#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
+static cl::opt<bool>
+AvoidSpeculation("avoid-speculation",
+ cl::desc("MachineLICM should avoid speculation"),
+ cl::init(true), cl::Hidden);
+
STATISTIC(NumHoisted,
"Number of machine instructions hoisted out of loops");
STATISTIC(NumLowRP,
namespace {
class MachineLICM : public MachineFunctionPass {
- bool PreRegAlloc;
-
const TargetMachine *TM;
const TargetInstrInfo *TII;
const TargetLowering *TLI;
const MachineFrameInfo *MFI;
MachineRegisterInfo *MRI;
const InstrItineraryData *InstrItins;
+ bool PreRegAlloc;
+ BitVector ReservedRegs;
// Various analyses that we use...
AliasAnalysis *AA; // Alias analysis info.
MachineLoop *CurLoop; // The current loop we are working on.
MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
- BitVector AllocatableSet;
-
// Track 'estimated' register pressure.
SmallSet<unsigned, 32> RegSeen;
SmallVector<unsigned, 8> RegPressure;
// For each opcode, keep a list of potential CSE instructions.
DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
+ enum {
+ SpeculateFalse = 0,
+ SpeculateTrue = 1,
+ SpeculateUnknown = 2
+ };
+
+ // If a MBB does not dominate loop exiting blocks then it may not safe
+ // to hoist loads from this block.
+ // Tri-state: 0 - false, 1 - true, 2 - unknown
+ unsigned SpeculationState;
+
public:
static char ID; // Pass identification, replacement for typeid
MachineLICM() :
virtual bool runOnMachineFunction(MachineFunction &MF);
- const char *getPassName() const { return "Machine Instruction LICM"; }
-
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<MachineLoopInfo>();
AU.addRequired<MachineDominatorTree>();
/// ProcessMI - Examine the instruction for potentai LICM candidate. Also
/// gather register def and frame object update information.
- void ProcessMI(MachineInstr *MI, unsigned *PhysRegDefs,
+ void ProcessMI(MachineInstr *MI,
+ BitVector &PhysRegDefs,
+ BitVector &PhysRegClobbers,
SmallSet<int, 32> &StoredFIs,
SmallVector<CandidateInfo, 32> &Candidates);
/// invariant. I.e., all virtual register operands are defined outside of
/// the loop, physical registers aren't accessed (explicitly or implicitly),
/// and the instruction is hoistable.
- ///
+ ///
bool IsLoopInvariantInst(MachineInstr &I);
/// HasAnyPHIUse - Return true if the specified register is used by any
/// hoist the given loop invariant.
bool IsProfitableToHoist(MachineInstr &MI);
- /// HoistRegion - Walk the specified region of the CFG (defined by all
- /// blocks dominated by the specified block, and that are in the current
- /// loop) in depth first order w.r.t the DominatorTree. This allows us to
- /// visit definitions before uses, allowing us to hoist a loop body in one
- /// pass without iteration.
+ /// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
+ /// If not then a load from this mbb may not be safe to hoist.
+ bool IsGuaranteedToExecute(MachineBasicBlock *BB);
+
+ void EnterScope(MachineBasicBlock *MBB);
+
+ void ExitScope(MachineBasicBlock *MBB);
+
+ /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to given
+ /// dominator tree node if its a leaf or all of its children are done. Walk
+ /// up the dominator tree to destroy ancestors which are now done.
+ void ExitScopeIfDone(MachineDomTreeNode *Node,
+ DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
+ DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap);
+
+ /// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all
+ /// blocks dominated by the specified header block, and that are in the
+ /// current loop) in depth first order w.r.t the DominatorTree. This allows
+ /// us to visit definitions before uses, allowing us to hoist a loop body in
+ /// one pass without iteration.
///
- void HoistRegion(MachineDomTreeNode *N, bool IsHeader = false);
+ void HoistOutOfLoop(MachineDomTreeNode *LoopHeaderNode);
+ void HoistRegion(MachineDomTreeNode *N, bool IsHeader);
/// getRegisterClassIDAndCost - For a given MI, register, and the operand
/// index, return the ID and cost of its representative register class by
bool EliminateCSE(MachineInstr *MI,
DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI);
+ /// MayCSE - Return true if the given instruction will be CSE'd if it's
+ /// hoisted out of the loop.
