void LazyLiveness::computeBackedgeChain(MachineFunction& mf,
MachineBasicBlock* MBB) {
SparseBitVector<128> tmp = rv[MBB];
tmp.set(preorder[MBB]);
tmp &= backedge_source;
calculated.set(preorder[MBB]);
for (SparseBitVector<128>::iterator I = tmp.begin(); I != tmp.end(); ++I) {
assert(rev_preorder.size() > *I && "Unknown block!");
MachineBasicBlock* SrcMBB = rev_preorder[*I];
for (MachineBasicBlock::succ_iterator SI = SrcMBB->succ_begin(),
SE = SrcMBB->succ_end(); SI != SE; ++SI) {
MachineBasicBlock* TgtMBB = *SI;
if (backedges.count(std::make_pair(SrcMBB, TgtMBB)) &&
!rv[MBB].test(preorder[TgtMBB])) {
if (!calculated.test(preorder[TgtMBB]))
computeBackedgeChain(mf, TgtMBB);
tv[MBB].set(preorder[TgtMBB]);
SparseBitVector<128> right = tv[TgtMBB];
tv[MBB] |= right;
}
}
tv[MBB].reset(preorder[MBB]);
}
}
static bool checkEFLAGSLive(MachineInstr *MI) {
if (MI->killsRegister(X86::EFLAGS))
return false;
// The EFLAGS operand of MI might be missing a kill marker.
// Figure out whether EFLAGS operand should LIVE after MI instruction.
MachineBasicBlock *BB = MI->getParent();
MachineBasicBlock::iterator ItrMI = MI;
// Scan forward through BB for a use/def of EFLAGS.
for (auto I = std::next(ItrMI), E = BB->end(); I != E; ++I) {
if (I->readsRegister(X86::EFLAGS))
return true;
if (I->definesRegister(X86::EFLAGS))
return false;
}
// We hit the end of the block, check whether EFLAGS is live into a successor.
for (auto I = BB->succ_begin(), E = BB->succ_end(); I != E; ++I) {
if ((*I)->isLiveIn(X86::EFLAGS))
return true;
}
return false;
}
void LiveIntervals::pruneValue(LiveRange &LR, SlotIndex Kill,
SmallVectorImpl<SlotIndex> *EndPoints) {
LiveQueryResult LRQ = LR.Query(Kill);
VNInfo *VNI = LRQ.valueOutOrDead();
if (!VNI)
return;
MachineBasicBlock *KillMBB = Indexes->getMBBFromIndex(Kill);
SlotIndex MBBEnd = Indexes->getMBBEndIdx(KillMBB);
// If VNI isn't live out from KillMBB, the value is trivially pruned.
if (LRQ.endPoint() < MBBEnd) {
LR.removeSegment(Kill, LRQ.endPoint());
if (EndPoints) EndPoints->push_back(LRQ.endPoint());
return;
}
// VNI is live out of KillMBB.
LR.removeSegment(Kill, MBBEnd);
if (EndPoints) EndPoints->push_back(MBBEnd);
// Find all blocks that are reachable from KillMBB without leaving VNI's live
// range. It is possible that KillMBB itself is reachable, so start a DFS
// from each successor.
typedef SmallPtrSet<MachineBasicBlock*, 9> VisitedTy;
VisitedTy Visited;
for (MachineBasicBlock::succ_iterator
SuccI = KillMBB->succ_begin(), SuccE = KillMBB->succ_end();
SuccI != SuccE; ++SuccI) {
for (df_ext_iterator<MachineBasicBlock*, VisitedTy>
I = df_ext_begin(*SuccI, Visited), E = df_ext_end(*SuccI, Visited);
I != E;) {
MachineBasicBlock *MBB = *I;
// Check if VNI is live in to MBB.
SlotIndex MBBStart, MBBEnd;
std::tie(MBBStart, MBBEnd) = Indexes->getMBBRange(MBB);
LiveQueryResult LRQ = LR.Query(MBBStart);
if (LRQ.valueIn() != VNI) {
// This block isn't part of the VNI segment. Prune the search.
I.skipChildren();
continue;
}
// Prune the search if VNI is killed in MBB.
if (LRQ.endPoint() < MBBEnd) {
LR.removeSegment(MBBStart, LRQ.endPoint());
if (EndPoints) EndPoints->push_back(LRQ.endPoint());
I.skipChildren();
continue;
}
// VNI is live through MBB.
LR.removeSegment(MBBStart, MBBEnd);
if (EndPoints) EndPoints->push_back(MBBEnd);
++I;
}
}
}
void RegDefsUses::addLiveOut(const MachineBasicBlock &MBB,
const MachineBasicBlock &SuccBB) {
for (MachineBasicBlock::const_succ_iterator SI = MBB.succ_begin(),
SE = MBB.succ_end(); SI != SE; ++SI)
if (*SI != &SuccBB)
for (const auto &LI : (*SI)->liveins())
Uses.set(LI.PhysReg);
}
bool PTXInstrInfo::
IsAnySuccessorAlsoLayoutSuccessor(const MachineBasicBlock& MBB) {
for (MachineBasicBlock::const_succ_iterator
i = MBB.succ_begin(), e = MBB.succ_end(); i != e; ++i)
if (MBB.isLayoutSuccessor((const MachineBasicBlock*) &*i))
return true;
return false;
}
static bool isCPSRLiveout(MachineBasicBlock &MBB) {
for (MachineBasicBlock::succ_iterator I = MBB.succ_begin(),
E = MBB.succ_end(); I != E; ++I) {
if ((*I)->isLiveIn(ARM::CPSR))
return true;
}
return false;
}
void RegDefsUses::addLiveOut(const MachineBasicBlock &MBB,
const MachineBasicBlock &SuccBB) {
for (MachineBasicBlock::const_succ_iterator SI = MBB.succ_begin(),
SE = MBB.succ_end(); SI != SE; ++SI)
if (*SI != &SuccBB)
for (MachineBasicBlock::livein_iterator LI = (*SI)->livein_begin(),
LE = (*SI)->livein_end(); LI != LE; ++LI)
Uses.set(*LI);
}
static bool
bothUsedInPHI(const MachineBasicBlock &A,
SmallPtrSet<MachineBasicBlock*, 8> SuccsB) {
for (MachineBasicBlock::const_succ_iterator SI = A.succ_begin(),
SE = A.succ_end(); SI != SE; ++SI) {
MachineBasicBlock *BB = *SI;
if (SuccsB.count(BB) && !BB->empty() && BB->begin()->isPHI())
return true;
}
return false;
}
void MIRPrinter::convert(ModuleSlotTracker &MST,
yaml::MachineBasicBlock &YamlMBB,
const MachineBasicBlock &MBB) {
assert(MBB.getNumber() >= 0 && "Invalid MBB number");
YamlMBB.ID = (unsigned)MBB.getNumber();
if (const auto *BB = MBB.getBasicBlock()) {
if (BB->hasName()) {
YamlMBB.Name.Value = BB->getName();
} else {
int Slot = MST.getLocalSlot(BB);
if (Slot == -1)
YamlMBB.IRBlock.Value = "<badref>";
else
YamlMBB.IRBlock.Value = (Twine("%ir-block.") + Twine(Slot)).str();
}
}
YamlMBB.Alignment = MBB.getAlignment();
YamlMBB.AddressTaken = MBB.hasAddressTaken();
YamlMBB.IsLandingPad = MBB.isLandingPad();
for (const auto *SuccMBB : MBB.successors()) {
std::string Str;
raw_string_ostream StrOS(Str);
MIPrinter(StrOS, MST, RegisterMaskIds, StackObjectOperandMapping)
.printMBBReference(*SuccMBB);
YamlMBB.Successors.push_back(StrOS.str());
}
if (MBB.hasSuccessorWeights()) {
for (auto I = MBB.succ_begin(), E = MBB.succ_end(); I != E; ++I)
YamlMBB.SuccessorWeights.push_back(
yaml::UnsignedValue(MBB.getSuccWeight(I)));
}
// Print the live in registers.
const auto *TRI = MBB.getParent()->getSubtarget().getRegisterInfo();
assert(TRI && "Expected target register info");
for (auto I = MBB.livein_begin(), E = MBB.livein_end(); I != E; ++I) {
std::string Str;
raw_string_ostream StrOS(Str);
printReg(*I, StrOS, TRI);
YamlMBB.LiveIns.push_back(StrOS.str());
}
// Print the machine instructions.
YamlMBB.Instructions.reserve(MBB.size());
std::string Str;
for (const auto &MI : MBB) {
raw_string_ostream StrOS(Str);
MIPrinter(StrOS, MST, RegisterMaskIds, StackObjectOperandMapping).print(MI);
YamlMBB.Instructions.push_back(StrOS.str());
Str.clear();
}
}
void SPScope::buildfcfg(void) {
std::set<const MachineBasicBlock *> body(++Blocks.begin(), Blocks.end());
std::vector<Edge> outedges;
for (unsigned i=0; i<Blocks.size(); i++) {
MachineBasicBlock *MBB = Blocks[i];
if (HeaderMap.count(MBB)) {
const SPScope *subloop = HeaderMap[MBB];
// successors of the loop
outedges.insert(outedges.end(),
subloop->ExitEdges.begin(),
subloop->ExitEdges.end());
} else {
// simple block
for (MachineBasicBlock::succ_iterator si = MBB->succ_begin(),
se = MBB->succ_end(); si != se; ++si) {
outedges.push_back(std::make_pair(MBB, *si));
}
}
Node &n = FCFG.getNodeFor(MBB);
for (unsigned i=0; i<outedges.size(); i++) {
const MachineBasicBlock *succ = outedges[i].second;
if (body.count(succ)) {
Node &ns = FCFG.getNodeFor(succ);
n.connect(ns, outedges[i]);
} else {
if (succ != getHeader()) {
// record exit edges
FCFG.toexit(n, outedges[i]);
} else {
// we don't need back edges recorded
FCFG.toexit(n);
}
}
}
// special case: top-level loop has no exit/backedge
if (outedges.empty()) {
assert(isTopLevel());
FCFG.toexit(n);
}
outedges.clear();
}
}
bool MIPrinter::canPredictSuccessors(const MachineBasicBlock &MBB) const {
SmallVector<MachineBasicBlock*,8> GuessedSuccs;
bool GuessedFallthrough;
guessSuccessors(MBB, GuessedSuccs, GuessedFallthrough);
if (GuessedFallthrough) {
const MachineFunction &MF = *MBB.getParent();
MachineFunction::const_iterator NextI = std::next(MBB.getIterator());
if (NextI != MF.end()) {
MachineBasicBlock *Next = const_cast<MachineBasicBlock*>(&*NextI);
if (!is_contained(GuessedSuccs, Next))
GuessedSuccs.push_back(Next);
}
}
if (GuessedSuccs.size() != MBB.succ_size())
return false;
return std::equal(MBB.succ_begin(), MBB.succ_end(), GuessedSuccs.begin());
}
bool LiveVariables::isLiveOut(unsigned Reg, const MachineBasicBlock &MBB) {
LiveVariables::VarInfo &VI = getVarInfo(Reg);
// 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.
