bool PHIElimination::SplitPHIEdges(MachineFunction &MF,
MachineBasicBlock &MBB,
LiveVariables &LV,
MachineLoopInfo *MLI) {
if (MBB.empty() || !MBB.front().isPHI() || MBB.isLandingPad())
return false; // Quick exit for basic blocks without PHIs.
bool Changed = false;
for (MachineBasicBlock::const_iterator BBI = MBB.begin(), BBE = MBB.end();
BBI != BBE && BBI->isPHI(); ++BBI) {
for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) {
unsigned Reg = BBI->getOperand(i).getReg();
MachineBasicBlock *PreMBB = BBI->getOperand(i+1).getMBB();
// We break edges when registers are live out from the predecessor block
// (not considering PHI nodes). If the register is live in to this block
// anyway, we would gain nothing from splitting.
// Avoid splitting backedges of loops. It would introduce small
// out-of-line blocks into the loop which is very bad for code placement.
if (PreMBB != &MBB &&
!LV.isLiveIn(Reg, MBB) && LV.isLiveOut(Reg, *PreMBB)) {
if (!MLI ||
!(MLI->getLoopFor(PreMBB) == MLI->getLoopFor(&MBB) &&
MLI->isLoopHeader(&MBB))) {
if (PreMBB->SplitCriticalEdge(&MBB, this)) {
Changed = true;
++NumCriticalEdgesSplit;
}
}
}
}
}
return Changed;
}
void clang::CheckDeadStores(CFG &cfg, LiveVariables &L, ParentMap &pmap,
BugReporter& BR) {
FindEscaped FS(&cfg);
FS.getCFG().VisitBlockStmts(FS);
DeadStoreObs A(BR.getContext(), BR, pmap, FS.Escaped);
L.runOnAllBlocks(cfg, &A);
}
bool llvm::PHIElimination::SplitPHIEdges(MachineFunction &MF,
MachineBasicBlock &MBB,
LiveVariables &LV) {
if (MBB.empty() || !MBB.front().isPHI() || MBB.isLandingPad())
return false; // Quick exit for basic blocks without PHIs.
for (MachineBasicBlock::const_iterator BBI = MBB.begin(), BBE = MBB.end();
BBI != BBE && BBI->isPHI(); ++BBI) {
for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) {
unsigned Reg = BBI->getOperand(i).getReg();
MachineBasicBlock *PreMBB = BBI->getOperand(i+1).getMBB();
// We break edges when registers are live out from the predecessor block
// (not considering PHI nodes). If the register is live in to this block
// anyway, we would gain nothing from splitting.
if (!LV.isLiveIn(Reg, MBB) && LV.isLiveOut(Reg, *PreMBB))
SplitCriticalEdge(PreMBB, &MBB);
}
}
return true;
}
/// LowerAtomicPHINode - Lower the PHI node at the top of the specified block,
/// under the assuption that it needs to be lowered in a way that supports
/// atomic execution of PHIs. This lowering method is always correct all of the
/// time.
///
void PHIElimination::LowerAtomicPHINode(
MachineBasicBlock &MBB,
MachineBasicBlock::iterator AfterPHIsIt) {
++NumAtomic;
// Unlink the PHI node from the basic block, but don't delete the PHI yet.
MachineInstr *MPhi = MBB.remove(MBB.begin());
unsigned NumSrcs = (MPhi->getNumOperands() - 1) / 2;
unsigned DestReg = MPhi->getOperand(0).getReg();
assert(MPhi->getOperand(0).getSubReg() == 0 && "Can't handle sub-reg PHIs");
bool isDead = MPhi->getOperand(0).isDead();
// Create a new register for the incoming PHI arguments.
MachineFunction &MF = *MBB.getParent();
unsigned IncomingReg = 0;
bool reusedIncoming = false; // Is IncomingReg reused from an earlier PHI?
// Insert a register to register copy at the top of the current block (but
// after any remaining phi nodes) which copies the new incoming register
// into the phi node destination.
const TargetInstrInfo *TII = MF.getTarget().getInstrInfo();
if (isSourceDefinedByImplicitDef(MPhi, MRI))
// If all sources of a PHI node are implicit_def, just emit an
// implicit_def instead of a copy.
BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(),
TII->get(TargetOpcode::IMPLICIT_DEF), DestReg);
else {
// Can we reuse an earlier PHI node? This only happens for critical edges,
// typically those created by tail duplication.
unsigned &entry = LoweredPHIs[MPhi];
if (entry) {
// An identical PHI node was already lowered. Reuse the incoming register.
IncomingReg = entry;
reusedIncoming = true;
++NumReused;
DEBUG(dbgs() << "Reusing " << PrintReg(IncomingReg) << " for " << *MPhi);
} else {
const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(DestReg);
entry = IncomingReg = MF.getRegInfo().createVirtualRegister(RC);
}
BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(),
TII->get(TargetOpcode::COPY), DestReg)
.addReg(IncomingReg);
}
// Update live variable information if there is any.
LiveVariables *LV = getAnalysisIfAvailable<LiveVariables>();
if (LV) {
MachineInstr *PHICopy = prior(AfterPHIsIt);
if (IncomingReg) {
LiveVariables::VarInfo &VI = LV->getVarInfo(IncomingReg);
// Increment use count of the newly created virtual register.