+ bool MayCSE(MachineInstr *MI);
+
/// Hoist - 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.
/// It returns true if the instruction is hoisted.
} // end anonymous namespace
char MachineLICM::ID = 0;
+char &llvm::MachineLICMID = MachineLICM::ID;
INITIALIZE_PASS_BEGIN(MachineLICM, "machinelicm",
"Machine Loop Invariant Code Motion", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
INITIALIZE_PASS_END(MachineLICM, "machinelicm",
"Machine Loop Invariant Code Motion", false, false)
-FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) {
- return new MachineLICM(PreRegAlloc);
-}
-
/// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most
/// loop that has a unique predecessor.
static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) {
}
bool MachineLICM::runOnMachineFunction(MachineFunction &MF) {
- if (PreRegAlloc)
- DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
- else
- DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
- DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n");
-
Changed = FirstInLoop = false;
TM = &MF.getTarget();
TII = TM->getInstrInfo();
MFI = MF.getFrameInfo();
MRI = &MF.getRegInfo();
InstrItins = TM->getInstrItineraryData();
- AllocatableSet = TRI->getAllocatableSet(MF);
+
+ PreRegAlloc = MRI->isSSA();
+
+ if (PreRegAlloc)
+ DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
+ else
+ DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
+ DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n");
if (PreRegAlloc) {
// Estimate register pressure during pre-regalloc pass.
for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
E = TRI->regclass_end(); I != E; ++I)
RegLimit[(*I)->getID()] = TRI->getRegPressureLimit(*I, MF);
+ } else {
+ ReservedRegs = TRI->getReservedRegs(MF);
}
// Get our Loop information...
// being hoisted.
MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
FirstInLoop = true;
- HoistRegion(N, true);
+ HoistOutOfLoop(N);
CSEMap.clear();
}
}
/// ProcessMI - Examine the instruction for potentai LICM candidate. Also
/// gather register def and frame object update information.
void MachineLICM::ProcessMI(MachineInstr *MI,
- unsigned *PhysRegDefs,
+ BitVector &PhysRegDefs,
+ BitVector &PhysRegClobbers,
SmallSet<int, 32> &StoredFIs,
SmallVector<CandidateInfo, 32> &Candidates) {
bool RuledOut = false;
continue;
}
+ // We can't hoist an instruction defining a physreg that is clobbered in
+ // the loop.
+ if (MO.isRegMask()) {
+ PhysRegClobbers.setBitsNotInMask(MO.getRegMask());
+ continue;
+ }
+
if (!MO.isReg())
continue;
unsigned Reg = MO.getReg();
"Not expecting virtual register!");
if (!MO.isDef()) {
- if (Reg && PhysRegDefs[Reg])
+ // Allow reserved register reads to be hoisted.
+ if (ReservedRegs.test(Reg))
+ continue;
+ if (Reg && (PhysRegDefs.test(Reg) || PhysRegClobbers.test(Reg)))
// If it's using a non-loop-invariant register, then it's obviously not
// safe to hoist.
HasNonInvariantUse = true;
}
if (MO.isImplicit()) {
- ++PhysRegDefs[Reg];
- for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
- ++PhysRegDefs[*AS];
+ for (const unsigned *AS = TRI->getOverlaps(Reg); *AS; ++AS)
+ PhysRegClobbers.set(*AS);
if (!MO.isDead())
// Non-dead implicit def? This cannot be hoisted.
RuledOut = true;
Def = Reg;
// If we have already seen another instruction that defines the same
- // register, then this is not safe.
- if (++PhysRegDefs[Reg] > 1)
- // MI defined register is seen defined by another instruction in
- // the loop, it cannot be a LICM candidate.
- RuledOut = true;
- for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
- if (++PhysRegDefs[*AS] > 1)
+ // register, then this is not safe. Two defs is indicated by setting a
+ // PhysRegClobbers bit.
+ for (const unsigned *AS = TRI->getOverlaps(Reg); *AS; ++AS) {
+ if (PhysRegDefs.test(*AS))
+ PhysRegClobbers.set(*AS);
+ if (PhysRegClobbers.test(*AS))
+ // MI defined register is seen defined by another instruction in
+ // the loop, it cannot be a LICM candidate.
