/// Merge an autorelease with a retain into a fused call.
bool
ObjCARCContract::ContractAutorelease(Function &F, Instruction *Autorelease,
InstructionClass Class,
SmallPtrSet<Instruction *, 4>
&DependingInstructions,
SmallPtrSet<const BasicBlock *, 4>
&Visited) {
const Value *Arg = GetObjCArg(Autorelease);
// Check that there are no instructions between the retain and the autorelease
// (such as an autorelease_pop) which may change the count.
CallInst *Retain = 0;
if (Class == IC_AutoreleaseRV)
FindDependencies(RetainAutoreleaseRVDep, Arg,
Autorelease->getParent(), Autorelease,
DependingInstructions, Visited, PA);
else
FindDependencies(RetainAutoreleaseDep, Arg,
Autorelease->getParent(), Autorelease,
DependingInstructions, Visited, PA);
Visited.clear();
if (DependingInstructions.size() != 1) {
DependingInstructions.clear();
return false;
}
Retain = dyn_cast_or_null<CallInst>(*DependingInstructions.begin());
DependingInstructions.clear();
if (!Retain ||
GetBasicInstructionClass(Retain) != IC_Retain ||
GetObjCArg(Retain) != Arg)
return false;
Changed = true;
++NumPeeps;
DEBUG(dbgs() << "ObjCARCContract::ContractAutorelease: Fusing "
"retain/autorelease. Erasing: " << *Autorelease << "\n"
" Old Retain: "
<< *Retain << "\n");
if (Class == IC_AutoreleaseRV)
Retain->setCalledFunction(getRetainAutoreleaseRVCallee(F.getParent()));
else
Retain->setCalledFunction(getRetainAutoreleaseCallee(F.getParent()));
DEBUG(dbgs() << " New Retain: "
<< *Retain << "\n");
EraseInstruction(Autorelease);
return true;
}
// Check PHI instructions at the beginning of MBB. It is assumed that
// calcRegsPassed has been run so BBInfo::isLiveOut is valid.
void MachineVerifier::checkPHIOps(const MachineBasicBlock *MBB) {
SmallPtrSet<const MachineBasicBlock*, 8> seen;
for (MachineBasicBlock::const_iterator BBI = MBB->begin(), BBE = MBB->end();
BBI != BBE && BBI->isPHI(); ++BBI) {
seen.clear();
for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) {
unsigned Reg = BBI->getOperand(i).getReg();
const MachineBasicBlock *Pre = BBI->getOperand(i + 1).getMBB();
if (!Pre->isSuccessor(MBB))
continue;
seen.insert(Pre);
BBInfo &PrInfo = MBBInfoMap[Pre];
if (PrInfo.reachable && !PrInfo.isLiveOut(Reg))
report("PHI operand is not live-out from predecessor",
&BBI->getOperand(i), i);
}
// Did we see all predecessors?
for (MachineBasicBlock::const_pred_iterator PrI = MBB->pred_begin(),
PrE = MBB->pred_end(); PrI != PrE; ++PrI) {
if (!seen.count(*PrI)) {
report("Missing PHI operand", BBI);
*OS << "BB#" << (*PrI)->getNumber()
<< " is a predecessor according to the CFG.\n";
}
}
}
}
/// FindSelectorAndURoR - Find the eh.selector call associated with the
/// eh.exception call. And indicate if there is a URoR "invoke" associated with
/// the eh.exception call. This recursively looks past instructions which don't
/// change the EH pointer value, like casts or PHI nodes.
bool
DwarfEHPrepare::FindSelectorAndURoR(Instruction *Inst, bool &URoRInvoke,
SmallPtrSet<IntrinsicInst*, 8> &SelCalls) {
SmallPtrSet<PHINode*, 32> SeenPHIs;
bool Changed = false;
restart:
for (Value::use_iterator
I = Inst->use_begin(), E = Inst->use_end(); I != E; ++I) {
Instruction *II = dyn_cast<Instruction>(*I);
if (!II || II->getParent()->getParent() != F) continue;
if (IntrinsicInst *Sel = dyn_cast<IntrinsicInst>(II)) {
if (Sel->getIntrinsicID() == Intrinsic::eh_selector)
SelCalls.insert(Sel);
} else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(II)) {
if (Invoke->getCalledFunction() == URoR)
URoRInvoke = true;
} else if (CastInst *CI = dyn_cast<CastInst>(II)) {
Changed |= FindSelectorAndURoR(CI, URoRInvoke, SelCalls);
} else if (StoreInst *SI = dyn_cast<StoreInst>(II)) {
if (!PromoteStoreInst(SI)) continue;
Changed = true;
SeenPHIs.clear();
goto restart; // Uses may have changed, restart loop.
} else if (PHINode *PN = dyn_cast<PHINode>(II)) {
if (SeenPHIs.insert(PN))
// Don't process a PHI node more than once.
Changed |= FindSelectorAndURoR(PN, URoRInvoke, SelCalls);
}
}
return Changed;
}
static void EliminateMultipleEntryLoops(MachineFunction &MF,
const MachineLoopInfo &MLI) {
SmallPtrSet<MachineBasicBlock *, 8> InSet;
for (scc_iterator<MachineFunction *> I = scc_begin(&MF), E = scc_end(&MF);
I != E; ++I) {
const std::vector<MachineBasicBlock *> &CurrentSCC = *I;
// Skip trivial SCCs.
if (CurrentSCC.size() == 1)
continue;
InSet.insert(CurrentSCC.begin(), CurrentSCC.end());
MachineBasicBlock *Header = nullptr;
for (MachineBasicBlock *MBB : CurrentSCC) {
for (MachineBasicBlock *Pred : MBB->predecessors()) {
if (InSet.count(Pred))
continue;
if (!Header) {
Header = MBB;
break;
}
// TODO: Implement multiple-entry loops.
report_fatal_error("multiple-entry loops are not supported yet");
}
}
assert(MLI.isLoopHeader(Header));
InSet.clear();
}
}
/// Return a set of basic blocks to insert sinked instructions.
///
/// The returned set of basic blocks (BBsToSinkInto) should satisfy:
///
/// * Inside the loop \p L
/// * For each UseBB in \p UseBBs, there is at least one BB in BBsToSinkInto
/// that domintates the UseBB
/// * Has minimum total frequency that is no greater than preheader frequency
///
/// The purpose of the function is to find the optimal sinking points to
/// minimize execution cost, which is defined as "sum of frequency of
/// BBsToSinkInto".
/// As a result, the returned BBsToSinkInto needs to have minimum total
/// frequency.
/// Additionally, if the total frequency of BBsToSinkInto exceeds preheader
/// frequency, the optimal solution is not sinking (return empty set).
///
/// \p ColdLoopBBs is used to help find the optimal sinking locations.
/// It stores a list of BBs that is:
///
/// * Inside the loop \p L
/// * Has a frequency no larger than the loop's preheader
/// * Sorted by BB frequency
///
/// The complexity of the function is O(UseBBs.size() * ColdLoopBBs.size()).
/// To avoid expensive computation, we cap the maximum UseBBs.size() in its
/// caller.
static SmallPtrSet<BasicBlock *, 2>
findBBsToSinkInto(const Loop &L, const SmallPtrSetImpl<BasicBlock *> &UseBBs,
const SmallVectorImpl<BasicBlock *> &ColdLoopBBs,
DominatorTree &DT, BlockFrequencyInfo &BFI) {
SmallPtrSet<BasicBlock *, 2> BBsToSinkInto;
if (UseBBs.size() == 0)
return BBsToSinkInto;
BBsToSinkInto.insert(UseBBs.begin(), UseBBs.end());
SmallPtrSet<BasicBlock *, 2> BBsDominatedByColdestBB;
// For every iteration:
// * Pick the ColdestBB from ColdLoopBBs
// * Find the set BBsDominatedByColdestBB that satisfy:
// - BBsDominatedByColdestBB is a subset of BBsToSinkInto
// - Every BB in BBsDominatedByColdestBB is dominated by ColdestBB
// * If Freq(ColdestBB) < Freq(BBsDominatedByColdestBB), remove
// BBsDominatedByColdestBB from BBsToSinkInto, add ColdestBB to
// BBsToSinkInto
for (BasicBlock *ColdestBB : ColdLoopBBs) {
BBsDominatedByColdestBB.clear();
for (BasicBlock *SinkedBB : BBsToSinkInto)
if (DT.dominates(ColdestBB, SinkedBB))
BBsDominatedByColdestBB.insert(SinkedBB);
if (BBsDominatedByColdestBB.size() == 0)
continue;
if (adjustedSumFreq(BBsDominatedByColdestBB, BFI) >
BFI.getBlockFreq(ColdestBB)) {
for (BasicBlock *DominatedBB : BBsDominatedByColdestBB) {
BBsToSinkInto.erase(DominatedBB);
}
BBsToSinkInto.insert(ColdestBB);
}
}
// If the total frequency of BBsToSinkInto is larger than preheader frequency,
// do not sink.