VI.NumUses++;
LV->setPHIJoin(IncomingReg);
// When we are reusing the incoming register, it may already have been
// killed in this block. The old kill will also have been inserted at
// AfterPHIsIt, so it appears before the current PHICopy.
if (reusedIncoming)
if (MachineInstr *OldKill = VI.findKill(&MBB)) {
DEBUG(dbgs() << "Remove old kill from " << *OldKill);
LV->removeVirtualRegisterKilled(IncomingReg, OldKill);
DEBUG(MBB.dump());
}
// Add information to LiveVariables to know that the incoming value is
// killed. Note that because the value is defined in several places (once
// each for each incoming block), the "def" block and instruction fields
// for the VarInfo is not filled in.
LV->addVirtualRegisterKilled(IncomingReg, PHICopy);
}
// Since we are going to be deleting the PHI node, if it is the last use of
// any registers, or if the value itself is dead, we need to move this
// information over to the new copy we just inserted.
LV->removeVirtualRegistersKilled(MPhi);
// If the result is dead, update LV.
if (isDead) {
LV->addVirtualRegisterDead(DestReg, PHICopy);
LV->removeVirtualRegisterDead(DestReg, MPhi);
}
}
// Adjust the VRegPHIUseCount map to account for the removal of this PHI node.
for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2)
--VRegPHIUseCount[BBVRegPair(MPhi->getOperand(i+1).getMBB()->getNumber(),
MPhi->getOperand(i).getReg())];
// Now loop over all of the incoming arguments, changing them to copy into the
// IncomingReg register in the corresponding predecessor basic block.
SmallPtrSet<MachineBasicBlock*, 8> MBBsInsertedInto;
for (int i = NumSrcs - 1; i >= 0; --i) {
unsigned SrcReg = MPhi->getOperand(i*2+1).getReg();
unsigned SrcSubReg = MPhi->getOperand(i*2+1).getSubReg();
assert(TargetRegisterInfo::isVirtualRegister(SrcReg) &&
"Machine PHI Operands must all be virtual registers!");
// Get the MachineBasicBlock equivalent of the BasicBlock that is the source
// path the PHI.
MachineBasicBlock &opBlock = *MPhi->getOperand(i*2+2).getMBB();
// If source is defined by an implicit def, there is no need to insert a
// copy.
MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
if (DefMI->isImplicitDef()) {
ImpDefs.insert(DefMI);
continue;
}
// Check to make sure we haven't already emitted the copy for this block.
// This can happen because PHI nodes may have multiple entries for the same
// basic block.
if (!MBBsInsertedInto.insert(&opBlock))
continue; // If the copy has already been emitted, we're done.
// Find a safe location to insert the copy, this may be the first terminator
// in the block (or end()).
MachineBasicBlock::iterator InsertPos =
findPHICopyInsertPoint(&opBlock, &MBB, SrcReg);
// Insert the copy.
if (!reusedIncoming && IncomingReg)
BuildMI(opBlock, InsertPos, MPhi->getDebugLoc(),
TII->get(TargetOpcode::COPY), IncomingReg).addReg(SrcReg, 0, SrcSubReg);
// Now update live variable information if we have it. Otherwise we're done
if (!LV) continue;
// We want to be able to insert a kill of the register if this PHI (aka, the
// copy we just inserted) is the last use of the source value. Live
// variable analysis conservatively handles this by saying that the value is
// live until the end of the block the PHI entry lives in. If the value
// really is dead at the PHI copy, there will be no successor blocks which
// have the value live-in.
// Also check to see if this register is in use by another PHI node which
// has not yet been eliminated. If so, it will be killed at an appropriate
// point later.
// Is it used by any PHI instructions in this block?
bool ValueIsUsed = VRegPHIUseCount[BBVRegPair(opBlock.getNumber(), SrcReg)];
// Okay, if we now know that the value is not live out of the block, we can
// add a kill marker in this block saying that it kills the incoming value!
if (!ValueIsUsed && !LV->isLiveOut(SrcReg, opBlock)) {
// In our final twist, we have to decide which instruction kills the
// register. In most cases this is the copy, however, the first
// terminator instruction at the end of the block may also use the value.
// In this case, we should mark *it* as being the killing block, not the
// copy.
MachineBasicBlock::iterator KillInst;
MachineBasicBlock::iterator Term = opBlock.getFirstTerminator();
if (Term != opBlock.end() && Term->readsRegister(SrcReg)) {
KillInst = Term;
// Check that no other terminators use values.
#ifndef NDEBUG
for (MachineBasicBlock::iterator TI = llvm::next(Term);
TI != opBlock.end(); ++TI) {
if (TI->isDebugValue())
continue;
assert(!TI->readsRegister(SrcReg) &&
"Terminator instructions cannot use virtual registers unless"
"they are the first terminator in a block!");
}
#endif
} else if (reusedIncoming || !IncomingReg) {
// We may have to rewind a bit if we didn't insert a copy this time.
KillInst = Term;
while (KillInst != opBlock.begin()) {
--KillInst;
if (KillInst->isDebugValue())
continue;
if (KillInst->readsRegister(SrcReg))
break;
}
} else {
// We just inserted this copy.
KillInst = prior(InsertPos);
}
assert(KillInst->readsRegister(SrcReg) && "Cannot find kill instruction");
// Finally, mark it killed.
LV->addVirtualRegisterKilled(SrcReg, KillInst);
// This vreg no longer lives all of the way through opBlock.
unsigned opBlockNum = opBlock.getNumber();
LV->getVarInfo(SrcReg).AliveBlocks.reset(opBlockNum);
}
}
// Really delete the PHI instruction now, if it is not in the LoweredPHIs map.
if (reusedIncoming || !IncomingReg)
MF.DeleteMachineInstr(MPhi);
}
/// LowerAtomicPHINode - Lower the PHI node at the top of the specified block,
/// under the assuption that it needs to be lowered in a way that supports
/// atomic execution of PHIs. This lowering method is always correct all of the
/// time.