SmallVector<MachineBasicBlock*, 8> OpSuccBlocks;
for (MachineBasicBlock::const_succ_iterator SI = MBB.succ_begin(),
E = MBB.succ_end(); SI != E; ++SI) {
MachineBasicBlock *SuccMBB = *SI;
// Is it alive in this successor?
unsigned SuccIdx = SuccMBB->getNumber();
if (VI.AliveBlocks.test(SuccIdx))
return true;
OpSuccBlocks.push_back(SuccMBB);
}
// Check to see if this value is live because there is a use in a successor
// that kills it.
switch (OpSuccBlocks.size()) {
case 1: {
MachineBasicBlock *SuccMBB = OpSuccBlocks[0];
for (unsigned i = 0, e = VI.Kills.size(); i != e; ++i)
if (VI.Kills[i]->getParent() == SuccMBB)
return true;
break;
}
case 2: {
MachineBasicBlock *SuccMBB1 = OpSuccBlocks[0], *SuccMBB2 = OpSuccBlocks[1];
for (unsigned i = 0, e = VI.Kills.size(); i != e; ++i)
if (VI.Kills[i]->getParent() == SuccMBB1 ||
VI.Kills[i]->getParent() == SuccMBB2)
return true;
break;
}
default:
std::sort(OpSuccBlocks.begin(), OpSuccBlocks.end());
for (unsigned i = 0, e = VI.Kills.size(); i != e; ++i)
if (std::binary_search(OpSuccBlocks.begin(), OpSuccBlocks.end(),
VI.Kills[i]->getParent()))
return true;
}
return false;
}
bool
TailDuplicatePass::canCompletelyDuplicateBB(MachineBasicBlock &BB) {
SmallPtrSet<MachineBasicBlock*, 8> Succs(BB.succ_begin(), BB.succ_end());
for (MachineBasicBlock::pred_iterator PI = BB.pred_begin(),
PE = BB.pred_end(); PI != PE; ++PI) {
MachineBasicBlock *PredBB = *PI;
if (PredBB->succ_size() > 1)
return false;
MachineBasicBlock *PredTBB = NULL, *PredFBB = NULL;
SmallVector<MachineOperand, 4> PredCond;
if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
return false;
if (!PredCond.empty())
return false;
}
return true;
}
void LoopSplitter::dumpOddTerminators() {
for (MachineFunction::iterator bbItr = mf->begin(), bbEnd = mf->end();
bbItr != bbEnd; ++bbItr) {
MachineBasicBlock *mbb = &*bbItr;
MachineBasicBlock *a = 0, *b = 0;
SmallVector<MachineOperand, 4> c;
if (tii->AnalyzeBranch(*mbb, a, b, c)) {
dbgs() << "MBB#" << mbb->getNumber() << " has multiway terminator.\n";
dbgs() << " Terminators:\n";
for (MachineBasicBlock::iterator iItr = mbb->begin(), iEnd = mbb->end();
iItr != iEnd; ++iItr) {
MachineInstr *instr= &*iItr;
dbgs() << " " << *instr << "";
}
dbgs() << "\n Listed successors: [ ";
for (MachineBasicBlock::succ_iterator sItr = mbb->succ_begin(), sEnd = mbb->succ_end();
sItr != sEnd; ++sItr) {
MachineBasicBlock *succMBB = *sItr;
dbgs() << succMBB->getNumber() << " ";
}
dbgs() << "]\n\n";
}
}
}
void StackColoring::calculateLocalLiveness() {
// Perform a standard reverse dataflow computation to solve for
// global liveness. The BEGIN set here is equivalent to KILL in the standard
// formulation, and END is equivalent to GEN. The result of this computation
// is a map from blocks to bitvectors where the bitvectors represent which
// allocas are live in/out of that block.
SmallPtrSet<MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(),
BasicBlockNumbering.end());
unsigned NumSSMIters = 0;
bool changed = true;
while (changed) {
changed = false;
++NumSSMIters;
SmallPtrSet<MachineBasicBlock*, 8> NextBBSet;
for (SmallVector<MachineBasicBlock*, 8>::iterator
PI = BasicBlockNumbering.begin(), PE = BasicBlockNumbering.end();
PI != PE; ++PI) {
MachineBasicBlock *BB = *PI;
if (!BBSet.count(BB)) continue;
BitVector LocalLiveIn;
BitVector LocalLiveOut;
// Forward propagation from begins to ends.
for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
PE = BB->pred_end(); PI != PE; ++PI)
LocalLiveIn |= BlockLiveness[*PI].LiveOut;
LocalLiveIn |= BlockLiveness[BB].End;
LocalLiveIn.reset(BlockLiveness[BB].Begin);
// Reverse propagation from ends to begins.
for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
SE = BB->succ_end(); SI != SE; ++SI)
LocalLiveOut |= BlockLiveness[*SI].LiveIn;
LocalLiveOut |= BlockLiveness[BB].Begin;
LocalLiveOut.reset(BlockLiveness[BB].End);
LocalLiveIn |= LocalLiveOut;
LocalLiveOut |= LocalLiveIn;
// After adopting the live bits, we need to turn-off the bits which
// are de-activated in this block.
LocalLiveOut.reset(BlockLiveness[BB].End);
LocalLiveIn.reset(BlockLiveness[BB].Begin);
// If we have both BEGIN and END markers in the same basic block then
// we know that the BEGIN marker comes after the END, because we already
// handle the case where the BEGIN comes before the END when collecting
// the markers (and building the BEGIN/END vectore).
// Want to enable the LIVE_IN and LIVE_OUT of slots that have both
// BEGIN and END because it means that the value lives before and after
// this basic block.
BitVector LocalEndBegin = BlockLiveness[BB].End;
LocalEndBegin &= BlockLiveness[BB].Begin;
LocalLiveIn |= LocalEndBegin;
LocalLiveOut |= LocalEndBegin;
if (LocalLiveIn.test(BlockLiveness[BB].LiveIn)) {
changed = true;
BlockLiveness[BB].LiveIn |= LocalLiveIn;
for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(),
PE = BB->pred_end(); PI != PE; ++PI)
NextBBSet.insert(*PI);
}
if (LocalLiveOut.test(BlockLiveness[BB].LiveOut)) {
changed = true;
BlockLiveness[BB].LiveOut |= LocalLiveOut;
for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
SE = BB->succ_end(); SI != SE; ++SI)
NextBBSet.insert(*SI);
}
}
BBSet = NextBBSet;
}// while changed.
}
bool UnreachableMachineBlockElim::runOnMachineFunction(MachineFunction &F) {
df_iterator_default_set<MachineBasicBlock*> Reachable;
bool ModifiedPHI = false;
MMI = getAnalysisIfAvailable<MachineModuleInfo>();
MachineDominatorTree *MDT = getAnalysisIfAvailable<MachineDominatorTree>();
MachineLoopInfo *MLI = getAnalysisIfAvailable<MachineLoopInfo>();
// Mark all reachable blocks.
for (MachineBasicBlock *BB : depth_first_ext(&F, Reachable))
(void)BB/* Mark all reachable blocks */;
// Loop over all dead blocks, remembering them and deleting all instructions
// in them.
std::vector<MachineBasicBlock*> DeadBlocks;
for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
MachineBasicBlock *BB = &*I;
// Test for deadness.
if (!Reachable.count(BB)) {
DeadBlocks.push_back(BB);
// Update dominator and loop info.
if (MLI) MLI->removeBlock(BB);
if (MDT && MDT->getNode(BB)) MDT->eraseNode(BB);
while (BB->succ_begin() != BB->succ_end()) {
MachineBasicBlock* succ = *BB->succ_begin();
MachineBasicBlock::iterator start = succ->begin();
while (start != succ->end() && start->isPHI()) {
for (unsigned i = start->getNumOperands() - 1; i >= 2; i-=2)
if (start->getOperand(i).isMBB() &&
start->getOperand(i).getMBB() == BB) {
start->RemoveOperand(i);
start->RemoveOperand(i-1);
}
start++;
}
BB->removeSuccessor(BB->succ_begin());
}
}
}
// Actually remove the blocks now.
for (unsigned i = 0, e = DeadBlocks.size(); i != e; ++i)
DeadBlocks[i]->eraseFromParent();
// Cleanup PHI nodes.
for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
MachineBasicBlock *BB = &*I;
// Prune unneeded PHI entries.
SmallPtrSet<MachineBasicBlock*, 8> preds(BB->pred_begin(),
BB->pred_end());
MachineBasicBlock::iterator phi = BB->begin();
while (phi != BB->end() && phi->isPHI()) {
for (unsigned i = phi->getNumOperands() - 1; i >= 2; i-=2)
if (!preds.count(phi->getOperand(i).getMBB())) {
phi->RemoveOperand(i);
phi->RemoveOperand(i-1);
ModifiedPHI = true;
}
if (phi->getNumOperands() == 3) {
const MachineOperand &Input = phi->getOperand(1);
const MachineOperand &Output = phi->getOperand(0);
unsigned InputReg = Input.getReg();
unsigned OutputReg = Output.getReg();
assert(Output.getSubReg() == 0 && "Cannot have output subregister");
ModifiedPHI = true;
if (InputReg != OutputReg) {
MachineRegisterInfo &MRI = F.getRegInfo();
unsigned InputSub = Input.getSubReg();
if (InputSub == 0 &&
MRI.constrainRegClass(InputReg, MRI.getRegClass(OutputReg))) {
MRI.replaceRegWith(OutputReg, InputReg);
} else {
// The input register to the PHI has a subregister or it can't be
// constrained to the proper register class:
// insert a COPY instead of simply replacing the output
// with the input.
const TargetInstrInfo *TII = F.getSubtarget().getInstrInfo();
BuildMI(*BB, BB->getFirstNonPHI(), phi->getDebugLoc(),
TII->get(TargetOpcode::COPY), OutputReg)
.addReg(InputReg, getRegState(Input), InputSub);
}
phi++->eraseFromParent();
}
continue;
}
++phi;
}
}
F.RenumberBlocks();
return (!DeadBlocks.empty() || ModifiedPHI);
}
// Return true if CC is live out of MBB.
static bool isCCLiveOut(MachineBasicBlock &MBB) {
for (auto SI = MBB.succ_begin(), SE = MBB.succ_end(); SI != SE; ++SI)
if ((*SI)->isLiveIn(SystemZ::CC))
return true;
return false;
}
bool LanaiInstrInfo::optimizeCompareInstr(
MachineInstr &CmpInstr, unsigned SrcReg, unsigned SrcReg2, int CmpMask,
int CmpValue, const MachineRegisterInfo *MRI) const {
// Get the unique definition of SrcReg.
MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
if (!MI)
return false;
// Get ready to iterate backward from CmpInstr.
MachineBasicBlock::iterator I = CmpInstr, E = MI,
B = CmpInstr.getParent()->begin();
// Early exit if CmpInstr is at the beginning of the BB.
if (I == B)
return false;
// There are two possible candidates which can be changed to set SR:
// One is MI, the other is a SUB instruction.
// * For SFSUB_F_RR(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
// * For SFSUB_F_RI(r1, CmpValue), we are looking for SUB(r1, CmpValue).
MachineInstr *Sub = nullptr;
if (SrcReg2 != 0)
// MI is not a candidate to transform into a flag setting instruction.
MI = nullptr;
else if (MI->getParent() != CmpInstr.getParent() || CmpValue != 0) {
// Conservatively refuse to convert an instruction which isn't in the same
// BB as the comparison. Don't return if SFSUB_F_RI and CmpValue != 0 as Sub
// may still be a candidate.
if (CmpInstr.getOpcode() == Lanai::SFSUB_F_RI_LO)
MI = nullptr;
else
return false;
}
// Check that SR isn't set between the comparison instruction and the
// instruction we want to change while searching for Sub.
const TargetRegisterInfo *TRI = &getRegisterInfo();
for (--I; I != E; --I) {
const MachineInstr &Instr = *I;
if (Instr.modifiesRegister(Lanai::SR, TRI) ||
Instr.readsRegister(Lanai::SR, TRI))
// This instruction modifies or uses SR after the one we want to change.
// We can't do this transformation.
return false;
// Check whether CmpInstr can be made redundant by the current instruction.
if (isRedundantFlagInstr(&CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
Sub = &*I;
break;
}
// Don't search outside the containing basic block.
if (I == B)
return false;
}
// Return false if no candidates exist.
if (!MI && !Sub)
return false;
// The single candidate is called MI.
if (!MI)
MI = Sub;
if (flagSettingOpcodeVariant(MI->getOpcode()) != Lanai::NOP) {
bool isSafe = false;
SmallVector<std::pair<MachineOperand *, LPCC::CondCode>, 4>
OperandsToUpdate;
I = CmpInstr;
E = CmpInstr.getParent()->end();
while (!isSafe && ++I != E) {
const MachineInstr &Instr = *I;
for (unsigned IO = 0, EO = Instr.getNumOperands(); !isSafe && IO != EO;
++IO) {
const MachineOperand &MO = Instr.getOperand(IO);
if (MO.isRegMask() && MO.clobbersPhysReg(Lanai::SR)) {
isSafe = true;
break;
}
if (!MO.isReg() || MO.getReg() != Lanai::SR)
continue;
if (MO.isDef()) {
isSafe = true;
break;
}
// Condition code is after the operand before SR.
LPCC::CondCode CC;
CC = (LPCC::CondCode)Instr.getOperand(IO - 1).getImm();
if (Sub) {
LPCC::CondCode NewCC = getOppositeCondition(CC);
if (NewCC == LPCC::ICC_T)
return false;
// If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on
// CMP needs to be updated to be based on SUB. Push the condition
// code operands to OperandsToUpdate. If it is safe to remove
// CmpInstr, the condition code of these operands will be modified.
if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
Sub->getOperand(2).getReg() == SrcReg) {
OperandsToUpdate.push_back(
std::make_pair(&((*I).getOperand(IO - 1)), NewCC));
}
} else {
// No Sub, so this is x = <op> y, z; cmp x, 0.
switch (CC) {
case LPCC::ICC_EQ: // Z
case LPCC::ICC_NE: // Z
case LPCC::ICC_MI: // N
case LPCC::ICC_PL: // N
case LPCC::ICC_F: // none
case LPCC::ICC_T: // none
// SR can be used multiple times, we should continue.
break;
case LPCC::ICC_CS: // C
case LPCC::ICC_CC: // C
case LPCC::ICC_VS: // V
case LPCC::ICC_VC: // V
case LPCC::ICC_HI: // C Z
case LPCC::ICC_LS: // C Z
case LPCC::ICC_GE: // N V
case LPCC::ICC_LT: // N V
case LPCC::ICC_GT: // Z N V
case LPCC::ICC_LE: // Z N V
// The instruction uses the V bit or C bit which is not safe.
return false;
case LPCC::UNKNOWN:
return false;
}
}
}
}
// If SR is not killed nor re-defined, we should check whether it is
// live-out. If it is live-out, do not optimize.
if (!isSafe) {
MachineBasicBlock *MBB = CmpInstr.getParent();
for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
SE = MBB->succ_end();
SI != SE; ++SI)
if ((*SI)->isLiveIn(Lanai::SR))
return false;
}
// Toggle the optional operand to SR.
MI->setDesc(get(flagSettingOpcodeVariant(MI->getOpcode())));
MI->addRegisterDefined(Lanai::SR);
CmpInstr.eraseFromParent();
return true;
}
return false;
}
bool LazyLiveness::runOnMachineFunction(MachineFunction &mf) {
rv.clear();
tv.clear();
backedges.clear();
backedge_source.clear();
backedge_target.clear();
calculated.clear();
preorder.clear();
rev_preorder.clear();
rv.resize(mf.size());
tv.resize(mf.size());
preorder.resize(mf.size());
rev_preorder.reserve(mf.size());
MRI = &mf.getRegInfo();
MachineDominatorTree& MDT = getAnalysis<MachineDominatorTree>();
// Step 0: Compute preorder numbering for all MBBs.
unsigned num = 0;
for (df_iterator<MachineDomTreeNode*> DI = df_begin(MDT.getRootNode()),
DE = df_end(MDT.getRootNode()); DI != DE; ++DI) {
preorder[(*DI)->getBlock()] = num++;
rev_preorder.push_back((*DI)->getBlock());
}
// Step 1: Compute the transitive closure of the CFG, ignoring backedges.
for (po_iterator<MachineBasicBlock*> POI = po_begin(&*mf.begin()),
POE = po_end(&*mf.begin()); POI != POE; ++POI) {
MachineBasicBlock* MBB = *POI;
SparseBitVector<128>& entry = rv[MBB];
entry.set(preorder[MBB]);
for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
SE = MBB->succ_end(); SI != SE; ++SI) {
DenseMap<MachineBasicBlock*, SparseBitVector<128> >::iterator SII =
rv.find(*SI);
// Because we're iterating in postorder, any successor that does not yet
// have an rv entry must be on a backedge.
if (SII != rv.end()) {
entry |= SII->second;
} else {
backedges.insert(std::make_pair(MBB, *SI));
backedge_source.set(preorder[MBB]);
backedge_target.set(preorder[*SI]);
}
}
}
for (SparseBitVector<128>::iterator I = backedge_source.begin();
I != backedge_source.end(); ++I)
computeBackedgeChain(mf, rev_preorder[*I]);
for (po_iterator<MachineBasicBlock*> POI = po_begin(&*mf.begin()),
POE = po_end(&*mf.begin()); POI != POE; ++POI)
if (!backedge_target.test(preorder[*POI]))
for (MachineBasicBlock::succ_iterator SI = (*POI)->succ_begin(),
SE = (*POI)->succ_end(); SI != SE; ++SI)
if (!backedges.count(std::make_pair(*POI, *SI)) && tv.count(*SI)) {
SparseBitVector<128> right = tv[*SI];
tv[*POI] |= right;
}
for (po_iterator<MachineBasicBlock*> POI = po_begin(&*mf.begin()),
POE = po_end(&*mf.begin()); POI != POE; ++POI)
tv[*POI].set(preorder[*POI]);
return false;
}
/// TailDuplicateBlocks - Look for small blocks that are unconditionally
/// branched to and do not fall through. Tail-duplicate their instructions
/// into their predecessors to eliminate (dynamic) branches.
bool TailDuplicatePass::TailDuplicateBlocks(MachineFunction &MF) {
bool MadeChange = false;
if (PreRegAlloc && TailDupVerify) {
DEBUG(dbgs() << "\n*** Before tail-duplicating\n");
VerifyPHIs(MF, true);
}
SmallVector<MachineInstr*, 8> NewPHIs;
MachineSSAUpdater SSAUpdate(MF, &NewPHIs);
for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
MachineBasicBlock *MBB = I++;
if (NumTails == TailDupLimit)
break;
// Save the successors list.
SmallSetVector<MachineBasicBlock*, 8> Succs(MBB->succ_begin(),
MBB->succ_end());
SmallVector<MachineBasicBlock*, 8> TDBBs;
SmallVector<MachineInstr*, 16> Copies;
if (TailDuplicate(MBB, MF, TDBBs, Copies)) {
++NumTails;
// TailBB's immediate successors are now successors of those predecessors
// which duplicated TailBB. Add the predecessors as sources to the PHI
// instructions.
bool isDead = MBB->pred_empty();
if (PreRegAlloc)
UpdateSuccessorsPHIs(MBB, isDead, TDBBs, Succs);
// If it is dead, remove it.
if (isDead) {
NumInstrDups -= MBB->size();
RemoveDeadBlock(MBB);
++NumDeadBlocks;
}
// Update SSA form.
if (!SSAUpdateVRs.empty()) {
for (unsigned i = 0, e = SSAUpdateVRs.size(); i != e; ++i) {
unsigned VReg = SSAUpdateVRs[i];
SSAUpdate.Initialize(VReg);
// If the original definition is still around, add it as an available
// value.
MachineInstr *DefMI = MRI->getVRegDef(VReg);
MachineBasicBlock *DefBB = 0;
if (DefMI) {
DefBB = DefMI->getParent();
SSAUpdate.AddAvailableValue(DefBB, VReg);
}
// Add the new vregs as available values.
