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 does not attempt to throttle itself to limit register pressure.
14 // The register allocation phases are expected to perform rematerialization
15 // to recover when register pressure is high.
17 // This pass is not intended to be a replacement or a complete alternative
18 // for the LLVM-IR-level LICM pass. It is only designed to hoist simple
19 // constructs that are not exposed before lowering and instruction selection.
21 //===----------------------------------------------------------------------===//
23 #define DEBUG_TYPE "machine-licm"
24 #include "llvm/CodeGen/Passes.h"
25 #include "llvm/CodeGen/MachineDominators.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineLoopInfo.h"
28 #include "llvm/CodeGen/MachineMemOperand.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/PseudoSourceValue.h"
31 #include "llvm/Target/TargetLowering.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
33 #include "llvm/Target/TargetInstrInfo.h"
34 #include "llvm/Target/TargetInstrItineraries.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/ADT/DenseMap.h"
38 #include "llvm/ADT/SmallSet.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/raw_ostream.h"
46 "Number of machine instructions hoisted out of loops");
48 "Number of instructions hoisted in low reg pressure situation");
49 STATISTIC(NumHighLatency,
50 "Number of high latency instructions hoisted");
52 "Number of hoisted machine instructions CSEed");
53 STATISTIC(NumPostRAHoisted,
54 "Number of machine instructions hoisted out of loops post regalloc");
57 class MachineLICM : public MachineFunctionPass {
60 const TargetMachine *TM;
61 const TargetInstrInfo *TII;
62 const TargetLowering *TLI;
63 const TargetRegisterInfo *TRI;
64 const MachineFrameInfo *MFI;
65 MachineRegisterInfo *MRI;
66 const InstrItineraryData *InstrItins;
68 // Various analyses that we use...
69 AliasAnalysis *AA; // Alias analysis info.
70 MachineLoopInfo *MLI; // Current MachineLoopInfo
71 MachineDominatorTree *DT; // Machine dominator tree for the cur loop
73 // State that is updated as we process loops
74 bool Changed; // True if a loop is changed.
75 bool FirstInLoop; // True if it's the first LICM in the loop.
76 MachineLoop *CurLoop; // The current loop we are working on.
77 MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
79 BitVector AllocatableSet;
81 // Track 'estimated' register pressure.
82 SmallSet<unsigned, 32> RegSeen;
83 SmallVector<unsigned, 8> RegPressure;
85 // Register pressure "limit" per register class. If the pressure
86 // is higher than the limit, then it's considered high.
87 SmallVector<unsigned, 8> RegLimit;
89 // Register pressure on path leading from loop preheader to current BB.
90 SmallVector<SmallVector<unsigned, 8>, 16> BackTrace;
92 // For each opcode, keep a list of potential CSE instructions.
93 DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
96 static char ID; // Pass identification, replacement for typeid
98 MachineFunctionPass(ID), PreRegAlloc(true) {
99 initializeMachineLICMPass(*PassRegistry::getPassRegistry());
102 explicit MachineLICM(bool PreRA) :
103 MachineFunctionPass(ID), PreRegAlloc(PreRA) {
104 initializeMachineLICMPass(*PassRegistry::getPassRegistry());
107 virtual bool runOnMachineFunction(MachineFunction &MF);
109 const char *getPassName() const { return "Machine Instruction LICM"; }
111 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
112 AU.setPreservesCFG();
113 AU.addRequired<MachineLoopInfo>();
114 AU.addRequired<MachineDominatorTree>();
115 AU.addRequired<AliasAnalysis>();
116 AU.addPreserved<MachineLoopInfo>();
117 AU.addPreserved<MachineDominatorTree>();
118 MachineFunctionPass::getAnalysisUsage(AU);
121 virtual void releaseMemory() {
126 for (DenseMap<unsigned,std::vector<const MachineInstr*> >::iterator
127 CI = CSEMap.begin(), CE = CSEMap.end(); CI != CE; ++CI)
133 /// CandidateInfo - Keep track of information about hoisting candidates.
134 struct CandidateInfo {
138 CandidateInfo(MachineInstr *mi, unsigned def, int fi)
139 : MI(mi), Def(def), FI(fi) {}
142 /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
143 /// invariants out to the preheader.
144 void HoistRegionPostRA();
146 /// HoistPostRA - When an instruction is found to only use loop invariant
147 /// operands that is safe to hoist, this instruction is called to do the
149 void HoistPostRA(MachineInstr *MI, unsigned Def);
151 /// ProcessMI - Examine the instruction for potentai LICM candidate. Also
152 /// gather register def and frame object update information.
