#define DEBUG_TYPE "machine-cse"
#include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/MachineDominators.h"
-#include "llvm/CodeGen/MachineInstr.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/Analysis/AliasAnalysis.h"
-#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/ScopedHashTable.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/RecyclingAllocator.h"
+#include "llvm/Target/TargetInstrInfo.h"
using namespace llvm;
STATISTIC(NumCoalesces, "Number of copies coalesced");
bool PerformTrivialCoalescing(MachineInstr *MI, MachineBasicBlock *MBB);
bool isPhysDefTriviallyDead(unsigned Reg,
MachineBasicBlock::const_iterator I,
- MachineBasicBlock::const_iterator E) const ;
+ MachineBasicBlock::const_iterator E) const;
bool hasLivePhysRegDefUses(const MachineInstr *MI,
const MachineBasicBlock *MBB,
SmallSet<unsigned,8> &PhysRefs,
- SmallVector<unsigned,2> &PhysDefs) const;
+ SmallVectorImpl<unsigned> &PhysDefs,
+ bool &PhysUseDef) const;
bool PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
SmallSet<unsigned,8> &PhysRefs,
- SmallVector<unsigned,2> &PhysDefs,
+ SmallVectorImpl<unsigned> &PhysDefs,
bool &NonLocal) const;
bool isCSECandidate(MachineInstr *MI);
bool isProfitableToCSE(unsigned CSReg, unsigned Reg,
void ExitScope(MachineBasicBlock *MBB);
bool ProcessBlock(MachineBasicBlock *MBB);
void ExitScopeIfDone(MachineDomTreeNode *Node,
- DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
- DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap);
+ DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren);
bool PerformCSE(MachineDomTreeNode *Node);
};
} // end anonymous namespace
// deleted.
continue;
MachineInstr *DefMI = MRI->getVRegDef(Reg);
- if (DefMI->getParent() != MBB)
- continue;
if (!DefMI->isCopy())
continue;
unsigned SrcReg = DefMI->getOperand(1).getReg();
bool SeenDef = false;
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = I->getOperand(i);
+ if (MO.isRegMask() && MO.clobbersPhysReg(Reg))
+ SeenDef = true;
if (!MO.isReg() || !MO.getReg())
continue;
if (!TRI->regsOverlap(MO.getReg(), Reg))
bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI,
const MachineBasicBlock *MBB,
SmallSet<unsigned,8> &PhysRefs,
- SmallVector<unsigned,2> &PhysDefs) const{
+ SmallVectorImpl<unsigned> &PhysDefs,
+ bool &PhysUseDef) const{
+ // First, add all uses to PhysRefs.
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg() || MO.isDef())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (TargetRegisterInfo::isVirtualRegister(Reg))
+ continue;
+ // Reading constant physregs is ok.
+ if (!MRI->isConstantPhysReg(Reg, *MBB->getParent()))
+ for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
+ PhysRefs.insert(*AI);
+ }
+
+ // Next, collect all defs into PhysDefs. If any is already in PhysRefs
+ // (which currently contains only uses), set the PhysUseDef flag.
+ PhysUseDef = false;
MachineBasicBlock::const_iterator I = MI; I = llvm::next(I);
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
- if (!MO.isReg())
+ if (!MO.isReg() || !MO.isDef())
continue;
unsigned Reg = MO.getReg();
if (!Reg)
continue;
if (TargetRegisterInfo::isVirtualRegister(Reg))
continue;
+ // Check against PhysRefs even if the def is "dead".
+ if (PhysRefs.count(Reg))
+ PhysUseDef = true;
// If the def is dead, it's ok. But the def may not marked "dead". That's
// common since this pass is run before livevariables. We can scan
// forward a few instructions and check if it is obviously dead.
- if (MO.isDef() &&
- (MO.isDead() || isPhysDefTriviallyDead(Reg, I, MBB->end())))
- continue;
- PhysRefs.insert(Reg);
- if (MO.isDef())
+ if (!MO.isDead() && !isPhysDefTriviallyDead(Reg, I, MBB->end()))
PhysDefs.push_back(Reg);
- for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias)
- PhysRefs.insert(*Alias);
}
+ // Finally, add all defs to PhysRefs as well.
