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/MC/MCInstrItineraries.h"
32 #include "llvm/Target/TargetLowering.h"
33 #include "llvm/Target/TargetRegisterInfo.h"
34 #include "llvm/Target/TargetInstrInfo.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/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
46 AvoidSpeculation("avoid-speculation",
47 cl::desc("MachineLICM should avoid speculation"),
48 cl::init(true), cl::Hidden);
51 "Number of machine instructions hoisted out of loops");
53 "Number of instructions hoisted in low reg pressure situation");
54 STATISTIC(NumHighLatency,
55 "Number of high latency instructions hoisted");
57 "Number of hoisted machine instructions CSEed");
58 STATISTIC(NumPostRAHoisted,
59 "Number of machine instructions hoisted out of loops post regalloc");
62 class MachineLICM : public MachineFunctionPass {
65 const TargetMachine *TM;
66 const TargetInstrInfo *TII;
67 const TargetLowering *TLI;
68 const TargetRegisterInfo *TRI;
69 const MachineFrameInfo *MFI;
70 MachineRegisterInfo *MRI;
71 const InstrItineraryData *InstrItins;
73 // Various analyses that we use...
74 AliasAnalysis *AA; // Alias analysis info.
75 MachineLoopInfo *MLI; // Current MachineLoopInfo
76 MachineDominatorTree *DT; // Machine dominator tree for the cur loop
78 // State that is updated as we process loops
79 bool Changed; // True if a loop is changed.
80 bool FirstInLoop; // True if it's the first LICM in the loop.
81 MachineLoop *CurLoop; // The current loop we are working on.
82 MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
84 // Track 'estimated' register pressure.
85 SmallSet<unsigned, 32> RegSeen;
86 SmallVector<unsigned, 8> RegPressure;
88 // Register pressure "limit" per register class. If the pressure
89 // is higher than the limit, then it's considered high.
90 SmallVector<unsigned, 8> RegLimit;
92 // Register pressure on path leading from loop preheader to current BB.
93 SmallVector<SmallVector<unsigned, 8>, 16> BackTrace;
95 // For each opcode, keep a list of potential CSE instructions.
96 DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
104 // If a MBB does not dominate loop exiting blocks then it may not safe
105 // to hoist loads from this block.
106 // Tri-state: 0 - false, 1 - true, 2 - unknown
107 unsigned SpeculationState;
110 static char ID; // Pass identification, replacement for typeid
112 MachineFunctionPass(ID), PreRegAlloc(true) {
113 initializeMachineLICMPass(*PassRegistry::getPassRegistry());
116 explicit MachineLICM(bool PreRA) :
117 MachineFunctionPass(ID), PreRegAlloc(PreRA) {
118 initializeMachineLICMPass(*PassRegistry::getPassRegistry());
121 virtual bool runOnMachineFunction(MachineFunction &MF);
123 const char *getPassName() const { return "Machine Instruction LICM"; }
125 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
126 AU.addRequired<MachineLoopInfo>();
127 AU.addRequired<MachineDominatorTree>();
128 AU.addRequired<AliasAnalysis>();
129 AU.addPreserved<MachineLoopInfo>();
130 AU.addPreserved<MachineDominatorTree>();
131 MachineFunctionPass::getAnalysisUsage(AU);
134 virtual void releaseMemory() {
139 for (DenseMap<unsigned,std::vector<const MachineInstr*> >::iterator
140 CI = CSEMap.begin(), CE = CSEMap.end(); CI != CE; ++CI)
146 /// CandidateInfo - Keep track of information about hoisting candidates.
147 struct CandidateInfo {
151 CandidateInfo(MachineInstr *mi, unsigned def, int fi)
152 : MI(mi), Def(def), FI(fi) {}
155 /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
156 /// invariants out to the preheader.
157 void HoistRegionPostRA();
159 /// HoistPostRA - When an instruction is found to only use loop invariant
160 /// operands that is safe to hoist, this instruction is called to do the
162 void HoistPostRA(MachineInstr *MI, unsigned Def);
164 /// ProcessMI - Examine the instruction for potentai LICM candidate. Also
165 /// gather register def and frame object update information.
166 void ProcessMI(MachineInstr *MI, unsigned *PhysRegDefs,
167 SmallSet<int, 32> &StoredFIs,
168 SmallVector<CandidateInfo, 32> &Candidates);
170 /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the
172 void AddToLiveIns(unsigned Reg);
174 /// IsLICMCandidate - Returns true if the instruction may be a suitable
175 /// candidate for LICM. e.g. If the instruction is a call, then it's
176 /// obviously not safe to hoist it.
177 bool IsLICMCandidate(MachineInstr &I);
179 /// IsLoopInvariantInst - Returns true if the instruction is loop
180 /// invariant. I.e., all virtual register operands are defined outside of
181 /// the loop, physical registers aren't accessed (explicitly or implicitly),
182 /// and the instruction is hoistable.
184 bool IsLoopInvariantInst(MachineInstr &I);
186 /// HasAnyPHIUse - Return true if the specified register is used by any
188 bool HasAnyPHIUse(unsigned Reg) const;
190 /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
191 /// and an use in the current loop, return true if the target considered
193 bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
196 bool IsCheapInstruction(MachineInstr &MI) const;
198 /// CanCauseHighRegPressure - Visit BBs from header to current BB,
199 /// check if hoisting an instruction of the given cost matrix can cause high
200 /// register pressure.