153 void ProcessMI(MachineInstr *MI, unsigned *PhysRegDefs,
154 SmallSet<int, 32> &StoredFIs,
155 SmallVector<CandidateInfo, 32> &Candidates);
157 /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the
159 void AddToLiveIns(unsigned Reg);
161 /// IsLICMCandidate - Returns true if the instruction may be a suitable
162 /// candidate for LICM. e.g. If the instruction is a call, then it's
163 /// obviously not safe to hoist it.
164 bool IsLICMCandidate(MachineInstr &I);
166 /// IsLoopInvariantInst - Returns true if the instruction is loop
167 /// invariant. I.e., all virtual register operands are defined outside of
168 /// the loop, physical registers aren't accessed (explicitly or implicitly),
169 /// and the instruction is hoistable.
171 bool IsLoopInvariantInst(MachineInstr &I);
173 /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
174 /// and an use in the current loop, return true if the target considered
176 bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
179 bool IsCheapInstruction(MachineInstr &MI) const;
181 /// CanCauseHighRegPressure - Visit BBs from header to current BB,
182 /// check if hoisting an instruction of the given cost matrix can cause high
183 /// register pressure.
184 bool CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost);
186 /// UpdateBackTraceRegPressure - Traverse the back trace from header to
187 /// the current block and update their register pressures to reflect the
188 /// effect of hoisting MI from the current block to the preheader.
189 void UpdateBackTraceRegPressure(const MachineInstr *MI);
191 /// IsProfitableToHoist - Return true if it is potentially profitable to
192 /// hoist the given loop invariant.
193 bool IsProfitableToHoist(MachineInstr &MI);
195 /// HoistRegion - Walk the specified region of the CFG (defined by all
196 /// blocks dominated by the specified block, and that are in the current
197 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
198 /// visit definitions before uses, allowing us to hoist a loop body in one
199 /// pass without iteration.
201 void HoistRegion(MachineDomTreeNode *N, bool IsHeader = false);
203 /// InitRegPressure - Find all virtual register references that are liveout
204 /// of the preheader to initialize the starting "register pressure". Note
205 /// this does not count live through (livein but not used) registers.
206 void InitRegPressure(MachineBasicBlock *BB);
208 /// UpdateRegPressure - Update estimate of register pressure after the
209 /// specified instruction.
210 void UpdateRegPressure(const MachineInstr *MI);
212 /// isLoadFromConstantMemory - Return true if the given instruction is a
213 /// load from constant memory.
214 bool isLoadFromConstantMemory(MachineInstr *MI);
216 /// ExtractHoistableLoad - Unfold a load from the given machineinstr if
217 /// the load itself could be hoisted. Return the unfolded and hoistable
218 /// load, or null if the load couldn't be unfolded or if it wouldn't
220 MachineInstr *ExtractHoistableLoad(MachineInstr *MI);
222 /// LookForDuplicate - Find an instruction amount PrevMIs that is a
223 /// duplicate of MI. Return this instruction if it's found.
224 const MachineInstr *LookForDuplicate(const MachineInstr *MI,
225 std::vector<const MachineInstr*> &PrevMIs);
227 /// EliminateCSE - Given a LICM'ed instruction, look for an instruction on
228 /// the preheader that compute the same value. If it's found, do a RAU on
229 /// with the definition of the existing instruction rather than hoisting
230 /// the instruction to the preheader.
231 bool EliminateCSE(MachineInstr *MI,
232 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI);
234 /// Hoist - When an instruction is found to only use loop invariant operands
235 /// that is safe to hoist, this instruction is called to do the dirty work.
236 /// It returns true if the instruction is hoisted.
237 bool Hoist(MachineInstr *MI, MachineBasicBlock *Preheader);
239 /// InitCSEMap - Initialize the CSE map with instructions that are in the
240 /// current loop preheader that may become duplicates of instructions that
241 /// are hoisted out of the loop.
242 void InitCSEMap(MachineBasicBlock *BB);
244 /// getCurPreheader - Get the preheader for the current loop, splitting
245 /// a critical edge if needed.
246 MachineBasicBlock *getCurPreheader();
248 } // end anonymous namespace
250 char MachineLICM::ID = 0;
251 INITIALIZE_PASS_BEGIN(MachineLICM, "machinelicm",
252 "Machine Loop Invariant Code Motion", false, false)
253 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
254 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
255 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
256 INITIALIZE_PASS_END(MachineLICM, "machinelicm",
257 "Machine Loop Invariant Code Motion", false, false)
259 FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) {
260 return new MachineLICM(PreRegAlloc);
263 /// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most
264 /// loop that has a unique predecessor.
265 static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) {
266 // Check whether this loop even has a unique predecessor.
267 if (!CurLoop->getLoopPredecessor())
269 // Ok, now check to see if any of its outer loops do.