if (adjustedSumFreq(BBsToSinkInto, BFI) >
BFI.getBlockFreq(L.getLoopPreheader()))
BBsToSinkInto.clear();
return BBsToSinkInto;
}
TEST(SmallPtrSetTest, GrowthTest) {
int i;
int buf[8];
for(i=0; i<8; ++i) buf[i]=0;
SmallPtrSet<int *, 4> s;
typedef SmallPtrSet<int *, 4>::iterator iter;
s.insert(&buf[0]);
s.insert(&buf[1]);
s.insert(&buf[2]);
s.insert(&buf[3]);
EXPECT_EQ(4U, s.size());
i = 0;
for(iter I=s.begin(), E=s.end(); I!=E; ++I, ++i)
(**I)++;
EXPECT_EQ(4, i);
for(i=0; i<8; ++i)
EXPECT_EQ(i<4?1:0,buf[i]);
s.insert(&buf[4]);
s.insert(&buf[5]);
s.insert(&buf[6]);
s.insert(&buf[7]);
i = 0;
for(iter I=s.begin(), E=s.end(); I!=E; ++I, ++i)
(**I)++;
EXPECT_EQ(8, i);
s.erase(&buf[4]);
s.erase(&buf[5]);
s.erase(&buf[6]);
s.erase(&buf[7]);
EXPECT_EQ(4U, s.size());
i = 0;
for(iter I=s.begin(), E=s.end(); I!=E; ++I, ++i)
(**I)++;
EXPECT_EQ(4, i);
for(i=0; i<8; ++i)
EXPECT_EQ(i<4?3:1,buf[i]);
s.clear();
for(i=0; i<8; ++i) buf[i]=0;
for(i=0; i<128; ++i) s.insert(&buf[i%8]); // test repeated entires
EXPECT_EQ(8U, s.size());
for(iter I=s.begin(), E=s.end(); I!=E; ++I, ++i)
(**I)++;
for(i=0; i<8; ++i)
EXPECT_EQ(1,buf[i]);
}
bool PropagateJuliaAddrspaces::runOnFunction(Function &F) {
visit(F);
for (auto it : ToInsert)
it.first->insertBefore(it.second);
for (Instruction *I : ToDelete)
I->eraseFromParent();
ToInsert.clear();
ToDelete.clear();
LiftingMap.clear();
Visited.clear();
return true;
}
static void determineMissingVNIs(const SlotIndexes &Indexes, LiveInterval &LI) {
SmallPtrSet<const MachineBasicBlock*, 5> Visited;
LiveRange::iterator OutIt;
VNInfo *PrevValNo = nullptr;
for (LiveRange::iterator I = LI.begin(), E = LI.end(); I != E; ++I) {
LiveRange::Segment &S = *I;
// Determine final VNI if necessary.
if (S.valno == nullptr) {
// This can only happen at the begin of a basic block.
assert(S.start.isBlock() && "valno should only be missing at block begin");
Visited.clear();
const MachineBasicBlock *MBB = Indexes.getMBBFromIndex(S.start);
for (const MachineBasicBlock *Pred : MBB->predecessors()) {
VNInfo *VNI = searchForVNI(Indexes, LI, Pred, Visited);
if (VNI != nullptr) {
S.valno = VNI;
break;
}
}
assert(S.valno != nullptr && "could not determine valno");
}
// Merge with previous segment if it has the same VNI.
if (PrevValNo == S.valno && OutIt->end == S.start) {
OutIt->end = S.end;
} else {
// Didn't merge. Move OutIt to next segment.
if (PrevValNo == nullptr)
OutIt = LI.begin();
else
++OutIt;
if (OutIt != I)
*OutIt = *I;
PrevValNo = S.valno;
}
}
// If we merged some segments chop off the end.
++OutIt;
LI.segments.erase(OutIt, LI.end());
}
static void determineMissingVNIs(const SlotIndexes &Indexes, LiveInterval &LI) {
SmallPtrSet<const MachineBasicBlock*, 5> Visited;
for (LiveRange::Segment &S : LI.segments) {
if (S.valno != nullptr)
continue;
// This can only happen at the begin of a basic block.
assert(S.start.isBlock() && "valno should only be missing at block begin");
Visited.clear();
const MachineBasicBlock *MBB = Indexes.getMBBFromIndex(S.start);
for (const MachineBasicBlock *Pred : MBB->predecessors()) {
VNInfo *VNI = searchForVNI(Indexes, LI, Pred, Visited);
if (VNI != nullptr) {
S.valno = VNI;
break;
}
}
assert(S.valno != nullptr && "could not determine valno");
}
}
// Given a list of loads that could be constant-folded (LoadBaseAddresses),
// estimate number of optimized instructions after substituting the concrete
// values for the given Iteration.
// Fill in SimplifiedInsns map for future use in DCE-estimation.
unsigned EstimateNumberOfSimplifiedInsns(unsigned Iteration) {
SmallVector<Instruction *, 8> Worklist;
SimplifiedValues.clear();
CountedInsns.clear();
NumberOfOptimizedInstructions = 0;
// We start by adding all loads to the worklist.
for (auto LoadDescr : LoadBaseAddresses) {
LoadInst *LI = LoadDescr.first;
SimplifiedValues[LI] = computeLoadValue(LI, Iteration);
if (CountedInsns.insert(LI).second)
NumberOfOptimizedInstructions += TTI.getUserCost(LI);
for (auto U : LI->users()) {
Instruction *UI = dyn_cast<Instruction>(U);
if (!UI)
continue;
if (!L->contains(UI))
continue;
Worklist.push_back(UI);
}
}
// And then we try to simplify every user of every instruction from the
// worklist. If we do simplify a user, add it to the worklist to process
// its users as well.
while (!Worklist.empty()) {
Instruction *I = Worklist.pop_back_val();
if (!visit(I))
continue;
for (auto U : I->users()) {
Instruction *UI = dyn_cast<Instruction>(U);
if (!UI)
continue;
if (!L->contains(UI))
continue;
Worklist.push_back(UI);
}
}
return NumberOfOptimizedInstructions;
}
bool OptimizeExts::runOnMachineFunction(MachineFunction &MF) {
TM = &MF.getTarget();
TII = TM->getInstrInfo();
MRI = &MF.getRegInfo();
DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : 0;
bool Changed = false;
SmallPtrSet<MachineInstr*, 8> LocalMIs;
for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
MachineBasicBlock *MBB = &*I;
LocalMIs.clear();
for (MachineBasicBlock::iterator MII = I->begin(), ME = I->end(); MII != ME;
++MII) {
MachineInstr *MI = &*MII;
Changed |= OptimizeInstr(MI, MBB, LocalMIs);
}
}
return Changed;
}
/// getMachineBasicBlocks - Populate given set using machine basic blocks which
/// have machine instructions that belong to lexical scope identified by
/// DebugLoc.
void LexicalScopes::
getMachineBasicBlocks(DebugLoc DL,
SmallPtrSet<const MachineBasicBlock*, 4> &MBBs) {
MBBs.clear();
LexicalScope *Scope = getOrCreateLexicalScope(DL);
if (!Scope)
return;
if (Scope == CurrentFnLexicalScope) {
for (MachineFunction::const_iterator I = MF->begin(), E = MF->end();
I != E; ++I)
MBBs.insert(I);
return;
}
SmallVector<InsnRange, 4> &InsnRanges = Scope->getRanges();
for (SmallVector<InsnRange, 4>::iterator I = InsnRanges.begin(),
E = InsnRanges.end(); I != E; ++I) {
InsnRange &R = *I;
MBBs.insert(R.first->getParent());
}
}
vector<const TargetRegisterInfo*> ipaFindUsedReturns(ParameterRegistry& registry, Function& function, const vector<const TargetRegisterInfo*>& returns)
{
// Excuse entry points from not having callers; use every return.
if (function.use_empty())
if (auto address = md::getVirtualAddress(function))
if (isEntryPoint(address->getLimitedValue()))
{
return returns;
}
// Otherwise, loop through callers and see which registers are used after the function call.