201 bool CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost);
203 /// UpdateBackTraceRegPressure - Traverse the back trace from header to
204 /// the current block and update their register pressures to reflect the
205 /// effect of hoisting MI from the current block to the preheader.
206 void UpdateBackTraceRegPressure(const MachineInstr *MI);
208 /// IsProfitableToHoist - Return true if it is potentially profitable to
209 /// hoist the given loop invariant.
210 bool IsProfitableToHoist(MachineInstr &MI);
212 /// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
213 /// If not then a load from this mbb may not be safe to hoist.
214 bool IsGuaranteedToExecute(MachineBasicBlock *BB);
216 void EnterScope(MachineBasicBlock *MBB);
218 void ExitScope(MachineBasicBlock *MBB);
220 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to given
221 /// dominator tree node if its a leaf or all of its children are done. Walk
222 /// up the dominator tree to destroy ancestors which are now done.
223 void ExitScopeIfDone(MachineDomTreeNode *Node,
224 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
225 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap);
227 /// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all
228 /// blocks dominated by the specified header block, and that are in the
229 /// current loop) in depth first order w.r.t the DominatorTree. This allows
230 /// us to visit definitions before uses, allowing us to hoist a loop body in
231 /// one pass without iteration.
233 void HoistOutOfLoop(MachineDomTreeNode *LoopHeaderNode);
234 void HoistRegion(MachineDomTreeNode *N, bool IsHeader);
236 /// getRegisterClassIDAndCost - For a given MI, register, and the operand
237 /// index, return the ID and cost of its representative register class by
239 void getRegisterClassIDAndCost(const MachineInstr *MI,
240 unsigned Reg, unsigned OpIdx,
241 unsigned &RCId, unsigned &RCCost) const;
243 /// InitRegPressure - Find all virtual register references that are liveout
244 /// of the preheader to initialize the starting "register pressure". Note
245 /// this does not count live through (livein but not used) registers.
246 void InitRegPressure(MachineBasicBlock *BB);
248 /// UpdateRegPressure - Update estimate of register pressure after the
249 /// specified instruction.
250 void UpdateRegPressure(const MachineInstr *MI);
252 /// ExtractHoistableLoad - Unfold a load from the given machineinstr if
253 /// the load itself could be hoisted. Return the unfolded and hoistable
254 /// load, or null if the load couldn't be unfolded or if it wouldn't
256 MachineInstr *ExtractHoistableLoad(MachineInstr *MI);
258 /// LookForDuplicate - Find an instruction amount PrevMIs that is a
259 /// duplicate of MI. Return this instruction if it's found.
260 const MachineInstr *LookForDuplicate(const MachineInstr *MI,
261 std::vector<const MachineInstr*> &PrevMIs);
263 /// EliminateCSE - Given a LICM'ed instruction, look for an instruction on
264 /// the preheader that compute the same value. If it's found, do a RAU on
265 /// with the definition of the existing instruction rather than hoisting
266 /// the instruction to the preheader.
267 bool EliminateCSE(MachineInstr *MI,
268 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI);
270 /// MayCSE - Return true if the given instruction will be CSE'd if it's
271 /// hoisted out of the loop.
272 bool MayCSE(MachineInstr *MI);
274 /// Hoist - When an instruction is found to only use loop invariant operands
275 /// that is safe to hoist, this instruction is called to do the dirty work.
276 /// It returns true if the instruction is hoisted.
277 bool Hoist(MachineInstr *MI, MachineBasicBlock *Preheader);
279 /// InitCSEMap - Initialize the CSE map with instructions that are in the
280 /// current loop preheader that may become duplicates of instructions that
281 /// are hoisted out of the loop.
282 void InitCSEMap(MachineBasicBlock *BB);
284 /// getCurPreheader - Get the preheader for the current loop, splitting
285 /// a critical edge if needed.
286 MachineBasicBlock *getCurPreheader();
288 } // end anonymous namespace
290 char MachineLICM::ID = 0;
291 INITIALIZE_PASS_BEGIN(MachineLICM, "machinelicm",
292 "Machine Loop Invariant Code Motion", false, false)
293 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
294 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
295 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
296 INITIALIZE_PASS_END(MachineLICM, "machinelicm",
297 "Machine Loop Invariant Code Motion", false, false)
299 FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) {
300 return new MachineLICM(PreRegAlloc);
303 /// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most
304 /// loop that has a unique predecessor.
305 static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) {
306 // Check whether this loop even has a unique predecessor.
307 if (!CurLoop->getLoopPredecessor())
309 // Ok, now check to see if any of its outer loops do.
310 for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop())
311 if (L->getLoopPredecessor())
313 // None of them did, so this is the outermost with a unique predecessor.
317 bool MachineLICM::runOnMachineFunction(MachineFunction &MF) {
319 DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
321 DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
322 DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n");
324 Changed = FirstInLoop = false;
325 TM = &MF.getTarget();
326 TII = TM->getInstrInfo();
327 TLI = TM->getTargetLowering();
328 TRI = TM->getRegisterInfo();
329 MFI = MF.getFrameInfo();
330 MRI = &MF.getRegInfo();
331 InstrItins = TM->getInstrItineraryData();
334 // Estimate register pressure during pre-regalloc pass.
335 unsigned NumRC = TRI->getNumRegClasses();
336 RegPressure.resize(NumRC);
337 std::fill(RegPressure.begin(), RegPressure.end(), 0);
338 RegLimit.resize(NumRC);
339 for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
340 E = TRI->regclass_end(); I != E; ++I)
341 RegLimit[(*I)->getID()] = TRI->getRegPressureLimit(*I, MF);
344 // Get our Loop information...
