/// 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
if (!CurPreheader)
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());
HoistRegion(N);
CSEMap.clear();
}
SmallSet<int, 32> &StoredFIs,
SmallVector<CandidateInfo, 32> &Candidates) {
bool RuledOut = false;
- bool HasRegFIUse = 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);
- HasRegFIUse = true;
+ HasNonInvariantUse = true;
continue;
}
"Not expecting virtual register!");
if (!MO.isDef()) {
- HasRegFIUse = true;
+ if (Reg && PhysRegDefs[Reg])
+ // If it's using a non-loop-invariant register, then it's obviously not
+ // safe to hoist.
+ HasNonInvariantUse = true;
continue;
}
// operands. FIXME: Consider unfold load folding instructions.
if (Def && !RuledOut) {
int FI = INT_MIN;
- // FIXME: Also hoist instructions if all source operands are live in
- // to the loop.
- if ((!HasRegFIUse && IsLICMCandidate(*MI)) ||
+ 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
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
MachineBasicBlock *MBB = MI->getParent();
CurPreheader->splice(CurPreheader->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;
/// 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) {
+ if (I.isImplicitDef())
+ return false;
+
const TargetInstrDesc &TID = I.getDesc();
// Ignore stuff that we obviously can't hoist.
/// 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.
"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;
// Otherwise, splice the instruction to the preheader.
CurPreheader->splice(CurPreheader->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())
CI->second.push_back(MI);