TargetInfo& targetInfo = registry.getTargetInfo();
SmallPtrSet<MemoryPhi*, 4> visited;
vector<const TargetRegisterInfo*> result;
for (auto& use : function.uses())
{
if (auto call = dyn_cast<CallInst>(use.getUser()))
{
auto parentFunction = call->getParent()->getParent();
if (parentFunction == &function)
{
// TODO: This isn't impossible to compute, just somewhat inconvenient.
continue;
}
auto parentArgs = static_cast<Argument*>(parentFunction->arg_begin());
auto pointerType = dyn_cast<PointerType>(parentArgs->getType());
assert(pointerType != nullptr && pointerType->getTypeAtIndex(int(0))->getStructName() == "struct.x86_regs");
(void) pointerType;
visited.clear();
MemorySSA& mssa = *registry.getMemorySSA(*parentFunction);
findUsedReturns(returns, targetInfo, mssa, visited, *mssa.getMemoryAccess(call), result);
}
}
return result;
}
bool LoopInstSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
if (skipOptnoneFunction(L))
return false;
DominatorTreeWrapperPass *DTWP =
getAnalysisIfAvailable<DominatorTreeWrapperPass>();
DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr;
LoopInfo *LI = &getAnalysis<LoopInfo>();
DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
AssumptionTracker *AT = &getAnalysis<AssumptionTracker>();
SmallVector<BasicBlock*, 8> ExitBlocks;
L->getUniqueExitBlocks(ExitBlocks);
array_pod_sort(ExitBlocks.begin(), ExitBlocks.end());
SmallPtrSet<const Instruction*, 8> S1, S2, *ToSimplify = &S1, *Next = &S2;
// The bit we are stealing from the pointer represents whether this basic
// block is the header of a subloop, in which case we only process its phis.
typedef PointerIntPair<BasicBlock*, 1> WorklistItem;
SmallVector<WorklistItem, 16> VisitStack;
SmallPtrSet<BasicBlock*, 32> Visited;
bool Changed = false;
bool LocalChanged;
do {
LocalChanged = false;
VisitStack.clear();
Visited.clear();
VisitStack.push_back(WorklistItem(L->getHeader(), false));
while (!VisitStack.empty()) {
WorklistItem Item = VisitStack.pop_back_val();
BasicBlock *BB = Item.getPointer();
bool IsSubloopHeader = Item.getInt();
// Simplify instructions in the current basic block.
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
Instruction *I = BI++;
// The first time through the loop ToSimplify is empty and we try to
// simplify all instructions. On later iterations ToSimplify is not
// empty and we only bother simplifying instructions that are in it.
if (!ToSimplify->empty() && !ToSimplify->count(I))
continue;
// Don't bother simplifying unused instructions.
if (!I->use_empty()) {
Value *V = SimplifyInstruction(I, DL, TLI, DT, AT);
if (V && LI->replacementPreservesLCSSAForm(I, V)) {
// Mark all uses for resimplification next time round the loop.
for (User *U : I->users())
Next->insert(cast<Instruction>(U));
I->replaceAllUsesWith(V);
LocalChanged = true;
++NumSimplified;
}
}
bool res = RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
if (res) {
// RecursivelyDeleteTriviallyDeadInstruction can remove
// more than one instruction, so simply incrementing the
// iterator does not work. When instructions get deleted
// re-iterate instead.
BI = BB->begin(); BE = BB->end();
LocalChanged |= res;
}
if (IsSubloopHeader && !isa<PHINode>(I))
break;
}
// Add all successors to the worklist, except for loop exit blocks and the
// bodies of subloops. We visit the headers of loops so that we can process
// their phis, but we contract the rest of the subloop body and only follow
// edges leading back to the original loop.
for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
++SI) {
BasicBlock *SuccBB = *SI;
if (!Visited.insert(SuccBB).second)
continue;
const Loop *SuccLoop = LI->getLoopFor(SuccBB);
if (SuccLoop && SuccLoop->getHeader() == SuccBB
&& L->contains(SuccLoop)) {
VisitStack.push_back(WorklistItem(SuccBB, true));
SmallVector<BasicBlock*, 8> SubLoopExitBlocks;
SuccLoop->getExitBlocks(SubLoopExitBlocks);
for (unsigned i = 0; i < SubLoopExitBlocks.size(); ++i) {
BasicBlock *ExitBB = SubLoopExitBlocks[i];
if (LI->getLoopFor(ExitBB) == L && Visited.insert(ExitBB).second)
VisitStack.push_back(WorklistItem(ExitBB, false));
}
continue;
}
bool IsExitBlock = std::binary_search(ExitBlocks.begin(),
ExitBlocks.end(), SuccBB);
if (IsExitBlock)
continue;
VisitStack.push_back(WorklistItem(SuccBB, false));
}
}
// Place the list of instructions to simplify on the next loop iteration
// into ToSimplify.
std::swap(ToSimplify, Next);
Next->clear();
Changed |= LocalChanged;
} while (LocalChanged);
return Changed;
}
/// processImplicitDefs - Process IMPLICIT_DEF instructions and make sure
/// there is one implicit_def for each use. Add isUndef marker to
/// implicit_def defs and their uses.
bool ProcessImplicitDefs::runOnMachineFunction(MachineFunction &fn) {
DEBUG(dbgs() << "********** PROCESS IMPLICIT DEFS **********\n"
<< "********** Function: "
<< ((Value*)fn.getFunction())->getName() << '\n');
bool Changed = false;
TII = fn.getTarget().getInstrInfo();
TRI = fn.getTarget().getRegisterInfo();
MRI = &fn.getRegInfo();
LV = &getAnalysis<LiveVariables>();
SmallSet<unsigned, 8> ImpDefRegs;
SmallVector<MachineInstr*, 8> ImpDefMIs;
SmallVector<MachineInstr*, 4> RUses;
SmallPtrSet<MachineBasicBlock*,16> Visited;
SmallPtrSet<MachineInstr*, 8> ModInsts;
MachineBasicBlock *Entry = fn.begin();
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;
for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
I != E; ) {
MachineInstr *MI = &*I;
++I;
if (MI->isImplicitDef()) {
ImpDefMIs.push_back(MI);
// Is this a sub-register read-modify-write?
if (MI->getOperand(0).readsReg())
continue;
unsigned Reg = MI->getOperand(0).getReg();
ImpDefRegs.insert(Reg);
if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
for (const unsigned *SS = TRI->getSubRegisters(Reg); *SS; ++SS)
ImpDefRegs.insert(*SS);
}
continue;
}
// Eliminate %reg1032:sub<def> = COPY undef.
if (MI->isCopy() && MI->getOperand(0).readsReg()) {
MachineOperand &MO = MI->getOperand(1);
if (MO.isUndef() || ImpDefRegs.count(MO.getReg())) {
if (MO.isKill()) {
LiveVariables::VarInfo& vi = LV->getVarInfo(MO.getReg());
vi.removeKill(MI);
}
unsigned Reg = MI->getOperand(0).getReg();
MI->eraseFromParent();
Changed = true;
// A REG_SEQUENCE may have been expanded into partial definitions.
// If this was the last one, mark Reg as implicitly defined.
if (TargetRegisterInfo::isVirtualRegister(Reg) && MRI->def_empty(Reg))
ImpDefRegs.insert(Reg);
continue;
}
}
bool ChangedToImpDef = false;
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand& MO = MI->getOperand(i);
if (!MO.isReg() || !MO.readsReg())
continue;
unsigned Reg = MO.getReg();
if (!Reg)
continue;
if (!ImpDefRegs.count(Reg))
continue;
// Use is a copy, just turn it into an implicit_def.
if (CanTurnIntoImplicitDef(MI, Reg, i, ImpDefRegs)) {
bool isKill = MO.isKill();
MI->setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
for (int j = MI->getNumOperands() - 1, ee = 0; j > ee; --j)
MI->RemoveOperand(j);
if (isKill) {
ImpDefRegs.erase(Reg);
LiveVariables::VarInfo& vi = LV->getVarInfo(Reg);
vi.removeKill(MI);
}
ChangedToImpDef = true;
Changed = true;
break;
}
Changed = true;
MO.setIsUndef();
// This is a partial register redef of an implicit def.