RuledOut = true;
+ PhysRegDefs.set(*AS);
+ }
}
// Only consider reloads for now and remats which do not have register
/// invariants out to the preheader.
void MachineLICM::HoistRegionPostRA() {
unsigned NumRegs = TRI->getNumRegs();
- unsigned *PhysRegDefs = new unsigned[NumRegs];
- std::fill(PhysRegDefs, PhysRegDefs + NumRegs, 0);
+ BitVector PhysRegDefs(NumRegs); // Regs defined once in the loop.
+ BitVector PhysRegClobbers(NumRegs); // Regs defined more than once.
SmallVector<CandidateInfo, 32> Candidates;
SmallSet<int, 32> StoredFIs;
const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
MachineBasicBlock *BB = Blocks[i];
+
+ // If the header of the loop containing this basic block is a landing pad,
+ // then don't try to hoist instructions out of this loop.
+ const MachineLoop *ML = MLI->getLoopFor(BB);
+ if (ML && ML->getHeader()->isLandingPad()) continue;
+
// 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.
for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
E = BB->livein_end(); I != E; ++I) {
unsigned Reg = *I;
- ++PhysRegDefs[Reg];
- for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS)
- ++PhysRegDefs[*AS];
+ for (const unsigned *AS = TRI->getOverlaps(Reg); *AS; ++AS)
+ PhysRegDefs.set(*AS);
}
+ SpeculationState = SpeculateUnknown;
for (MachineBasicBlock::iterator
MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
MachineInstr *MI = &*MII;
- ProcessMI(MI, PhysRegDefs, StoredFIs, Candidates);
+ ProcessMI(MI, PhysRegDefs, PhysRegClobbers, StoredFIs, Candidates);
}
}
StoredFIs.count(Candidates[i].FI))
continue;
- if (PhysRegDefs[Candidates[i].Def] == 1) {
+ if (!PhysRegClobbers.test(Candidates[i].Def)) {
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()]) {
+ // Allow hoisting of reserved register reads that aren't call preserved.
+ // For example %rip.
+ // IsLoopInvariantInst() already checks MRI->isConstantPhysReg().
+ if (ReservedRegs.test(MO.getReg()))
+ continue;
+ if (PhysRegDefs.test(MO.getReg()) ||
+ PhysRegClobbers.test(MO.getReg())) {
// If it's using a non-loop-invariant register, then it's obviously
// not safe to hoist.
Safe = false;
HoistPostRA(MI, Candidates[i].Def);
}
}
-
- delete[] PhysRegDefs;
}
/// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the current
// Now move the instructions to the predecessor, inserting it before any
// terminator instructions.
- DEBUG({
- dbgs() << "Hoisting " << *MI;
- if (Preheader->getBasicBlock())
- dbgs() << " to MachineBasicBlock "
- << Preheader->getName();
- if (MI->getParent()->getBasicBlock())
- dbgs() << " from MachineBasicBlock "
- << MI->getParent()->getName();
- dbgs() << "\n";
- });
+ DEBUG(dbgs() << "Hoisting to BB#" << Preheader->getNumber() << " from BB#"
+ << MI->getParent()->getNumber() << ": " << *MI);
// Splice the instruction to the preheader.
MachineBasicBlock *MBB = MI->getParent();
Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
- // Add register to livein list to all the BBs in the current loop since a
+ // 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);
Changed = true;
}
-/// HoistRegion - Walk the specified region of the CFG (defined by all blocks
-/// dominated by the specified block, and that are in the current loop) in depth
-/// first order w.r.t the DominatorTree. This allows us to visit definitions
-/// before uses, allowing us to hoist a loop body in one pass without iteration.
-///
-void MachineLICM::HoistRegion(MachineDomTreeNode *N, bool IsHeader) {
- assert(N != 0 && "Null dominator tree node?");
- MachineBasicBlock *BB = N->getBlock();
+// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
+// If not then a load from this mbb may not be safe to hoist.
+bool MachineLICM::IsGuaranteedToExecute(MachineBasicBlock *BB) {
+ if (SpeculationState != SpeculateUnknown)
+ return SpeculationState == SpeculateFalse;
+
+ if (BB != CurLoop->getHeader()) {
+ // Check loop exiting blocks.
+ SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks;
+ CurLoop->getExitingBlocks(CurrentLoopExitingBlocks);
+ for (unsigned i = 0, e = CurrentLoopExitingBlocks.size(); i != e; ++i)
+ if (!DT->dominates(BB, CurrentLoopExitingBlocks[i])) {
+ SpeculationState = SpeculateTrue;
+ return false;
+ }
+ }
- // If this subregion is not in the top level loop at all, exit.
- if (!CurLoop->contains(BB)) return;
+ SpeculationState = SpeculateFalse;
+ return true;
+}
- MachineBasicBlock *Preheader = getCurPreheader();
- if (!Preheader)
+void MachineLICM::EnterScope(MachineBasicBlock *MBB) {
+ DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
+
+ // Remember livein register pressure.
+ BackTrace.push_back(RegPressure);
+}
+
+void MachineLICM::ExitScope(MachineBasicBlock *MBB) {
+ DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
+ BackTrace.pop_back();
+}
+
+/// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
+/// dominator tree node if its a leaf or all of its children are done. Walk
+/// up the dominator tree to destroy ancestors which are now done.
+void MachineLICM::ExitScopeIfDone(MachineDomTreeNode *Node,
+ DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
+ DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
+ if (OpenChildren[Node])
return;
- if (IsHeader) {
+ // Pop scope.
+ ExitScope(Node->getBlock());
+
+ // Now traverse upwards to pop ancestors whose offsprings are all done.
+ while (MachineDomTreeNode *Parent = ParentMap[Node]) {
+ unsigned Left = --OpenChildren[Parent];
+ if (Left != 0)
+ break;
+ ExitScope(Parent->getBlock());
+ Node = Parent;
+ }
+}
+
+/// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all
+/// blocks dominated by the specified header block, and that are in the
+/// current loop) in depth first order w.r.t the DominatorTree. This allows
+/// us to visit definitions before uses, allowing us to hoist a loop body in
+/// one pass without iteration.
+///
+void MachineLICM::HoistOutOfLoop(MachineDomTreeNode *HeaderN) {
+ SmallVector<MachineDomTreeNode*, 32> Scopes;
+ SmallVector<MachineDomTreeNode*, 8> WorkList;
+ DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
+ DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
+
+ // Perform a DFS walk to determine the order of visit.
+ WorkList.push_back(HeaderN);
+ do {
+ MachineDomTreeNode *Node = WorkList.pop_back_val();
+ assert(Node != 0 && "Null dominator tree node?");
+ MachineBasicBlock *BB = Node->getBlock();
+
+ // If the header of the loop containing this basic block is a landing pad,
+ // then don't try to hoist instructions out of this loop.
+ const MachineLoop *ML = MLI->getLoopFor(BB);
+ if (ML && ML->getHeader()->isLandingPad())
+ continue;
+
+ // If this subregion is not in the top level loop at all, exit.
+ if (!CurLoop->contains(BB))
+ continue;
+
+ Scopes.push_back(Node);
+ const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
+ unsigned NumChildren = Children.size();
+
+ // 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)
+ NumChildren = 0;
+
+ OpenChildren[Node] = NumChildren;
+ // Add children in reverse order as then the next popped worklist node is
+ // the first child of this node. This means we ultimately traverse the
+ // DOM tree in exactly the same order as if we'd recursed.
+ for (int i = (int)NumChildren-1; i >= 0; --i) {
+ MachineDomTreeNode *Child = Children[i];
+ ParentMap[Child] = Node;
+ WorkList.push_back(Child);
+ }
+ } while (!WorkList.empty());
+
+ if (Scopes.size() != 0) {
+ MachineBasicBlock *Preheader = getCurPreheader();
+ if (!Preheader)
+ return;
+
// Compute registers which are livein into the loop headers.
RegSeen.clear();
BackTrace.clear();
InitRegPressure(Preheader);
}
- // Remember livein register pressure.
- BackTrace.push_back(RegPressure);
+ // Now perform LICM.