///
void llvm::PHIElimination::LowerAtomicPHINode(
MachineBasicBlock &MBB,
MachineBasicBlock::iterator AfterPHIsIt) {
// Unlink the PHI node from the basic block, but don't delete the PHI yet.
MachineInstr *MPhi = MBB.remove(MBB.begin());
unsigned NumSrcs = (MPhi->getNumOperands() - 1) / 2;
unsigned DestReg = MPhi->getOperand(0).getReg();
bool isDead = MPhi->getOperand(0).isDead();
// Create a new register for the incoming PHI arguments.
MachineFunction &MF = *MBB.getParent();
const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(DestReg);
unsigned IncomingReg = 0;
// Insert a register to register copy at the top of the current block (but
// after any remaining phi nodes) which copies the new incoming register
// into the phi node destination.
const TargetInstrInfo *TII = MF.getTarget().getInstrInfo();
if (isSourceDefinedByImplicitDef(MPhi, MRI))
// If all sources of a PHI node are implicit_def, just emit an
// implicit_def instead of a copy.
BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(),
TII->get(TargetInstrInfo::IMPLICIT_DEF), DestReg);
else {
IncomingReg = MF.getRegInfo().createVirtualRegister(RC);
TII->copyRegToReg(MBB, AfterPHIsIt, DestReg, IncomingReg, RC, RC);
}
// Record PHI def.
assert(!hasPHIDef(DestReg) && "Vreg has multiple phi-defs?");
PHIDefs[DestReg] = &MBB;
// Update live variable information if there is any.
LiveVariables *LV = getAnalysisIfAvailable<LiveVariables>();
if (LV) {
MachineInstr *PHICopy = prior(AfterPHIsIt);
if (IncomingReg) {
// Increment use count of the newly created virtual register.
LV->getVarInfo(IncomingReg).NumUses++;
// Add information to LiveVariables to know that the incoming value is
// killed. Note that because the value is defined in several places (once
// each for each incoming block), the "def" block and instruction fields
// for the VarInfo is not filled in.
LV->addVirtualRegisterKilled(IncomingReg, PHICopy);
}
// Since we are going to be deleting the PHI node, if it is the last use of
// any registers, or if the value itself is dead, we need to move this
// information over to the new copy we just inserted.
LV->removeVirtualRegistersKilled(MPhi);
// If the result is dead, update LV.
if (isDead) {
LV->addVirtualRegisterDead(DestReg, PHICopy);
LV->removeVirtualRegisterDead(DestReg, MPhi);
}
}
// Adjust the VRegPHIUseCount map to account for the removal of this PHI node.
for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2)
--VRegPHIUseCount[BBVRegPair(MPhi->getOperand(i + 1).getMBB(),
MPhi->getOperand(i).getReg())];
// Now loop over all of the incoming arguments, changing them to copy into the
// IncomingReg register in the corresponding predecessor basic block.
SmallPtrSet<MachineBasicBlock*, 8> MBBsInsertedInto;
for (int i = NumSrcs - 1; i >= 0; --i) {
unsigned SrcReg = MPhi->getOperand(i*2+1).getReg();
assert(TargetRegisterInfo::isVirtualRegister(SrcReg) &&
"Machine PHI Operands must all be virtual registers!");
// Get the MachineBasicBlock equivalent of the BasicBlock that is the source
// path the PHI.
MachineBasicBlock &opBlock = *MPhi->getOperand(i*2+2).getMBB();
// Record the kill.
PHIKills[SrcReg].insert(&opBlock);
// If source is defined by an implicit def, there is no need to insert a
// copy.
MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
if (DefMI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF) {
ImpDefs.insert(DefMI);
continue;
}
// Check to make sure we haven't already emitted the copy for this block.
// This can happen because PHI nodes may have multiple entries for the same
// basic block.
if (!MBBsInsertedInto.insert(&opBlock))
continue; // If the copy has already been emitted, we're done.
// Find a safe location to insert the copy, this may be the first terminator
// in the block (or end()).
MachineBasicBlock::iterator InsertPos = FindCopyInsertPoint(opBlock, SrcReg);
// Insert the copy.
TII->copyRegToReg(opBlock, InsertPos, IncomingReg, SrcReg, RC, RC);
// Now update live variable information if we have it. Otherwise we're done
if (!LV) continue;
// We want to be able to insert a kill of the register if this PHI (aka, the
// copy we just inserted) is the last use of the source value. Live
// variable analysis conservatively handles this by saying that the value is
// live until the end of the block the PHI entry lives in. If the value
// really is dead at the PHI copy, there will be no successor blocks which
// have the value live-in.
//
// Check to see if the copy is the last use, and if so, update the live
// variables information so that it knows the copy source instruction kills
// the incoming value.
LiveVariables::VarInfo &InRegVI = LV->getVarInfo(SrcReg);
// Loop over all of the successors of the basic block, checking to see if
// the value is either live in the block, or if it is killed in the block.
// Also check to see if this register is in use by another PHI node which
// has not yet been eliminated. If so, it will be killed at an appropriate
// point later.
