// State that is updated as we process loops
bool Changed; // True if a loop is changed.
+ bool FirstInLoop; // True if it's the first LICM in the loop.
MachineLoop *CurLoop; // The current loop we are working on.
MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
BitVector AllocatableSet;
- // For each opcode, keep a list of potentail CSE instructions.
+ // For each opcode, keep a list of potential CSE instructions.
DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
public:
static char ID; // Pass identification, replacement for typeid
MachineLICM() :
- MachineFunctionPass(&ID), PreRegAlloc(true) {}
+ MachineFunctionPass(ID), PreRegAlloc(true) {}
explicit MachineLICM(bool PreRA) :
- MachineFunctionPass(&ID), PreRegAlloc(PreRA) {}
+ MachineFunctionPass(ID), PreRegAlloc(PreRA) {}
virtual bool runOnMachineFunction(MachineFunction &MF);
const char *getPassName() const { return "Machine Instruction LICM"; }
- // FIXME: Loop preheaders?
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
AU.addRequired<MachineLoopInfo>();
/// CandidateInfo - Keep track of information about hoisting candidates.
struct CandidateInfo {
MachineInstr *MI;
- int FI;
unsigned Def;
- CandidateInfo(MachineInstr *mi, int fi, unsigned def)
- : MI(mi), FI(fi), Def(def) {}
+ int FI;
+ CandidateInfo(MachineInstr *mi, unsigned def, int fi)
+ : MI(mi), Def(def), FI(fi) {}
};
/// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
/// invariants out to the preheader.
- void HoistRegionPostRA(MachineDomTreeNode *N);
+ void HoistRegionPostRA();
/// HoistPostRA - When an instruction is found to only use loop invariant
/// operands that is safe to hoist, this instruction is called to do the
SmallSet<int, 32> &StoredFIs,
SmallVector<CandidateInfo, 32> &Candidates);
- /// AddToLiveIns - Add 'Reg' to the livein sets of BBs in the backedge path
- /// from MBB to LoopHeader (inclusive).
- void AddToLiveIns(unsigned Reg,
- MachineBasicBlock *MBB, MachineBasicBlock *LoopHeader);
+ /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the
+ /// current loop.
+ void AddToLiveIns(unsigned Reg);
+
+ /// IsLICMCandidate - Returns true if the instruction may be a suitable
+ /// candidate for LICM. e.g. If the instruction is a call, then it's
+ /// obviously not safe to hoist it.
+ bool IsLICMCandidate(MachineInstr &I);
/// IsLoopInvariantInst - Returns true if the instruction is loop
/// invariant. I.e., all virtual register operands are defined outside of
/// current loop preheader that may become duplicates of instructions that
/// are hoisted out of the loop.
void InitCSEMap(MachineBasicBlock *BB);
+
+ /// getCurPreheader - Get the preheader for the current loop, splitting
+ /// a critical edge if needed.
+ MachineBasicBlock *getCurPreheader();
};
} // end anonymous namespace
char MachineLICM::ID = 0;
-static RegisterPass<MachineLICM>
-X("machinelicm", "Machine Loop Invariant Code Motion");
+INITIALIZE_PASS(MachineLICM, "machinelicm",
+ "Machine Loop Invariant Code Motion", false, false);
FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) {
return new MachineLICM(PreRegAlloc);
}
-/// LoopIsOuterMostWithPreheader - Test if the given loop is the outer-most
-/// loop that has a preheader.
-static bool LoopIsOuterMostWithPreheader(MachineLoop *CurLoop) {
+/// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most
+/// loop that has a unique predecessor.
+static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) {
+ // Check whether this loop even has a unique predecessor.
+ if (!CurLoop->getLoopPredecessor())
+ return false;
+ // Ok, now check to see if any of its outer loops do.
for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop())
- if (L->getLoopPreheader())
+ if (L->getLoopPredecessor())
return false;
+ // None of them did, so this is the outermost with a unique predecessor.
return true;
}
else
DEBUG(dbgs() << "******** Post-regalloc Machine LICM ********\n");
- Changed = false;
+ Changed = FirstInLoop = false;
TM = &MF.getTarget();
TII = TM->getInstrInfo();
TRI = TM->getRegisterInfo();
DT = &getAnalysis<MachineDominatorTree>();
AA = &getAnalysis<AliasAnalysis>();
- for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I){
- CurLoop = *I;
+ SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
+ while (!Worklist.empty()) {
+ CurLoop = Worklist.pop_back_val();
+ CurPreheader = 0;
// If this is done before regalloc, only visit outer-most preheader-sporting
// loops.