270 for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop())
271 if (L->getLoopPredecessor())
273 // None of them did, so this is the outermost with a unique predecessor.
277 bool MachineLICM::runOnMachineFunction(MachineFunction &MF) {
279 DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
281 DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
282 DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n");
284 Changed = FirstInLoop = false;
285 TM = &MF.getTarget();
286 TII = TM->getInstrInfo();
287 TLI = TM->getTargetLowering();
288 TRI = TM->getRegisterInfo();
289 MFI = MF.getFrameInfo();
290 MRI = &MF.getRegInfo();
291 InstrItins = TM->getInstrItineraryData();
292 AllocatableSet = TRI->getAllocatableSet(MF);
295 // Estimate register pressure during pre-regalloc pass.
296 unsigned NumRC = TRI->getNumRegClasses();
297 RegPressure.resize(NumRC);
298 std::fill(RegPressure.begin(), RegPressure.end(), 0);
299 RegLimit.resize(NumRC);
300 for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
301 E = TRI->regclass_end(); I != E; ++I)
302 RegLimit[(*I)->getID()] = TLI->getRegPressureLimit(*I, MF);
305 // Get our Loop information...
306 MLI = &getAnalysis<MachineLoopInfo>();
307 DT = &getAnalysis<MachineDominatorTree>();
308 AA = &getAnalysis<AliasAnalysis>();
310 SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
311 while (!Worklist.empty()) {
312 CurLoop = Worklist.pop_back_val();
315 // If this is done before regalloc, only visit outer-most preheader-sporting
317 if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) {
318 Worklist.append(CurLoop->begin(), CurLoop->end());
325 // CSEMap is initialized for loop header when the first instruction is
327 MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
329 HoistRegion(N, true);
337 /// InstructionStoresToFI - Return true if instruction stores to the
339 static bool InstructionStoresToFI(const MachineInstr *MI, int FI) {
340 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
341 oe = MI->memoperands_end(); o != oe; ++o) {
342 if (!(*o)->isStore() || !(*o)->getValue())
344 if (const FixedStackPseudoSourceValue *Value =
345 dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) {
346 if (Value->getFrameIndex() == FI)
353 /// ProcessMI - Examine the instruction for potentai LICM candidate. Also
354 /// gather register def and frame object update information.
355 void MachineLICM::ProcessMI(MachineInstr *MI,
356 unsigned *PhysRegDefs,
357 SmallSet<int, 32> &StoredFIs,
358 SmallVector<CandidateInfo, 32> &Candidates) {
359 bool RuledOut = false;
360 bool HasNonInvariantUse = false;
362 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
363 const MachineOperand &MO = MI->getOperand(i);
365 // Remember if the instruction stores to the frame index.
366 int FI = MO.getIndex();
367 if (!StoredFIs.count(FI) &&
368 MFI->isSpillSlotObjectIndex(FI) &&
369 InstructionStoresToFI(MI, FI))
370 StoredFIs.insert(FI);
371 HasNonInvariantUse = true;
377 unsigned Reg = MO.getReg();
380 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
381 "Not expecting virtual register!");
384 if (Reg && PhysRegDefs[Reg])
385 // If it's using a non-loop-invariant register, then it's obviously not
387 HasNonInvariantUse = true;
391 if (MO.isImplicit()) {
393 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
396 // Non-dead implicit def? This cannot be hoisted.
398 // No need to check if a dead implicit def is also defined by
399 // another instruction.
403 // FIXME: For now, avoid instructions with multiple defs, unless
404 // it's a dead implicit def.
410 // If we have already seen another instruction that defines the same
411 // register, then this is not safe.
412 if (++PhysRegDefs[Reg] > 1)
413 // MI defined register is seen defined by another instruction in
414 // the loop, it cannot be a LICM candidate.
416 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
417 if (++PhysRegDefs[*AS] > 1)
421 // Only consider reloads for now and remats which do not have register
422 // operands. FIXME: Consider unfold load folding instructions.
423 if (Def && !RuledOut) {
425 if ((!HasNonInvariantUse && IsLICMCandidate(*MI)) ||
426 (TII->isLoadFromStackSlot(MI, FI) && MFI->isSpillSlotObjectIndex(FI)))
427 Candidates.push_back(CandidateInfo(MI, Def, FI));
431 /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
432 /// invariants out to the preheader.
433 void MachineLICM::HoistRegionPostRA() {
434 unsigned NumRegs = TRI->getNumRegs();
435 unsigned *PhysRegDefs = new unsigned[NumRegs];
436 std::fill(PhysRegDefs, PhysRegDefs + NumRegs, 0);
438 SmallVector<CandidateInfo, 32> Candidates;
439 SmallSet<int, 32> StoredFIs;
441 // Walk the entire region, count number of defs for each register, and
442 // collect potential LICM candidates.