// Is it used by any PHI instructions in this block?
bool ValueIsLive = VRegPHIUseCount[BBVRegPair(&opBlock, SrcReg)] != 0;
std::vector<MachineBasicBlock*> OpSuccBlocks;
// Otherwise, scan successors, including the BB the PHI node lives in.
for (MachineBasicBlock::succ_iterator SI = opBlock.succ_begin(),
E = opBlock.succ_end(); SI != E && !ValueIsLive; ++SI) {
MachineBasicBlock *SuccMBB = *SI;
// Is it alive in this successor?
unsigned SuccIdx = SuccMBB->getNumber();
if (InRegVI.AliveBlocks.test(SuccIdx)) {
ValueIsLive = true;
break;
}
OpSuccBlocks.push_back(SuccMBB);
}
// Check to see if this value is live because there is a use in a successor
// that kills it.
if (!ValueIsLive) {
switch (OpSuccBlocks.size()) {
case 1: {
MachineBasicBlock *MBB = OpSuccBlocks[0];
for (unsigned i = 0, e = InRegVI.Kills.size(); i != e; ++i)
if (InRegVI.Kills[i]->getParent() == MBB) {
ValueIsLive = true;
break;
}
break;
}
case 2: {
MachineBasicBlock *MBB1 = OpSuccBlocks[0], *MBB2 = OpSuccBlocks[1];
for (unsigned i = 0, e = InRegVI.Kills.size(); i != e; ++i)
if (InRegVI.Kills[i]->getParent() == MBB1 ||
InRegVI.Kills[i]->getParent() == MBB2) {
ValueIsLive = true;
break;
}
break;
}
default:
std::sort(OpSuccBlocks.begin(), OpSuccBlocks.end());
for (unsigned i = 0, e = InRegVI.Kills.size(); i != e; ++i)
if (std::binary_search(OpSuccBlocks.begin(), OpSuccBlocks.end(),
InRegVI.Kills[i]->getParent())) {
ValueIsLive = true;
break;
}
}
}
// Okay, if we now know that the value is not live out of the block, we can
// add a kill marker in this block saying that it kills the incoming value!
if (!ValueIsLive) {
// In our final twist, we have to decide which instruction kills the
// register. In most cases this is the copy, however, the first
// terminator instruction at the end of the block may also use the value.
// In this case, we should mark *it* as being the killing block, not the
// copy.
MachineBasicBlock::iterator KillInst = prior(InsertPos);
MachineBasicBlock::iterator Term = opBlock.getFirstTerminator();
if (Term != opBlock.end()) {
if (Term->readsRegister(SrcReg))
KillInst = Term;
// Check that no other terminators use values.
#ifndef NDEBUG
for (MachineBasicBlock::iterator TI = next(Term); TI != opBlock.end();
++TI) {
assert(!TI->readsRegister(SrcReg) &&
"Terminator instructions cannot use virtual registers unless"
"they are the first terminator in a block!");
}
#endif
}
// Finally, mark it killed.
LV->addVirtualRegisterKilled(SrcReg, KillInst);
// This vreg no longer lives all of the way through opBlock.
unsigned opBlockNum = opBlock.getNumber();
InRegVI.AliveBlocks.reset(opBlockNum);
}
}
// Really delete the PHI instruction now!
MF.DeleteMachineInstr(MPhi);
++NumAtomic;
}
MachineBasicBlock *
MachineBasicBlock::SplitCriticalEdge(MachineBasicBlock *Succ, Pass *P) {
// Splitting the critical edge to a landing pad block is non-trivial. Don't do
// it in this generic function.
if (Succ->isLandingPad())
return nullptr;
MachineFunction *MF = getParent();
DebugLoc dl; // FIXME: this is nowhere
// Performance might be harmed on HW that implements branching using exec mask
// where both sides of the branches are always executed.
if (MF->getTarget().requiresStructuredCFG())
return nullptr;
// We may need to update this's terminator, but we can't do that if
// AnalyzeBranch fails. If this uses a jump table, we won't touch it.
const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
SmallVector<MachineOperand, 4> Cond;
if (TII->AnalyzeBranch(*this, TBB, FBB, Cond))
return nullptr;
// Avoid bugpoint weirdness: A block may end with a conditional branch but
// jumps to the same MBB is either case. We have duplicate CFG edges in that
// case that we can't handle. Since this never happens in properly optimized
// code, just skip those edges.
if (TBB && TBB == FBB) {
DEBUG(dbgs() << "Won't split critical edge after degenerate BB#"
<< getNumber() << '\n');
return nullptr;
}
MachineBasicBlock *NMBB = MF->CreateMachineBasicBlock();
MF->insert(std::next(MachineFunction::iterator(this)), NMBB);
DEBUG(dbgs() << "Splitting critical edge:"
" BB#" << getNumber()
<< " -- BB#" << NMBB->getNumber()
<< " -- BB#" << Succ->getNumber() << '\n');
LiveIntervals *LIS = P->getAnalysisIfAvailable<LiveIntervals>();
SlotIndexes *Indexes = P->getAnalysisIfAvailable<SlotIndexes>();
if (LIS)
LIS->insertMBBInMaps(NMBB);
else if (Indexes)
Indexes->insertMBBInMaps(NMBB);
// On some targets like Mips, branches may kill virtual registers. Make sure
// that LiveVariables is properly updated after updateTerminator replaces the
// terminators.
LiveVariables *LV = P->getAnalysisIfAvailable<LiveVariables>();
// Collect a list of virtual registers killed by the terminators.