// Make sure the whole register is defined by the instruction.
if (MO.isDef()) {
MI->addRegisterDefined(Reg);
continue;
}
if (MO.isKill() || MI->isRegTiedToDefOperand(i)) {
// Make sure other reads of Reg are also marked <undef>.
for (unsigned j = i+1; j != e; ++j) {
MachineOperand &MOJ = MI->getOperand(j);
if (MOJ.isReg() && MOJ.getReg() == Reg && MOJ.readsReg())
MOJ.setIsUndef();
}
ImpDefRegs.erase(Reg);
}
}
if (ChangedToImpDef) {
// Backtrack to process this new implicit_def.
--I;
} else {
for (unsigned i = 0; i != MI->getNumOperands(); ++i) {
MachineOperand& MO = MI->getOperand(i);
if (!MO.isReg() || !MO.isDef())
continue;
ImpDefRegs.erase(MO.getReg());
}
}
}
// Any outstanding liveout implicit_def's?
for (unsigned i = 0, e = ImpDefMIs.size(); i != e; ++i) {
MachineInstr *MI = ImpDefMIs[i];
unsigned Reg = MI->getOperand(0).getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg) ||
!ImpDefRegs.count(Reg)) {
// Delete all "local" implicit_def's. That include those which define
// physical registers since they cannot be liveout.
MI->eraseFromParent();
Changed = true;
continue;
}
// If there are multiple defs of the same register and at least one
// is not an implicit_def, do not insert implicit_def's before the
// uses.
bool Skip = false;
SmallVector<MachineInstr*, 4> DeadImpDefs;
for (MachineRegisterInfo::def_iterator DI = MRI->def_begin(Reg),
DE = MRI->def_end(); DI != DE; ++DI) {
MachineInstr *DeadImpDef = &*DI;
if (!DeadImpDef->isImplicitDef()) {
Skip = true;
break;
}
DeadImpDefs.push_back(DeadImpDef);
}
if (Skip)
continue;
// The only implicit_def which we want to keep are those that are live
// out of its block.
for (unsigned j = 0, ee = DeadImpDefs.size(); j != ee; ++j)
DeadImpDefs[j]->eraseFromParent();
Changed = true;
// Process each use instruction once.
for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
UE = MRI->use_end(); UI != UE; ++UI) {
if (UI.getOperand().isUndef())
continue;
MachineInstr *RMI = &*UI;
if (ModInsts.insert(RMI))
RUses.push_back(RMI);
}
for (unsigned i = 0, e = RUses.size(); i != e; ++i) {
MachineInstr *RMI = RUses[i];
// Turn a copy use into an implicit_def.
if (isUndefCopy(RMI, Reg, ImpDefRegs)) {
RMI->setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
bool isKill = false;
SmallVector<unsigned, 4> Ops;
for (unsigned j = 0, ee = RMI->getNumOperands(); j != ee; ++j) {
MachineOperand &RRMO = RMI->getOperand(j);
if (RRMO.isReg() && RRMO.getReg() == Reg) {
Ops.push_back(j);
if (RRMO.isKill())
isKill = true;
}
}
// Leave the other operands along.
for (unsigned j = 0, ee = Ops.size(); j != ee; ++j) {
unsigned OpIdx = Ops[j];
RMI->RemoveOperand(OpIdx-j);
}
// Update LiveVariables varinfo if the instruction is a kill.
if (isKill) {
LiveVariables::VarInfo& vi = LV->getVarInfo(Reg);
vi.removeKill(RMI);
}
continue;
}
// Replace Reg with a new vreg that's marked implicit.
const TargetRegisterClass* RC = MRI->getRegClass(Reg);
unsigned NewVReg = MRI->createVirtualRegister(RC);
bool isKill = true;
for (unsigned j = 0, ee = RMI->getNumOperands(); j != ee; ++j) {
MachineOperand &RRMO = RMI->getOperand(j);
if (RRMO.isReg() && RRMO.getReg() == Reg) {
RRMO.setReg(NewVReg);
RRMO.setIsUndef();
if (isKill) {
// Only the first operand of NewVReg is marked kill.
RRMO.setIsKill();
isKill = false;
}
}
}
}
RUses.clear();
ModInsts.clear();
}
ImpDefRegs.clear();
ImpDefMIs.clear();
}
return Changed;
}
int Compilation::performJobsImpl() {
// Create a TaskQueue for execution.
std::unique_ptr<TaskQueue> TQ;
if (SkipTaskExecution)
TQ.reset(new DummyTaskQueue(NumberOfParallelCommands));
else
TQ.reset(new TaskQueue(NumberOfParallelCommands));
PerformJobsState State;
using DependencyGraph = DependencyGraph<const Job *>;
DependencyGraph DepGraph;
SmallPtrSet<const Job *, 16> DeferredCommands;
SmallVector<const Job *, 16> InitialOutOfDateCommands;
DependencyGraph::MarkTracer ActualIncrementalTracer;
DependencyGraph::MarkTracer *IncrementalTracer = nullptr;
if (ShowIncrementalBuildDecisions)
IncrementalTracer = &ActualIncrementalTracer;
auto noteBuilding = [&] (const Job *cmd, StringRef reason) {
if (!ShowIncrementalBuildDecisions)
return;
if (State.ScheduledCommands.count(cmd))
return;
llvm::outs() << "Queuing "
<< llvm::sys::path::filename(cmd->getOutput().getBaseInput(0))
<< " " << reason << "\n";
IncrementalTracer->printPath(llvm::outs(), cmd,
[](raw_ostream &out, const Job *base) {
out << llvm::sys::path::filename(base->getOutput().getBaseInput(0));
});
};
// Set up scheduleCommandIfNecessaryAndPossible.
// This will only schedule the given command if it has not been scheduled
// and if all of its inputs are in FinishedCommands.
auto scheduleCommandIfNecessaryAndPossible = [&] (const Job *Cmd) {
if (State.ScheduledCommands.count(Cmd))
return;
if (auto Blocking = findUnfinishedJob(Cmd->getInputs(),
State.FinishedCommands)) {
State.BlockingCommands[Blocking].push_back(Cmd);
return;
}
// FIXME: Failing here should not take down the whole process.
bool success = writeFilelistIfNecessary(Cmd, Diags);
assert(success && "failed to write filelist");
(void)success;
assert(Cmd->getExtraEnvironment().empty() &&
"not implemented for compilations with multiple jobs");
State.ScheduledCommands.insert(Cmd);
TQ->addTask(Cmd->getExecutable(), Cmd->getArguments(), llvm::None,
(void *)Cmd);
};
// When a task finishes, we need to reevaluate the other commands that
// might have been blocked.
auto markFinished = [&] (const Job *Cmd) {
State.FinishedCommands.insert(Cmd);
auto BlockedIter = State.BlockingCommands.find(Cmd);
if (BlockedIter != State.BlockingCommands.end()) {
auto AllBlocked = std::move(BlockedIter->second);
State.BlockingCommands.erase(BlockedIter);
for (auto *Blocked : AllBlocked)
scheduleCommandIfNecessaryAndPossible(Blocked);
}
};
// Schedule all jobs we can.
for (const Job *Cmd : getJobs()) {
if (!getIncrementalBuildEnabled()) {
scheduleCommandIfNecessaryAndPossible(Cmd);
continue;
}
// Try to load the dependencies file for this job. If there isn't one, we
// always have to run the job, but it doesn't affect any other jobs. If
// there should be one but it's not present or can't be loaded, we have to
// run all the jobs.
// FIXME: We can probably do better here!