345 MLI = &getAnalysis<MachineLoopInfo>();
346 DT = &getAnalysis<MachineDominatorTree>();
347 AA = &getAnalysis<AliasAnalysis>();
349 SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
350 while (!Worklist.empty()) {
351 CurLoop = Worklist.pop_back_val();
354 // If this is done before regalloc, only visit outer-most preheader-sporting
356 if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) {
357 Worklist.append(CurLoop->begin(), CurLoop->end());
364 // CSEMap is initialized for loop header when the first instruction is
366 MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
376 /// InstructionStoresToFI - Return true if instruction stores to the
378 static bool InstructionStoresToFI(const MachineInstr *MI, int FI) {
379 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
380 oe = MI->memoperands_end(); o != oe; ++o) {
381 if (!(*o)->isStore() || !(*o)->getValue())
383 if (const FixedStackPseudoSourceValue *Value =
384 dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) {
385 if (Value->getFrameIndex() == FI)
392 /// ProcessMI - Examine the instruction for potentai LICM candidate. Also
393 /// gather register def and frame object update information.
394 void MachineLICM::ProcessMI(MachineInstr *MI,
395 unsigned *PhysRegDefs,
396 SmallSet<int, 32> &StoredFIs,
397 SmallVector<CandidateInfo, 32> &Candidates) {
398 bool RuledOut = false;
399 bool HasNonInvariantUse = false;
401 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
402 const MachineOperand &MO = MI->getOperand(i);
404 // Remember if the instruction stores to the frame index.
405 int FI = MO.getIndex();
406 if (!StoredFIs.count(FI) &&
407 MFI->isSpillSlotObjectIndex(FI) &&
408 InstructionStoresToFI(MI, FI))
409 StoredFIs.insert(FI);
410 HasNonInvariantUse = true;
416 unsigned Reg = MO.getReg();
419 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
420 "Not expecting virtual register!");
423 if (Reg && PhysRegDefs[Reg])
424 // If it's using a non-loop-invariant register, then it's obviously not
426 HasNonInvariantUse = true;
430 if (MO.isImplicit()) {
432 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
435 // Non-dead implicit def? This cannot be hoisted.
437 // No need to check if a dead implicit def is also defined by
438 // another instruction.
442 // FIXME: For now, avoid instructions with multiple defs, unless
443 // it's a dead implicit def.
449 // If we have already seen another instruction that defines the same
450 // register, then this is not safe.
451 if (++PhysRegDefs[Reg] > 1)
452 // MI defined register is seen defined by another instruction in
453 // the loop, it cannot be a LICM candidate.
455 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
456 if (++PhysRegDefs[*AS] > 1)
460 // Only consider reloads for now and remats which do not have register
461 // operands. FIXME: Consider unfold load folding instructions.
462 if (Def && !RuledOut) {
464 if ((!HasNonInvariantUse && IsLICMCandidate(*MI)) ||
465 (TII->isLoadFromStackSlot(MI, FI) && MFI->isSpillSlotObjectIndex(FI)))
466 Candidates.push_back(CandidateInfo(MI, Def, FI));
470 /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
471 /// invariants out to the preheader.
472 void MachineLICM::HoistRegionPostRA() {
473 unsigned NumRegs = TRI->getNumRegs();
474 unsigned *PhysRegDefs = new unsigned[NumRegs];
475 std::fill(PhysRegDefs, PhysRegDefs + NumRegs, 0);
477 SmallVector<CandidateInfo, 32> Candidates;
478 SmallSet<int, 32> StoredFIs;
480 // Walk the entire region, count number of defs for each register, and
481 // collect potential LICM candidates.
482 const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
483 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
484 MachineBasicBlock *BB = Blocks[i];
486 // If the header of the loop containing this basic block is a landing pad,
487 // then don't try to hoist instructions out of this loop.
488 const MachineLoop *ML = MLI->getLoopFor(BB);
489 if (ML && ML->getHeader()->isLandingPad()) continue;
491 // Conservatively treat live-in's as an external def.
492 // FIXME: That means a reload that're reused in successor block(s) will not
494 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
495 E = BB->livein_end(); I != E; ++I) {
498 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
502 SpeculationState = SpeculateUnknown;
503 for (MachineBasicBlock::iterator
504 MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
505 MachineInstr *MI = &*MII;
506 ProcessMI(MI, PhysRegDefs, StoredFIs, Candidates);
510 // Now evaluate whether the potential candidates qualify.
511 // 1. Check if the candidate defined register is defined by another
512 // instruction in the loop.
513 // 2. If the candidate is a load from stack slot (always true for now),
514 // check if the slot is stored anywhere in the loop.
515 for (unsigned i = 0, e = Candidates.size(); i != e; ++i) {
516 if (Candidates[i].FI != INT_MIN &&
517 StoredFIs.count(Candidates[i].FI))
520 if (PhysRegDefs[Candidates[i].Def] == 1) {
522 MachineInstr *MI = Candidates[i].MI;
523 for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) {
524 const MachineOperand &MO = MI->getOperand(j);
525 if (!MO.isReg() || MO.isDef() || !MO.getReg())
527 if (PhysRegDefs[MO.getReg()]) {
528 // If it's using a non-loop-invariant register, then it's obviously
529 // not safe to hoist.
535 HoistPostRA(MI, Candidates[i].Def);
539 delete[] PhysRegDefs;
542 /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the current
543 /// loop, and make sure it is not killed by any instructions in the loop.