DenseMap<unsigned, AvailableValsTy>::iterator LI =
SSAUpdateVals.find(VReg);
for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
MachineBasicBlock *SrcBB = LI->second[j].first;
unsigned SrcReg = LI->second[j].second;
SSAUpdate.AddAvailableValue(SrcBB, SrcReg);
}
// Rewrite uses that are outside of the original def's block.
MachineRegisterInfo::use_iterator UI = MRI->use_begin(VReg);
while (UI != MRI->use_end()) {
MachineOperand &UseMO = UI.getOperand();
MachineInstr *UseMI = &*UI;
++UI;
if (UseMI->isDebugValue()) {
// SSAUpdate can replace the use with an undef. That creates
// a debug instruction that is a kill.
// FIXME: Should it SSAUpdate job to delete debug instructions
// instead of replacing the use with undef?
UseMI->eraseFromParent();
continue;
}
if (UseMI->getParent() == DefBB && !UseMI->isPHI())
continue;
SSAUpdate.RewriteUse(UseMO);
}
}
SSAUpdateVRs.clear();
SSAUpdateVals.clear();
}
// Eliminate some of the copies inserted by tail duplication to maintain
// SSA form.
for (unsigned i = 0, e = Copies.size(); i != e; ++i) {
MachineInstr *Copy = Copies[i];
if (!Copy->isCopy())
continue;
unsigned Dst = Copy->getOperand(0).getReg();
unsigned Src = Copy->getOperand(1).getReg();
MachineRegisterInfo::use_iterator UI = MRI->use_begin(Src);
if (++UI == MRI->use_end()) {
// Copy is the only use. Do trivial copy propagation here.
MRI->replaceRegWith(Dst, Src);
Copy->eraseFromParent();
}
}
if (PreRegAlloc && TailDupVerify)
VerifyPHIs(MF, false);
MadeChange = true;
}
}
NumAddedPHIs += NewPHIs.size();
return MadeChange;
}
/// SinkInstruction - Determine whether it is safe to sink the specified machine
/// instruction out of its current block into a successor.
bool MachineSinking::SinkInstruction(MachineInstr *MI, bool &SawStore) {
// Check if it's safe to move the instruction.
if (!MI->isSafeToMove(TII, AA, SawStore))
return false;
// FIXME: This should include support for sinking instructions within the
// block they are currently in to shorten the live ranges. We often get
// instructions sunk into the top of a large block, but it would be better to
// also sink them down before their first use in the block. This xform has to
// be careful not to *increase* register pressure though, e.g. sinking
// "x = y + z" down if it kills y and z would increase the live ranges of y
// and z and only shrink the live range of x.
// Loop over all the operands of the specified instruction. If there is
// anything we can't handle, bail out.
MachineBasicBlock *ParentBlock = MI->getParent();
// SuccToSinkTo - This is the successor to sink this instruction to, once we
// decide.
MachineBasicBlock *SuccToSinkTo = 0;
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue; // Ignore non-register operands.
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
if (MO.isUse()) {
// 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 (!RegInfo->def_empty(Reg))
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 (!RegInfo->def_empty(AliasReg))
return false;
if (AllocatableSet.test(AliasReg))
return false;
}
} else if (!MO.isDead()) {
// A def that isn't dead. We can't move it.
return false;
}
} else {
// Virtual register uses are always safe to sink.
if (MO.isUse()) continue;
// If it's not safe to move defs of the register class, then abort.
if (!TII->isSafeToMoveRegClassDefs(RegInfo->getRegClass(Reg)))
return false;
// FIXME: This picks a successor to sink into based on having one
// successor that dominates all the uses. However, there are cases where
// sinking can happen but where the sink point isn't a successor. For
// example:
// x = computation
// if () {} else {}
// use x
// the instruction could be sunk over the whole diamond for the
// if/then/else (or loop, etc), allowing it to be sunk into other blocks
// after that.
// Virtual register defs can only be sunk if all their uses are in blocks
// dominated by one of the successors.
if (SuccToSinkTo) {
// If a previous operand picked a block to sink to, then this operand
// must be sinkable to the same block.
if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo))
return false;
continue;
}
// Otherwise, we should look at all the successors and decide which one
// we should sink to.
for (MachineBasicBlock::succ_iterator SI = ParentBlock->succ_begin(),
E = ParentBlock->succ_end(); SI != E; ++SI) {
if (AllUsesDominatedByBlock(Reg, *SI)) {
SuccToSinkTo = *SI;
break;
}
}
// If we couldn't find a block to sink to, ignore this instruction.
if (SuccToSinkTo == 0)
return false;
}
}
// If there are no outputs, it must have side-effects.
if (SuccToSinkTo == 0)
return false;
// It's not safe to sink instructions to EH landing pad. Control flow into
// landing pad is implicitly defined.
if (SuccToSinkTo->isLandingPad())
return false;
// It is not possible to sink an instruction into its own block. This can
// happen with loops.
if (MI->getParent() == SuccToSinkTo)
return false;
DEBUG(dbgs() << "Sink instr " << *MI);
DEBUG(dbgs() << "to block " << *SuccToSinkTo);
// If the block has multiple predecessors, this would introduce computation on
// a path that it doesn't already exist. We could split the critical edge,
// but for now we just punt.
// FIXME: Split critical edges if not backedges.
if (SuccToSinkTo->pred_size() > 1) {
DEBUG(dbgs() << " *** PUNTING: Critical edge found\n");
return false;
}
// Determine where to insert into. Skip phi nodes.
MachineBasicBlock::iterator InsertPos = SuccToSinkTo->begin();
while (InsertPos != SuccToSinkTo->end() && InsertPos->isPHI())
++InsertPos;
// Move the instruction.
SuccToSinkTo->splice(InsertPos, ParentBlock, MI,
++MachineBasicBlock::iterator(MI));
return true;
}
/// SplitInterferencesForBasicBlock - traverses a basic block, splitting any
/// interferences found between registers in the same congruence class. It
/// takes two DenseMaps as arguments that it also updates:
///
/// 1) CurrentDominatingParent, which maps a color to the register in that
/// congruence class whose definition was most recently seen.
///
/// 2) ImmediateDominatingParent, which maps a register to the register in the
/// same congruence class that most immediately dominates it.
///
/// This function assumes that it is being called in a depth-first traversal
/// of the dominator tree.
///
/// The algorithm used here is a generalization of the dominance-based SSA test
/// for two variables. If there are variables a_1, ..., a_n such that
///
/// def(a_1) dom ... dom def(a_n),
///
/// then we can test for an interference between any two a_i by only using O(n)
/// interference tests between pairs of variables. If i < j and a_i and a_j
/// interfere, then a_i is alive at def(a_j), so it is also alive at def(a_i+1).
/// Thus, in order to test for an interference involving a_i, we need only check
/// for a potential interference with a_i+1.
///
/// This method can be generalized to arbitrary sets of variables by performing
/// a depth-first traversal of the dominator tree. As we traverse down a branch
/// of the dominator tree, we keep track of the current dominating variable and
/// only perform an interference test with that variable. However, when we go to
/// another branch of the dominator tree, the definition of the current dominating
/// variable may no longer dominate the current block. In order to correct this,
/// we need to use a stack of past choices of the current dominating variable
/// and pop from this stack until we find a variable whose definition actually
/// dominates the current block.
///
/// There will be one push on this stack for each variable that has become the
/// current dominating variable, so instead of using an explicit stack we can
/// simply associate the previous choice for a current dominating variable with
/// the new choice. This works better in our implementation, where we test for
/// interference in multiple distinct sets at once.
void
StrongPHIElimination::SplitInterferencesForBasicBlock(
MachineBasicBlock &MBB,
DenseMap<unsigned, unsigned> &CurrentDominatingParent,
DenseMap<unsigned, unsigned> &ImmediateDominatingParent) {
// Sort defs by their order in the original basic block, as the code below
// assumes that it is processing definitions in dominance order.
std::vector<MachineInstr*> &DefInstrs = PHISrcDefs[&MBB];
std::sort(DefInstrs.begin(), DefInstrs.end(), MIIndexCompare(LI));
for (std::vector<MachineInstr*>::const_iterator BBI = DefInstrs.begin(),
BBE = DefInstrs.end(); BBI != BBE; ++BBI) {
for (MachineInstr::const_mop_iterator I = (*BBI)->operands_begin(),
E = (*BBI)->operands_end(); I != E; ++I) {
const MachineOperand &MO = *I;
// FIXME: This would be faster if it were possible to bail out of checking
// an instruction's operands after the explicit defs, but this is incorrect
// for variadic instructions, which may appear before register allocation
// in the future.
if (!MO.isReg() || !MO.isDef())
continue;
unsigned DestReg = MO.getReg();
if (!DestReg || !TargetRegisterInfo::isVirtualRegister(DestReg))
continue;
// If the virtual register being defined is not used in any PHI or has
// already been isolated, then there are no more interferences to check.
unsigned DestColor = getRegColor(DestReg);
if (!DestColor)
continue;
// The input to this pass sometimes is not in SSA form in every basic
// block, as some virtual registers have redefinitions. We could eliminate
// this by fixing the passes that generate the non-SSA code, or we could
// handle it here by tracking defining machine instructions rather than
// virtual registers. For now, we just handle the situation conservatively
// in a way that will possibly lead to false interferences.
unsigned &CurrentParent = CurrentDominatingParent[DestColor];
unsigned NewParent = CurrentParent;
if (NewParent == DestReg)
continue;
// Pop registers from the stack represented by ImmediateDominatingParent
// until we find a parent that dominates the current instruction.
while (NewParent && (!DT->dominates(MRI->getVRegDef(NewParent), *BBI)
|| !getRegColor(NewParent)))
NewParent = ImmediateDominatingParent[NewParent];
// If NewParent is nonzero, then its definition dominates the current
// instruction, so it is only necessary to check for the liveness of
// NewParent in order to check for an interference.
if (NewParent
&& LI->getInterval(NewParent).liveAt(LI->getInstructionIndex(*BBI))) {
// If there is an interference, always isolate the new register. This
// could be improved by using a heuristic that decides which of the two
// registers to isolate.
isolateReg(DestReg);
CurrentParent = NewParent;
} else {
// If there is no interference, update ImmediateDominatingParent and set
// the CurrentDominatingParent for this color to the current register.
ImmediateDominatingParent[DestReg] = NewParent;
CurrentParent = DestReg;
}
}
}
// We now walk the PHIs in successor blocks and check for interferences. This
// is necessary because the use of a PHI's operands are logically contained in
// the predecessor block. The def of a PHI's destination register is processed
// along with the other defs in a basic block.