+ for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i)
+ for (MCRegAliasIterator AI(PhysDefs[i], TRI, true); AI.isValid(); ++AI)
+ PhysRefs.insert(*AI);
+
return !PhysRefs.empty();
}
bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
SmallSet<unsigned,8> &PhysRefs,
- SmallVector<unsigned,2> &PhysDefs,
+ SmallVectorImpl<unsigned> &PhysDefs,
bool &NonLocal) const {
// For now conservatively returns false if the common subexpression is
// not in the same basic block as the given instruction. The only exception
return false;
for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i) {
- if (TRI->isInAllocatableClass(PhysDefs[i]))
- // Avoid extending live range of physical registers unless
- // they are unallocatable.
+ if (MRI->isAllocatable(PhysDefs[i]) || MRI->isReserved(PhysDefs[i]))
+ // Avoid extending live range of physical registers if they are
+ //allocatable or reserved.
return false;
}
CrossMBB = true;
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = I->getOperand(i);
+ // RegMasks go on instructions like calls that clobber lots of physregs.
+ // Don't attempt to CSE across such an instruction.
+ if (MO.isRegMask())
+ return false;
if (!MO.isReg() || !MO.isDef())
continue;
unsigned MOReg = MO.getReg();
MachineInstr *CSMI, MachineInstr *MI) {
// FIXME: Heuristics that works around the lack the live range splitting.
+ // If CSReg is used at all uses of Reg, CSE should not increase register
+ // pressure of CSReg.
+ bool MayIncreasePressure = true;
+ if (TargetRegisterInfo::isVirtualRegister(CSReg) &&
+ TargetRegisterInfo::isVirtualRegister(Reg)) {
+ MayIncreasePressure = false;
+ SmallPtrSet<MachineInstr*, 8> CSUses;
+ for (MachineRegisterInfo::use_nodbg_iterator I =MRI->use_nodbg_begin(CSReg),
+ E = MRI->use_nodbg_end(); I != E; ++I) {
+ MachineInstr *Use = &*I;
+ CSUses.insert(Use);
+ }
+ for (MachineRegisterInfo::use_nodbg_iterator I = MRI->use_nodbg_begin(Reg),
+ E = MRI->use_nodbg_end(); I != E; ++I) {
+ MachineInstr *Use = &*I;
+ if (!CSUses.count(Use)) {
+ MayIncreasePressure = true;
+ break;
+ }
+ }
+ }
+ if (!MayIncreasePressure) return true;
+
// Heuristics #1: Don't CSE "cheap" computation if the def is not local or in
// an immediate predecessor. We don't want to increase register pressure and
// end up causing other computation to be spilled.
DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(MBB);
assert(SI != ScopeMap.end());
- ScopeMap.erase(SI);
delete SI->second;
+ ScopeMap.erase(SI);
}
bool MachineCSE::ProcessBlock(MachineBasicBlock *MBB) {
bool Changed = false;
SmallVector<std::pair<unsigned, unsigned>, 8> CSEPairs;
+ SmallVector<unsigned, 2> ImplicitDefsToUpdate;
for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ) {
MachineInstr *MI = &*I;
++I;
// used, then it's not safe to replace it with a common subexpression.
// It's also not safe if the instruction uses physical registers.
bool CrossMBBPhysDef = false;
- SmallSet<unsigned,8> PhysRefs;
+ SmallSet<unsigned, 8> PhysRefs;
SmallVector<unsigned, 2> PhysDefs;
- if (FoundCSE && hasLivePhysRegDefUses(MI, MBB, PhysRefs, PhysDefs)) {
+ bool PhysUseDef = false;
+ if (FoundCSE && hasLivePhysRegDefUses(MI, MBB, PhysRefs,
+ PhysDefs, PhysUseDef)) {
FoundCSE = false;
// ... Unless the CS is local or is in the sole predecessor block
// and it also defines the physical register which is not clobbered
// in between and the physical register uses were not clobbered.