Job::Condition Condition = Job::Condition::Always;
StringRef DependenciesFile =
Cmd->getOutput().getAdditionalOutputForType(types::TY_SwiftDeps);
if (!DependenciesFile.empty()) {
if (Cmd->getCondition() == Job::Condition::NewlyAdded) {
DepGraph.addIndependentNode(Cmd);
} else {
switch (DepGraph.loadFromPath(Cmd, DependenciesFile)) {
case DependencyGraphImpl::LoadResult::HadError:
disableIncrementalBuild();
for (const Job *Cmd : DeferredCommands)
scheduleCommandIfNecessaryAndPossible(Cmd);
DeferredCommands.clear();
break;
case DependencyGraphImpl::LoadResult::UpToDate:
Condition = Cmd->getCondition();
break;
case DependencyGraphImpl::LoadResult::AffectsDownstream:
llvm_unreachable("we haven't marked anything in this graph yet");
}
}
}
switch (Condition) {
case Job::Condition::Always:
if (getIncrementalBuildEnabled() && !DependenciesFile.empty()) {
InitialOutOfDateCommands.push_back(Cmd);
DepGraph.markIntransitive(Cmd);
}
SWIFT_FALLTHROUGH;
case Job::Condition::RunWithoutCascading:
noteBuilding(Cmd, "(initial)");
scheduleCommandIfNecessaryAndPossible(Cmd);
break;
case Job::Condition::CheckDependencies:
DeferredCommands.insert(Cmd);
break;
case Job::Condition::NewlyAdded:
llvm_unreachable("handled above");
}
}
if (getIncrementalBuildEnabled()) {
SmallVector<const Job *, 16> AdditionalOutOfDateCommands;
// We scheduled all of the files that have actually changed. Now add the
// files that haven't changed, so that they'll get built in parallel if
// possible and after the first set of files if it's not.
for (auto *Cmd : InitialOutOfDateCommands) {
DepGraph.markTransitive(AdditionalOutOfDateCommands, Cmd,
IncrementalTracer);
}
for (auto *transitiveCmd : AdditionalOutOfDateCommands)
noteBuilding(transitiveCmd, "because of the initial set:");
size_t firstSize = AdditionalOutOfDateCommands.size();
// Check all cross-module dependencies as well.
for (StringRef dependency : DepGraph.getExternalDependencies()) {
llvm::sys::fs::file_status depStatus;
if (!llvm::sys::fs::status(dependency, depStatus))
if (depStatus.getLastModificationTime() < LastBuildTime)
continue;
// If the dependency has been modified since the oldest built file,
// or if we can't stat it for some reason (perhaps it's been deleted?),
// trigger rebuilds through the dependency graph.
DepGraph.markExternal(AdditionalOutOfDateCommands, dependency);
}
for (auto *externalCmd :
llvm::makeArrayRef(AdditionalOutOfDateCommands).slice(firstSize)) {
noteBuilding(externalCmd, "because of external dependencies");
}
for (auto *AdditionalCmd : AdditionalOutOfDateCommands) {
if (!DeferredCommands.count(AdditionalCmd))
continue;
scheduleCommandIfNecessaryAndPossible(AdditionalCmd);
DeferredCommands.erase(AdditionalCmd);
}
}
int Result = EXIT_SUCCESS;
// Set up a callback which will be called immediately after a task has
// started. This callback may be used to provide output indicating that the
// task began.
auto taskBegan = [this] (ProcessId Pid, void *Context) {
// TODO: properly handle task began.
const Job *BeganCmd = (const Job *)Context;
// For verbose output, print out each command as it begins execution.
if (Level == OutputLevel::Verbose)
BeganCmd->printCommandLine(llvm::errs());
else if (Level == OutputLevel::Parseable)
parseable_output::emitBeganMessage(llvm::errs(), *BeganCmd, Pid);
};
// Set up a callback which will be called immediately after a task has
// finished execution. This callback should determine if execution should
// continue (if execution should stop, this callback should return true), and
// it should also schedule any additional commands which we now know need
// to run.
auto taskFinished = [&] (ProcessId Pid, int ReturnCode, StringRef Output,
void *Context) -> TaskFinishedResponse {
const Job *FinishedCmd = (const Job *)Context;
if (Level == OutputLevel::Parseable) {
// Parseable output was requested.
parseable_output::emitFinishedMessage(llvm::errs(), *FinishedCmd, Pid,
ReturnCode, Output);
} else {
// Otherwise, send the buffered output to stderr, though only if we
// support getting buffered output.
if (TaskQueue::supportsBufferingOutput())
llvm::errs() << Output;
}
if (ReturnCode != EXIT_SUCCESS) {
// The task failed, so return true without performing any further
// dependency analysis.
// Store this task's ReturnCode as our Result if we haven't stored
// anything yet.
if (Result == EXIT_SUCCESS)
Result = ReturnCode;
if (!isa<CompileJobAction>(FinishedCmd->getSource()) ||
ReturnCode != EXIT_FAILURE) {
Diags.diagnose(SourceLoc(), diag::error_command_failed,
FinishedCmd->getSource().getClassName(),
ReturnCode);
}
return ContinueBuildingAfterErrors ?
TaskFinishedResponse::ContinueExecution :
TaskFinishedResponse::StopExecution;
}
// When a task finishes, we need to reevaluate the other commands that
// might have been blocked.
markFinished(FinishedCmd);
// In order to handle both old dependencies that have disappeared and new
// dependencies that have arisen, we need to reload the dependency file.
if (getIncrementalBuildEnabled()) {
const CommandOutput &Output = FinishedCmd->getOutput();
StringRef DependenciesFile =
Output.getAdditionalOutputForType(types::TY_SwiftDeps);
if (!DependenciesFile.empty()) {
SmallVector<const Job *, 16> Dependents;
bool wasCascading = DepGraph.isMarked(FinishedCmd);
switch (DepGraph.loadFromPath(FinishedCmd, DependenciesFile)) {
case DependencyGraphImpl::LoadResult::HadError:
disableIncrementalBuild();
for (const Job *Cmd : DeferredCommands)
scheduleCommandIfNecessaryAndPossible(Cmd);
DeferredCommands.clear();
Dependents.clear();
break;
case DependencyGraphImpl::LoadResult::UpToDate:
if (!wasCascading)
break;
SWIFT_FALLTHROUGH;
case DependencyGraphImpl::LoadResult::AffectsDownstream:
DepGraph.markTransitive(Dependents, FinishedCmd);
break;
}
for (const Job *Cmd : Dependents) {
DeferredCommands.erase(Cmd);
noteBuilding(Cmd, "because of dependencies discovered later");
scheduleCommandIfNecessaryAndPossible(Cmd);
}
}
}
return TaskFinishedResponse::ContinueExecution;
};
auto taskSignalled = [&] (ProcessId Pid, StringRef ErrorMsg, StringRef Output,
void *Context) -> TaskFinishedResponse {
const Job *SignalledCmd = (const Job *)Context;
if (Level == OutputLevel::Parseable) {
// Parseable output was requested.
parseable_output::emitSignalledMessage(llvm::errs(), *SignalledCmd, Pid,
ErrorMsg, Output);
} else {
// Otherwise, send the buffered output to stderr, though only if we
// support getting buffered output.
if (TaskQueue::supportsBufferingOutput())
llvm::errs() << Output;
}
if (!ErrorMsg.empty())
Diags.diagnose(SourceLoc(), diag::error_unable_to_execute_command,
ErrorMsg);
Diags.diagnose(SourceLoc(), diag::error_command_signalled,
SignalledCmd->getSource().getClassName());
// Since the task signalled, unconditionally set result to -2.
Result = -2;
return TaskFinishedResponse::StopExecution;
};
do {
// Ask the TaskQueue to execute.
TQ->execute(taskBegan, taskFinished, taskSignalled);
// Mark all remaining deferred commands as skipped.
for (const Job *Cmd : DeferredCommands) {
if (Level == OutputLevel::Parseable) {
// Provide output indicating this command was skipped if parseable output
// was requested.
parseable_output::emitSkippedMessage(llvm::errs(), *Cmd);
}
State.ScheduledCommands.insert(Cmd);
markFinished(Cmd);
}
// ...which may allow us to go on and do later tasks.
} while (Result == 0 && TQ->hasRemainingTasks());
if (Result == 0) {
assert(State.BlockingCommands.empty() &&
"some blocking commands never finished properly");
} else {
// Make sure we record any files that still need to be rebuilt.
for (const Job *Cmd : getJobs()) {
// Skip files that don't use dependency analysis.