+ for (unsigned i = 0, e = Scopes.size(); i != e; ++i) {
+ MachineDomTreeNode *Node = Scopes[i];
+ MachineBasicBlock *MBB = Node->getBlock();
- for (MachineBasicBlock::iterator
- MII = BB->begin(), E = BB->end(); MII != E; ) {
- MachineBasicBlock::iterator NextMII = MII; ++NextMII;
- MachineInstr *MI = &*MII;
- if (!Hoist(MI, Preheader))
- UpdateRegPressure(MI);
- MII = NextMII;
- }
+ MachineBasicBlock *Preheader = getCurPreheader();
+ if (!Preheader)
+ continue;
- // 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]);
- }
+ EnterScope(MBB);
- BackTrace.pop_back();
+ // Process the block
+ SpeculationState = SpeculateUnknown;
+ for (MachineBasicBlock::iterator
+ MII = MBB->begin(), E = MBB->end(); MII != E; ) {
+ MachineBasicBlock::iterator NextMII = MII; ++NextMII;
+ MachineInstr *MI = &*MII;
+ if (!Hoist(MI, Preheader))
+ UpdateRegPressure(MI);
+ MII = NextMII;
+ }
+
+ // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
+ ExitScopeIfDone(Node, OpenChildren, ParentMap);
+ }
}
static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
unsigned &RCId, unsigned &RCCost) const {
const TargetRegisterClass *RC = MRI->getRegClass(Reg);
EVT VT = *RC->vt_begin();
- if (VT == MVT::untyped) {
+ if (VT == MVT::Untyped) {
RCId = RC->getID();
RCCost = 1;
} else {
RCCost = TLI->getRepRegClassCostFor(VT);
}
}
-
+
/// InitRegPressure - Find all virtual register references that are liveout of
/// the preheader to initialize the starting "register pressure". Note this
/// does not count live through (livein but not used) registers.
}
}
+/// isLoadFromGOTOrConstantPool - Return true if this machine instruction
+/// loads from global offset table or constant pool.
+static bool isLoadFromGOTOrConstantPool(MachineInstr &MI) {
+ assert (MI.mayLoad() && "Expected MI that loads!");
+ for (MachineInstr::mmo_iterator I = MI.memoperands_begin(),
+ E = MI.memoperands_end(); I != E; ++I) {
+ if (const Value *V = (*I)->getValue()) {
+ if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V))
+ if (PSV == PSV->getGOT() || PSV == PSV->getConstantPool())
+ return true;
+ }
+ }
+ return false;
+}
+
/// 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 DontMoveAcrossStore = true;
if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore))
return false;
-
+
+ // If it is load then check if it is guaranteed to execute by making sure that
+ // it dominates all exiting blocks. If it doesn't, then there is a path out of
+ // the loop which does not execute this load, so we can't hoist it. Loads
+ // from constant memory are not safe to speculate all the time, for example
+ // indexed load from a jump table.
+ // Stores and side effects are already checked by isSafeToMove.
+ if (I.mayLoad() && !isLoadFromGOTOrConstantPool(I) &&
+ !IsGuaranteedToExecute(I.getParent()))
+ return false;
+
return true;
}
/// invariant. I.e., all virtual register operands are defined outside of the
/// loop, physical registers aren't accessed explicitly, and there are no side
/// effects that aren't captured by the operands or other flags.
-///
+///
bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) {
if (!IsLICMCandidate(I))
return false;
// If the physreg has no defs anywhere, it's just an ambient register
// and we can freely move its uses. Alternatively, if it's allocatable,
// it could get allocated to something with a def during allocation.
- if (!MRI->def_empty(Reg))
+ if (!MRI->isConstantPhysReg(Reg, *I.getParent()->getParent()))
return false;
- if (AllocatableSet.test(Reg))
- return false;
- // Check for a def among the register's aliases too.