544 void MachineLICM::AddToLiveIns(unsigned Reg) {
545 const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
546 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
547 MachineBasicBlock *BB = Blocks[i];
548 if (!BB->isLiveIn(Reg))
550 for (MachineBasicBlock::iterator
551 MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
552 MachineInstr *MI = &*MII;
553 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
554 MachineOperand &MO = MI->getOperand(i);
555 if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue;
556 if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg()))
563 /// HoistPostRA - When an instruction is found to only use loop invariant
564 /// operands that is safe to hoist, this instruction is called to do the
566 void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) {
567 MachineBasicBlock *Preheader = getCurPreheader();
568 if (!Preheader) return;
570 // Now move the instructions to the predecessor, inserting it before any
571 // terminator instructions.
573 dbgs() << "Hoisting " << *MI;
574 if (Preheader->getBasicBlock())
575 dbgs() << " to MachineBasicBlock "
576 << Preheader->getName();
577 if (MI->getParent()->getBasicBlock())
578 dbgs() << " from MachineBasicBlock "
579 << MI->getParent()->getName();
583 // Splice the instruction to the preheader.
584 MachineBasicBlock *MBB = MI->getParent();
585 Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
587 // Add register to livein list to all the BBs in the current loop since a
588 // loop invariant must be kept live throughout the whole loop. This is
589 // important to ensure later passes do not scavenge the def register.
596 // IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
597 // If not then a load from this mbb may not be safe to hoist.
598 bool MachineLICM::IsGuaranteedToExecute(MachineBasicBlock *BB) {
599 if (SpeculationState != SpeculateUnknown)
600 return SpeculationState == SpeculateFalse;
602 if (BB != CurLoop->getHeader()) {
603 // Check loop exiting blocks.
604 SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks;
605 CurLoop->getExitingBlocks(CurrentLoopExitingBlocks);
606 for (unsigned i = 0, e = CurrentLoopExitingBlocks.size(); i != e; ++i)
607 if (!DT->dominates(BB, CurrentLoopExitingBlocks[i])) {
608 SpeculationState = SpeculateTrue;
613 SpeculationState = SpeculateFalse;
617 void MachineLICM::EnterScope(MachineBasicBlock *MBB) {
618 DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
620 // Remember livein register pressure.
621 BackTrace.push_back(RegPressure);
624 void MachineLICM::ExitScope(MachineBasicBlock *MBB) {
625 DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
626 BackTrace.pop_back();
629 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
630 /// dominator tree node if its a leaf or all of its children are done. Walk
631 /// up the dominator tree to destroy ancestors which are now done.
632 void MachineLICM::ExitScopeIfDone(MachineDomTreeNode *Node,
633 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
634 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
635 if (OpenChildren[Node])
639 ExitScope(Node->getBlock());
641 // Now traverse upwards to pop ancestors whose offsprings are all done.
642 while (MachineDomTreeNode *Parent = ParentMap[Node]) {
643 unsigned Left = --OpenChildren[Parent];
646 ExitScope(Parent->getBlock());
651 /// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all
652 /// blocks dominated by the specified header block, and that are in the
653 /// current loop) in depth first order w.r.t the DominatorTree. This allows
654 /// us to visit definitions before uses, allowing us to hoist a loop body in
655 /// one pass without iteration.
657 void MachineLICM::HoistOutOfLoop(MachineDomTreeNode *HeaderN) {
658 SmallVector<MachineDomTreeNode*, 32> Scopes;
659 SmallVector<MachineDomTreeNode*, 8> WorkList;
660 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
661 DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
663 // Perform a DFS walk to determine the order of visit.
664 WorkList.push_back(HeaderN);
666 MachineDomTreeNode *Node = WorkList.pop_back_val();
667 assert(Node != 0 && "Null dominator tree node?");
668 MachineBasicBlock *BB = Node->getBlock();
670 // If the header of the loop containing this basic block is a landing pad,
671 // then don't try to hoist instructions out of this loop.
672 const MachineLoop *ML = MLI->getLoopFor(BB);
673 if (ML && ML->getHeader()->isLandingPad())
676 // If this subregion is not in the top level loop at all, exit.
677 if (!CurLoop->contains(BB))
680 Scopes.push_back(Node);
681 const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
682 unsigned NumChildren = Children.size();
684 // Don't hoist things out of a large switch statement. This often causes
685 // code to be hoisted that wasn't going to be executed, and increases
686 // register pressure in a situation where it's likely to matter.
687 if (BB->succ_size() >= 25)
690 OpenChildren[Node] = NumChildren;
691 // Add children in reverse order as then the next popped worklist node is
692 // the first child of this node. This means we ultimately traverse the
693 // DOM tree in exactly the same order as if we'd recursed.
694 for (int i = (int)NumChildren-1; i >= 0; --i) {
695 MachineDomTreeNode *Child = Children[i];
696 ParentMap[Child] = Node;
697 WorkList.push_back(Child);
699 } while (!WorkList.empty());
701 if (Scopes.size() != 0) {
702 MachineBasicBlock *Preheader = getCurPreheader();
706 // Compute registers which are livein into the loop headers.