CurrentPHIForColor.clear();
for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
SE = MBB.succ_end(); SI != SE; ++SI) {
for (MachineBasicBlock::iterator BBI = (*SI)->begin(), BBE = (*SI)->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
MachineInstr *PHI = BBI;
// If a PHI is already isolated, either by being isolated directly or
// having all of its operands isolated, ignore it.
unsigned Color = getPHIColor(PHI);
if (!Color)
continue;
// Find the index of the PHI operand that corresponds to this basic block.
unsigned PredIndex;
for (PredIndex = 1; PredIndex < PHI->getNumOperands(); PredIndex += 2) {
if (PHI->getOperand(PredIndex + 1).getMBB() == &MBB)
break;
}
assert(PredIndex < PHI->getNumOperands());
unsigned PredOperandReg = PHI->getOperand(PredIndex).getReg();
// Pop registers from the stack represented by ImmediateDominatingParent
// until we find a parent that dominates the current instruction.
unsigned &CurrentParent = CurrentDominatingParent[Color];
unsigned NewParent = CurrentParent;
while (NewParent
&& (!DT->dominates(MRI->getVRegDef(NewParent)->getParent(), &MBB)
|| !getRegColor(NewParent)))
NewParent = ImmediateDominatingParent[NewParent];
CurrentParent = NewParent;
// If there is an interference with a register, always isolate the
// register rather than the PHI. It is also possible to isolate the
// PHI, but that introduces copies for all of the registers involved
// in that PHI.
if (NewParent && LI->isLiveOutOfMBB(LI->getInterval(NewParent), &MBB)
&& NewParent != PredOperandReg)
isolateReg(NewParent);
std::pair<MachineInstr*, unsigned>
&CurrentPHI = CurrentPHIForColor[Color];
// If two PHIs have the same operand from every shared predecessor, then
// they don't actually interfere. Otherwise, isolate the current PHI. This
// could possibly be improved, e.g. we could isolate the PHI with the
// fewest operands.
if (CurrentPHI.first && CurrentPHI.second != PredOperandReg)
isolatePHI(PHI);
else
CurrentPHI = std::make_pair(PHI, PredOperandReg);
}
}
}
bool StrongPHIElimination::runOnMachineFunction(MachineFunction &MF) {
MRI = &MF.getRegInfo();
TII = MF.getTarget().getInstrInfo();
DT = &getAnalysis<MachineDominatorTree>();
LI = &getAnalysis<LiveIntervals>();
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
for (MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
unsigned DestReg = BBI->getOperand(0).getReg();
addReg(DestReg);
PHISrcDefs[I].push_back(BBI);
for (unsigned i = 1; i < BBI->getNumOperands(); i += 2) {
MachineOperand &SrcMO = BBI->getOperand(i);
unsigned SrcReg = SrcMO.getReg();
addReg(SrcReg);
unionRegs(DestReg, SrcReg);
MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
if (DefMI)
PHISrcDefs[DefMI->getParent()].push_back(DefMI);
}
}
}
// Perform a depth-first traversal of the dominator tree, splitting
// interferences amongst PHI-congruence classes.
DenseMap<unsigned, unsigned> CurrentDominatingParent;
DenseMap<unsigned, unsigned> ImmediateDominatingParent;
for (df_iterator<MachineDomTreeNode*> DI = df_begin(DT->getRootNode()),
DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
SplitInterferencesForBasicBlock(*DI->getBlock(),
CurrentDominatingParent,
ImmediateDominatingParent);
}
// Insert copies for all PHI source and destination registers.
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
for (MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
InsertCopiesForPHI(BBI, I);
}
}
// FIXME: Preserve the equivalence classes during copy insertion and use
// the preversed equivalence classes instead of recomputing them.
RegNodeMap.clear();
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
for (MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
unsigned DestReg = BBI->getOperand(0).getReg();
addReg(DestReg);
for (unsigned i = 1; i < BBI->getNumOperands(); i += 2) {
unsigned SrcReg = BBI->getOperand(i).getReg();
addReg(SrcReg);
unionRegs(DestReg, SrcReg);
}
}
}
DenseMap<unsigned, unsigned> RegRenamingMap;
bool Changed = false;
for (MachineFunction::iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
MachineBasicBlock::iterator BBI = I->begin(), BBE = I->end();
while (BBI != BBE && BBI->isPHI()) {
MachineInstr *PHI = BBI;
assert(PHI->getNumOperands() > 0);
unsigned SrcReg = PHI->getOperand(1).getReg();
unsigned SrcColor = getRegColor(SrcReg);
unsigned NewReg = RegRenamingMap[SrcColor];
if (!NewReg) {
NewReg = SrcReg;
RegRenamingMap[SrcColor] = SrcReg;
}
MergeLIsAndRename(SrcReg, NewReg);
unsigned DestReg = PHI->getOperand(0).getReg();
if (!InsertedDestCopies.count(DestReg))
MergeLIsAndRename(DestReg, NewReg);
for (unsigned i = 3; i < PHI->getNumOperands(); i += 2) {
unsigned SrcReg = PHI->getOperand(i).getReg();
MergeLIsAndRename(SrcReg, NewReg);
}
++BBI;
LI->RemoveMachineInstrFromMaps(PHI);
PHI->eraseFromParent();
Changed = true;
}
}
// Due to the insertion of copies to split live ranges, the live intervals are
// guaranteed to not overlap, except in one case: an original PHI source and a
// PHI destination copy. In this case, they have the same value and thus don't
// truly intersect, so we merge them into the value live at that point.
// FIXME: Is there some better way we can handle this?
for (DestCopyMap::iterator I = InsertedDestCopies.begin(),
E = InsertedDestCopies.end(); I != E; ++I) {
unsigned DestReg = I->first;
unsigned DestColor = getRegColor(DestReg);
unsigned NewReg = RegRenamingMap[DestColor];
LiveInterval &DestLI = LI->getInterval(DestReg);
LiveInterval &NewLI = LI->getInterval(NewReg);
assert(DestLI.ranges.size() == 1
&& "PHI destination copy's live interval should be a single live "
"range from the beginning of the BB to the copy instruction.");
LiveRange *DestLR = DestLI.begin();
VNInfo *NewVNI = NewLI.getVNInfoAt(DestLR->start);
if (!NewVNI) {
NewVNI = NewLI.createValueCopy(DestLR->valno, LI->getVNInfoAllocator());
MachineInstr *CopyInstr = I->second;
CopyInstr->getOperand(1).setIsKill(true);
}
LiveRange NewLR(DestLR->start, DestLR->end, NewVNI);
NewLI.addRange(NewLR);
LI->removeInterval(DestReg);
MRI->replaceRegWith(DestReg, NewReg);
}
// Adjust the live intervals of all PHI source registers to handle the case
// where the PHIs in successor blocks were the only later uses of the source
// register.
for (SrcCopySet::iterator I = InsertedSrcCopySet.begin(),
E = InsertedSrcCopySet.end(); I != E; ++I) {
MachineBasicBlock *MBB = I->first;
unsigned SrcReg = I->second;
if (unsigned RenamedRegister = RegRenamingMap[getRegColor(SrcReg)])
SrcReg = RenamedRegister;
LiveInterval &SrcLI = LI->getInterval(SrcReg);
bool isLiveOut = false;
for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
SE = MBB->succ_end(); SI != SE; ++SI) {
if (SrcLI.liveAt(LI->getMBBStartIdx(*SI))) {
isLiveOut = true;
break;
}
}
if (isLiveOut)
continue;
MachineOperand *LastUse = findLastUse(MBB, SrcReg);
assert(LastUse);
SlotIndex LastUseIndex = LI->getInstructionIndex(LastUse->getParent());
SrcLI.removeRange(LastUseIndex.getRegSlot(), LI->getMBBEndIdx(MBB));
LastUse->setIsKill(true);
}
Allocator.Reset();
RegNodeMap.clear();
PHISrcDefs.clear();
InsertedSrcCopySet.clear();
InsertedSrcCopyMap.clear();
InsertedDestCopies.clear();
return Changed;
}
bool DeadMachineInstructionElim::runOnMachineFunction(MachineFunction &MF) {
if (skipOptnoneFunction(*MF.getFunction()))
return false;
bool AnyChanges = false;
MRI = &MF.getRegInfo();
TRI = MF.getSubtarget().getRegisterInfo();
TII = MF.getSubtarget().getInstrInfo();
// Loop over all instructions in all blocks, from bottom to top, so that it's
// more likely that chains of dependent but ultimately dead instructions will
// be cleaned up.
for (MachineFunction::reverse_iterator I = MF.rbegin(), E = MF.rend();
I != E; ++I) {
MachineBasicBlock *MBB = &*I;
// Start out assuming that reserved registers are live out of this block.
LivePhysRegs = MRI->getReservedRegs();
// Add live-ins from sucessors to LivePhysRegs. Normally, physregs are not
// live across blocks, but some targets (x86) can have flags live out of a
// block.
for (MachineBasicBlock::succ_iterator S = MBB->succ_begin(),
E = MBB->succ_end(); S != E; S++)
for (MachineBasicBlock::livein_iterator LI = (*S)->livein_begin();
LI != (*S)->livein_end(); LI++)
LivePhysRegs.set(*LI);
// Now scan the instructions and delete dead ones, tracking physreg
// liveness as we go.
for (MachineBasicBlock::reverse_iterator MII = MBB->rbegin(),
MIE = MBB->rend(); MII != MIE; ) {
MachineInstr *MI = &*MII;
// If the instruction is dead, delete it!
if (isDead(MI)) {
DEBUG(dbgs() << "DeadMachineInstructionElim: DELETING: " << *MI);
// It is possible that some DBG_VALUE instructions refer to this
// instruction. They get marked as undef and will be deleted
// in the live debug variable analysis.
MI->eraseFromParentAndMarkDBGValuesForRemoval();
AnyChanges = true;
++NumDeletes;
MIE = MBB->rend();
// MII is now pointing to the next instruction to process,
// so don't increment it.
continue;
}
// Record the physreg defs.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
if (MO.isReg() && MO.isDef()) {
unsigned Reg = MO.getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
// Check the subreg set, not the alias set, because a def
// of a super-register may still be partially live after
// this def.
for (MCSubRegIterator SR(Reg, TRI,/*IncludeSelf=*/true);
SR.isValid(); ++SR)
LivePhysRegs.reset(*SR);
}
} else if (MO.isRegMask()) {
// Register mask of preserved registers. All clobbers are dead.