443 const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
444 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
445 MachineBasicBlock *BB = Blocks[i];
446 // Conservatively treat live-in's as an external def.
447 // FIXME: That means a reload that're reused in successor block(s) will not
449 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
450 E = BB->livein_end(); I != E; ++I) {
453 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
457 for (MachineBasicBlock::iterator
458 MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
459 MachineInstr *MI = &*MII;
460 ProcessMI(MI, PhysRegDefs, StoredFIs, Candidates);
464 // Now evaluate whether the potential candidates qualify.
465 // 1. Check if the candidate defined register is defined by another
466 // instruction in the loop.
467 // 2. If the candidate is a load from stack slot (always true for now),
468 // check if the slot is stored anywhere in the loop.
469 for (unsigned i = 0, e = Candidates.size(); i != e; ++i) {
470 if (Candidates[i].FI != INT_MIN &&
471 StoredFIs.count(Candidates[i].FI))
474 if (PhysRegDefs[Candidates[i].Def] == 1) {
476 MachineInstr *MI = Candidates[i].MI;
477 for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) {
478 const MachineOperand &MO = MI->getOperand(j);
479 if (!MO.isReg() || MO.isDef() || !MO.getReg())
481 if (PhysRegDefs[MO.getReg()]) {
482 // If it's using a non-loop-invariant register, then it's obviously
483 // not safe to hoist.
489 HoistPostRA(MI, Candidates[i].Def);
493 delete[] PhysRegDefs;
496 /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the current
497 /// loop, and make sure it is not killed by any instructions in the loop.
498 void MachineLICM::AddToLiveIns(unsigned Reg) {
499 const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
500 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
501 MachineBasicBlock *BB = Blocks[i];
502 if (!BB->isLiveIn(Reg))
504 for (MachineBasicBlock::iterator
505 MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
506 MachineInstr *MI = &*MII;
507 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
508 MachineOperand &MO = MI->getOperand(i);
509 if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue;
510 if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg()))
517 /// HoistPostRA - When an instruction is found to only use loop invariant
518 /// operands that is safe to hoist, this instruction is called to do the
520 void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) {
521 MachineBasicBlock *Preheader = getCurPreheader();
522 if (!Preheader) return;
524 // Now move the instructions to the predecessor, inserting it before any
525 // terminator instructions.
527 dbgs() << "Hoisting " << *MI;
528 if (Preheader->getBasicBlock())
529 dbgs() << " to MachineBasicBlock "
530 << Preheader->getName();
531 if (MI->getParent()->getBasicBlock())
532 dbgs() << " from MachineBasicBlock "
533 << MI->getParent()->getName();
537 // Splice the instruction to the preheader.
538 MachineBasicBlock *MBB = MI->getParent();
539 Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
541 // Add register to livein list to all the BBs in the current loop since a
542 // loop invariant must be kept live throughout the whole loop. This is
543 // important to ensure later passes do not scavenge the def register.
550 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
551 /// dominated by the specified block, and that are in the current loop) in depth
552 /// first order w.r.t the DominatorTree. This allows us to visit definitions
553 /// before uses, allowing us to hoist a loop body in one pass without iteration.
555 void MachineLICM::HoistRegion(MachineDomTreeNode *N, bool IsHeader) {
556 assert(N != 0 && "Null dominator tree node?");
557 MachineBasicBlock *BB = N->getBlock();
559 // If this subregion is not in the top level loop at all, exit.
560 if (!CurLoop->contains(BB)) return;
562 MachineBasicBlock *Preheader = getCurPreheader();
567 // Compute registers which are livein into the loop headers.
570 InitRegPressure(Preheader);
573 // Remember livein register pressure.
574 BackTrace.push_back(RegPressure);
576 for (MachineBasicBlock::iterator
577 MII = BB->begin(), E = BB->end(); MII != E; ) {
578 MachineBasicBlock::iterator NextMII = MII; ++NextMII;
579 MachineInstr *MI = &*MII;
580 if (!Hoist(MI, Preheader))
581 UpdateRegPressure(MI);
585 // Don't hoist things out of a large switch statement. This often causes
586 // code to be hoisted that wasn't going to be executed, and increases
587 // register pressure in a situation where it's likely to matter.
588 if (BB->succ_size() < 25) {
589 const std::vector<MachineDomTreeNode*> &Children = N->getChildren();
590 for (unsigned I = 0, E = Children.size(); I != E; ++I)
591 HoistRegion(Children[I]);
594 BackTrace.pop_back();
597 static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
598 return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
601 /// InitRegPressure - Find all virtual register references that are liveout of
602 /// the preheader to initialize the starting "register pressure". Note this
603 /// does not count live through (livein but not used) registers.