709 InitRegPressure(Preheader);
713 for (unsigned i = 0, e = Scopes.size(); i != e; ++i) {
714 MachineDomTreeNode *Node = Scopes[i];
715 MachineBasicBlock *MBB = Node->getBlock();
717 MachineBasicBlock *Preheader = getCurPreheader();
724 SpeculationState = SpeculateUnknown;
725 for (MachineBasicBlock::iterator
726 MII = MBB->begin(), E = MBB->end(); MII != E; ) {
727 MachineBasicBlock::iterator NextMII = MII; ++NextMII;
728 MachineInstr *MI = &*MII;
729 if (!Hoist(MI, Preheader))
730 UpdateRegPressure(MI);
734 // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
735 ExitScopeIfDone(Node, OpenChildren, ParentMap);
739 static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
740 return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
743 /// getRegisterClassIDAndCost - For a given MI, register, and the operand
744 /// index, return the ID and cost of its representative register class.
746 MachineLICM::getRegisterClassIDAndCost(const MachineInstr *MI,
747 unsigned Reg, unsigned OpIdx,
748 unsigned &RCId, unsigned &RCCost) const {
749 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
750 EVT VT = *RC->vt_begin();
751 if (VT == MVT::Untyped) {
755 RCId = TLI->getRepRegClassFor(VT)->getID();
756 RCCost = TLI->getRepRegClassCostFor(VT);
760 /// InitRegPressure - Find all virtual register references that are liveout of
761 /// the preheader to initialize the starting "register pressure". Note this
762 /// does not count live through (livein but not used) registers.
763 void MachineLICM::InitRegPressure(MachineBasicBlock *BB) {
764 std::fill(RegPressure.begin(), RegPressure.end(), 0);
766 // If the preheader has only a single predecessor and it ends with a
767 // fallthrough or an unconditional branch, then scan its predecessor for live
768 // defs as well. This happens whenever the preheader is created by splitting
769 // the critical edge from the loop predecessor to the loop header.
770 if (BB->pred_size() == 1) {
771 MachineBasicBlock *TBB = 0, *FBB = 0;
772 SmallVector<MachineOperand, 4> Cond;
773 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty())
774 InitRegPressure(*BB->pred_begin());
777 for (MachineBasicBlock::iterator MII = BB->begin(), E = BB->end();
779 MachineInstr *MI = &*MII;
780 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
781 const MachineOperand &MO = MI->getOperand(i);
782 if (!MO.isReg() || MO.isImplicit())
784 unsigned Reg = MO.getReg();
785 if (!TargetRegisterInfo::isVirtualRegister(Reg))
788 bool isNew = RegSeen.insert(Reg);
789 unsigned RCId, RCCost;
790 getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost);
792 RegPressure[RCId] += RCCost;
794 bool isKill = isOperandKill(MO, MRI);
795 if (isNew && !isKill)
796 // Haven't seen this, it must be a livein.
797 RegPressure[RCId] += RCCost;
798 else if (!isNew && isKill)
799 RegPressure[RCId] -= RCCost;
805 /// UpdateRegPressure - Update estimate of register pressure after the
806 /// specified instruction.
807 void MachineLICM::UpdateRegPressure(const MachineInstr *MI) {
808 if (MI->isImplicitDef())
811 SmallVector<unsigned, 4> Defs;
812 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
813 const MachineOperand &MO = MI->getOperand(i);
814 if (!MO.isReg() || MO.isImplicit())
816 unsigned Reg = MO.getReg();
817 if (!TargetRegisterInfo::isVirtualRegister(Reg))
820 bool isNew = RegSeen.insert(Reg);
823 else if (!isNew && isOperandKill(MO, MRI)) {
824 unsigned RCId, RCCost;
825 getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost);
826 if (RCCost > RegPressure[RCId])
827 RegPressure[RCId] = 0;
829 RegPressure[RCId] -= RCCost;
834 while (!Defs.empty()) {
835 unsigned Reg = Defs.pop_back_val();
836 unsigned RCId, RCCost;
837 getRegisterClassIDAndCost(MI, Reg, Idx, RCId, RCCost);
838 RegPressure[RCId] += RCCost;
843 /// isLoadFromGOTOrConstantPool - Return true if this machine instruction
844 /// loads from global offset table or constant pool.
845 static bool isLoadFromGOTOrConstantPool(MachineInstr &MI) {
846 assert (MI.mayLoad() && "Expected MI that loads!");
847 for (MachineInstr::mmo_iterator I = MI.memoperands_begin(),
848 E = MI.memoperands_end(); I != E; ++I) {
849 if (const Value *V = (*I)->getValue()) {
850 if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V))
851 if (PSV == PSV->getGOT() || PSV == PSV->getConstantPool())
858 /// IsLICMCandidate - Returns true if the instruction may be a suitable
859 /// candidate for LICM. e.g. If the instruction is a call, then it's obviously
860 /// not safe to hoist it.
861 bool MachineLICM::IsLICMCandidate(MachineInstr &I) {
862 // Check if it's safe to move the instruction.
863 bool DontMoveAcrossStore = true;
864 if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore))
867 // If it is load then check if it is guaranteed to execute by making sure that
868 // it dominates all exiting blocks. If it doesn't, then there is a path out of
869 // the loop which does not execute this load, so we can't hoist it. Loads
870 // from constant memory are not safe to speculate all the time, for example
871 // indexed load from a jump table.
872 // Stores and side effects are already checked by isSafeToMove.
873 if (I.mayLoad() && !isLoadFromGOTOrConstantPool(I) &&
874 !IsGuaranteedToExecute(I.getParent()))
880 /// IsLoopInvariantInst - Returns true if the instruction is loop
881 /// invariant. I.e., all virtual register operands are defined outside of the
882 /// loop, physical registers aren't accessed explicitly, and there are no side
883 /// effects that aren't captured by the operands or other flags.
