#define DEBUG_TYPE "regalloc"
#include "Spiller.h"
-#include "LiveRangeEdit.h"
-#include "VirtRegMap.h"
+#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/TinyPtrVector.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/LiveRangeEdit.h"
#include "llvm/CodeGen/LiveStackAnalysis.h"
#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineInstrBundle.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/Target/TargetMachine.h"
-#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/CodeGen/VirtRegMap.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetMachine.h"
using namespace llvm;
STATISTIC(NumSpilledRanges, "Number of spilled live ranges");
-STATISTIC(NumSnippets, "Number of snippets included in spills");
+STATISTIC(NumSnippets, "Number of spilled snippets");
STATISTIC(NumSpills, "Number of spills inserted");
+STATISTIC(NumSpillsRemoved, "Number of spills removed");
STATISTIC(NumReloads, "Number of reloads inserted");
+STATISTIC(NumReloadsRemoved, "Number of reloads removed");
STATISTIC(NumFolded, "Number of folded stack accesses");
STATISTIC(NumFoldedLoads, "Number of folded loads");
STATISTIC(NumRemats, "Number of rematerialized defs for spilling");
-STATISTIC(NumOmitReloadSpill, "Number of omitted spills after reloads");
-STATISTIC(NumHoistLocal, "Number of locally hoisted spills");
-STATISTIC(NumHoistGlobal, "Number of globally hoisted spills");
-STATISTIC(NumRedundantSpills, "Number of redundant spills identified");
+STATISTIC(NumOmitReloadSpill, "Number of omitted spills of reloads");
+STATISTIC(NumHoists, "Number of hoisted spills");
+
+static cl::opt<bool> DisableHoisting("disable-spill-hoist", cl::Hidden,
+ cl::desc("Disable inline spill hoisting"));
namespace {
class InlineSpiller : public Spiller {
- MachineFunctionPass &Pass;
MachineFunction &MF;
LiveIntervals &LIS;
LiveStacks &LSS;
MachineRegisterInfo &MRI;
const TargetInstrInfo &TII;
const TargetRegisterInfo &TRI;
+ const MachineBlockFrequencyInfo &MBFI;
// Variables that are valid during spill(), but used by multiple methods.
LiveRangeEdit *Edit;
// True when value is defined by an original PHI not from splitting.
bool DefByOrigPHI;
+ // True when the COPY defining this value killed its source.
+ bool KillsSource;
+
// The preferred register to spill.
unsigned SpillReg;
TinyPtrVector<VNInfo*> Deps;
SibValueInfo(unsigned Reg, VNInfo *VNI)
- : AllDefsAreReloads(true), DefByOrigPHI(false),
+ : AllDefsAreReloads(true), DefByOrigPHI(false), KillsSource(false),
SpillReg(Reg), SpillVNI(VNI), SpillMBB(0), DefMI(0) {}
// Returns true when a def has been found.
typedef DenseMap<VNInfo*, SibValueInfo> SibValueMap;
SibValueMap SibValues;
- // Values live-out from basic blocks. This is the same as
- // LI.getVNInfoAt(LIS.getMBBEndIdx(MBB).getPrevSlot())
- typedef DenseMap<MachineBasicBlock*, VNInfo*> LiveOutMap;
- LiveOutMap LiveOutValues;
-
// Dead defs generated during spilling.
SmallVector<MachineInstr*, 8> DeadDefs;
InlineSpiller(MachineFunctionPass &pass,
MachineFunction &mf,
VirtRegMap &vrm)
- : Pass(pass),
- MF(mf),
+ : MF(mf),
LIS(pass.getAnalysis<LiveIntervals>()),
LSS(pass.getAnalysis<LiveStacks>()),
AA(&pass.getAnalysis<AliasAnalysis>()),
MFI(*mf.getFrameInfo()),
MRI(mf.getRegInfo()),
TII(*mf.getTarget().getInstrInfo()),
- TRI(*mf.getTarget().getRegisterInfo()) {}
+ TRI(*mf.getTarget().getRegisterInfo()),
+ MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()) {}
void spill(LiveRangeEdit &);
void reMaterializeAll();
bool coalesceStackAccess(MachineInstr *MI, unsigned Reg);
- bool foldMemoryOperand(MachineBasicBlock::iterator MI,
- const SmallVectorImpl<unsigned> &Ops,
+ bool foldMemoryOperand(ArrayRef<std::pair<MachineInstr*, unsigned> >,
MachineInstr *LoadMI = 0);
void insertReload(LiveInterval &NewLI, SlotIndex,
MachineBasicBlock::iterator MI);
OS << " all-reloads";
if (SVI.DefByOrigPHI)
OS << " orig-phi";
+ if (SVI.KillsSource)
+ OS << " kill";
OS << " deps[";
for (unsigned i = 0, e = SVI.Deps.size(); i != e; ++i)
OS << ' ' << SVI.Deps[i]->id << '@' << SVI.Deps[i]->def;
/// propagateSiblingValue - Propagate the value in SVI to dependents if it is
/// known. Otherwise remember the dependency for later.