LivePhysRegs.clearBitsNotInMask(MO.getRegMask());
}
}
// Record the physreg uses, after the defs, in case a physreg is
// both defined and used in the same instruction.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
if (MO.isReg() && MO.isUse()) {
unsigned Reg = MO.getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
LivePhysRegs.set(*AI);
}
}
}
// We didn't delete the current instruction, so increment MII to
// the next one.
++MII;
}
}
LivePhysRegs.clear();
return AnyChanges;
}
bool LiveVariables::runOnMachineFunction(MachineFunction &mf) {
MF = &mf;
MRI = &mf.getRegInfo();
TRI = MF->getTarget().getRegisterInfo();
ReservedRegisters = TRI->getReservedRegs(mf);
unsigned NumRegs = TRI->getNumRegs();
PhysRegDef = new MachineInstr*[NumRegs];
PhysRegUse = new MachineInstr*[NumRegs];
PHIVarInfo = new SmallVector<unsigned, 4>[MF->getNumBlockIDs()];
std::fill(PhysRegDef, PhysRegDef + NumRegs, (MachineInstr*)0);
std::fill(PhysRegUse, PhysRegUse + NumRegs, (MachineInstr*)0);
PHIJoins.clear();
// FIXME: LiveIntervals will be updated to remove its dependence on
// LiveVariables to improve compilation time and eliminate bizarre pass
// dependencies. Until then, we can't change much in -O0.
if (!MRI->isSSA())
report_fatal_error("regalloc=... not currently supported with -O0");
analyzePHINodes(mf);
// Calculate live variable information in depth first order on the CFG of the
// function. This guarantees that we will see the definition of a virtual
// register before its uses due to dominance properties of SSA (except for PHI
// nodes, which are treated as a special case).
MachineBasicBlock *Entry = MF->begin();
SmallPtrSet<MachineBasicBlock*,16> Visited;
for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*,16> >
DFI = df_ext_begin(Entry, Visited), E = df_ext_end(Entry, Visited);
DFI != E; ++DFI) {
MachineBasicBlock *MBB = *DFI;
// Mark live-in registers as live-in.
SmallVector<unsigned, 4> Defs;
for (MachineBasicBlock::livein_iterator II = MBB->livein_begin(),
EE = MBB->livein_end(); II != EE; ++II) {
assert(TargetRegisterInfo::isPhysicalRegister(*II) &&
"Cannot have a live-in virtual register!");
HandlePhysRegDef(*II, 0, Defs);
}
// Loop over all of the instructions, processing them.
DistanceMap.clear();
unsigned Dist = 0;
for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
I != E; ++I) {
MachineInstr *MI = I;
if (MI->isDebugValue())
continue;
DistanceMap.insert(std::make_pair(MI, Dist++));
// Process all of the operands of the instruction...
unsigned NumOperandsToProcess = MI->getNumOperands();
// Unless it is a PHI node. In this case, ONLY process the DEF, not any
// of the uses. They will be handled in other basic blocks.
if (MI->isPHI())
NumOperandsToProcess = 1;
// Clear kill and dead markers. LV will recompute them.
SmallVector<unsigned, 4> UseRegs;
SmallVector<unsigned, 4> DefRegs;
SmallVector<unsigned, 1> RegMasks;
for (unsigned i = 0; i != NumOperandsToProcess; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (MO.isRegMask()) {
RegMasks.push_back(i);
continue;
}
if (!MO.isReg() || MO.getReg() == 0)
continue;
unsigned MOReg = MO.getReg();
if (MO.isUse()) {
MO.setIsKill(false);
UseRegs.push_back(MOReg);
} else /*MO.isDef()*/ {
MO.setIsDead(false);
DefRegs.push_back(MOReg);
}
}
// Process all uses.
for (unsigned i = 0, e = UseRegs.size(); i != e; ++i) {
unsigned MOReg = UseRegs[i];
if (TargetRegisterInfo::isVirtualRegister(MOReg))
HandleVirtRegUse(MOReg, MBB, MI);
else if (!ReservedRegisters[MOReg])
HandlePhysRegUse(MOReg, MI);
}
// Process all masked registers. (Call clobbers).
for (unsigned i = 0, e = RegMasks.size(); i != e; ++i)
HandleRegMask(MI->getOperand(RegMasks[i]));
// Process all defs.
for (unsigned i = 0, e = DefRegs.size(); i != e; ++i) {
unsigned MOReg = DefRegs[i];
if (TargetRegisterInfo::isVirtualRegister(MOReg))
HandleVirtRegDef(MOReg, MI);
else if (!ReservedRegisters[MOReg])
HandlePhysRegDef(MOReg, MI, Defs);
}
UpdatePhysRegDefs(MI, Defs);
}
// Handle any virtual assignments from PHI nodes which might be at the
// bottom of this basic block. We check all of our successor blocks to see
// if they have PHI nodes, and if so, we simulate an assignment at the end
// of the current block.
if (!PHIVarInfo[MBB->getNumber()].empty()) {
SmallVector<unsigned, 4>& VarInfoVec = PHIVarInfo[MBB->getNumber()];
for (SmallVector<unsigned, 4>::iterator I = VarInfoVec.begin(),
E = VarInfoVec.end(); I != E; ++I)
// Mark it alive only in the block we are representing.
MarkVirtRegAliveInBlock(getVarInfo(*I),MRI->getVRegDef(*I)->getParent(),
MBB);
}
// Finally, if the last instruction in the block is a return, make sure to
// mark it as using all of the live-out values in the function.
// Things marked both call and return are tail calls; do not do this for
// them. The tail callee need not take the same registers as input
// that it produces as output, and there are dependencies for its input
// registers elsewhere.
if (!MBB->empty() && MBB->back().isReturn()
&& !MBB->back().isCall()) {
MachineInstr *Ret = &MBB->back();
for (MachineRegisterInfo::liveout_iterator
I = MF->getRegInfo().liveout_begin(),
E = MF->getRegInfo().liveout_end(); I != E; ++I) {
assert(TargetRegisterInfo::isPhysicalRegister(*I) &&
"Cannot have a live-out virtual register!");
HandlePhysRegUse(*I, Ret);
// Add live-out registers as implicit uses.
if (!Ret->readsRegister(*I))
Ret->addOperand(MachineOperand::CreateReg(*I, false, true));
}
}
// MachineCSE may CSE instructions which write to non-allocatable physical
// registers across MBBs. Remember if any reserved register is liveout.
SmallSet<unsigned, 4> LiveOuts;
for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
SE = MBB->succ_end(); SI != SE; ++SI) {
MachineBasicBlock *SuccMBB = *SI;
if (SuccMBB->isLandingPad())
continue;
for (MachineBasicBlock::livein_iterator LI = SuccMBB->livein_begin(),
LE = SuccMBB->livein_end(); LI != LE; ++LI) {
unsigned LReg = *LI;
if (!TRI->isInAllocatableClass(LReg))
// Ignore other live-ins, e.g. those that are live into landing pads.
LiveOuts.insert(LReg);
}
}
// Loop over PhysRegDef / PhysRegUse, killing any registers that are
// available at the end of the basic block.
for (unsigned i = 0; i != NumRegs; ++i)
if ((PhysRegDef[i] || PhysRegUse[i]) && !LiveOuts.count(i))
HandlePhysRegDef(i, 0, Defs);
std::fill(PhysRegDef, PhysRegDef + NumRegs, (MachineInstr*)0);
std::fill(PhysRegUse, PhysRegUse + NumRegs, (MachineInstr*)0);
}
// Convert and transfer the dead / killed information we have gathered into
// VirtRegInfo onto MI's.
for (unsigned i = 0, e1 = VirtRegInfo.size(); i != e1; ++i) {
const unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
for (unsigned j = 0, e2 = VirtRegInfo[Reg].Kills.size(); j != e2; ++j)
if (VirtRegInfo[Reg].Kills[j] == MRI->getVRegDef(Reg))
VirtRegInfo[Reg].Kills[j]->addRegisterDead(Reg, TRI);
else
VirtRegInfo[Reg].Kills[j]->addRegisterKilled(Reg, TRI);
}
// Check to make sure there are no unreachable blocks in the MC CFG for the
// function. If so, it is due to a bug in the instruction selector or some
// other part of the code generator if this happens.
#ifndef NDEBUG
for(MachineFunction::iterator i = MF->begin(), e = MF->end(); i != e; ++i)
assert(Visited.count(&*i) != 0 && "unreachable basic block found");
#endif
delete[] PhysRegDef;
delete[] PhysRegUse;
delete[] PHIVarInfo;
return false;
}
bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) {
DEBUG(dbgs() << "********** Hexagon New Value Jump **********\n"
<< "********** Function: "
<< MF.getName() << "\n");
if (skipFunction(*MF.getFunction()))
return false;
// If we move NewValueJump before register allocation we'll need live variable
// analysis here too.
QII = static_cast<const HexagonInstrInfo *>(MF.getSubtarget().getInstrInfo());
QRI = static_cast<const HexagonRegisterInfo *>(
MF.getSubtarget().getRegisterInfo());
MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
if (DisableNewValueJumps) {
return false;
}
int nvjCount = DbgNVJCount;
int nvjGenerated = 0;
// Loop through all the bb's of the function
for (MachineFunction::iterator MBBb = MF.begin(), MBBe = MF.end();
MBBb != MBBe; ++MBBb) {
MachineBasicBlock *MBB = &*MBBb;
DEBUG(dbgs() << "** dumping bb ** "
<< MBB->getNumber() << "\n");
DEBUG(MBB->dump());
DEBUG(dbgs() << "\n" << "********** dumping instr bottom up **********\n");
bool foundJump = false;
bool foundCompare = false;
bool invertPredicate = false;
unsigned predReg = 0; // predicate reg of the jump.
unsigned cmpReg1 = 0;
int cmpOp2 = 0;
bool MO1IsKill = false;
bool MO2IsKill = false;
MachineBasicBlock::iterator jmpPos;
MachineBasicBlock::iterator cmpPos;
MachineInstr *cmpInstr = nullptr, *jmpInstr = nullptr;
MachineBasicBlock *jmpTarget = nullptr;
bool afterRA = false;
bool isSecondOpReg = false;
bool isSecondOpNewified = false;
// Traverse the basic block - bottom up
for (MachineBasicBlock::iterator MII = MBB->end(), E = MBB->begin();
MII != E;) {
MachineInstr &MI = *--MII;
if (MI.isDebugValue()) {
continue;
}
if ((nvjCount == 0) || (nvjCount > -1 && nvjCount <= nvjGenerated))
break;
DEBUG(dbgs() << "Instr: "; MI.dump(); dbgs() << "\n");
if (!foundJump && (MI.getOpcode() == Hexagon::J2_jumpt ||
MI.getOpcode() == Hexagon::J2_jumpf ||
MI.getOpcode() == Hexagon::J2_jumptnewpt ||
MI.getOpcode() == Hexagon::J2_jumptnew ||
MI.getOpcode() == Hexagon::J2_jumpfnewpt ||
MI.getOpcode() == Hexagon::J2_jumpfnew)) {
// This is where you would insert your compare and
// instr that feeds compare
jmpPos = MII;
jmpInstr = &MI;
predReg = MI.getOperand(0).getReg();
afterRA = TargetRegisterInfo::isPhysicalRegister(predReg);
// If ifconverter had not messed up with the kill flags of the
// operands, the following check on the kill flag would suffice.