885 bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) {
886 if (!IsLICMCandidate(I))
889 // The instruction is loop invariant if all of its operands are.
890 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
891 const MachineOperand &MO = I.getOperand(i);
896 unsigned Reg = MO.getReg();
897 if (Reg == 0) continue;
899 // Don't hoist an instruction that uses or defines a physical register.
900 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
902 // If the physreg has no defs anywhere, it's just an ambient register
903 // and we can freely move its uses. Alternatively, if it's allocatable,
904 // it could get allocated to something with a def during allocation.
905 if (!MRI->isConstantPhysReg(Reg, *I.getParent()->getParent()))
907 // Otherwise it's safe to move.
909 } else if (!MO.isDead()) {
910 // A def that isn't dead. We can't move it.
912 } else if (CurLoop->getHeader()->isLiveIn(Reg)) {
913 // If the reg is live into the loop, we can't hoist an instruction
914 // which would clobber it.
922 assert(MRI->getVRegDef(Reg) &&
923 "Machine instr not mapped for this vreg?!");
925 // If the loop contains the definition of an operand, then the instruction
926 // isn't loop invariant.
927 if (CurLoop->contains(MRI->getVRegDef(Reg)))
931 // If we got this far, the instruction is loop invariant!
936 /// HasAnyPHIUse - Return true if the specified register is used by any
938 bool MachineLICM::HasAnyPHIUse(unsigned Reg) const {
939 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
940 UE = MRI->use_end(); UI != UE; ++UI) {
941 MachineInstr *UseMI = &*UI;
944 // Look pass copies as well.
945 if (UseMI->isCopy()) {
946 unsigned Def = UseMI->getOperand(0).getReg();
947 if (TargetRegisterInfo::isVirtualRegister(Def) &&
955 /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
956 /// and an use in the current loop, return true if the target considered
958 bool MachineLICM::HasHighOperandLatency(MachineInstr &MI,
959 unsigned DefIdx, unsigned Reg) const {
960 if (!InstrItins || InstrItins->isEmpty() || MRI->use_nodbg_empty(Reg))
963 for (MachineRegisterInfo::use_nodbg_iterator I = MRI->use_nodbg_begin(Reg),
964 E = MRI->use_nodbg_end(); I != E; ++I) {
965 MachineInstr *UseMI = &*I;
966 if (UseMI->isCopyLike())
968 if (!CurLoop->contains(UseMI->getParent()))
970 for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
971 const MachineOperand &MO = UseMI->getOperand(i);
972 if (!MO.isReg() || !MO.isUse())
974 unsigned MOReg = MO.getReg();
978 if (TII->hasHighOperandLatency(InstrItins, MRI, &MI, DefIdx, UseMI, i))
982 // Only look at the first in loop use.
989 /// IsCheapInstruction - Return true if the instruction is marked "cheap" or
990 /// the operand latency between its def and a use is one or less.
991 bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const {
992 if (MI.isAsCheapAsAMove() || MI.isCopyLike())
994 if (!InstrItins || InstrItins->isEmpty())
997 bool isCheap = false;
998 unsigned NumDefs = MI.getDesc().getNumDefs();
999 for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
1000 MachineOperand &DefMO = MI.getOperand(i);
1001 if (!DefMO.isReg() || !DefMO.isDef())
1004 unsigned Reg = DefMO.getReg();
1005 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1008 if (!TII->hasLowDefLatency(InstrItins, &MI, i))
1016 /// CanCauseHighRegPressure - Visit BBs from header to current BB, check
1017 /// if hoisting an instruction of the given cost matrix can cause high
1018 /// register pressure.
1019 bool MachineLICM::CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost) {
1020 for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
1022 if (CI->second <= 0)
1025 unsigned RCId = CI->first;
1026 unsigned Limit = RegLimit[RCId];
1027 int Cost = CI->second;
1028 for (unsigned i = BackTrace.size(); i != 0; --i) {
1029 SmallVector<unsigned, 8> &RP = BackTrace[i-1];
1030 if (RP[RCId] + Cost >= Limit)
1038 /// UpdateBackTraceRegPressure - Traverse the back trace from header to the
1039 /// current block and update their register pressures to reflect the effect
1040 /// of hoisting MI from the current block to the preheader.
1041 void MachineLICM::UpdateBackTraceRegPressure(const MachineInstr *MI) {
1042 if (MI->isImplicitDef())
1045 // First compute the 'cost' of the instruction, i.e. its contribution
1046 // to register pressure.
1047 DenseMap<unsigned, int> Cost;
1048 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
1049 const MachineOperand &MO = MI->getOperand(i);
1050 if (!MO.isReg() || MO.isImplicit())
1052 unsigned Reg = MO.getReg();
1053 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1056 unsigned RCId, RCCost;
1057 getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost);
1059 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
1060 if (CI != Cost.end())
1061 CI->second += RCCost;
1063 Cost.insert(std::make_pair(RCId, RCCost));
1064 } else if (isOperandKill(MO, MRI)) {
1065 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
1066 if (CI != Cost.end())
1067 CI->second -= RCCost;
1069 Cost.insert(std::make_pair(RCId, -RCCost));
1073 // Update register pressure of blocks from loop header to current block.