///
-/// @param SVI SibValues entry to propagate.
+/// @param SVIIter SibValues entry to propagate.
/// @param VNI Dependent value, or NULL to propagate to all saved dependents.
-void InlineSpiller::propagateSiblingValue(SibValueMap::iterator SVI,
+void InlineSpiller::propagateSiblingValue(SibValueMap::iterator SVIIter,
VNInfo *VNI) {
+ SibValueMap::value_type *SVI = &*SVIIter;
+
// When VNI is non-NULL, add it to SVI's deps, and only propagate to that.
TinyPtrVector<VNInfo*> FirstDeps;
if (VNI) {
if (!SVI->second.hasDef())
return;
- // Work list of values to propagate. It would be nice to use a SetVector
- // here, but then we would be forced to use a SmallSet.
- SmallVector<SibValueMap::iterator, 8> WorkList(1, SVI);
- SmallPtrSet<VNInfo*, 8> WorkSet;
+ // Work list of values to propagate.
+ SmallSetVector<SibValueMap::value_type *, 8> WorkList;
+ WorkList.insert(SVI);
do {
SVI = WorkList.pop_back_val();
- WorkSet.erase(SVI->first);
TinyPtrVector<VNInfo*> *Deps = VNI ? &FirstDeps : &SVI->second.Deps;
VNI = 0;
// Should this value be propagated as a preferred spill candidate? We don't
// propagate values of registers that are about to spill.
- bool PropSpill = !isRegToSpill(SV.SpillReg);
+ bool PropSpill = !DisableHoisting && !isRegToSpill(SV.SpillReg);
unsigned SpillDepth = ~0u;
for (TinyPtrVector<VNInfo*>::iterator DepI = Deps->begin(),
if (PropSpill && SV.SpillVNI != DepSV.SpillVNI) {
if (SV.SpillMBB == DepSV.SpillMBB) {
// DepSV is in the same block. Hoist when dominated.
- if (SV.SpillVNI->def < DepSV.SpillVNI->def) {
+ if (DepSV.KillsSource && SV.SpillVNI->def < DepSV.SpillVNI->def) {
// This is an alternative def earlier in the same MBB.
// Hoist the spill as far as possible in SpillMBB. This can ease
// register pressure:
// spill x
// y = use x<kill>
//
+ // This hoist only helps when the DepSV copy kills its source.
Changed = true;
DepSV.SpillReg = SV.SpillReg;
DepSV.SpillVNI = SV.SpillVNI;
continue;
// Something changed in DepSVI. Propagate to dependents.
- if (WorkSet.insert(DepSVI->first))
- WorkList.push_back(DepSVI);
+ WorkList.insert(&*DepSVI);
DEBUG(dbgs() << " update " << DepSVI->first->id << '@'
<< DepSVI->first->def << " to:\t" << DepSV);
// Trace through PHI-defs created by live range splitting.
if (VNI->isPHIDef()) {
+ // Stop at original PHIs. We don't know the value at the predecessors.
if (VNI->def == OrigVNI->def) {
DEBUG(dbgs() << "orig phi value\n");
SVI->second.DefByOrigPHI = true;
propagateSiblingValue(SVI);
continue;
}
- // Get values live-out of predecessors.
+
+ // This is a PHI inserted by live range splitting. We could trace the
+ // live-out value from predecessor blocks, but that search can be very
+ // expensive if there are many predecessors and many more PHIs as
+ // generated by tail-dup when it sees an indirectbr. Instead, look at
+ // all the non-PHI defs that have the same value as OrigVNI. They must
+ // jointly dominate VNI->def. This is not optimal since VNI may actually
+ // be jointly dominated by a smaller subset of defs, so there is a change
+ // we will miss a AllDefsAreReloads optimization.
+
+ // Separate all values dominated by OrigVNI into PHIs and non-PHIs.
+ SmallVector<VNInfo*, 8> PHIs, NonPHIs;
LiveInterval &LI = LIS.getInterval(Reg);
- MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def);
- DEBUG(dbgs() << "split phi value, check " << MBB->pred_size()
- << " preds\n");
- for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
- PE = MBB->pred_end(); PI != PE; ++PI) {
- // Use a cache of block live-out values. This is faster than using
- // getVNInfoAt on complex intervals.