StringRef DependenciesFile =
Cmd->getOutput().getAdditionalOutputForType(types::TY_SwiftDeps);
if (DependenciesFile.empty())
continue;
// Don't worry about commands that finished or weren't going to run.
if (State.FinishedCommands.count(Cmd))
continue;
if (!State.ScheduledCommands.count(Cmd))
continue;
bool isCascading = true;
if (getIncrementalBuildEnabled())
isCascading = DepGraph.isMarked(Cmd);
State.UnfinishedCommands.insert({Cmd, isCascading});
}
}
if (!CompilationRecordPath.empty() && !SkipTaskExecution) {
InputInfoMap InputInfo;
populateInputInfoMap(InputInfo, State);
checkForOutOfDateInputs(Diags, InputInfo);
writeCompilationRecord(CompilationRecordPath, ArgsHash, BuildStartTime,
InputInfo);
}
if (Result == 0)
Result = Diags.hadAnyError();
return Result;
}
bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
if (DisablePeephole)
return false;
TM = &MF.getTarget();
TII = TM->getInstrInfo();
MRI = &MF.getRegInfo();
DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : 0;
bool Changed = false;
SmallPtrSet<MachineInstr*, 8> LocalMIs;
SmallSet<unsigned, 4> ImmDefRegs;
DenseMap<unsigned, MachineInstr*> ImmDefMIs;
unsigned FoldAsLoadDefReg;
for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
MachineBasicBlock *MBB = &*I;
bool SeenMoveImm = false;
LocalMIs.clear();
ImmDefRegs.clear();
ImmDefMIs.clear();
FoldAsLoadDefReg = 0;
for (MachineBasicBlock::iterator
MII = I->begin(), MIE = I->end(); MII != MIE; ) {
MachineInstr *MI = &*MII;
// We may be erasing MI below, increment MII now.
++MII;
LocalMIs.insert(MI);
// If there exists an instruction which belongs to the following
// categories, we will discard the load candidate.
if (MI->isLabel() || MI->isPHI() || MI->isImplicitDef() ||
MI->isKill() || MI->isInlineAsm() || MI->isDebugValue() ||
MI->hasUnmodeledSideEffects()) {
FoldAsLoadDefReg = 0;
continue;
}
if (MI->mayStore() || MI->isCall())
FoldAsLoadDefReg = 0;
if ((MI->isBitcast() && optimizeBitcastInstr(MI, MBB)) ||
(MI->isCompare() && optimizeCmpInstr(MI, MBB)) ||
(MI->isSelect() && optimizeSelect(MI))) {
// MI is deleted.
LocalMIs.erase(MI);
Changed = true;
continue;
}
if (isMoveImmediate(MI, ImmDefRegs, ImmDefMIs)) {
SeenMoveImm = true;
} else {
Changed |= optimizeExtInstr(MI, MBB, LocalMIs);
// optimizeExtInstr might have created new instructions after MI
// and before the already incremented MII. Adjust MII so that the
// next iteration sees the new instructions.
MII = MI;
++MII;
if (SeenMoveImm)
Changed |= foldImmediate(MI, MBB, ImmDefRegs, ImmDefMIs);
}
// Check whether MI is a load candidate for folding into a later
// instruction. If MI is not a candidate, check whether we can fold an
// earlier load into MI.
if (!isLoadFoldable(MI, FoldAsLoadDefReg) && FoldAsLoadDefReg) {
// We need to fold load after optimizeCmpInstr, since optimizeCmpInstr
// can enable folding by converting SUB to CMP.
MachineInstr *DefMI = 0;
MachineInstr *FoldMI = TII->optimizeLoadInstr(MI, MRI,
FoldAsLoadDefReg, DefMI);
if (FoldMI) {
// Update LocalMIs since we replaced MI with FoldMI and deleted DefMI.
LocalMIs.erase(MI);
LocalMIs.erase(DefMI);
LocalMIs.insert(FoldMI);
MI->eraseFromParent();
DefMI->eraseFromParent();
++NumLoadFold;
// MI is replaced with FoldMI.
Changed = true;
continue;
}
}
}
}
return Changed;
}
/// ClusterNeighboringLoads - Force nearby loads together by "flagging" them.
/// This function finds loads of the same base and different offsets. If the
/// offsets are not far apart (target specific), it add MVT::Flag inputs and
/// outputs to ensure they are scheduled together and in order. This
/// optimization may benefit some targets by improving cache locality.
void ScheduleDAGSDNodes::ClusterNeighboringLoads() {
SmallPtrSet<SDNode*, 16> Visited;
SmallVector<int64_t, 4> Offsets;
DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
SDNode *Node = &*NI;
if (!Node || !Node->isMachineOpcode())
continue;
unsigned Opc = Node->getMachineOpcode();
const TargetInstrDesc &TID = TII->get(Opc);
if (!TID.mayLoad())
continue;
SDNode *Chain = 0;
unsigned NumOps = Node->getNumOperands();
if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
Chain = Node->getOperand(NumOps-1).getNode();
if (!Chain)
continue;
// Look for other loads of the same chain. Find loads that are loading from
// the same base pointer and different offsets.
Visited.clear();
Offsets.clear();
O2SMap.clear();
bool Cluster = false;
SDNode *Base = Node;
int64_t BaseOffset;
for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
I != E; ++I) {
SDNode *User = *I;
if (User == Node || !Visited.insert(User))
continue;
int64_t Offset1, Offset2;
if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
Offset1 == Offset2)
// FIXME: Should be ok if they addresses are identical. But earlier
// optimizations really should have eliminated one of the loads.
continue;
if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
Offsets.push_back(Offset1);
O2SMap.insert(std::make_pair(Offset2, User));
Offsets.push_back(Offset2);
if (Offset2 < Offset1) {
Base = User;
BaseOffset = Offset2;
} else {
BaseOffset = Offset1;
}
Cluster = true;
}
if (!Cluster)
continue;
// Sort them in increasing order.
std::sort(Offsets.begin(), Offsets.end());
// Check if the loads are close enough.
SmallVector<SDNode*, 4> Loads;
unsigned NumLoads = 0;
int64_t BaseOff = Offsets[0];
SDNode *BaseLoad = O2SMap[BaseOff];
Loads.push_back(BaseLoad);
for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
int64_t Offset = Offsets[i];
SDNode *Load = O2SMap[Offset];
if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,
NumLoads))
break; // Stop right here. Ignore loads that are further away.
Loads.push_back(Load);
++NumLoads;
}
if (NumLoads == 0)
continue;
// Cluster loads by adding MVT::Flag outputs and inputs. This also
// ensure they are scheduled in order of increasing addresses.
SDNode *Lead = Loads[0];
AddFlags(Lead, SDValue(0,0), true, DAG);
SDValue InFlag = SDValue(Lead, Lead->getNumValues()-1);
for (unsigned i = 1, e = Loads.size(); i != e; ++i) {
bool OutFlag = i < e-1;
SDNode *Load = Loads[i];
AddFlags(Load, InFlag, OutFlag, DAG);
if (OutFlag)
InFlag = SDValue(Load, Load->getNumValues()-1);
++LoadsClustered;
}
}
}
/// Given \p BBs as input, find another set of BBs which collectively
/// dominates \p BBs and have the minimal sum of frequencies. Return the BB
/// set found in \p BBs.
static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI,
BasicBlock *Entry,
SmallPtrSet<BasicBlock *, 8> &BBs) {
assert(!BBs.count(Entry) && "Assume Entry is not in BBs");
// Nodes on the current path to the root.
SmallPtrSet<BasicBlock *, 8> Path;
// Candidates includes any block 'BB' in set 'BBs' that is not strictly
// dominated by any other blocks in set 'BBs', and all nodes in the path
// in the dominator tree from Entry to 'BB'.
SmallPtrSet<BasicBlock *, 16> Candidates;
for (auto BB : BBs) {
// Ignore unreachable basic blocks.
if (!DT.isReachableFromEntry(BB))
continue;
Path.clear();
// Walk up the dominator tree until Entry or another BB in BBs
// is reached. Insert the nodes on the way to the Path.
BasicBlock *Node = BB;
// The "Path" is a candidate path to be added into Candidates set.
bool isCandidate = false;
do {
Path.insert(Node);
if (Node == Entry || Candidates.count(Node)) {
isCandidate = true;
break;
}
assert(DT.getNode(Node)->getIDom() &&
"Entry doens't dominate current Node");
Node = DT.getNode(Node)->getIDom()->getBlock();
} while (!BBs.count(Node));
// If isCandidate is false, Node is another Block in BBs dominating
// current 'BB'. Drop the nodes on the Path.
if (!isCandidate)
continue;
// Add nodes on the Path into Candidates.