SmallVector<unsigned, 4> KilledRegs;
if (LV)
for (instr_iterator I = getFirstInstrTerminator(), E = instr_end();
I != E; ++I) {
MachineInstr *MI = I;
for (MachineInstr::mop_iterator OI = MI->operands_begin(),
OE = MI->operands_end(); OI != OE; ++OI) {
if (!OI->isReg() || OI->getReg() == 0 ||
!OI->isUse() || !OI->isKill() || OI->isUndef())
continue;
unsigned Reg = OI->getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg) ||
LV->getVarInfo(Reg).removeKill(MI)) {
KilledRegs.push_back(Reg);
DEBUG(dbgs() << "Removing terminator kill: " << *MI);
OI->setIsKill(false);
}
}
}
SmallVector<unsigned, 4> UsedRegs;
if (LIS) {
for (instr_iterator I = getFirstInstrTerminator(), E = instr_end();
I != E; ++I) {
MachineInstr *MI = I;
for (MachineInstr::mop_iterator OI = MI->operands_begin(),
OE = MI->operands_end(); OI != OE; ++OI) {
if (!OI->isReg() || OI->getReg() == 0)
continue;
unsigned Reg = OI->getReg();
if (std::find(UsedRegs.begin(), UsedRegs.end(), Reg) == UsedRegs.end())
UsedRegs.push_back(Reg);
}
}
}
ReplaceUsesOfBlockWith(Succ, NMBB);
// If updateTerminator() removes instructions, we need to remove them from
// SlotIndexes.
SmallVector<MachineInstr*, 4> Terminators;
if (Indexes) {
for (instr_iterator I = getFirstInstrTerminator(), E = instr_end();
I != E; ++I)
Terminators.push_back(I);
}
updateTerminator();
if (Indexes) {
SmallVector<MachineInstr*, 4> NewTerminators;
for (instr_iterator I = getFirstInstrTerminator(), E = instr_end();
I != E; ++I)
NewTerminators.push_back(I);
for (SmallVectorImpl<MachineInstr*>::iterator I = Terminators.begin(),
E = Terminators.end(); I != E; ++I) {
if (std::find(NewTerminators.begin(), NewTerminators.end(), *I) ==
NewTerminators.end())
Indexes->removeMachineInstrFromMaps(*I);
}
}
// Insert unconditional "jump Succ" instruction in NMBB if necessary.
NMBB->addSuccessor(Succ);
if (!NMBB->isLayoutSuccessor(Succ)) {
Cond.clear();
MF->getSubtarget().getInstrInfo()->InsertBranch(*NMBB, Succ, nullptr, Cond,
dl);
if (Indexes) {
for (instr_iterator I = NMBB->instr_begin(), E = NMBB->instr_end();
I != E; ++I) {
// Some instructions may have been moved to NMBB by updateTerminator(),
// so we first remove any instruction that already has an index.
if (Indexes->hasIndex(I))
Indexes->removeMachineInstrFromMaps(I);
Indexes->insertMachineInstrInMaps(I);
}
}
}
// Fix PHI nodes in Succ so they refer to NMBB instead of this
for (MachineBasicBlock::instr_iterator
i = Succ->instr_begin(),e = Succ->instr_end();
i != e && i->isPHI(); ++i)
for (unsigned ni = 1, ne = i->getNumOperands(); ni != ne; ni += 2)
if (i->getOperand(ni+1).getMBB() == this)
i->getOperand(ni+1).setMBB(NMBB);
// Inherit live-ins from the successor
for (MachineBasicBlock::livein_iterator I = Succ->livein_begin(),
E = Succ->livein_end(); I != E; ++I)
NMBB->addLiveIn(*I);
// Update LiveVariables.
const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
if (LV) {
// Restore kills of virtual registers that were killed by the terminators.
while (!KilledRegs.empty()) {
unsigned Reg = KilledRegs.pop_back_val();
for (instr_iterator I = instr_end(), E = instr_begin(); I != E;) {
if (!(--I)->addRegisterKilled(Reg, TRI, /* addIfNotFound= */ false))
continue;
if (TargetRegisterInfo::isVirtualRegister(Reg))
LV->getVarInfo(Reg).Kills.push_back(I);
DEBUG(dbgs() << "Restored terminator kill: " << *I);
break;
}
}
// Update relevant live-through information.
LV->addNewBlock(NMBB, this, Succ);
}
if (LIS) {
// After splitting the edge and updating SlotIndexes, live intervals may be
// in one of two situations, depending on whether this block was the last in
// the function. If the original block was the last in the function, all live
// intervals will end prior to the beginning of the new split block. If the
// original block was not at the end of the function, all live intervals will
// extend to the end of the new split block.
bool isLastMBB =
std::next(MachineFunction::iterator(NMBB)) == getParent()->end();
SlotIndex StartIndex = Indexes->getMBBEndIdx(this);
SlotIndex PrevIndex = StartIndex.getPrevSlot();
SlotIndex EndIndex = Indexes->getMBBEndIdx(NMBB);
// Find the registers used from NMBB in PHIs in Succ.
SmallSet<unsigned, 8> PHISrcRegs;
for (MachineBasicBlock::instr_iterator
I = Succ->instr_begin(), E = Succ->instr_end();
I != E && I->isPHI(); ++I) {
for (unsigned ni = 1, ne = I->getNumOperands(); ni != ne; ni += 2) {
if (I->getOperand(ni+1).getMBB() == NMBB) {
MachineOperand &MO = I->getOperand(ni);
unsigned Reg = MO.getReg();
PHISrcRegs.insert(Reg);
if (MO.isUndef())
continue;
LiveInterval &LI = LIS->getInterval(Reg);
VNInfo *VNI = LI.getVNInfoAt(PrevIndex);
assert(VNI && "PHI sources should be live out of their predecessors.");
LI.addSegment(LiveInterval::Segment(StartIndex, EndIndex, VNI));
}
}
}
MachineRegisterInfo *MRI = &getParent()->getRegInfo();
for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
if (PHISrcRegs.count(Reg) || !LIS->hasInterval(Reg))
continue;
LiveInterval &LI = LIS->getInterval(Reg);
if (!LI.liveAt(PrevIndex))
continue;
bool isLiveOut = LI.liveAt(LIS->getMBBStartIdx(Succ));
if (isLiveOut && isLastMBB) {
VNInfo *VNI = LI.getVNInfoAt(PrevIndex);
assert(VNI && "LiveInterval should have VNInfo where it is live.");
LI.addSegment(LiveInterval::Segment(StartIndex, EndIndex, VNI));
} else if (!isLiveOut && !isLastMBB) {
LI.removeSegment(StartIndex, EndIndex);
}
}
// Update all intervals for registers whose uses may have been modified by
// updateTerminator().