- VNInfo *&PVNI = LiveOutValues[*PI];
- if (!PVNI)
- PVNI = LI.getVNInfoAt(LIS.getMBBEndIdx(*PI).getPrevSlot());
- if (!PVNI)
+ LiveInterval &OrigLI = LIS.getInterval(Original);
+
+ for (LiveInterval::vni_iterator VI = LI.vni_begin(), VE = LI.vni_end();
+ VI != VE; ++VI) {
+ VNInfo *VNI2 = *VI;
+ if (VNI2->isUnused())
+ continue;
+ if (!OrigLI.containsOneValue() &&
+ OrigLI.getVNInfoAt(VNI2->def) != OrigVNI)
continue;
- // Known predecessor value? Try an insertion.
+ if (VNI2->isPHIDef() && VNI2->def != OrigVNI->def)
+ PHIs.push_back(VNI2);
+ else
+ NonPHIs.push_back(VNI2);
+ }
+ DEBUG(dbgs() << "split phi value, checking " << PHIs.size()
+ << " phi-defs, and " << NonPHIs.size()
+ << " non-phi/orig defs\n");
+
+ // Create entries for all the PHIs. Don't add them to the worklist, we
+ // are processing all of them in one go here.
+ for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
+ SibValues.insert(std::make_pair(PHIs[i], SibValueInfo(Reg, PHIs[i])));
+
+ // Add every PHI as a dependent of all the non-PHIs.
+ for (unsigned i = 0, e = NonPHIs.size(); i != e; ++i) {
+ VNInfo *NonPHI = NonPHIs[i];
+ // Known value? Try an insertion.
tie(SVI, Inserted) =
- SibValues.insert(std::make_pair(PVNI, SibValueInfo(Reg, PVNI)));
- // This is the first time we see PVNI, add it to the worklist.
+ SibValues.insert(std::make_pair(NonPHI, SibValueInfo(Reg, NonPHI)));
+ // Add all the PHIs as dependents of NonPHI.
+ for (unsigned pi = 0, pe = PHIs.size(); pi != pe; ++pi)
+ SVI->second.Deps.push_back(PHIs[pi]);
+ // This is the first time we see NonPHI, add it to the worklist.
if (Inserted)
- WorkList.push_back(std::make_pair(Reg, PVNI));
- propagateSiblingValue(SVI, VNI);
+ WorkList.push_back(std::make_pair(Reg, NonPHI));
+ else
+ // Propagate to all inserted PHIs, not just VNI.
+ propagateSiblingValue(SVI);
}
+
// Next work list item.
continue;
}
if (unsigned SrcReg = isFullCopyOf(MI, Reg)) {
if (isSibling(SrcReg)) {
LiveInterval &SrcLI = LIS.getInterval(SrcReg);
- VNInfo *SrcVNI = SrcLI.getVNInfoAt(VNI->def.getUseIndex());
- assert(SrcVNI && "Copy from non-existing value");
+ LiveRangeQuery SrcQ(SrcLI, VNI->def);
+ assert(SrcQ.valueIn() && "Copy from non-existing value");
+ // Check if this COPY kills its source.
+ SVI->second.KillsSource = SrcQ.isKill();
+ VNInfo *SrcVNI = SrcQ.valueIn();
DEBUG(dbgs() << "copy of " << PrintReg(SrcReg) << ':'
- << SrcVNI->id << '@' << SrcVNI->def << '\n');
+ << SrcVNI->id << '@' << SrcVNI->def
+ << " kill=" << unsigned(SVI->second.KillsSource) << '\n');
// Known sibling source value? Try an insertion.
tie(SVI, Inserted) = SibValues.insert(std::make_pair(SrcVNI,
SibValueInfo(SrcReg, SrcVNI)));
propagateSiblingValue(SVI);
} while (!WorkList.empty());
- // Look up the value we were looking for. We already did this lokup at the
+ // Look up the value we were looking for. We already did this lookup at the
// top of the function, but SibValues may have been invalidated.
SVI = SibValues.find(UseVNI);
assert(SVI != SibValues.end() && "Didn't compute requested info");
/// Keep track of values that may be rematerializable.
void InlineSpiller::analyzeSiblingValues() {
SibValues.clear();
- LiveOutValues.clear();
// No siblings at all?
if (Edit->getReg() == Original)
if (VNI->isUnused())
continue;
MachineInstr *DefMI = 0;
+ if (!VNI->isPHIDef()) {
+ DefMI = LIS.getInstructionFromIndex(VNI->def);
+ assert(DefMI && "No defining instruction");
+ }
// Check possible sibling copies.