604 void MachineLICM::InitRegPressure(MachineBasicBlock *BB) {
605 std::fill(RegPressure.begin(), RegPressure.end(), 0);
607 // If the preheader has only a single predecessor and it ends with a
608 // fallthrough or an unconditional branch, then scan its predecessor for live
609 // defs as well. This happens whenever the preheader is created by splitting
610 // the critical edge from the loop predecessor to the loop header.
611 if (BB->pred_size() == 1) {
612 MachineBasicBlock *TBB = 0, *FBB = 0;
613 SmallVector<MachineOperand, 4> Cond;
614 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty())
615 InitRegPressure(*BB->pred_begin());
618 for (MachineBasicBlock::iterator MII = BB->begin(), E = BB->end();
620 MachineInstr *MI = &*MII;
621 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
622 const MachineOperand &MO = MI->getOperand(i);
623 if (!MO.isReg() || MO.isImplicit())
625 unsigned Reg = MO.getReg();
626 if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg))
629 bool isNew = RegSeen.insert(Reg);
630 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
631 EVT VT = *RC->vt_begin();
632 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
634 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
636 bool isKill = isOperandKill(MO, MRI);
637 if (isNew && !isKill)
638 // Haven't seen this, it must be a livein.
639 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
640 else if (!isNew && isKill)
641 RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
647 /// UpdateRegPressure - Update estimate of register pressure after the
648 /// specified instruction.
649 void MachineLICM::UpdateRegPressure(const MachineInstr *MI) {
650 if (MI->isImplicitDef())
653 SmallVector<unsigned, 4> Defs;
654 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
655 const MachineOperand &MO = MI->getOperand(i);
656 if (!MO.isReg() || MO.isImplicit())
658 unsigned Reg = MO.getReg();
659 if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg))
662 bool isNew = RegSeen.insert(Reg);
665 else if (!isNew && isOperandKill(MO, MRI)) {
666 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
667 EVT VT = *RC->vt_begin();
668 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
669 unsigned RCCost = TLI->getRepRegClassCostFor(VT);
671 if (RCCost > RegPressure[RCId])
672 RegPressure[RCId] = 0;
674 RegPressure[RCId] -= RCCost;
678 while (!Defs.empty()) {
679 unsigned Reg = Defs.pop_back_val();
680 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
681 EVT VT = *RC->vt_begin();
682 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
683 unsigned RCCost = TLI->getRepRegClassCostFor(VT);
684 RegPressure[RCId] += RCCost;
688 /// IsLICMCandidate - Returns true if the instruction may be a suitable
689 /// candidate for LICM. e.g. If the instruction is a call, then it's obviously
690 /// not safe to hoist it.
691 bool MachineLICM::IsLICMCandidate(MachineInstr &I) {
692 // Check if it's safe to move the instruction.
693 bool DontMoveAcrossStore = true;
694 if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore))
700 /// IsLoopInvariantInst - Returns true if the instruction is loop
701 /// invariant. I.e., all virtual register operands are defined outside of the
702 /// loop, physical registers aren't accessed explicitly, and there are no side
703 /// effects that aren't captured by the operands or other flags.
705 bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) {
706 if (!IsLICMCandidate(I))
709 // The instruction is loop invariant if all of its operands are.
710 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
711 const MachineOperand &MO = I.getOperand(i);
716 unsigned Reg = MO.getReg();
717 if (Reg == 0) continue;
719 // Don't hoist an instruction that uses or defines a physical register.
720 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
722 // If the physreg has no defs anywhere, it's just an ambient register
723 // and we can freely move its uses. Alternatively, if it's allocatable,
724 // it could get allocated to something with a def during allocation.
725 if (!MRI->def_empty(Reg))
727 if (AllocatableSet.test(Reg))
729 // Check for a def among the register's aliases too.
730 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
731 unsigned AliasReg = *Alias;
732 if (!MRI->def_empty(AliasReg))
734 if (AllocatableSet.test(AliasReg))
737 // Otherwise it's safe to move.
739 } else if (!MO.isDead()) {
740 // A def that isn't dead. We can't move it.
742 } else if (CurLoop->getHeader()->isLiveIn(Reg)) {
743 // If the reg is live into the loop, we can't hoist an instruction
744 // which would clobber it.
752 assert(MRI->getVRegDef(Reg) &&
753 "Machine instr not mapped for this vreg?!");
755 // If the loop contains the definition of an operand, then the instruction
756 // isn't loop invariant.
757 if (CurLoop->contains(MRI->getVRegDef(Reg)))
761 // If we got this far, the instruction is loop invariant!
766 /// HasPHIUses - Return true if the specified register has any PHI use.