Candidates.insert(Path.begin(), Path.end());
}
// Sort the nodes in Candidates in top-down order and save the nodes
// in Orders.
unsigned Idx = 0;
SmallVector<BasicBlock *, 16> Orders;
Orders.push_back(Entry);
while (Idx != Orders.size()) {
BasicBlock *Node = Orders[Idx++];
for (auto ChildDomNode : DT.getNode(Node)->getChildren()) {
if (Candidates.count(ChildDomNode->getBlock()))
Orders.push_back(ChildDomNode->getBlock());
}
}
// Visit Orders in bottom-up order.
using InsertPtsCostPair =
std::pair<SmallPtrSet<BasicBlock *, 16>, BlockFrequency>;
// InsertPtsMap is a map from a BB to the best insertion points for the
// subtree of BB (subtree not including the BB itself).
DenseMap<BasicBlock *, InsertPtsCostPair> InsertPtsMap;
InsertPtsMap.reserve(Orders.size() + 1);
for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) {
BasicBlock *Node = *RIt;
bool NodeInBBs = BBs.count(Node);
SmallPtrSet<BasicBlock *, 16> &InsertPts = InsertPtsMap[Node].first;
BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second;
// Return the optimal insert points in BBs.
if (Node == Entry) {
BBs.clear();
if (InsertPtsFreq > BFI.getBlockFreq(Node) ||
(InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1))
BBs.insert(Entry);
else
BBs.insert(InsertPts.begin(), InsertPts.end());
break;
}
BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock();
// Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child
// will update its parent's ParentInsertPts and ParentPtsFreq.
SmallPtrSet<BasicBlock *, 16> &ParentInsertPts = InsertPtsMap[Parent].first;
BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second;
// Choose to insert in Node or in subtree of Node.
// Don't hoist to EHPad because we may not find a proper place to insert
// in EHPad.
// If the total frequency of InsertPts is the same as the frequency of the
// target Node, and InsertPts contains more than one nodes, choose hoisting
// to reduce code size.
if (NodeInBBs ||
(!Node->isEHPad() &&
(InsertPtsFreq > BFI.getBlockFreq(Node) ||
(InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) {
ParentInsertPts.insert(Node);
ParentPtsFreq += BFI.getBlockFreq(Node);
} else {
ParentInsertPts.insert(InsertPts.begin(), InsertPts.end());
ParentPtsFreq += InsertPtsFreq;
}
}
}
static bool simplifyLoopInst(Loop &L, DominatorTree &DT, LoopInfo &LI,
AssumptionCache &AC, const TargetLibraryInfo &TLI,
MemorySSAUpdater *MSSAU) {
const DataLayout &DL = L.getHeader()->getModule()->getDataLayout();
SimplifyQuery SQ(DL, &TLI, &DT, &AC);
// On the first pass over the loop body we try to simplify every instruction.
// On subsequent passes, we can restrict this to only simplifying instructions
// where the inputs have been updated. We end up needing two sets: one
// containing the instructions we are simplifying in *this* pass, and one for
// the instructions we will want to simplify in the *next* pass. We use
// pointers so we can swap between two stably allocated sets.
SmallPtrSet<const Instruction *, 8> S1, S2, *ToSimplify = &S1, *Next = &S2;
// Track the PHI nodes that have already been visited during each iteration so
// that we can identify when it is necessary to iterate.
SmallPtrSet<PHINode *, 4> VisitedPHIs;
// While simplifying we may discover dead code or cause code to become dead.
// Keep track of all such instructions and we will delete them at the end.
SmallVector<Instruction *, 8> DeadInsts;
// First we want to create an RPO traversal of the loop body. By processing in
// RPO we can ensure that definitions are processed prior to uses (for non PHI
// uses) in all cases. This ensures we maximize the simplifications in each
// iteration over the loop and minimizes the possible causes for continuing to
// iterate.
LoopBlocksRPO RPOT(&L);
RPOT.perform(&LI);
MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr;
bool Changed = false;
for (;;) {
if (MSSAU && VerifyMemorySSA)
MSSA->verifyMemorySSA();
for (BasicBlock *BB : RPOT) {
for (Instruction &I : *BB) {
if (auto *PI = dyn_cast<PHINode>(&I))
VisitedPHIs.insert(PI);
if (I.use_empty()) {
if (isInstructionTriviallyDead(&I, &TLI))
DeadInsts.push_back(&I);
continue;
}
// We special case the first iteration which we can detect due to the
// empty `ToSimplify` set.
bool IsFirstIteration = ToSimplify->empty();
if (!IsFirstIteration && !ToSimplify->count(&I))
continue;
Value *V = SimplifyInstruction(&I, SQ.getWithInstruction(&I));
if (!V || !LI.replacementPreservesLCSSAForm(&I, V))
continue;
for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
UI != UE;) {
Use &U = *UI++;
auto *UserI = cast<Instruction>(U.getUser());
U.set(V);
// If the instruction is used by a PHI node we have already processed
// we'll need to iterate on the loop body to converge, so add it to
// the next set.
if (auto *UserPI = dyn_cast<PHINode>(UserI))
if (VisitedPHIs.count(UserPI)) {
Next->insert(UserPI);
continue;
}
// If we are only simplifying targeted instructions and the user is an
// instruction in the loop body, add it to our set of targeted
// instructions. Because we process defs before uses (outside of PHIs)
// we won't have visited it yet.
//
// We also skip any uses outside of the loop being simplified. Those
// should always be PHI nodes due to LCSSA form, and we don't want to
// try to simplify those away.
assert((L.contains(UserI) || isa<PHINode>(UserI)) &&
"Uses outside the loop should be PHI nodes due to LCSSA!");
if (!IsFirstIteration && L.contains(UserI))
ToSimplify->insert(UserI);
}
if (MSSAU)
if (Instruction *SimpleI = dyn_cast_or_null<Instruction>(V))
if (MemoryAccess *MA = MSSA->getMemoryAccess(&I))
if (MemoryAccess *ReplacementMA = MSSA->getMemoryAccess(SimpleI))
MA->replaceAllUsesWith(ReplacementMA);
assert(I.use_empty() && "Should always have replaced all uses!");
if (isInstructionTriviallyDead(&I, &TLI))
DeadInsts.push_back(&I);
++NumSimplified;
Changed = true;
}
}
// Delete any dead instructions found thus far now that we've finished an
// iteration over all instructions in all the loop blocks.
if (!DeadInsts.empty()) {
Changed = true;
RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, &TLI, MSSAU);
}
if (MSSAU && VerifyMemorySSA)
MSSA->verifyMemorySSA();
// If we never found a PHI that needs to be simplified in the next
// iteration, we're done.
if (Next->empty())
break;
// Otherwise, put the next set in place for the next iteration and reset it
// and the visited PHIs for that iteration.
std::swap(Next, ToSimplify);
Next->clear();
VisitedPHIs.clear();
DeadInsts.clear();
}
return Changed;
}
bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
if (DisablePeephole)
return false;
TM = &MF.getTarget();
TII = TM->getInstrInfo();
MRI = &MF.getRegInfo();
DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : 0;
bool Changed = false;
SmallPtrSet<MachineInstr*, 8> LocalMIs;
SmallSet<unsigned, 4> ImmDefRegs;
DenseMap<unsigned, MachineInstr*> ImmDefMIs;
for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
MachineBasicBlock *MBB = &*I;
bool SeenMoveImm = false;
LocalMIs.clear();
ImmDefRegs.clear();
ImmDefMIs.clear();
bool First = true;
MachineBasicBlock::iterator PMII;
for (MachineBasicBlock::iterator
MII = I->begin(), MIE = I->end(); MII != MIE; ) {
MachineInstr *MI = &*MII;
LocalMIs.insert(MI);
if (MI->isLabel() || MI->isPHI() || MI->isImplicitDef() ||
MI->isKill() || MI->isInlineAsm() || MI->isDebugValue() ||
MI->hasUnmodeledSideEffects()) {
++MII;
continue;
}
if (MI->isBitcast()) {
if (optimizeBitcastInstr(MI, MBB)) {
// MI is deleted.
LocalMIs.erase(MI);
Changed = true;
MII = First ? I->begin() : llvm::next(PMII);
continue;
}
} else if (MI->isCompare()) {
if (optimizeCmpInstr(MI, MBB)) {
// MI is deleted.
LocalMIs.erase(MI);
Changed = true;
MII = First ? I->begin() : llvm::next(PMII);
continue;
}
}
if (isMoveImmediate(MI, ImmDefRegs, ImmDefMIs)) {
SeenMoveImm = true;
} else {
Changed |= optimizeExtInstr(MI, MBB, LocalMIs);
if (SeenMoveImm)
Changed |= foldImmediate(MI, MBB, ImmDefRegs, ImmDefMIs);
}
First = false;
PMII = MII;
++MII;
}
}
return Changed;
}
/// runOnMachineFunction - Reduce two-address instructions to two operands.