- if (VNI->isPHIDef() || VNI->getCopy()) {
+ if (VNI->isPHIDef() || DefMI->isCopy()) {
VNInfo *OrigVNI = OrigLI.getVNInfoAt(VNI->def);
assert(OrigVNI && "Def outside original live range");
if (OrigVNI->def != VNI->def)
DefMI = traceSiblingValue(Reg, VNI, OrigVNI);
}
- if (!DefMI && !VNI->isPHIDef())
- DefMI = LIS.getInstructionFromIndex(VNI->def);
- if (DefMI && Edit->checkRematerializable(VNI, DefMI, TII, AA)) {
+ if (DefMI && Edit->checkRematerializable(VNI, DefMI, AA)) {
DEBUG(dbgs() << "Value " << PrintReg(Reg) << ':' << VNI->id << '@'
<< VNI->def << " may remat from " << *DefMI);
}
/// a spill at a better location.
bool InlineSpiller::hoistSpill(LiveInterval &SpillLI, MachineInstr *CopyMI) {
SlotIndex Idx = LIS.getInstructionIndex(CopyMI);
- VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getDefIndex());
- assert(VNI && VNI->def == Idx.getDefIndex() && "Not defined by copy");
+ VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot());
+ assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy");
SibValueMap::iterator I = SibValues.find(VNI);
if (I == SibValues.end())
return false;
MRI.getRegClass(SVI.SpillReg), &TRI);
--MII; // Point to store instruction.
LIS.InsertMachineInstrInMaps(MII);
- VRM.addSpillSlotUse(StackSlot, MII);
DEBUG(dbgs() << "\thoisted: " << SVI.SpillVNI->def << '\t' << *MII);
- if (MBB == CopyMI->getParent())
- ++NumHoistLocal;
- else
- ++NumHoistGlobal;
+ ++NumSpills;
+ ++NumHoists;
return true;
}
// Find all spills and copies of VNI.
for (MachineRegisterInfo::use_nodbg_iterator UI = MRI.use_nodbg_begin(Reg);
MachineInstr *MI = UI.skipInstruction();) {
- if (!MI->isCopy() && !MI->getDesc().mayStore())
+ if (!MI->isCopy() && !MI->mayStore())
continue;
SlotIndex Idx = LIS.getInstructionIndex(MI);
if (LI->getVNInfoAt(Idx) != VNI)
if (unsigned DstReg = isFullCopyOf(MI, Reg)) {
if (isSibling(DstReg)) {
LiveInterval &DstLI = LIS.getInterval(DstReg);
- VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getDefIndex());
+ VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot());
assert(DstVNI && "Missing defined value");
- assert(DstVNI->def == Idx.getDefIndex() && "Wrong copy def slot");
+ assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot");
WorkList.push_back(std::make_pair(&DstLI, DstVNI));
}
continue;
// eliminateDeadDefs won't normally remove stores, so switch opcode.
MI->setDesc(TII.get(TargetOpcode::KILL));
DeadDefs.push_back(MI);
- ++NumRedundantSpills;
+ ++NumSpillsRemoved;
+ --NumSpills;
}
}
} while (!WorkList.empty());
MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def);
for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
PE = MBB->pred_end(); PI != PE; ++PI) {
- VNInfo *PVNI = LI->getVNInfoAt(LIS.getMBBEndIdx(*PI).getPrevSlot());
+ VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(*PI));
if (PVNI)
WorkList.push_back(std::make_pair(LI, PVNI));
}
continue;
LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg());
assert(isRegToSpill(SnipLI.reg) && "Unexpected register in copy");
- VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getUseIndex());
+ VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true));
assert(SnipVNI && "Snippet undefined before copy");
WorkList.push_back(std::make_pair(&SnipLI, SnipVNI));
} while (!WorkList.empty());
/// reMaterializeFor - Attempt to rematerialize before MI instead of reloading.
bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg,
MachineBasicBlock::iterator MI) {
- SlotIndex UseIdx = LIS.getInstructionIndex(MI).getUseIndex();
+ SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true);
VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex());
if (!ParentVNI) {
SibValueMap::const_iterator SibI = SibValues.find(ParentVNI);
if (SibI != SibValues.end())
RM.OrigMI = SibI->second.DefMI;
- if (!Edit->canRematerializeAt(RM, UseIdx, false, LIS)) {
+ if (!Edit->canRematerializeAt(RM, UseIdx, false)) {
markValueUsed(&VirtReg, ParentVNI);
DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << *MI);
return false;
// If the instruction also writes VirtReg.reg, it had better not require the
// same register for uses and defs.