767 static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *MRI) {
768 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
769 UE = MRI->use_end(); UI != UE; ++UI) {
770 MachineInstr *UseMI = &*UI;
777 /// isLoadFromConstantMemory - Return true if the given instruction is a
778 /// load from constant memory. Machine LICM will hoist these even if they are
779 /// not re-materializable.
780 bool MachineLICM::isLoadFromConstantMemory(MachineInstr *MI) {
781 if (!MI->getDesc().mayLoad()) return false;
782 if (!MI->hasOneMemOperand()) return false;
783 MachineMemOperand *MMO = *MI->memoperands_begin();
784 if (MMO->isVolatile()) return false;
785 if (!MMO->getValue()) return false;
786 const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(MMO->getValue());
788 MachineFunction &MF = *MI->getParent()->getParent();
789 return PSV->isConstant(MF.getFrameInfo());
791 return AA->pointsToConstantMemory(AliasAnalysis::Location(MMO->getValue(),
793 MMO->getTBAAInfo()));
797 /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
798 /// and an use in the current loop, return true if the target considered
800 bool MachineLICM::HasHighOperandLatency(MachineInstr &MI,
801 unsigned DefIdx, unsigned Reg) const {
802 if (!InstrItins || InstrItins->isEmpty() || MRI->use_nodbg_empty(Reg))
805 for (MachineRegisterInfo::use_nodbg_iterator I = MRI->use_nodbg_begin(Reg),
806 E = MRI->use_nodbg_end(); I != E; ++I) {
807 MachineInstr *UseMI = &*I;
808 if (UseMI->isCopyLike())
810 if (!CurLoop->contains(UseMI->getParent()))
812 for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
813 const MachineOperand &MO = UseMI->getOperand(i);
814 if (!MO.isReg() || !MO.isUse())
816 unsigned MOReg = MO.getReg();
820 if (TII->hasHighOperandLatency(InstrItins, MRI, &MI, DefIdx, UseMI, i))
824 // Only look at the first in loop use.
831 /// IsCheapInstruction - Return true if the instruction is marked "cheap" or
832 /// the operand latency between its def and a use is one or less.
833 bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const {
834 if (MI.getDesc().isAsCheapAsAMove() || MI.isCopyLike())
836 if (!InstrItins || InstrItins->isEmpty())
839 bool isCheap = false;
840 unsigned NumDefs = MI.getDesc().getNumDefs();
841 for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
842 MachineOperand &DefMO = MI.getOperand(i);
843 if (!DefMO.isReg() || !DefMO.isDef())
846 unsigned Reg = DefMO.getReg();
847 if (TargetRegisterInfo::isPhysicalRegister(Reg))
850 if (!TII->hasLowDefLatency(InstrItins, &MI, i))
858 /// CanCauseHighRegPressure - Visit BBs from header to current BB, check
859 /// if hoisting an instruction of the given cost matrix can cause high
860 /// register pressure.
861 bool MachineLICM::CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost) {
862 for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
867 unsigned RCId = CI->first;
868 for (unsigned i = BackTrace.size(); i != 0; --i) {
869 SmallVector<unsigned, 8> &RP = BackTrace[i-1];
870 if (RP[RCId] + CI->second >= RegLimit[RCId])
878 /// UpdateBackTraceRegPressure - Traverse the back trace from header to the
879 /// current block and update their register pressures to reflect the effect
880 /// of hoisting MI from the current block to the preheader.
881 void MachineLICM::UpdateBackTraceRegPressure(const MachineInstr *MI) {
882 if (MI->isImplicitDef())
885 // First compute the 'cost' of the instruction, i.e. its contribution
886 // to register pressure.
887 DenseMap<unsigned, int> Cost;
888 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
889 const MachineOperand &MO = MI->getOperand(i);
890 if (!MO.isReg() || MO.isImplicit())
892 unsigned Reg = MO.getReg();
893 if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg))
896 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
897 EVT VT = *RC->vt_begin();
898 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
899 unsigned RCCost = TLI->getRepRegClassCostFor(VT);
901 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
902 if (CI != Cost.end())
903 CI->second += RCCost;
905 Cost.insert(std::make_pair(RCId, RCCost));
906 } else if (isOperandKill(MO, MRI)) {
907 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
908 if (CI != Cost.end())
909 CI->second -= RCCost;
911 Cost.insert(std::make_pair(RCId, -RCCost));
915 // Update register pressure of blocks from loop header to current block.
916 for (unsigned i = 0, e = BackTrace.size(); i != e; ++i) {
917 SmallVector<unsigned, 8> &RP = BackTrace[i];
918 for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
920 unsigned RCId = CI->first;
921 RP[RCId] += CI->second;
926 /// IsProfitableToHoist - Return true if it is potentially profitable to hoist
927 /// the given loop invariant.