// if(!jmpInstr->getOperand(0).isKill()) break;
// This predicate register is live out out of BB
// this would only work if we can actually use Live
// variable analysis on phy regs - but LLVM does not
// provide LV analysis on phys regs.
//if(LVs.isLiveOut(predReg, *MBB)) break;
// Get all the successors of this block - which will always
// be 2. Check if the predicate register is live in in those
// successor. If yes, we can not delete the predicate -
// I am doing this only because LLVM does not provide LiveOut
// at the BB level.
bool predLive = false;
for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
SIE = MBB->succ_end(); SI != SIE; ++SI) {
MachineBasicBlock* succMBB = *SI;
if (succMBB->isLiveIn(predReg)) {
predLive = true;
}
}
if (predLive)
break;
if (!MI.getOperand(1).isMBB())
continue;
jmpTarget = MI.getOperand(1).getMBB();
foundJump = true;
if (MI.getOpcode() == Hexagon::J2_jumpf ||
MI.getOpcode() == Hexagon::J2_jumpfnewpt ||
MI.getOpcode() == Hexagon::J2_jumpfnew) {
invertPredicate = true;
}
continue;
}
// No new value jump if there is a barrier. A barrier has to be in its
// own packet. A barrier has zero operands. We conservatively bail out
// here if we see any instruction with zero operands.
if (foundJump && MI.getNumOperands() == 0)
break;
if (foundJump && !foundCompare && MI.getOperand(0).isReg() &&
MI.getOperand(0).getReg() == predReg) {
// Not all compares can be new value compare. Arch Spec: 7.6.1.1
if (isNewValueJumpCandidate(MI)) {
assert(
(MI.getDesc().isCompare()) &&
"Only compare instruction can be collapsed into New Value Jump");
isSecondOpReg = MI.getOperand(2).isReg();
if (!canCompareBeNewValueJump(QII, QRI, MII, predReg, isSecondOpReg,
afterRA, jmpPos, MF))
break;
cmpInstr = &MI;
cmpPos = MII;
foundCompare = true;
// We need cmpReg1 and cmpOp2(imm or reg) while building
// new value jump instruction.
cmpReg1 = MI.getOperand(1).getReg();
if (MI.getOperand(1).isKill())
MO1IsKill = true;
if (isSecondOpReg) {
cmpOp2 = MI.getOperand(2).getReg();
if (MI.getOperand(2).isKill())
MO2IsKill = true;
} else
cmpOp2 = MI.getOperand(2).getImm();
continue;
}
}
if (foundCompare && foundJump) {
// If "common" checks fail, bail out on this BB.
if (!commonChecksToProhibitNewValueJump(afterRA, MII))
break;
bool foundFeeder = false;
MachineBasicBlock::iterator feederPos = MII;
if (MI.getOperand(0).isReg() && MI.getOperand(0).isDef() &&
(MI.getOperand(0).getReg() == cmpReg1 ||
(isSecondOpReg &&
MI.getOperand(0).getReg() == (unsigned)cmpOp2))) {
unsigned feederReg = MI.getOperand(0).getReg();
// First try to see if we can get the feeder from the first operand
// of the compare. If we can not, and if secondOpReg is true
// (second operand of the compare is also register), try that one.
// TODO: Try to come up with some heuristic to figure out which
// feeder would benefit.
if (feederReg == cmpReg1) {
if (!canBeFeederToNewValueJump(QII, QRI, MII, jmpPos, cmpPos, MF)) {
if (!isSecondOpReg)
break;
else
continue;
} else
foundFeeder = true;
}
if (!foundFeeder &&
isSecondOpReg &&
feederReg == (unsigned) cmpOp2)
if (!canBeFeederToNewValueJump(QII, QRI, MII, jmpPos, cmpPos, MF))
break;
if (isSecondOpReg) {
// In case of CMPLT, or CMPLTU, or EQ with the second register
// to newify, swap the operands.
if (cmpInstr->getOpcode() == Hexagon::C2_cmpeq &&
feederReg == (unsigned) cmpOp2) {
unsigned tmp = cmpReg1;
bool tmpIsKill = MO1IsKill;
cmpReg1 = cmpOp2;
MO1IsKill = MO2IsKill;
cmpOp2 = tmp;
MO2IsKill = tmpIsKill;
}
// Now we have swapped the operands, all we need to check is,
// if the second operand (after swap) is the feeder.
// And if it is, make a note.
if (feederReg == (unsigned)cmpOp2)
isSecondOpNewified = true;
}
// Now that we are moving feeder close the jump,
// make sure we are respecting the kill values of
// the operands of the feeder.
bool updatedIsKill = false;
for (unsigned i = 0; i < MI.getNumOperands(); i++) {
MachineOperand &MO = MI.getOperand(i);
if (MO.isReg() && MO.isUse()) {
unsigned feederReg = MO.getReg();
for (MachineBasicBlock::iterator localII = feederPos,
end = jmpPos; localII != end; localII++) {
MachineInstr &localMI = *localII;
for (unsigned j = 0; j < localMI.getNumOperands(); j++) {
MachineOperand &localMO = localMI.getOperand(j);
if (localMO.isReg() && localMO.isUse() &&
localMO.isKill() && feederReg == localMO.getReg()) {
// We found that there is kill of a use register
// Set up a kill flag on the register
localMO.setIsKill(false);
MO.setIsKill();
updatedIsKill = true;
break;
}
}
if (updatedIsKill) break;
}
}
if (updatedIsKill) break;
}
MBB->splice(jmpPos, MI.getParent(), MI);
MBB->splice(jmpPos, MI.getParent(), cmpInstr);
DebugLoc dl = MI.getDebugLoc();
MachineInstr *NewMI;
assert((isNewValueJumpCandidate(*cmpInstr)) &&
"This compare is not a New Value Jump candidate.");
unsigned opc = getNewValueJumpOpcode(cmpInstr, cmpOp2,
isSecondOpNewified,
jmpTarget, MBPI);
if (invertPredicate)
opc = QII->getInvertedPredicatedOpcode(opc);
if (isSecondOpReg)
NewMI = BuildMI(*MBB, jmpPos, dl,
QII->get(opc))
.addReg(cmpReg1, getKillRegState(MO1IsKill))
.addReg(cmpOp2, getKillRegState(MO2IsKill))
.addMBB(jmpTarget);
else if ((cmpInstr->getOpcode() == Hexagon::C2_cmpeqi ||
cmpInstr->getOpcode() == Hexagon::C2_cmpgti) &&
cmpOp2 == -1 )
// Corresponding new-value compare jump instructions don't have the
// operand for -1 immediate value.
NewMI = BuildMI(*MBB, jmpPos, dl,
QII->get(opc))
.addReg(cmpReg1, getKillRegState(MO1IsKill))
.addMBB(jmpTarget);
else
NewMI = BuildMI(*MBB, jmpPos, dl,
QII->get(opc))
.addReg(cmpReg1, getKillRegState(MO1IsKill))
.addImm(cmpOp2)
.addMBB(jmpTarget);
assert(NewMI && "New Value Jump Instruction Not created!");
(void)NewMI;
if (cmpInstr->getOperand(0).isReg() &&
cmpInstr->getOperand(0).isKill())
cmpInstr->getOperand(0).setIsKill(false);
if (cmpInstr->getOperand(1).isReg() &&
cmpInstr->getOperand(1).isKill())
cmpInstr->getOperand(1).setIsKill(false);
cmpInstr->eraseFromParent();
jmpInstr->eraseFromParent();
++nvjGenerated;
++NumNVJGenerated;
break;
}
}
}
}
return true;
}
bool DeadMachineInstructionElim::runOnMachineFunction(MachineFunction &MF) {
bool AnyChanges = false;
MRI = &MF.getRegInfo();
TRI = MF.getTarget().getRegisterInfo();
TII = MF.getTarget().getInstrInfo();
// Treat reserved registers as always live.
BitVector ReservedRegs = TRI->getReservedRegs(MF);
// Loop over all instructions in all blocks, from bottom to top, so that it's
// more likely that chains of dependent but ultimately dead instructions will
// be cleaned up.
for (MachineFunction::reverse_iterator I = MF.rbegin(), E = MF.rend();
I != E; ++I) {
MachineBasicBlock *MBB = &*I;
// Start out assuming that reserved registers are live out of this block.
LivePhysRegs = ReservedRegs;
// Also add any explicit live-out physregs for this block.
if (!MBB->empty() && MBB->back().isReturn())
for (MachineRegisterInfo::liveout_iterator LOI = MRI->liveout_begin(),
LOE = MRI->liveout_end(); LOI != LOE; ++LOI) {
unsigned Reg = *LOI;
if (TargetRegisterInfo::isPhysicalRegister(Reg))
LivePhysRegs.set(Reg);
}
// Add live-ins from sucessors to LivePhysRegs. Normally, physregs are not
// live across blocks, but some targets (x86) can have flags live out of a
// block.
for (MachineBasicBlock::succ_iterator S = MBB->succ_begin(),
E = MBB->succ_end(); S != E; S++)
for (MachineBasicBlock::livein_iterator LI = (*S)->livein_begin();
LI != (*S)->livein_end(); LI++)
LivePhysRegs.set(*LI);
// Now scan the instructions and delete dead ones, tracking physreg
// liveness as we go.
for (MachineBasicBlock::reverse_iterator MII = MBB->rbegin(),
MIE = MBB->rend(); MII != MIE; ) {
MachineInstr *MI = &*MII;
// If the instruction is dead, delete it!
if (isDead(MI)) {
DEBUG(dbgs() << "DeadMachineInstructionElim: DELETING: " << *MI);
// It is possible that some DBG_VALUE instructions refer to this
// instruction. Examine each def operand for such references;
// if found, mark the DBG_VALUE as undef (but don't delete it).