- bool Reads, Writes;
- SmallVector<unsigned, 8> Ops;
- tie(Reads, Writes) = MI->readsWritesVirtualRegister(VirtReg.reg, &Ops);
- if (Writes) {
- for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
- MachineOperand &MO = MI->getOperand(Ops[i]);
- if (MO.isUse() ? MI->isRegTiedToDefOperand(Ops[i]) : MO.getSubReg()) {
- markValueUsed(&VirtReg, ParentVNI);
- DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << *MI);
- return false;
- }
- }
+ SmallVector<std::pair<MachineInstr*, unsigned>, 8> Ops;
+ MIBundleOperands::VirtRegInfo RI =
+ MIBundleOperands(MI).analyzeVirtReg(VirtReg.reg, &Ops);
+ if (RI.Tied) {
+ markValueUsed(&VirtReg, ParentVNI);
+ DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << *MI);
+ return false;
}
// Before rematerializing into a register for a single instruction, try to
// fold a load into the instruction. That avoids allocating a new register.
- if (RM.OrigMI->getDesc().canFoldAsLoad() &&
- foldMemoryOperand(MI, Ops, RM.OrigMI)) {
+ if (RM.OrigMI->canFoldAsLoad() &&
+ foldMemoryOperand(Ops, RM.OrigMI)) {
Edit->markRematerialized(RM.ParentVNI);
++NumFoldedLoads;
return true;
}
// Alocate a new register for the remat.
- LiveInterval &NewLI = Edit->createFrom(Original, LIS, VRM);
+ LiveInterval &NewLI = Edit->createFrom(Original);
NewLI.markNotSpillable();
// Finally we can rematerialize OrigMI before MI.
SlotIndex DefIdx = Edit->rematerializeAt(*MI->getParent(), MI, NewLI.reg, RM,
- LIS, TII, TRI);
+ TRI);
DEBUG(dbgs() << "\tremat: " << DefIdx << '\t'
<< *LIS.getInstructionFromIndex(DefIdx));
// Replace operands
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
- MachineOperand &MO = MI->getOperand(Ops[i]);
+ MachineOperand &MO = MI->getOperand(Ops[i].second);
if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg) {
MO.setReg(NewLI.reg);
MO.setIsKill();
}
DEBUG(dbgs() << "\t " << UseIdx << '\t' << *MI);
- VNInfo *DefVNI = NewLI.getNextValue(DefIdx, 0, LIS.getVNInfoAllocator());
- NewLI.addRange(LiveRange(DefIdx, UseIdx.getDefIndex(), DefVNI));
+ VNInfo *DefVNI = NewLI.getNextValue(DefIdx, LIS.getVNInfoAllocator());
+ NewLI.addRange(LiveRange(DefIdx, UseIdx.getRegSlot(), DefVNI));
DEBUG(dbgs() << "\tinterval: " << NewLI << '\n');
++NumRemats;
return true;
/// and trim the live ranges after.
void InlineSpiller::reMaterializeAll() {
// analyzeSiblingValues has already tested all relevant defining instructions.
- if (!Edit->anyRematerializable(LIS, TII, AA))
+ if (!Edit->anyRematerializable(AA))
return;
UsedValues.clear();
LiveInterval &LI = LIS.getInterval(Reg);
for (MachineRegisterInfo::use_nodbg_iterator
RI = MRI.use_nodbg_begin(Reg);
- MachineInstr *MI = RI.skipInstruction();)
+ MachineInstr *MI = RI.skipBundle();)
anyRemat |= reMaterializeFor(LI, MI);
}
if (!anyRemat)
if (DeadDefs.empty())
return;
DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n");
- Edit->eliminateDeadDefs(DeadDefs, LIS, VRM, TII);
+ Edit->eliminateDeadDefs(DeadDefs, RegsToSpill);
// Get rid of deleted and empty intervals.
- for (unsigned i = RegsToSpill.size(); i != 0; --i) {
- unsigned Reg = RegsToSpill[i-1];
- if (!LIS.hasInterval(Reg)) {
- RegsToSpill.erase(RegsToSpill.begin() + (i - 1));
+ unsigned ResultPos = 0;
+ for (unsigned i = 0, e = RegsToSpill.size(); i != e; ++i) {
+ unsigned Reg = RegsToSpill[i];
+ if (!LIS.hasInterval(Reg))
continue;
- }
+
LiveInterval &LI = LIS.getInterval(Reg);
- if (!LI.empty())
+ if (LI.empty()) {
+ Edit->eraseVirtReg(Reg);
continue;
- Edit->eraseVirtReg(Reg, LIS);
- RegsToSpill.erase(RegsToSpill.begin() + (i - 1));
+ }
+
+ RegsToSpill[ResultPos++] = Reg;
}
+ RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end());
DEBUG(dbgs() << RegsToSpill.size() << " registers to spill after remat.\n");
}
/// If MI is a load or store of StackSlot, it can be removed.
bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, unsigned Reg) {
int FI = 0;
- unsigned InstrReg;
- if (!(InstrReg = TII.isLoadFromStackSlot(MI, FI)) &&
- !(InstrReg = TII.isStoreToStackSlot(MI, FI)))
- return false;
+ unsigned InstrReg = TII.isLoadFromStackSlot(MI, FI);
+ bool IsLoad = InstrReg;
+ if (!IsLoad)
+ InstrReg = TII.isStoreToStackSlot(MI, FI);
// We have a stack access. Is it the right register and slot?
if (InstrReg != Reg || FI != StackSlot)
DEBUG(dbgs() << "Coalescing stack access: " << *MI);
LIS.RemoveMachineInstrFromMaps(MI);
MI->eraseFromParent();
+
+ if (IsLoad) {
+ ++NumReloadsRemoved;
+ --NumReloads;
+ } else {
+ ++NumSpillsRemoved;
+ --NumSpills;
+ }
+
return true;
}
-/// foldMemoryOperand - Try folding stack slot references in Ops into MI.
-/// @param MI Instruction using or defining the current register.
-/// @param Ops Operand indices from readsWritesVirtualRegister().
+/// foldMemoryOperand - Try folding stack slot references in Ops into their
+/// instructions.
+///
+/// @param Ops Operand indices from analyzeVirtReg().
/// @param LoadMI Load instruction to use instead of stack slot when non-null.
-/// @return True on success, and MI will be erased.
-bool InlineSpiller::foldMemoryOperand(MachineBasicBlock::iterator MI,
- const SmallVectorImpl<unsigned> &Ops,
- MachineInstr *LoadMI) {
+/// @return True on success.
+bool InlineSpiller::
+foldMemoryOperand(ArrayRef<std::pair<MachineInstr*, unsigned> > Ops,
+ MachineInstr *LoadMI) {
+ if (Ops.empty())
+ return false;
+ // Don't attempt folding in bundles.
+ MachineInstr *MI = Ops.front().first;
+ if (Ops.back().first != MI || MI->isBundled())
+ return false;
+
+ bool WasCopy = MI->isCopy();
+ unsigned ImpReg = 0;
+
// TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied
// operands.
SmallVector<unsigned, 8> FoldOps;
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
- unsigned Idx = Ops[i];
+ unsigned Idx = Ops[i].second;
MachineOperand &MO = MI->getOperand(Idx);
- if (MO.isImplicit())
+ if (MO.isImplicit()) {
+ ImpReg = MO.getReg();
continue;
+ }
// FIXME: Teach targets to deal with subregs.
if (MO.getSubReg())
return false;
: TII.foldMemoryOperand(MI, FoldOps, StackSlot);
if (!FoldMI)
return false;
+
+ // Remove LIS for any dead defs in the original MI not in FoldMI.
+ for (MIBundleOperands MO(MI); MO.isValid(); ++MO) {
+ if (!MO->isReg())
+ continue;
+ unsigned Reg = MO->getReg();
+ if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg) ||
+ MRI.isReserved(Reg)) {
+ continue;
+ }
+ MIBundleOperands::PhysRegInfo RI =
+ MIBundleOperands(FoldMI).analyzePhysReg(Reg, &TRI);
+ if (MO->readsReg()) {
+ assert(RI.Reads && "Cannot fold physreg reader");
+ continue;
+ }
+ if (RI.Defines)
+ continue;
+ // FoldMI does not define this physreg. Remove the LI segment.
+ assert(MO->isDead() && "Cannot fold physreg def");
+ for (MCRegUnitIterator Units(Reg, &TRI); Units.isValid(); ++Units) {
+ if (LiveInterval *LI = LIS.getCachedRegUnit(*Units)) {
+ SlotIndex Idx = LIS.getInstructionIndex(MI).getRegSlot();
+ if (VNInfo *VNI = LI->getVNInfoAt(Idx))
+ LI->removeValNo(VNI);
+ }
+ }
+ }
LIS.ReplaceMachineInstrInMaps(MI, FoldMI);
- if (!LoadMI)
- VRM.addSpillSlotUse(StackSlot, FoldMI);
MI->eraseFromParent();
- DEBUG(dbgs() << "\tfolded: " << *FoldMI);
- ++NumFolded;
+
+ // TII.foldMemoryOperand may have left some implicit operands on the
+ // instruction. Strip them.
