex;
1248 TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(),
1249 /*UnfoldLoad=*/true,
1250 /*UnfoldStore=*/false,
1252 if (NewOpc == 0) return nullptr;
1253 const MCInstrDesc &MID = TII->get(NewOpc);
1254 if (MID.getNumDefs() != 1) return nullptr;
1255 MachineFunction &MF = *MI->getParent()->getParent();
1256 const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI, MF);
1257 // Ok, we're unfolding. Create a temporary register and do the unfold.
1258 unsigned Reg = MRI->createVirtualRegister(RC);
1260 SmallVector<MachineInstr *, 2> NewMIs;
1262 TII->unfoldMemoryOperand(MF, MI, Reg,
1263 /*UnfoldLoad=*/true, /*UnfoldStore=*/false,
1267 "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold "
1269 assert(NewMIs.size() == 2 &&
1270 "Unfolded a load into multiple instructions!");
1271 MachineBasicBlock *MBB = MI->getParent();
1272 MachineBasicBlock::iterator Pos = MI;
1273 MBB->insert(Pos, NewMIs[0]);
1274 MBB->insert(Pos, NewMIs[1]);
1275 // If unfolding produced a load that wasn't loop-invariant or profitable to
1276 // hoist, discard the new instructions and bail.
1277 if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) {
1278 NewMIs[0]->eraseFromParent();
1279 NewMIs[1]->eraseFromParent();
1283 // Update register pressure for the unfolded instruction.
1284 UpdateRegPressure(NewMIs[1]);
1286 // Otherwise we successfully unfolded a load that we can hoist.
1287 MI->eraseFromParent();
1291 void MachineLICM::InitCSEMap(MachineBasicBlock *BB) {
1292 for (MachineBasicBlock::iterator I = BB->begin(),E = BB->end(); I != E; ++I) {
1293 const MachineInstr *MI = &*I;
1294 unsigned Opcode = MI->getOpcode();
1295 CSEMap[Opcode].push_back(MI);
1300 MachineLICM::LookForDuplicate(const MachineInstr *MI,
1301 std::vector<const MachineInstr*> &PrevMIs) {
1302 for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) {
1303 const MachineInstr *PrevMI = PrevMIs[i];
1304 if (TII->produceSameValue(MI, PrevMI, (PreRegAlloc ? MRI : nullptr)))
1310 bool MachineLICM::EliminateCSE(MachineInstr *MI,
1311 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI) {
1312 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1313 // the undef property onto uses.
1314 if (CI == CSEMap.end() || MI->isImplicitDef())
1317 if (const MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
1318 DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup);
1320 // Replace virtual registers defined by MI by their counterparts defined
1322 SmallVector<unsigned, 2> Defs;
1323 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1324 const MachineOperand &MO = MI->getOperand(i);
1326 // Physical registers may not differ here.
1327 assert((!MO.isReg() || MO.getReg() == 0 ||
1328 !TargetRegisterInfo::isPhysicalRegister(MO.getReg()) ||
1329 MO.getReg() == Dup->getOperand(i).getReg()) &&
1330 "Instructions with different phys regs are not identical!");
1332 if (MO.isReg() && MO.isDef() &&
1333 !TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1337 SmallVector<const TargetRegisterClass*, 2> OrigRCs;
1338 for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1339 unsigned Idx = Defs[i];
1340 unsigned Reg = MI->getOperand(Idx).getReg();
1341 unsigned DupReg = Dup->getOperand(Idx).getReg();
1342 OrigRCs.push_back(MRI->getRegClass(DupReg));
1344 if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) {
1345 // Restore old RCs if more than one defs.
1346 for (unsigned j = 0; j != i; ++j)
1347 MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]);
1352 for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1353 unsigned Idx = Defs[i];
1354 unsigned Reg = MI->getOperand(Idx).getReg();
1355 unsigned DupReg = Dup->getOperand(Idx).getReg();
1356 MRI->replaceRegWith(Reg, DupReg);
1357 MRI->clearKillFlags(DupReg);
1360 MI->eraseFromParent();
1367 /// MayCSE - Return true if the given instruction will be CSE'd if it's
1368 /// hoisted out of the loop.
1369 bool MachineLICM::MayCSE(MachineInstr *MI) {
1370 unsigned Opcode = MI->getOpcode();
1371 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
1372 CI = CSEMap.find(Opcode);
1373 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1374 // the undef property onto uses.
1375 if (CI == CSEMap.end() || MI->isImplicitDef())
1378 return LookForDuplicate(MI, CI->second) != nullptr;
1381 /// Hoist - When an instruction is found to use only loop invariant operands
1382 /// that are safe to hoist, this instruction is called to do the dirty work.
1384 bool MachineLICM::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader) {
1385 // First check whether we should hoist this instruction.
1386 if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) {
1387 // If not, try unfolding a hoistable load.
1388 MI = ExtractHoistableLoad(MI);
1389 if (!MI) return false;
1392 // Now move the instructions to the predecessor, inserting it before any
1393 // terminator instructions.
1395 dbgs() << "Hoisting " << *MI;
1396 if (Preheader->getBasicBlock())
1397 dbgs() << " to MachineBasicBlock "
1398 << Preheader->getName();
1399 if (MI->getParent()->getBasicBlock())
1400 dbgs() << " from MachineBasicBlock "
1401 << MI->getParent()->getName();
1405 // If this is the first instruction being hoisted to the preheader,
1406 // initialize the CSE map with potential common expressions.
1408 InitCSEMap(Preheader);
1409 FirstInLoop = false;
1412 // Look for opportunity to CSE the hoisted instruction.
1413 unsigned Opcode = MI->getOpcode();
1414 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
1415 CI = CSEMap.find(Opcode);
1416 if (!EliminateCSE(MI, CI)) {
1417 // Otherwise, splice the instruction to the preheader.
1418 Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
1420 // Update register pressure for BBs from header to this block.
1421 UpdateBackTraceRegPressure(MI);
1423 // Clear the kill flags of any register this instruction defines,
1424 // since they may need to be live throughout the entire loop
1425 // rather than just live for part of it.
1426 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1427 MachineOperand &MO = MI->getOperand(i);
1428 if (MO.isReg() && MO.isDef() && !MO.isDead())
1429 MRI->clearKillFlags(MO.getReg());
1432 // Add to the CSE map.
1433 if (CI != CSEMap.end())
1434 CI->second.push_back(MI);
1436 CSEMap[Opcode].push_back(MI);
1445 MachineBasicBlock *MachineLICM::getCurPreheader() {
1446 // Determine the block to which to hoist instructions. If we can't find a
1447 // suitable loop predecessor, we can't do any hoisting.
1449 // If we've tried to get a preheader and failed, don't try again.
1450 if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
1453 if (!CurPreheader) {
1454 CurPreheader = CurLoop->getLoopPreheader();
1455 if (!CurPreheader) {
1456 MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
1458 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1462 CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), this);
1463 if (!CurPreheader) {
1464 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1469 return CurPreheader;