LIS->repairIntervalsInRange(this, getFirstTerminator(), end(), UsedRegs);
}
if (MachineDominatorTree *MDT =
P->getAnalysisIfAvailable<MachineDominatorTree>())
MDT->recordSplitCriticalEdge(this, Succ, NMBB);
if (MachineLoopInfo *MLI = P->getAnalysisIfAvailable<MachineLoopInfo>())
if (MachineLoop *TIL = MLI->getLoopFor(this)) {
// If one or the other blocks were not in a loop, the new block is not
// either, and thus LI doesn't need to be updated.
if (MachineLoop *DestLoop = MLI->getLoopFor(Succ)) {
if (TIL == DestLoop) {
// Both in the same loop, the NMBB joins loop.
DestLoop->addBasicBlockToLoop(NMBB, MLI->getBase());
} else if (TIL->contains(DestLoop)) {
// Edge from an outer loop to an inner loop. Add to the outer loop.
TIL->addBasicBlockToLoop(NMBB, MLI->getBase());
} else if (DestLoop->contains(TIL)) {
// Edge from an inner loop to an outer loop. Add to the outer loop.
DestLoop->addBasicBlockToLoop(NMBB, MLI->getBase());
} else {
// Edge from two loops with no containment relation. Because these
// are natural loops, we know that the destination block must be the
// header of its loop (adding a branch into a loop elsewhere would
// create an irreducible loop).
assert(DestLoop->getHeader() == Succ &&
"Should not create irreducible loops!");
if (MachineLoop *P = DestLoop->getParentLoop())
P->addBasicBlockToLoop(NMBB, MLI->getBase());
}
}
}
return NMBB;
}
bool SIFixSGPRLiveRanges::runOnMachineFunction(MachineFunction &MF) {
MachineRegisterInfo &MRI = MF.getRegInfo();
const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
const SIRegisterInfo *TRI = static_cast<const SIRegisterInfo *>(
MF.getSubtarget().getRegisterInfo());
bool MadeChange = false;
MachinePostDominatorTree *PDT = &getAnalysis<MachinePostDominatorTree>();
std::vector<std::pair<unsigned, LiveRange *>> SGPRLiveRanges;
LiveIntervals *LIS = &getAnalysis<LiveIntervals>();
LiveVariables *LV = getAnalysisIfAvailable<LiveVariables>();
MachineBasicBlock *Entry = MF.begin();
// Use a depth first order so that in SSA, we encounter all defs before
// uses. Once the defs of the block have been found, attempt to insert
// SGPR_USE instructions in successor blocks if required.
for (MachineBasicBlock *MBB : depth_first(Entry)) {
for (const MachineInstr &MI : *MBB) {
for (const MachineOperand &MO : MI.defs()) {
if (MO.isImplicit())
continue;
unsigned Def = MO.getReg();
if (TargetRegisterInfo::isVirtualRegister(Def)) {
if (TRI->isSGPRClass(MRI.getRegClass(Def))) {
// Only consider defs that are live outs. We don't care about def /
// use within the same block.
LiveRange &LR = LIS->getInterval(Def);
if (LIS->isLiveOutOfMBB(LR, MBB))
SGPRLiveRanges.push_back(std::make_pair(Def, &LR));
}
} else if (TRI->isSGPRClass(TRI->getPhysRegClass(Def))) {
SGPRLiveRanges.push_back(std::make_pair(Def, &LIS->getRegUnit(Def)));
}
}
}
if (MBB->succ_size() < 2)
continue;
// We have structured control flow, so the number of successors should be
// two.
assert(MBB->succ_size() == 2);
MachineBasicBlock *SuccA = *MBB->succ_begin();
MachineBasicBlock *SuccB = *(++MBB->succ_begin());
MachineBasicBlock *NCD = PDT->findNearestCommonDominator(SuccA, SuccB);
if (!NCD)
continue;
MachineBasicBlock::iterator NCDTerm = NCD->getFirstTerminator();
if (NCDTerm != NCD->end() && NCDTerm->getOpcode() == AMDGPU::SI_ELSE) {
assert(NCD->succ_size() == 2);
// We want to make sure we insert the Use after the ENDIF, not after
// the ELSE.
NCD = PDT->findNearestCommonDominator(*NCD->succ_begin(),
*(++NCD->succ_begin()));
}
for (std::pair<unsigned, LiveRange*> RegLR : SGPRLiveRanges) {
unsigned Reg = RegLR.first;
LiveRange *LR = RegLR.second;
// FIXME: We could be smarter here. If the register is Live-In to one
// block, but the other doesn't have any SGPR defs, then there won't be a
// conflict. Also, if the branch condition is uniform then there will be
// no conflict.
bool LiveInToA = LIS->isLiveInToMBB(*LR, SuccA);
bool LiveInToB = LIS->isLiveInToMBB(*LR, SuccB);
if (!LiveInToA && !LiveInToB) {
DEBUG(dbgs() << PrintReg(Reg, TRI, 0)
<< " is live into neither successor\n");
continue;
}
if (LiveInToA && LiveInToB) {
DEBUG(dbgs() << PrintReg(Reg, TRI, 0)
<< " is live into both successors\n");
continue;
}
// This interval is live in to one successor, but not the other, so
// we need to update its range so it is live in to both.