+ if (ImpReg)
+ for (unsigned i = FoldMI->getNumOperands(); i; --i) {
+ MachineOperand &MO = FoldMI->getOperand(i - 1);
+ if (!MO.isReg() || !MO.isImplicit())
+ break;
+ if (MO.getReg() == ImpReg)
+ FoldMI->RemoveOperand(i - 1);
+ }
+
+ DEBUG(dbgs() << "\tfolded: " << LIS.getInstructionIndex(FoldMI) << '\t'
+ << *FoldMI);
+ if (!WasCopy)
+ ++NumFolded;
+ else if (Ops.front().second == 0)
+ ++NumSpills;
+ else
+ ++NumReloads;
return true;
}
TII.loadRegFromStackSlot(MBB, MI, NewLI.reg, StackSlot,
MRI.getRegClass(NewLI.reg), &TRI);
--MI; // Point to load instruction.
- SlotIndex LoadIdx = LIS.InsertMachineInstrInMaps(MI).getDefIndex();
- VRM.addSpillSlotUse(StackSlot, MI);
+ SlotIndex LoadIdx = LIS.InsertMachineInstrInMaps(MI).getRegSlot();
+ // Some (out-of-tree) targets have EC reload instructions.
+ if (MachineOperand *MO = MI->findRegisterDefOperand(NewLI.reg))
+ if (MO->isEarlyClobber())
+ LoadIdx = LoadIdx.getRegSlot(true);
DEBUG(dbgs() << "\treload: " << LoadIdx << '\t' << *MI);
- VNInfo *LoadVNI = NewLI.getNextValue(LoadIdx, 0,
- LIS.getVNInfoAllocator());
+ VNInfo *LoadVNI = NewLI.getNextValue(LoadIdx, LIS.getVNInfoAllocator());
NewLI.addRange(LiveRange(LoadIdx, Idx, LoadVNI));
++NumReloads;
}
TII.storeRegToStackSlot(MBB, ++MI, NewLI.reg, true, StackSlot,
MRI.getRegClass(NewLI.reg), &TRI);
--MI; // Point to store instruction.
- SlotIndex StoreIdx = LIS.InsertMachineInstrInMaps(MI).getDefIndex();
- VRM.addSpillSlotUse(StackSlot, MI);
+ SlotIndex StoreIdx = LIS.InsertMachineInstrInMaps(MI).getRegSlot();
DEBUG(dbgs() << "\tspilled: " << StoreIdx << '\t' << *MI);
- VNInfo *StoreVNI = NewLI.getNextValue(Idx, 0, LIS.getVNInfoAllocator());
+ VNInfo *StoreVNI = NewLI.getNextValue(Idx, LIS.getVNInfoAllocator());
NewLI.addRange(LiveRange(Idx, StoreIdx, StoreVNI));
++NumSpills;
}
LiveInterval &OldLI = LIS.getInterval(Reg);
// Iterate over instructions using Reg.
- for (MachineRegisterInfo::reg_iterator RI = MRI.reg_begin(Reg);
- MachineInstr *MI = RI.skipInstruction();) {
+ for (MachineRegisterInfo::reg_iterator RegI = MRI.reg_begin(Reg);
+ MachineInstr *MI = RegI.skipBundle();) {
// Debug values are not allowed to affect codegen.
if (MI->isDebugValue()) {
uint64_t Offset = MI->getOperand(1).getImm();
const MDNode *MDPtr = MI->getOperand(2).getMetadata();
DebugLoc DL = MI->getDebugLoc();
- if (MachineInstr *NewDV = TII.emitFrameIndexDebugValue(MF, StackSlot,
- Offset, MDPtr, DL)) {
- DEBUG(dbgs() << "Modifying debug info due to spill:" << "\t" << *MI);
- MachineBasicBlock *MBB = MI->getParent();
- MBB->insert(MBB->erase(MI), NewDV);
- } else {
- DEBUG(dbgs() << "Removing debug info due to spill:" << "\t" << *MI);
- MI->eraseFromParent();
- }
+ DEBUG(dbgs() << "Modifying debug info due to spill:" << "\t" << *MI);
+ MachineBasicBlock *MBB = MI->getParent();
+ BuildMI(*MBB, MBB->erase(MI), DL, TII.get(TargetOpcode::DBG_VALUE))
+ .addFrameIndex(StackSlot).addImm(Offset).addMetadata(MDPtr);
continue;
}
continue;
// Analyze instruction.
- bool Reads, Writes;
- SmallVector<unsigned, 8> Ops;
- tie(Reads, Writes) = MI->readsWritesVirtualRegister(Reg, &Ops);
+ SmallVector<std::pair<MachineInstr*, unsigned>, 8> Ops;
+ MIBundleOperands::VirtRegInfo RI =
+ MIBundleOperands(MI).analyzeVirtReg(Reg, &Ops);
// Find the slot index where this instruction reads and writes OldLI.