1074 for (unsigned i = 0, e = BackTrace.size(); i != e; ++i) {
1075 SmallVector<unsigned, 8> &RP = BackTrace[i];
1076 for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
1078 unsigned RCId = CI->first;
1079 RP[RCId] += CI->second;
1084 /// IsProfitableToHoist - Return true if it is potentially profitable to hoist
1085 /// the given loop invariant.
1086 bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
1087 if (MI.isImplicitDef())
1090 // If the instruction is cheap, only hoist if it is re-materilizable. LICM
1091 // will increase register pressure. It's probably not worth it if the
1092 // instruction is cheap.
1093 // Also hoist loads from constant memory, e.g. load from stubs, GOT. Hoisting
1094 // these tend to help performance in low register pressure situation. The
1095 // trade off is it may cause spill in high pressure situation. It will end up
1096 // adding a store in the loop preheader. But the reload is no more expensive.
1097 // The side benefit is these loads are frequently CSE'ed.
1098 if (IsCheapInstruction(MI)) {
1099 if (!TII->isTriviallyReMaterializable(&MI, AA))
1102 // Estimate register pressure to determine whether to LICM the instruction.
1103 // In low register pressure situation, we can be more aggressive about
1104 // hoisting. Also, favors hoisting long latency instructions even in
1105 // moderately high pressure situation.
1106 // FIXME: If there are long latency loop-invariant instructions inside the
1107 // loop at this point, why didn't the optimizer's LICM hoist them?
1108 DenseMap<unsigned, int> Cost;
1109 for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
1110 const MachineOperand &MO = MI.getOperand(i);
1111 if (!MO.isReg() || MO.isImplicit())
1113 unsigned Reg = MO.getReg();
1114 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1117 unsigned RCId, RCCost;
1118 getRegisterClassIDAndCost(&MI, Reg, i, RCId, RCCost);
1120 if (HasHighOperandLatency(MI, i, Reg)) {
1125 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
1126 if (CI != Cost.end())
1127 CI->second += RCCost;
1129 Cost.insert(std::make_pair(RCId, RCCost));
1130 } else if (isOperandKill(MO, MRI)) {
1131 // Is a virtual register use is a kill, hoisting it out of the loop
1132 // may actually reduce register pressure or be register pressure
1134 DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
1135 if (CI != Cost.end())
1136 CI->second -= RCCost;
1138 Cost.insert(std::make_pair(RCId, -RCCost));
1142 // Visit BBs from header to current BB, if hoisting this doesn't cause
1143 // high register pressure, then it's safe to proceed.
1144 if (!CanCauseHighRegPressure(Cost)) {
1149 // Do not "speculate" in high register pressure situation. If an
1150 // instruction is not guaranteed to be executed in the loop, it's best to be
1152 if (AvoidSpeculation &&
1153 (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI)))
1156 // High register pressure situation, only hoist if the instruction is going to
1158 if (!TII->isTriviallyReMaterializable(&MI, AA) &&
1159 !MI.isInvariantLoad(AA))
1163 // If result(s) of this instruction is used by PHIs outside of the loop, then
1164 // don't hoist it if the instruction because it will introduce an extra copy.
1165 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1166 const MachineOperand &MO = MI.getOperand(i);
1167 if (!MO.isReg() || !MO.isDef())
1169 if (HasAnyPHIUse(MO.getReg()))
1176 MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) {
1177 // Don't unfold simple loads.
1178 if (MI->canFoldAsLoad())
1181 // If not, we may be able to unfold a load and hoist that.
1182 // First test whether the instruction is loading from an amenable
1184 if (!MI->isInvariantLoad(AA))
1187 // Next determine the register class for a temporary register.
1188 unsigned LoadRegIndex;
1190 TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(),
1191 /*UnfoldLoad=*/true,
1192 /*UnfoldStore=*/false,
1194 if (NewOpc == 0) return 0;
1195 const MCInstrDesc &MID = TII->get(NewOpc);
1196 if (MID.getNumDefs() != 1) return 0;
1197 const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI);
1198 // Ok, we're unfolding. Create a temporary register and do the unfold.
1199 unsigned Reg = MRI->createVirtualRegister(RC);
1201 MachineFunction &MF = *MI->getParent()->getParent();
1202 SmallVector<MachineInstr *, 2> NewMIs;
1204 TII->unfoldMemoryOperand(MF, MI, Reg,
1205 /*UnfoldLoad=*/true, /*UnfoldStore=*/false,
1209 "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold "
1211 assert(NewMIs.size() == 2 &&
1212 "Unfolded a load into multiple instructions!");
1213 MachineBasicBlock *MBB = MI->getParent();
1214 MachineBasicBlock::iterator Pos = MI;
1215 MBB->insert(Pos, NewMIs[0]);
1216 MBB->insert(Pos, NewMIs[1]);
1217 // If unfolding produced a load that wasn't loop-invariant or profitable to
1218 // hoist, discard the new instructions and bail.
1219 if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) {
1220 NewMIs[0]->eraseFromParent();
1221 NewMIs[1]->eraseFromParent();
1225 // Update register pressure for the unfolded instruction.
1226 UpdateRegPressure(NewMIs[1]);
1228 // Otherwise we successfully unfolded a load that we can hoist.