DEBUG(dbgs() << "Possible SGPR conflict detected for "
<< PrintReg(Reg, TRI, 0) << " in " << *LR
<< " BB#" << SuccA->getNumber() << ", BB#"
<< SuccB->getNumber()
<< " with NCD = BB#" << NCD->getNumber() << '\n');
assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
"Not expecting to extend live range of physreg");
// FIXME: Need to figure out how to update LiveRange here so this pass
// will be able to preserve LiveInterval analysis.
MachineInstr *NCDSGPRUse =
BuildMI(*NCD, NCD->getFirstNonPHI(), DebugLoc(),
TII->get(AMDGPU::SGPR_USE))
.addReg(Reg, RegState::Implicit);
MadeChange = true;
SlotIndex SI = LIS->InsertMachineInstrInMaps(NCDSGPRUse);
LIS->extendToIndices(*LR, SI.getRegSlot());
if (LV) {
// TODO: This won't work post-SSA
LV->HandleVirtRegUse(Reg, NCD, NCDSGPRUse);
}
DEBUG(NCDSGPRUse->dump());
}
}
return MadeChange;
}
MachineBasicBlock *
MachineBasicBlock::SplitCriticalEdge(MachineBasicBlock *Succ, Pass *P) {
MachineFunction *MF = getParent();
DebugLoc dl; // FIXME: this is nowhere
// We may need to update this's terminator, but we can't do that if
// AnalyzeBranch fails. If this uses a jump table, we won't touch it.
const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
MachineBasicBlock *TBB = 0, *FBB = 0;
SmallVector<MachineOperand, 4> Cond;
if (TII->AnalyzeBranch(*this, TBB, FBB, Cond))
return NULL;
// Avoid bugpoint weirdness: A block may end with a conditional branch but
// jumps to the same MBB is either case. We have duplicate CFG edges in that
// case that we can't handle. Since this never happens in properly optimized
// code, just skip those edges.
if (TBB && TBB == FBB) {
DEBUG(dbgs() << "Won't split critical edge after degenerate BB#"
<< getNumber() << '\n');
return NULL;
}
MachineBasicBlock *NMBB = MF->CreateMachineBasicBlock();
MF->insert(llvm::next(MachineFunction::iterator(this)), NMBB);
DEBUG(dbgs() << "Splitting critical edge:"
" BB#" << getNumber()
<< " -- BB#" << NMBB->getNumber()
<< " -- BB#" << Succ->getNumber() << '\n');
// On some targets like Mips, branches may kill virtual registers. Make sure
// that LiveVariables is properly updated after updateTerminator replaces the
// terminators.
LiveVariables *LV = P->getAnalysisIfAvailable<LiveVariables>();
// Collect a list of virtual registers killed by the terminators.
SmallVector<unsigned, 4> KilledRegs;
if (LV)
for (iterator I = getFirstTerminator(), E = end(); I != E; ++I) {
MachineInstr *MI = I;
for (MachineInstr::mop_iterator OI = MI->operands_begin(),
OE = MI->operands_end(); OI != OE; ++OI) {
if (!OI->isReg() || !OI->isUse() || !OI->isKill() || OI->isUndef())
continue;
unsigned Reg = OI->getReg();
if (TargetRegisterInfo::isVirtualRegister(Reg) &&
LV->getVarInfo(Reg).removeKill(MI)) {
KilledRegs.push_back(Reg);
DEBUG(dbgs() << "Removing terminator kill: " << *MI);
OI->setIsKill(false);
}
}
}
ReplaceUsesOfBlockWith(Succ, NMBB);
updateTerminator();
// Insert unconditional "jump Succ" instruction in NMBB if necessary.
NMBB->addSuccessor(Succ);
if (!NMBB->isLayoutSuccessor(Succ)) {
Cond.clear();
MF->getTarget().getInstrInfo()->InsertBranch(*NMBB, Succ, NULL, Cond, dl);
}
// Fix PHI nodes in Succ so they refer to NMBB instead of this
for (MachineBasicBlock::iterator i = Succ->begin(), e = Succ->end();
i != e && i->isPHI(); ++i)
for (unsigned ni = 1, ne = i->getNumOperands(); ni != ne; ni += 2)
if (i->getOperand(ni+1).getMBB() == this)
i->getOperand(ni+1).setMBB(NMBB);
// Inherit live-ins from the successor
for (MachineBasicBlock::livein_iterator I = Succ->livein_begin(),
E = Succ->livein_end(); I != E; ++I)
NMBB->addLiveIn(*I);
// Update LiveVariables.
if (LV) {
// Restore kills of virtual registers that were killed by the terminators.
while (!KilledRegs.empty()) {
unsigned Reg = KilledRegs.pop_back_val();
for (iterator I = end(), E = begin(); I != E;) {
if (!(--I)->addRegisterKilled(Reg, NULL, /* addIfNotFound= */ false))
continue;
LV->getVarInfo(Reg).Kills.push_back(I);
DEBUG(dbgs() << "Restored terminator kill: " << *I);
break;
}
}
// Update relevant live-through information.