// This is usually the def slot, except for tied early clobbers.
- SlotIndex Idx = LIS.getInstructionIndex(MI).getDefIndex();
- if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getUseIndex()))
+ SlotIndex Idx = LIS.getInstructionIndex(MI).getRegSlot();
+ if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true)))
if (SlotIndex::isSameInstr(Idx, VNI->def))
Idx = VNI->def;
SnippetCopies.insert(MI);
continue;
}
- if (Writes) {
+ if (RI.Writes) {
// Hoist the spill of a sib-reg copy.
if (hoistSpill(OldLI, MI)) {
// This COPY is now dead, the value is already in the stack slot.
}
// Attempt to fold memory ops.
- if (foldMemoryOperand(MI, Ops))
+ if (foldMemoryOperand(Ops))
continue;
// Allocate interval around instruction.
// FIXME: Infer regclass from instruction alone.
- LiveInterval &NewLI = Edit->createFrom(Reg, LIS, VRM);
+ LiveInterval &NewLI = Edit->createFrom(Reg);
NewLI.markNotSpillable();
- if (Reads)
+ if (RI.Reads)
insertReload(NewLI, Idx, MI);
// Rewrite instruction operands.
bool hasLiveDef = false;
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
- MachineOperand &MO = MI->getOperand(Ops[i]);
+ MachineOperand &MO = Ops[i].first->getOperand(Ops[i].second);
MO.setReg(NewLI.reg);
if (MO.isUse()) {
- if (!MI->isRegTiedToDefOperand(Ops[i]))
+ if (!Ops[i].first->isRegTiedToDefOperand(Ops[i].second))
MO.setIsKill();
} else {
if (!MO.isDead())
DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << *MI);
// FIXME: Use a second vreg if instruction has no tied ops.
- if (Writes && hasLiveDef)
- insertSpill(NewLI, OldLI, Idx, MI);
+ if (RI.Writes) {
+ if (hasLiveDef)
+ insertSpill(NewLI, OldLI, Idx, MI);
+ else {
+ // This instruction defines a dead value. We don't need to spill it,
+ // but do create a live range for the dead value.
+ VNInfo *VNI = NewLI.getNextValue(Idx, LIS.getVNInfoAllocator());
+ NewLI.addRange(LiveRange(Idx, Idx.getDeadSlot(), VNI));
+ }
+ }
DEBUG(dbgs() << "\tinterval: " << NewLI << '\n');
}
if (StackSlot == VirtRegMap::NO_STACK_SLOT) {
StackSlot = VRM.assignVirt2StackSlot(Original);
StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original));
- StackInt->getNextValue(SlotIndex(), 0, LSS.getVNInfoAllocator());
+ StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator());
} else
StackInt = &LSS.getInterval(StackSlot);
// Hoisted spills may cause dead code.
if (!DeadDefs.empty()) {
DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n");
- Edit->eliminateDeadDefs(DeadDefs, LIS, VRM, TII);
+ Edit->eliminateDeadDefs(DeadDefs, RegsToSpill);
}
// Finally delete the SnippetCopies.
MachineInstr *MI = RI.skipInstruction();) {
assert(SnippetCopies.count(MI) && "Remaining use wasn't a snippet copy");
// FIXME: Do this with a LiveRangeEdit callback.
- VRM.RemoveMachineInstrFromMaps(MI);
LIS.RemoveMachineInstrFromMaps(MI);
MI->eraseFromParent();
}
// Delete all spilled registers.
for (unsigned i = 0, e = RegsToSpill.size(); i != e; ++i)
- Edit->eraseVirtReg(RegsToSpill[i], LIS);
+ Edit->eraseVirtReg(RegsToSpill[i]);
}
void InlineSpiller::spill(LiveRangeEdit &edit) {
DEBUG(dbgs() << "Inline spilling "
<< MRI.getRegClass(edit.getReg())->getName()
- << ':' << edit.getParent() << "\nFrom original "
- << LIS.getInterval(Original) << '\n');
+ << ':' << PrintReg(edit.getReg()) << ' ' << edit.getParent()
+ << "\nFrom original " << LIS.getInterval(Original) << '\n');
assert(edit.getParent().isSpillable() &&
"Attempting to spill already spilled value.");
assert(DeadDefs.empty() && "Previous spill didn't remove dead defs");
if (!RegsToSpill.empty())
spillAll();
- Edit->calculateRegClassAndHint(MF, LIS, Loops);
+ Edit->calculateRegClassAndHint(MF, Loops, MBFI);
}
projects / oota-llvm.git / blobdiff