- if (PreRegAlloc && !LoopIsOuterMostWithPreheader(CurLoop))
- continue;
-
- // Determine the block to which to hoist instructions. If we can't find a
- // suitable loop preheader, we can't do any hoisting.
- //
- // FIXME: We are only hoisting if the basic block coming into this loop
- // has only one successor. This isn't the case in general because we haven't
- // broken critical edges or added preheaders.
- CurPreheader = CurLoop->getLoopPreheader();
- if (!CurPreheader)
+ if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) {
+ Worklist.append(CurLoop->begin(), CurLoop->end());
continue;
+ }
- // CSEMap is initialized for loop header when the first instruction is
- // being hoisted.
- MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
if (!PreRegAlloc)
- HoistRegionPostRA(N);
+ HoistRegionPostRA();
else {
+ // CSEMap is initialized for loop header when the first instruction is
+ // being hoisted.
+ MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
+ FirstInLoop = true;
HoistRegion(N);
CSEMap.clear();
}
SmallSet<int, 32> &StoredFIs,
SmallVector<CandidateInfo, 32> &Candidates) {
bool RuledOut = false;
+ bool HasNonInvariantUse = false;
unsigned Def = 0;
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
MFI->isSpillSlotObjectIndex(FI) &&
InstructionStoresToFI(MI, FI))
StoredFIs.insert(FI);
+ HasNonInvariantUse = true;
continue;
}
assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
"Not expecting virtual register!");
- if (!MO.isDef())
+ if (!MO.isDef()) {
+ if (Reg && PhysRegDefs[Reg])
+ // If it's using a non-loop-invariant register, then it's obviously not
+ // safe to hoist.
+ HasNonInvariantUse = true;
continue;
+ }
if (MO.isImplicit()) {
++PhysRegDefs[Reg];
RuledOut = true;
}
- // FIXME: Only consider reloads for now. We should be able to handle
- // remats which does not have register operands.
+ // Only consider reloads for now and remats which do not have register
+ // operands. FIXME: Consider unfold load folding instructions.
if (Def && !RuledOut) {
- int FI;
- if (TII->isLoadFromStackSlot(MI, FI) &&
- MFI->isSpillSlotObjectIndex(FI))
- Candidates.push_back(CandidateInfo(MI, FI, Def));
+ int FI = INT_MIN;
+ if ((!HasNonInvariantUse && IsLICMCandidate(*MI)) ||
+ (TII->isLoadFromStackSlot(MI, FI) && MFI->isSpillSlotObjectIndex(FI)))
+ Candidates.push_back(CandidateInfo(MI, Def, FI));
}
}
/// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop
/// invariants out to the preheader.
-void MachineLICM::HoistRegionPostRA(MachineDomTreeNode *N) {
- assert(N != 0 && "Null dominator tree node?");
-
+void MachineLICM::HoistRegionPostRA() {
unsigned NumRegs = TRI->getNumRegs();
unsigned *PhysRegDefs = new unsigned[NumRegs];
std::fill(PhysRegDefs, PhysRegDefs + NumRegs, 0);
SmallSet<int, 32> StoredFIs;
// Walk the entire region, count number of defs for each register, and
- // return potential LICM candidates.
- SmallVector<MachineDomTreeNode*, 8> WorkList;
- WorkList.push_back(N);
- do {
- N = WorkList.pop_back_val();
- MachineBasicBlock *BB = N->getBlock();
-
- if (!CurLoop->contains(MLI->getLoopFor(BB)))
- continue;
+ // collect potential LICM candidates.
+ const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
+ for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
+ MachineBasicBlock *BB = Blocks[i];
// Conservatively treat live-in's as an external def.
// FIXME: That means a reload that're reused in successor block(s) will not
// be LICM'ed.
MachineInstr *MI = &*MII;
ProcessMI(MI, PhysRegDefs, StoredFIs, Candidates);
}
-
- const std::vector<MachineDomTreeNode*> &Children = N->getChildren();
- for (unsigned I = 0, E = Children.size(); I != E; ++I)
- WorkList.push_back(Children[I]);
- } while (!WorkList.empty());
+ }
// Now evaluate whether the potential candidates qualify.