928 bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
929 if (MI.isImplicitDef())
932 // If the instruction is cheap, only hoist if it is re-materilizable. LICM
933 // will increase register pressure. It's probably not worth it if the
934 // instruction is cheap.
935 // Also hoist loads from constant memory, e.g. load from stubs, GOT. Hoisting
936 // these tend to help performance in low register pressure situation. The
937 // trade off is it may cause spill in high pressure situation. It will end up
938 // adding a store in the loop preheader. But the reload is no more expensive.
939 // The side benefit is these loads are frequently CSE'ed.
940 if (IsCheapInstruction(MI)) {
941 if (!TII->isTriviallyReMaterializable(&MI, AA))
944 // Estimate register pressure to determine whether to LICM the instruction.
945 // In low register pressure situation, we can be more aggressive about
946 // hoisting. Also, favors hoisting long latency instructions even in
947 // moderately high pressure situation.
948 DenseMap<unsigned, int> Cost;
949 for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
950 const MachineOperand &MO = MI.getOperand(i);
951 if (!MO.isReg() || MO.isImplicit())
953 unsigned Reg = MO.getReg();
954 if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg))
957 if (HasHighOperandLatency(MI, i, Reg)) {
962 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
963 EVT VT = *RC->vt_begin();
964 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
965 unsigned RCCost = TLI->getRepRegClassCostFor(VT);
966 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
967 if (CI != Cost.end())
968 CI->second += RCCost;
970 Cost.insert(std::make_pair(RCId, RCCost));
971 } else if (isOperandKill(MO, MRI)) {
972 // Is a virtual register use is a kill, hoisting it out of the loop
973 // may actually reduce register pressure or be register pressure
975 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
976 EVT VT = *RC->vt_begin();
977 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
978 unsigned RCCost = TLI->getRepRegClassCostFor(VT);
979 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
980 if (CI != Cost.end())
981 CI->second -= RCCost;
983 Cost.insert(std::make_pair(RCId, -RCCost));
987 // Visit BBs from header to current BB, if hoisting this doesn't cause
988 // high register pressure, then it's safe to proceed.
989 if (!CanCauseHighRegPressure(Cost)) {
994 // High register pressure situation, only hoist if the instruction is going to
996 if (!TII->isTriviallyReMaterializable(&MI, AA) &&
997 !isLoadFromConstantMemory(&MI))
1001 // If result(s) of this instruction is used by PHIs, then don't hoist it.
1002 // The presence of joins makes it difficult for current register allocator
1003 // implementation to perform remat.
1004 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1005 const MachineOperand &MO = MI.getOperand(i);
1006 if (!MO.isReg() || !MO.isDef())
1008 if (HasPHIUses(MO.getReg(), MRI))
1015 MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) {
1016 // Don't unfold simple loads.
1017 if (MI->getDesc().canFoldAsLoad())
1020 // If not, we may be able to unfold a load and hoist that.
1021 // First test whether the instruction is loading from an amenable
1023 if (!isLoadFromConstantMemory(MI))
1026 // Next determine the register class for a temporary register.
1027 unsigned LoadRegIndex;
1029 TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(),
1030 /*UnfoldLoad=*/true,
1031 /*UnfoldStore=*/false,
1033 if (NewOpc == 0) return 0;
1034 const TargetInstrDesc &TID = TII->get(NewOpc);
1035 if (TID.getNumDefs() != 1) return 0;
1036 const TargetRegisterClass *RC = TID.OpInfo[LoadRegIndex].getRegClass(TRI);
1037 // Ok, we're unfolding. Create a temporary register and do the unfold.
1038 unsigned Reg = MRI->createVirtualRegister(RC);
1040 MachineFunction &MF = *MI->getParent()->getParent();
1041 SmallVector<MachineInstr *, 2> NewMIs;
1043 TII->unfoldMemoryOperand(MF, MI, Reg,
1044 /*UnfoldLoad=*/true, /*UnfoldStore=*/false,
1048 "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold "
1050 assert(NewMIs.size() == 2 &&
1051 "Unfolded a load into multiple instructions!");
1052 MachineBasicBlock *MBB = MI->getParent();
1053 MBB->insert(MI, NewMIs[0]);
1054 MBB->insert(MI, NewMIs[1]);
1055 // If unfolding produced a load that wasn't loop-invariant or profitable to
1056 // hoist, discard the new instructions and bail.
1057 if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) {
1058 NewMIs[0]->eraseFromParent();
1059 NewMIs[1]->eraseFromParent();
1063 // Update register pressure for the unfolded instruction.
1064 UpdateRegPressure(NewMIs[1]);
1066 // Otherwise we successfully unfolded a load that we can hoist.