LV->addNewBlock(NMBB, this, Succ);
}
if (MachineDominatorTree *MDT =
P->getAnalysisIfAvailable<MachineDominatorTree>()) {
// Update dominator information.
MachineDomTreeNode *SucccDTNode = MDT->getNode(Succ);
bool IsNewIDom = true;
for (const_pred_iterator PI = Succ->pred_begin(), E = Succ->pred_end();
PI != E; ++PI) {
MachineBasicBlock *PredBB = *PI;
if (PredBB == NMBB)
continue;
if (!MDT->dominates(SucccDTNode, MDT->getNode(PredBB))) {
IsNewIDom = false;
break;
}
}
// We know "this" dominates the newly created basic block.
MachineDomTreeNode *NewDTNode = MDT->addNewBlock(NMBB, this);
// If all the other predecessors of "Succ" are dominated by "Succ" itself
// then the new block is the new immediate dominator of "Succ". Otherwise,
// the new block doesn't dominate anything.
if (IsNewIDom)
MDT->changeImmediateDominator(SucccDTNode, NewDTNode);
}
if (MachineLoopInfo *MLI = P->getAnalysisIfAvailable<MachineLoopInfo>())
if (MachineLoop *TIL = MLI->getLoopFor(this)) {
// If one or the other blocks were not in a loop, the new block is not
// either, and thus LI doesn't need to be updated.
if (MachineLoop *DestLoop = MLI->getLoopFor(Succ)) {
if (TIL == DestLoop) {
// Both in the same loop, the NMBB joins loop.
DestLoop->addBasicBlockToLoop(NMBB, MLI->getBase());
} else if (TIL->contains(DestLoop)) {
// Edge from an outer loop to an inner loop. Add to the outer loop.
TIL->addBasicBlockToLoop(NMBB, MLI->getBase());
} else if (DestLoop->contains(TIL)) {
// Edge from an inner loop to an outer loop. Add to the outer loop.
DestLoop->addBasicBlockToLoop(NMBB, MLI->getBase());
} else {
// Edge from two loops with no containment relation. Because these
// are natural loops, we know that the destination block must be the
// header of its loop (adding a branch into a loop elsewhere would
// create an irreducible loop).
assert(DestLoop->getHeader() == Succ &&
"Should not create irreducible loops!");
if (MachineLoop *P = DestLoop->getParentLoop())
P->addBasicBlockToLoop(NMBB, MLI->getBase());
}
}
}
return NMBB;
}
bool PHIElimination::SplitPHIEdges(MachineFunction &MF,
MachineBasicBlock &MBB,
LiveVariables &LV,
MachineLoopInfo *MLI) {
if (MBB.empty() || !MBB.front().isPHI() || MBB.isLandingPad())
return false; // Quick exit for basic blocks without PHIs.
const MachineLoop *CurLoop = MLI ? MLI->getLoopFor(&MBB) : 0;
bool IsLoopHeader = CurLoop && &MBB == CurLoop->getHeader();
bool Changed = false;
for (MachineBasicBlock::iterator BBI = MBB.begin(), BBE = MBB.end();
BBI != BBE && BBI->isPHI(); ++BBI) {
for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) {
unsigned Reg = BBI->getOperand(i).getReg();
MachineBasicBlock *PreMBB = BBI->getOperand(i+1).getMBB();
// Is there a critical edge from PreMBB to MBB?
if (PreMBB->succ_size() == 1)
continue;
// Avoid splitting backedges of loops. It would introduce small
// out-of-line blocks into the loop which is very bad for code placement.
if (PreMBB == &MBB)
continue;
const MachineLoop *PreLoop = MLI ? MLI->getLoopFor(PreMBB) : 0;
if (IsLoopHeader && PreLoop == CurLoop)
continue;
// LV doesn't consider a phi use live-out, so isLiveOut only returns true
// when the source register is live-out for some other reason than a phi
// use. That means the copy we will insert in PreMBB won't be a kill, and
// there is a risk it may not be coalesced away.
//
// If the copy would be a kill, there is no need to split the edge.
if (!LV.isLiveOut(Reg, *PreMBB))
continue;
DEBUG(dbgs() << PrintReg(Reg) << " live-out before critical edge BB#"
<< PreMBB->getNumber() << " -> BB#" << MBB.getNumber()
<< ": " << *BBI);
// If Reg is not live-in to MBB, it means it must be live-in to some
// other PreMBB successor, and we can avoid the interference by splitting
// the edge.
//
// If Reg *is* live-in to MBB, the interference is inevitable and a copy
// is likely to be left after coalescing. If we are looking at a loop
// exiting edge, split it so we won't insert code in the loop, otherwise
// don't bother.
bool ShouldSplit = !LV.isLiveIn(Reg, MBB);
// Check for a loop exiting edge.
if (!ShouldSplit && CurLoop != PreLoop) {
DEBUG({
dbgs() << "Split wouldn't help, maybe avoid loop copies?\n";
if (PreLoop) dbgs() << "PreLoop: " << *PreLoop;
if (CurLoop) dbgs() << "CurLoop: " << *CurLoop;
});
// This edge could be entering a loop, exiting a loop, or it could be
// both: Jumping directly form one loop to the header of a sibling
// loop.
// Split unless this edge is entering CurLoop from an outer loop.
ShouldSplit = PreLoop && !PreLoop->contains(CurLoop);
}
if (!ShouldSplit)
continue;
if (!PreMBB->SplitCriticalEdge(&MBB, this)) {
DEBUG(dbgs() << "Failed to split ciritcal edge.\n");
continue;
}
Changed = true;
++NumCriticalEdgesSplit;
}