1229 MI->eraseFromParent();
1233 void MachineLICM::InitCSEMap(MachineBasicBlock *BB) {
1234 for (MachineBasicBlock::iterator I = BB->begin(),E = BB->end(); I != E; ++I) {
1235 const MachineInstr *MI = &*I;
1236 unsigned Opcode = MI->getOpcode();
1237 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
1238 CI = CSEMap.find(Opcode);
1239 if (CI != CSEMap.end())
1240 CI->second.push_back(MI);
1242 std::vector<const MachineInstr*> CSEMIs;
1243 CSEMIs.push_back(MI);
1244 CSEMap.insert(std::make_pair(Opcode, CSEMIs));
1250 MachineLICM::LookForDuplicate(const MachineInstr *MI,
1251 std::vector<const MachineInstr*> &PrevMIs) {
1252 for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) {
1253 const MachineInstr *PrevMI = PrevMIs[i];
1254 if (TII->produceSameValue(MI, PrevMI, (PreRegAlloc ? MRI : 0)))
1260 bool MachineLICM::EliminateCSE(MachineInstr *MI,
1261 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI) {
1262 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1263 // the undef property onto uses.
1264 if (CI == CSEMap.end() || MI->isImplicitDef())
1267 if (const MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
1268 DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup);
1270 // Replace virtual registers defined by MI by their counterparts defined
1272 SmallVector<unsigned, 2> Defs;
1273 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1274 const MachineOperand &MO = MI->getOperand(i);
1276 // Physical registers may not differ here.
1277 assert((!MO.isReg() || MO.getReg() == 0 ||
1278 !TargetRegisterInfo::isPhysicalRegister(MO.getReg()) ||
1279 MO.getReg() == Dup->getOperand(i).getReg()) &&
1280 "Instructions with different phys regs are not identical!");
1282 if (MO.isReg() && MO.isDef() &&
1283 !TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1287 SmallVector<const TargetRegisterClass*, 2> OrigRCs;
1288 for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1289 unsigned Idx = Defs[i];
1290 unsigned Reg = MI->getOperand(Idx).getReg();
1291 unsigned DupReg = Dup->getOperand(Idx).getReg();
1292 OrigRCs.push_back(MRI->getRegClass(DupReg));
1294 if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) {
1295 // Restore old RCs if more than one defs.
1296 for (unsigned j = 0; j != i; ++j)
1297 MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]);
1302 for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
1303 unsigned Idx = Defs[i];
1304 unsigned Reg = MI->getOperand(Idx).getReg();
1305 unsigned DupReg = Dup->getOperand(Idx).getReg();
1306 MRI->replaceRegWith(Reg, DupReg);
1307 MRI->clearKillFlags(DupReg);
1310 MI->eraseFromParent();
1317 /// MayCSE - Return true if the given instruction will be CSE'd if it's
1318 /// hoisted out of the loop.
1319 bool MachineLICM::MayCSE(MachineInstr *MI) {
1320 unsigned Opcode = MI->getOpcode();
1321 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
1322 CI = CSEMap.find(Opcode);
1323 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
1324 // the undef property onto uses.
1325 if (CI == CSEMap.end() || MI->isImplicitDef())
1328 return LookForDuplicate(MI, CI->second) != 0;
1331 /// Hoist - When an instruction is found to use only loop invariant operands
1332 /// that are safe to hoist, this instruction is called to do the dirty work.
1334 bool MachineLICM::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader) {
1335 // First check whether we should hoist this instruction.
1336 if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) {
1337 // If not, try unfolding a hoistable load.
1338 MI = ExtractHoistableLoad(MI);
1339 if (!MI) return false;
1342 // Now move the instructions to the predecessor, inserting it before any
1343 // terminator instructions.
1345 dbgs() << "Hoisting " << *MI;
1346 if (Preheader->getBasicBlock())
1347 dbgs() << " to MachineBasicBlock "
1348 << Preheader->getName();
1349 if (MI->getParent()->getBasicBlock())
1350 dbgs() << " from MachineBasicBlock "
1351 << MI->getParent()->getName();
1355 // If this is the first instruction being hoisted to the preheader,
1356 // initialize the CSE map with potential common expressions.
1358 InitCSEMap(Preheader);
1359 FirstInLoop = false;
1362 // Look for opportunity to CSE the hoisted instruction.
1363 unsigned Opcode = MI->getOpcode();
1364 DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
1365 CI = CSEMap.find(Opcode);
1366 if (!EliminateCSE(MI, CI)) {
1367 // Otherwise, splice the instruction to the preheader.
1368 Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
1370 // Update register pressure for BBs from header to this block.
1371 UpdateBackTraceRegPressure(MI);
1373 // Clear the kill flags of any register this instruction defines,
1374 // since they may need to be live throughout the entire loop
1375 // rather than just live for part of it.
1376 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1377 MachineOperand &MO = MI->getOperand(i);
1378 if (MO.isReg() && MO.isDef() && !MO.isDead())
1379 MRI->clearKillFlags(MO.getReg());
1382 // Add to the CSE map.
1383 if (CI != CSEMap.end())
1384 CI->second.push_back(MI);
1386 std::vector<const MachineInstr*> CSEMIs;
1387 CSEMIs.push_back(MI);
1388 CSEMap.insert(std::make_pair(Opcode, CSEMIs));
1398 MachineBasicBlock *MachineLICM::getCurPreheader() {
1399 // Determine the block to which to hoist instructions. If we can't find a
1400 // suitable loop predecessor, we can't do any hoisting.
1402 // If we've tried to get a preheader and failed, don't try again.
1403 if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
1406 if (!CurPreheader) {
1407 CurPreheader = CurLoop->getLoopPreheader();
1408 if (!CurPreheader) {
1409 MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
1411 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1415 CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), this);
1416 if (!CurPreheader) {
1417 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
1422 return CurPreheader;