- unsigned CSVN = VNT.lookup(MI);
- MachineInstr *CSMI = Exps[CSVN];
- if (PhysRegDefsReach(CSMI, MI, PhysRefs, PhysDefs, CrossMBBPhysDef))
- FoundCSE = true;
+ // This can never be the case if the instruction both uses and
+ // defines the same physical register, which was detected above.
+ if (!PhysUseDef) {
+ unsigned CSVN = VNT.lookup(MI);
+ MachineInstr *CSMI = Exps[CSVN];
+ if (PhysRegDefsReach(CSMI, MI, PhysRefs, PhysDefs, CrossMBBPhysDef))
+ FoundCSE = true;
+ }
}
if (!FoundCSE) {
// Check if it's profitable to perform this CSE.
bool DoCSE = true;
- unsigned NumDefs = MI->getDesc().getNumDefs();
+ unsigned NumDefs = MI->getDesc().getNumDefs() +
+ MI->getDesc().getNumImplicitDefs();
+
for (unsigned i = 0, e = MI->getNumOperands(); NumDefs && i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg() || !MO.isDef())
continue;
unsigned OldReg = MO.getReg();
unsigned NewReg = CSMI->getOperand(i).getReg();
- if (OldReg == NewReg)
+
+ // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
+ // we should make sure it is not dead at CSMI.
+ if (MO.isImplicit() && !MO.isDead() && CSMI->getOperand(i).isDead())
+ ImplicitDefsToUpdate.push_back(i);
+ if (OldReg == NewReg) {
+ --NumDefs;
continue;
+ }
assert(TargetRegisterInfo::isVirtualRegister(OldReg) &&
TargetRegisterInfo::isVirtualRegister(NewReg) &&
"Do not CSE physical register defs!");
if (!isProfitableToCSE(NewReg, OldReg, CSMI, MI)) {
+ DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n");
DoCSE = false;
break;
}
// within the register class of the new instruction.
const TargetRegisterClass *OldRC = MRI->getRegClass(OldReg);
if (!MRI->constrainRegClass(NewReg, OldRC)) {
+ DEBUG(dbgs() << "*** Not the same register class, avoid CSE!\n");
DoCSE = false;
break;
}
MRI->clearKillFlags(CSEPairs[i].second);
}
+ // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
+ // we should make sure it is not dead at CSMI.
+ for (unsigned i = 0, e = ImplicitDefsToUpdate.size(); i != e; ++i)
+ CSMI->getOperand(ImplicitDefsToUpdate[i]).setIsDead(false);
+
if (CrossMBBPhysDef) {
// Add physical register defs now coming in from a predecessor to MBB
// livein list.
++NumCommutes;
Changed = true;
} else {
- DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n");
VNT.insert(MI, CurrVN++);
Exps.push_back(MI);
}
CSEPairs.clear();
+ ImplicitDefsToUpdate.clear();
}
return Changed;
/// up the dominator tree to destroy ancestors which are now done.
void
MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node,
- DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
- DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
+ DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren) {
if (OpenChildren[Node])
return;
ExitScope(Node->getBlock());
// Now traverse upwards to pop ancestors whose offsprings are all done.
- while (MachineDomTreeNode *Parent = ParentMap[Node]) {
+ while (MachineDomTreeNode *Parent = Node->getIDom()) {
unsigned Left = --OpenChildren[Parent];
if (Left != 0)
break;
bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) {
SmallVector<MachineDomTreeNode*, 32> Scopes;
SmallVector<MachineDomTreeNode*, 8> WorkList;
- DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
CurrVN = 0;
OpenChildren[Node] = NumChildren;
for (unsigned i = 0; i != NumChildren; ++i) {
MachineDomTreeNode *Child = Children[i];
- ParentMap[Child] = Node;
WorkList.push_back(Child);
}
} while (!WorkList.empty());
EnterScope(MBB);
Changed |= ProcessBlock(MBB);
// If it's a leaf node, it's done. Traverse upwards to pop ancestors.
- ExitScopeIfDone(Node, OpenChildren, ParentMap);
+ ExitScopeIfDone(Node, OpenChildren);
}
return Changed;