// 1. Check if the candidate defined register is defined by another
// 2. If the candidate is a load from stack slot (always true for now),
// check if the slot is stored anywhere in the loop.
for (unsigned i = 0, e = Candidates.size(); i != e; ++i) {
- if (StoredFIs.count(Candidates[i].FI))
+ if (Candidates[i].FI != INT_MIN &&
+ StoredFIs.count(Candidates[i].FI))
continue;
- if (PhysRegDefs[Candidates[i].Def] == 1)
- HoistPostRA(Candidates[i].MI, Candidates[i].Def);
+ if (PhysRegDefs[Candidates[i].Def] == 1) {
+ bool Safe = true;
+ MachineInstr *MI = Candidates[i].MI;
+ for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) {
+ const MachineOperand &MO = MI->getOperand(j);
+ if (!MO.isReg() || MO.isDef() || !MO.getReg())
+ continue;
+ if (PhysRegDefs[MO.getReg()]) {
+ // If it's using a non-loop-invariant register, then it's obviously
+ // not safe to hoist.
+ Safe = false;
+ break;
+ }
+ }
+ if (Safe)
+ HoistPostRA(MI, Candidates[i].Def);
+ }
}
delete[] PhysRegDefs;
}
-/// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the
-/// backedge path from MBB to LoopHeader.
-void MachineLICM::AddToLiveIns(unsigned Reg, MachineBasicBlock *MBB,
- MachineBasicBlock *LoopHeader) {
- SmallPtrSet<MachineBasicBlock*, 4> Visited;
- SmallVector<MachineBasicBlock*, 4> WorkList;
- WorkList.push_back(MBB);
- do {
- MBB = WorkList.pop_back_val();
- if (!Visited.insert(MBB))
- continue;
- MBB->addLiveIn(Reg);
- if (MBB == LoopHeader)
- continue;
- for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
- E = MBB->pred_end(); PI != E; ++PI)
- WorkList.push_back(*PI);
- } while (!WorkList.empty());
+/// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the current
+/// loop, and make sure it is not killed by any instructions in the loop.
+void MachineLICM::AddToLiveIns(unsigned Reg) {
+ const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
+ for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
+ MachineBasicBlock *BB = Blocks[i];
+ if (!BB->isLiveIn(Reg))
+ BB->addLiveIn(Reg);
+ for (MachineBasicBlock::iterator
+ MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
+ MachineInstr *MI = &*MII;
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue;
+ if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg()))
+ MO.setIsKill(false);
+ }
+ }
+ }
}
/// HoistPostRA - When an instruction is found to only use loop invariant
/// operands that is safe to hoist, this instruction is called to do the
/// dirty work.
void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) {
+ MachineBasicBlock *Preheader = getCurPreheader();
+ if (!Preheader) return;
+
// Now move the instructions to the predecessor, inserting it before any
// terminator instructions.
DEBUG({
dbgs() << "Hoisting " << *MI;
- if (CurPreheader->getBasicBlock())
+ if (Preheader->getBasicBlock())
dbgs() << " to MachineBasicBlock "
- << CurPreheader->getName();
+ << Preheader->getName();
if (MI->getParent()->getBasicBlock())
dbgs() << " from MachineBasicBlock "
<< MI->getParent()->getName();
// Splice the instruction to the preheader.
MachineBasicBlock *MBB = MI->getParent();
- CurPreheader->splice(CurPreheader->getFirstTerminator(), MBB, MI);
+ Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
- // Add register to livein list to BBs in the path from loop header to original
- // BB. Note, currently it's not necessary to worry about adding it to all BB's
- // with uses. Reload that're reused in successor block(s) are not being
- // hoisted.
- AddToLiveIns(Def, MBB, CurLoop->getHeader());
+ // Add register to livein list to all the BBs in the current loop since a
+ // loop invariant must be kept live throughout the whole loop. This is
+ // important to ensure later passes do not scavenge the def register.
+ AddToLiveIns(Def);
++NumPostRAHoisted;
Changed = true;
MII = NextMII;
}
- const std::vector<MachineDomTreeNode*> &Children = N->getChildren();
- for (unsigned I = 0, E = Children.size(); I != E; ++I)
- HoistRegion(Children[I]);
+ // Don't hoist things out of a large switch statement. This often causes
+ // code to be hoisted that wasn't going to be executed, and increases
+ // register pressure in a situation where it's likely to matter.
+ if (BB->succ_size() < 25) {
+ const std::vector<MachineDomTreeNode*> &Children = N->getChildren();
+ for (unsigned I = 0, E = Children.size(); I != E; ++I)
+ HoistRegion(Children[I]);
+ }
+}
+
+/// IsLICMCandidate - Returns true if the instruction may be a suitable
+/// candidate for LICM. e.g. If the instruction is a call, then it's obviously
+/// not safe to hoist it.