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 (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
MachineRegisterInfo::use_iterator nextI;
for (MachineRegisterInfo::use_iterator I = MRI->use_begin(Reg),
E = MRI->use_end(); I!=E; I=nextI) {
nextI = llvm::next(I); // I is invalidated by the setReg
MachineOperand& Use = I.getOperand();
MachineInstr *UseMI = Use.getParent();
if (UseMI==MI)
continue;
assert(Use.isDebug());
UseMI->getOperand(0).setReg(0U);
}
}
AnyChanges = true;
MI->eraseFromParent();
++NumDeletes;
MIE = MBB->rend();
// MII is now pointing to the next instruction to process,
// so don't increment it.
continue;
}
// Record the physreg defs.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
if (MO.isReg() && MO.isDef()) {
unsigned Reg = MO.getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
LivePhysRegs.reset(Reg);
// Check the subreg set, not the alias set, because a def
// of a super-register may still be partially live after
// this def.
for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
*SubRegs; ++SubRegs)
LivePhysRegs.reset(*SubRegs);
}
} else if (MO.isRegMask()) {
// Register mask of preserved registers. All clobbers are dead.
if (const uint32_t *Mask = MO.getRegMask())
LivePhysRegs.clearBitsNotInMask(Mask);
else
LivePhysRegs.reset();
LivePhysRegs |= ReservedRegs;
}
}
// Record the physreg uses, after the defs, in case a physreg is
// both defined and used in the same instruction.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
if (MO.isReg() && MO.isUse()) {
unsigned Reg = MO.getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
LivePhysRegs.set(Reg);
for (const unsigned *AliasSet = TRI->getAliasSet(Reg);
*AliasSet; ++AliasSet)
LivePhysRegs.set(*AliasSet);
}
}
}
// We didn't delete the current instruction, so increment MII to
// the next one.
++MII;
}
}
LivePhysRegs.clear();
return AnyChanges;
}
void MIPrinter::print(const MachineBasicBlock &MBB) {
assert(MBB.getNumber() >= 0 && "Invalid MBB number");
OS << "bb." << MBB.getNumber();
bool HasAttributes = false;
if (const auto *BB = MBB.getBasicBlock()) {
if (BB->hasName()) {
OS << "." << BB->getName();
} else {
HasAttributes = true;
OS << " (";
int Slot = MST.getLocalSlot(BB);
if (Slot == -1)
OS << "<ir-block badref>";
else
OS << (Twine("%ir-block.") + Twine(Slot)).str();
}
}
if (MBB.hasAddressTaken()) {
OS << (HasAttributes ? ", " : " (");
OS << "address-taken";
HasAttributes = true;
}
if (MBB.isEHPad()) {
OS << (HasAttributes ? ", " : " (");
OS << "landing-pad";
HasAttributes = true;
}
if (MBB.getAlignment()) {
OS << (HasAttributes ? ", " : " (");
OS << "align " << MBB.getAlignment();
HasAttributes = true;
}
if (HasAttributes)
OS << ")";
OS << ":\n";
bool HasLineAttributes = false;
// Print the successors
bool canPredictProbs = canPredictBranchProbabilities(MBB);
// Even if the list of successors is empty, if we cannot guess it,
// we need to print it to tell the parser that the list is empty.
// This is needed, because MI model unreachable as empty blocks
// with an empty successor list. If the parser would see that
// without the successor list, it would guess the code would
// fallthrough.
if ((!MBB.succ_empty() && !SimplifyMIR) || !canPredictProbs ||
!canPredictSuccessors(MBB)) {
OS.indent(2) << "successors: ";
for (auto I = MBB.succ_begin(), E = MBB.succ_end(); I != E; ++I) {
if (I != MBB.succ_begin())
OS << ", ";
OS << printMBBReference(**I);
if (!SimplifyMIR || !canPredictProbs)
OS << '('
<< format("0x%08" PRIx32, MBB.getSuccProbability(I).getNumerator())
<< ')';
}
OS << "\n";
HasLineAttributes = true;
}
// Print the live in registers.
const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
if (MRI.tracksLiveness() && !MBB.livein_empty()) {
const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
OS.indent(2) << "liveins: ";
bool First = true;
for (const auto &LI : MBB.liveins()) {
if (!First)
OS << ", ";
First = false;
OS << printReg(LI.PhysReg, &TRI);
if (!LI.LaneMask.all())
OS << ":0x" << PrintLaneMask(LI.LaneMask);
}
OS << "\n";
HasLineAttributes = true;
}
if (HasLineAttributes)
OS << "\n";
bool IsInBundle = false;
for (auto I = MBB.instr_begin(), E = MBB.instr_end(); I != E; ++I) {
const MachineInstr &MI = *I;
if (IsInBundle && !MI.isInsideBundle()) {
OS.indent(2) << "}\n";
IsInBundle = false;
}
OS.indent(IsInBundle ? 4 : 2);
print(MI);
if (!IsInBundle && MI.getFlag(MachineInstr::BundledSucc)) {
OS << " {";
IsInBundle = true;
}
OS << "\n";
}
if (IsInBundle)
OS.indent(2) << "}\n";
}
void MIPrinter::print(const MachineBasicBlock &MBB) {
assert(MBB.getNumber() >= 0 && "Invalid MBB number");
OS << "bb." << MBB.getNumber();
bool HasAttributes = false;
if (const auto *BB = MBB.getBasicBlock()) {
if (BB->hasName()) {
OS << "." << BB->getName();
} else {
HasAttributes = true;
OS << " (";
int Slot = MST.getLocalSlot(BB);
if (Slot == -1)
OS << "<ir-block badref>";
else
OS << (Twine("%ir-block.") + Twine(Slot)).str();
}
}
if (MBB.hasAddressTaken()) {
OS << (HasAttributes ? ", " : " (");
OS << "address-taken";
HasAttributes = true;
}
if (MBB.isEHPad()) {
OS << (HasAttributes ? ", " : " (");
OS << "landing-pad";
HasAttributes = true;
}
if (MBB.getAlignment()) {
OS << (HasAttributes ? ", " : " (");
OS << "align " << MBB.getAlignment();
HasAttributes = true;
}
if (HasAttributes)
OS << ")";
OS << ":\n";
bool HasLineAttributes = false;
// Print the successors
if (!MBB.succ_empty()) {
OS.indent(2) << "successors: ";
for (auto I = MBB.succ_begin(), E = MBB.succ_end(); I != E; ++I) {
if (I != MBB.succ_begin())
OS << ", ";
printMBBReference(**I);
if (MBB.hasSuccessorWeights())
OS << '(' << MBB.getSuccWeight(I) << ')';
}
OS << "\n";
HasLineAttributes = true;
}
// Print the live in registers.
const auto *TRI = MBB.getParent()->getSubtarget().getRegisterInfo();
assert(TRI && "Expected target register info");
if (!MBB.livein_empty()) {
OS.indent(2) << "liveins: ";
bool First = true;
for (unsigned LI : MBB.liveins()) {
if (!First)
OS << ", ";
First = false;
printReg(LI, OS, TRI);
}
OS << "\n";
HasLineAttributes = true;
}
if (HasLineAttributes)
OS << "\n";
bool IsInBundle = false;
for (auto I = MBB.instr_begin(), E = MBB.instr_end(); I != E; ++I) {
const MachineInstr &MI = *I;
if (IsInBundle && !MI.isInsideBundle()) {
OS.indent(2) << "}\n";
IsInBundle = false;
}
OS.indent(IsInBundle ? 4 : 2);
print(MI);
if (!IsInBundle && MI.getFlag(MachineInstr::BundledSucc)) {
OS << " {";
IsInBundle = true;
}
OS << "\n";
}
if (IsInBundle)
OS.indent(2) << "}\n";
}
bool UnreachableMachineBlockElim::runOnMachineFunction(MachineFunction &F) {
SmallPtrSet<MachineBasicBlock*, 8> Reachable;
MMI = getAnalysisIfAvailable<MachineModuleInfo>();
MachineDominatorTree *MDT = getAnalysisIfAvailable<MachineDominatorTree>();
MachineLoopInfo *MLI = getAnalysisIfAvailable<MachineLoopInfo>();
// Mark all reachable blocks.
for (df_ext_iterator<MachineFunction*, SmallPtrSet<MachineBasicBlock*, 8> >
I = df_ext_begin(&F, Reachable), E = df_ext_end(&F, Reachable);
I != E; ++I)
/* Mark all reachable blocks */;
// Loop over all dead blocks, remembering them and deleting all instructions
// in them.
std::vector<MachineBasicBlock*> DeadBlocks;
for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
MachineBasicBlock *BB = I;
// Test for deadness.
if (!Reachable.count(BB)) {
DeadBlocks.push_back(BB);
// Update dominator and loop info.
if (MLI) MLI->removeBlock(BB);
if (MDT && MDT->getNode(BB)) MDT->eraseNode(BB);
while (BB->succ_begin() != BB->succ_end()) {
MachineBasicBlock* succ = *BB->succ_begin();
MachineBasicBlock::iterator start = succ->begin();
while (start != succ->end() && start->isPHI()) {
for (unsigned i = start->getNumOperands() - 1; i >= 2; i-=2)
if (start->getOperand(i).isMBB() &&
start->getOperand(i).getMBB() == BB) {
start->RemoveOperand(i);
start->RemoveOperand(i-1);
}
start++;
}
BB->removeSuccessor(BB->succ_begin());
}
}
}
// Actually remove the blocks now.
for (unsigned i = 0, e = DeadBlocks.size(); i != e; ++i)
DeadBlocks[i]->eraseFromParent();
// Cleanup PHI nodes.
for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
MachineBasicBlock *BB = I;
// Prune unneeded PHI entries.
SmallPtrSet<MachineBasicBlock*, 8> preds(BB->pred_begin(),
BB->pred_end());
MachineBasicBlock::iterator phi = BB->begin();
while (phi != BB->end() && phi->isPHI()) {
for (unsigned i = phi->getNumOperands() - 1; i >= 2; i-=2)
if (!preds.count(phi->getOperand(i).getMBB())) {
phi->RemoveOperand(i);
phi->RemoveOperand(i-1);
}
if (phi->getNumOperands() == 3) {
unsigned Input = phi->getOperand(1).getReg();
unsigned Output = phi->getOperand(0).getReg();
MachineInstr* temp = phi;
++phi;
temp->eraseFromParent();
if (Input != Output)
F.getRegInfo().replaceRegWith(Output, Input);
continue;
}
++phi;
}
}
F.RenumberBlocks();
return DeadBlocks.size();
}