1067 MI->eraseFromParent();
1071 void MachineLICM::InitCSEMap(MachineBasicBlock *BB) {
1072 for (MachineBasicBlock::iterator I = BB->begin(),E = BB->end(); I != E; ++I) {
1073 const MachineInstr *MI = &*I;
1074 // FIXME: For now, only hoist re-materilizable instructions. LICM will
1075 // increase register pressure. We want to make sure it doesn't increase
1077 if (TII->isTriviallyReMaterializable(MI, AA)) {
1078 unsigned Opcode = MI->getOpcode();
1079 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
1080 CI = CSEMap.find(Opcode);
1081 if (CI != CSEMap.end())
1082 CI->second.push_back(MI);
1084 std::vector<const MachineInstr*> CSEMIs;
1085 CSEMIs.push_back(MI);
1086 CSEMap.insert(std::make_pair(Opcode, CSEMIs));
1093 MachineLICM::LookForDuplicate(const MachineInstr *MI,
1094 std::vector<const MachineInstr*> &PrevMIs) {
1095 for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) {
1096 const MachineInstr *PrevMI = PrevMIs[i];
1097 if (TII->produceSameValue(MI, PrevMI))
1103 bool MachineLICM::EliminateCSE(MachineInstr *MI,
1104 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI) {
1105 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1106 // the undef property onto uses.
1107 if (CI == CSEMap.end() || MI->isImplicitDef())
1110 if (const MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
1111 DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup);
1113 // Replace virtual registers defined by MI by their counterparts defined
1115 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1116 const MachineOperand &MO = MI->getOperand(i);
1118 // Physical registers may not differ here.
1119 assert((!MO.isReg() || MO.getReg() == 0 ||
1120 !TargetRegisterInfo::isPhysicalRegister(MO.getReg()) ||
1121 MO.getReg() == Dup->getOperand(i).getReg()) &&
1122 "Instructions with different phys regs are not identical!");
1124 if (MO.isReg() && MO.isDef() &&
1125 !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
1126 MRI->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg());
1127 MRI->clearKillFlags(Dup->getOperand(i).getReg());
1130 MI->eraseFromParent();
1137 /// Hoist - When an instruction is found to use only loop invariant operands
1138 /// that are safe to hoist, this instruction is called to do the dirty work.
1140 bool MachineLICM::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader) {
1141 // First check whether we should hoist this instruction.
1142 if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) {
1143 // If not, try unfolding a hoistable load.
1144 MI = ExtractHoistableLoad(MI);
1145 if (!MI) return false;
1148 // Now move the instructions to the predecessor, inserting it before any
1149 // terminator instructions.
1151 dbgs() << "Hoisting " << *MI;
1152 if (Preheader->getBasicBlock())
1153 dbgs() << " to MachineBasicBlock "
1154 << Preheader->getName();
1155 if (MI->getParent()->getBasicBlock())
1156 dbgs() << " from MachineBasicBlock "
1157 << MI->getParent()->getName();
1161 // If this is the first instruction being hoisted to the preheader,
1162 // initialize the CSE map with potential common expressions.
1164 InitCSEMap(Preheader);
1165 FirstInLoop = false;
1168 // Look for opportunity to CSE the hoisted instruction.
1169 unsigned Opcode = MI->getOpcode();
1170 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
1171 CI = CSEMap.find(Opcode);
1172 if (!EliminateCSE(MI, CI)) {
1173 // Otherwise, splice the instruction to the preheader.
1174 Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
1176 // Update register pressure for BBs from header to this block.
1177 UpdateBackTraceRegPressure(MI);
1179 // Clear the kill flags of any register this instruction defines,
1180 // since they may need to be live throughout the entire loop
1181 // rather than just live for part of it.
1182 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1183 MachineOperand &MO = MI->getOperand(i);
1184 if (MO.isReg() && MO.isDef() && !MO.isDead())
1185 MRI->clearKillFlags(MO.getReg());
1188 // Add to the CSE map.
1189 if (CI != CSEMap.end())
1190 CI->second.push_back(MI);
1192 std::vector<const MachineInstr*> CSEMIs;
1193 CSEMIs.push_back(MI);
1194 CSEMap.insert(std::make_pair(Opcode, CSEMIs));
1204 MachineBasicBlock *MachineLICM::getCurPreheader() {
1205 // Determine the block to which to hoist instructions. If we can't find a
1206 // suitable loop predecessor, we can't do any hoisting.
1208 // If we've tried to get a preheader and failed, don't try again.
1209 if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
1212 if (!CurPreheader) {
1213 CurPreheader = CurLoop->getLoopPreheader();
1214 if (!CurPreheader) {
1215 MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
1217 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1221 CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), this);
1222 if (!CurPreheader) {
1223 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1228 return CurPreheader;