- for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
- unsigned AliasReg = *Alias;
- if (!MRI->def_empty(AliasReg))
- return false;
- if (AllocatableSet.test(AliasReg))
- return false;
- }
// Otherwise it's safe to move.
continue;
} else if (!MO.isDead()) {
/// IsCheapInstruction - Return true if the instruction is marked "cheap" or
/// the operand latency between its def and a use is one or less.
bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const {
- if (MI.getDesc().isAsCheapAsAMove() || MI.isCopyLike())
+ if (MI.isAsCheapAsAMove() || MI.isCopyLike())
return true;
if (!InstrItins || InstrItins->isEmpty())
return false;
bool MachineLICM::CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost) {
for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
CI != CE; ++CI) {
- if (CI->second <= 0)
+ if (CI->second <= 0)
continue;
unsigned RCId = CI->first;
+ unsigned Limit = RegLimit[RCId];
+ int Cost = CI->second;
for (unsigned i = BackTrace.size(); i != 0; --i) {
SmallVector<unsigned, 8> &RP = BackTrace[i-1];
- if (RP[RCId] + CI->second >= RegLimit[RCId])
+ if (RP[RCId] + Cost >= Limit)
return true;
}
}
return false;
} else {
// Estimate register pressure to determine whether to LICM the instruction.
- // In low register pressure situation, we can be more aggressive about
+ // In low register pressure situation, we can be more aggressive about
// hoisting. Also, favors hoisting long latency instructions even in
// moderately high pressure situation.
// FIXME: If there are long latency loop-invariant instructions inside the
return true;
}
+ // Do not "speculate" in high register pressure situation. If an
+ // instruction is not guaranteed to be executed in the loop, it's best to be
+ // conservative.
+ if (AvoidSpeculation &&
+ (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI)))
+ return false;
+
// High register pressure situation, only hoist if the instruction is going to
// be remat'ed.
if (!TII->isTriviallyReMaterializable(&MI, AA) &&
MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) {
// Don't unfold simple loads.
- if (MI->getDesc().canFoldAsLoad())
+ if (MI->canFoldAsLoad())
return 0;
// If not, we may be able to unfold a load and hoist that.
assert(NewMIs.size() == 2 &&
"Unfolded a load into multiple instructions!");
MachineBasicBlock *MBB = MI->getParent();
- MBB->insert(MI, NewMIs[0]);
- MBB->insert(MI, NewMIs[1]);
+ MachineBasicBlock::iterator Pos = MI;
+ MBB->insert(Pos, NewMIs[0]);
+ MBB->insert(Pos, NewMIs[1]);
// If unfolding produced a load that wasn't loop-invariant or profitable to
// hoist, discard the new instructions and bail.
if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) {
// Replace virtual registers defined by MI by their counterparts defined
// by Dup.
+ SmallVector<unsigned, 2> Defs;
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
"Instructions with different phys regs are not identical!");
if (MO.isReg() && MO.isDef() &&
- !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
- MRI->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg());
- MRI->clearKillFlags(Dup->getOperand(i).getReg());
+ !TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
+ Defs.push_back(i);
+ }
+
+ SmallVector<const TargetRegisterClass*, 2> OrigRCs;
+ for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
+ unsigned Idx = Defs[i];
+ unsigned Reg = MI->getOperand(Idx).getReg();
+ unsigned DupReg = Dup->getOperand(Idx).getReg();
+ OrigRCs.push_back(MRI->getRegClass(DupReg));
+
+ if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) {
+ // Restore old RCs if more than one defs.
+ for (unsigned j = 0; j != i; ++j)
+ MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]);
+ return false;
}
}
+
+ for (unsigned i = 0, e = Defs.size(); i != e; ++i) {
+ unsigned Idx = Defs[i];
+ unsigned Reg = MI->getOperand(Idx).getReg();
+ unsigned DupReg = Dup->getOperand(Idx).getReg();
+ MRI->replaceRegWith(Reg, DupReg);
+ MRI->clearKillFlags(DupReg);
+ }
+
MI->eraseFromParent();
++NumCSEed;
return true;
return false;
}
+/// MayCSE - Return true if the given instruction will be CSE'd if it's
+/// hoisted out of the loop.
+bool MachineLICM::MayCSE(MachineInstr *MI) {
+ unsigned Opcode = MI->getOpcode();
+ DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
+ CI = CSEMap.find(Opcode);
+ // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
+ // the undef property onto uses.
+ if (CI == CSEMap.end() || MI->isImplicitDef())
+ return false;
+
+ return LookForDuplicate(MI, CI->second) != 0;
+}
+
/// Hoist - When an instruction is found to use only loop invariant operands
/// that are safe to hoist, this instruction is called to do the dirty work.
///