///
bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
DEBUG(errs() << "Machine Function\n");
const TargetMachine &TM = MF.getTarget();
MRI = &MF.getRegInfo();
TII = TM.getInstrInfo();
TRI = TM.getRegisterInfo();
LV = getAnalysisIfAvailable<LiveVariables>();
AA = &getAnalysis<AliasAnalysis>();
bool MadeChange = false;
DEBUG(errs() << "********** REWRITING TWO-ADDR INSTRS **********\n");
DEBUG(errs() << "********** Function: "
<< MF.getFunction()->getName() << '\n');
// ReMatRegs - Keep track of the registers whose def's are remat'ed.
BitVector ReMatRegs;
ReMatRegs.resize(MRI->getLastVirtReg()+1);
typedef DenseMap<unsigned, SmallVector<std::pair<unsigned, unsigned>, 4> >
TiedOperandMap;
TiedOperandMap TiedOperands(4);
SmallPtrSet<MachineInstr*, 8> Processed;
for (MachineFunction::iterator mbbi = MF.begin(), mbbe = MF.end();
mbbi != mbbe; ++mbbi) {
unsigned Dist = 0;
DistanceMap.clear();
SrcRegMap.clear();
DstRegMap.clear();
Processed.clear();
for (MachineBasicBlock::iterator mi = mbbi->begin(), me = mbbi->end();
mi != me; ) {
MachineBasicBlock::iterator nmi = next(mi);
const TargetInstrDesc &TID = mi->getDesc();
bool FirstTied = true;
DistanceMap.insert(std::make_pair(mi, ++Dist));
ProcessCopy(&*mi, &*mbbi, Processed);
// First scan through all the tied register uses in this instruction
// and record a list of pairs of tied operands for each register.
unsigned NumOps = (mi->getOpcode() == TargetInstrInfo::INLINEASM)
? mi->getNumOperands() : TID.getNumOperands();
for (unsigned SrcIdx = 0; SrcIdx < NumOps; ++SrcIdx) {
unsigned DstIdx = 0;
if (!mi->isRegTiedToDefOperand(SrcIdx, &DstIdx))
continue;
if (FirstTied) {
FirstTied = false;
++NumTwoAddressInstrs;
DEBUG(errs() << '\t' << *mi);
}
assert(mi->getOperand(SrcIdx).isReg() &&
mi->getOperand(SrcIdx).getReg() &&
mi->getOperand(SrcIdx).isUse() &&
"two address instruction invalid");
unsigned regB = mi->getOperand(SrcIdx).getReg();
TiedOperandMap::iterator OI = TiedOperands.find(regB);
if (OI == TiedOperands.end()) {
SmallVector<std::pair<unsigned, unsigned>, 4> TiedPair;
OI = TiedOperands.insert(std::make_pair(regB, TiedPair)).first;
}
OI->second.push_back(std::make_pair(SrcIdx, DstIdx));
}
// Now iterate over the information collected above.
for (TiedOperandMap::iterator OI = TiedOperands.begin(),
OE = TiedOperands.end(); OI != OE; ++OI) {
SmallVector<std::pair<unsigned, unsigned>, 4> &TiedPairs = OI->second;
// If the instruction has a single pair of tied operands, try some
// transformations that may either eliminate the tied operands or
// improve the opportunities for coalescing away the register copy.
if (TiedOperands.size() == 1 && TiedPairs.size() == 1) {
unsigned SrcIdx = TiedPairs[0].first;
unsigned DstIdx = TiedPairs[0].second;
// If the registers are already equal, nothing needs to be done.
if (mi->getOperand(SrcIdx).getReg() ==
mi->getOperand(DstIdx).getReg())
break; // Done with this instruction.
if (TryInstructionTransform(mi, nmi, mbbi, SrcIdx, DstIdx, Dist))
break; // The tied operands have been eliminated.
}
bool RemovedKillFlag = false;
bool AllUsesCopied = true;
unsigned LastCopiedReg = 0;
unsigned regB = OI->first;
for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
unsigned SrcIdx = TiedPairs[tpi].first;
unsigned DstIdx = TiedPairs[tpi].second;
unsigned regA = mi->getOperand(DstIdx).getReg();
// Grab regB from the instruction because it may have changed if the
// instruction was commuted.
regB = mi->getOperand(SrcIdx).getReg();
if (regA == regB) {
// The register is tied to multiple destinations (or else we would
// not have continued this far), but this use of the register
// already matches the tied destination. Leave it.
AllUsesCopied = false;
continue;
}
LastCopiedReg = regA;
assert(TargetRegisterInfo::isVirtualRegister(regB) &&
"cannot make instruction into two-address form");
#ifndef NDEBUG
// First, verify that we don't have a use of "a" in the instruction
// (a = b + a for example) because our transformation will not
// work. This should never occur because we are in SSA form.
for (unsigned i = 0; i != mi->getNumOperands(); ++i)
assert(i == DstIdx ||
!mi->getOperand(i).isReg() ||
mi->getOperand(i).getReg() != regA);
#endif
// Emit a copy or rematerialize the definition.
const TargetRegisterClass *rc = MRI->getRegClass(regB);
MachineInstr *DefMI = MRI->getVRegDef(regB);
// If it's safe and profitable, remat the definition instead of
// copying it.
if (DefMI &&
DefMI->getDesc().isAsCheapAsAMove() &&
DefMI->isSafeToReMat(TII, regB, AA) &&
isProfitableToReMat(regB, rc, mi, DefMI, mbbi, Dist)){
DEBUG(errs() << "2addr: REMATTING : " << *DefMI << "\n");
unsigned regASubIdx = mi->getOperand(DstIdx).getSubReg();
TII->reMaterialize(*mbbi, mi, regA, regASubIdx, DefMI, TRI);
ReMatRegs.set(regB);
++NumReMats;
} else {
bool Emitted = TII->copyRegToReg(*mbbi, mi, regA, regB, rc, rc);
(void)Emitted;
assert(Emitted && "Unable to issue a copy instruction!\n");
}
MachineBasicBlock::iterator prevMI = prior(mi);
// Update DistanceMap.
DistanceMap.insert(std::make_pair(prevMI, Dist));
DistanceMap[mi] = ++Dist;
DEBUG(errs() << "\t\tprepend:\t" << *prevMI);
MachineOperand &MO = mi->getOperand(SrcIdx);
assert(MO.isReg() && MO.getReg() == regB && MO.isUse() &&
"inconsistent operand info for 2-reg pass");
if (MO.isKill()) {
MO.setIsKill(false);
RemovedKillFlag = true;
}
MO.setReg(regA);
}
if (AllUsesCopied) {
// Replace other (un-tied) uses of regB with LastCopiedReg.
for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) {
MachineOperand &MO = mi->getOperand(i);
if (MO.isReg() && MO.getReg() == regB && MO.isUse()) {
if (MO.isKill()) {
MO.setIsKill(false);
RemovedKillFlag = true;
}
MO.setReg(LastCopiedReg);
}
}
// Update live variables for regB.
if (RemovedKillFlag && LV && LV->getVarInfo(regB).removeKill(mi))
LV->addVirtualRegisterKilled(regB, prior(mi));
} else if (RemovedKillFlag) {
// Some tied uses of regB matched their destination registers, so
// regB is still used in this instruction, but a kill flag was
// removed from a different tied use of regB, so now we need to add
// a kill flag to one of the remaining uses of regB.
for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) {
MachineOperand &MO = mi->getOperand(i);
if (MO.isReg() && MO.getReg() == regB && MO.isUse()) {
MO.setIsKill(true);
break;
}
}
}
MadeChange = true;
DEBUG(errs() << "\t\trewrite to:\t" << *mi);
}
// Clear TiedOperands here instead of at the top of the loop
// since most instructions do not have tied operands.
TiedOperands.clear();
mi = nmi;
}
}
// Some remat'ed instructions are dead.
int VReg = ReMatRegs.find_first();
while (VReg != -1) {
if (MRI->use_empty(VReg)) {
MachineInstr *DefMI = MRI->getVRegDef(VReg);
DefMI->eraseFromParent();
}
VReg = ReMatRegs.find_next(VReg);
}
return MadeChange;
}