+bool MachineLICM::IsLICMCandidate(MachineInstr &I) {
+ // Check if it's safe to move the instruction.
+ bool DontMoveAcrossStore = true;
+ if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore))
+ return false;
+
+ return true;
}
/// IsLoopInvariantInst - Returns true if the instruction is loop
/// effects that aren't captured by the operands or other flags.
///
bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) {
- const TargetInstrDesc &TID = I.getDesc();
-
- // Ignore stuff that we obviously can't hoist.
- if (TID.mayStore() || TID.isCall() || TID.isTerminator() ||
- TID.hasUnmodeledSideEffects())
+ if (!IsLICMCandidate(I))
return false;
- if (TID.mayLoad()) {
- // Okay, this instruction does a load. As a refinement, we allow the target
- // to decide whether the loaded value is actually a constant. If so, we can
- // actually use it as a load.
- if (!I.isInvariantLoad(AA))
- // FIXME: we should be able to hoist loads with no other side effects if
- // there are no other instructions which can change memory in this loop.
- // This is a trivial form of alias analysis.
- return false;
- }
-
// The instruction is loop invariant if all of its operands are.
for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
const MachineOperand &MO = I.getOperand(i);
/// IsProfitableToHoist - Return true if it is potentially profitable to hoist
/// the given loop invariant.
bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
- if (MI.isImplicitDef())
- return false;
-
// FIXME: For now, only hoist re-materilizable instructions. LICM will
// increase register pressure. We want to make sure it doesn't increase
// spilling.
bool MachineLICM::EliminateCSE(MachineInstr *MI,
DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI) {
- if (CI == CSEMap.end())
+ // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
+ // the undef property onto uses.
+ if (CI == CSEMap.end() || MI->isImplicitDef())
return false;
if (const MachineInstr *Dup = LookForDuplicate(MI, CI->second)) {
"Instructions with different phys regs are not identical!");
if (MO.isReg() && MO.isDef() &&
- !TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
+ !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
RegInfo->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg());
+ RegInfo->clearKillFlags(Dup->getOperand(i).getReg());
+ }
}
MI->eraseFromParent();
++NumCSEed;
/// that are safe to hoist, this instruction is called to do the dirty work.
///
void MachineLICM::Hoist(MachineInstr *MI) {
+ MachineBasicBlock *Preheader = getCurPreheader();
+ if (!Preheader) return;
+
// First check whether we should hoist this instruction.
if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) {
// If not, try unfolding a hoistable load.
// terminator instructions.
DEBUG({
dbgs() << "Hoisting " << *MI;
- if (CurPreheader->getBasicBlock())
+ if (Preheader->getBasicBlock())
dbgs() << " to MachineBasicBlock "
- << CurPreheader->getName();
+ << Preheader->getName();
if (MI->getParent()->getBasicBlock())
dbgs() << " from MachineBasicBlock "
<< MI->getParent()->getName();
// If this is the first instruction being hoisted to the preheader,
// initialize the CSE map with potential common expressions.
- InitCSEMap(CurPreheader);
+ if (FirstInLoop) {
+ InitCSEMap(Preheader);
+ FirstInLoop = false;
+ }
// Look for opportunity to CSE the hoisted instruction.
unsigned Opcode = MI->getOpcode();
CI = CSEMap.find(Opcode);
if (!EliminateCSE(MI, CI)) {
// Otherwise, splice the instruction to the preheader.
- CurPreheader->splice(CurPreheader->getFirstTerminator(),MI->getParent(),MI);
+ Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
+
+ // Clear the kill flags of any register this instruction defines,
+ // since they may need to be live throughout the entire loop
+ // rather than just live for part of it.
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = MI->getOperand(i);
+ if (MO.isReg() && MO.isDef() && !MO.isDead())
+ RegInfo->clearKillFlags(MO.getReg());
+ }
// Add to the CSE map.
if (CI != CSEMap.end())
++NumHoisted;
Changed = true;
}
+
+MachineBasicBlock *MachineLICM::getCurPreheader() {
+ // Determine the block to which to hoist instructions. If we can't find a
+ // suitable loop predecessor, we can't do any hoisting.
+
+ // If we've tried to get a preheader and failed, don't try again.
+ if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
+ return 0;
+
+ if (!CurPreheader) {
+ CurPreheader = CurLoop->getLoopPreheader();
+ if (!CurPreheader) {
+ MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
+ if (!Pred) {
+ CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
+ return 0;
+ }
+
+ CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), this);
+ if (!CurPreheader) {
+ CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
+ return 0;
+ }
+ }
+ }
+ return CurPreheader;
+}