#include "SpillPlacement.h"
#include "SplitKit.h"
#include "VirtRegMap.h"
-#include "RegisterCoalescer.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Function.h"
STATISTIC(NumLocalSplits, "Number of split local live ranges");
STATISTIC(NumEvicted, "Number of interferences evicted");
-cl::opt<bool> CompactRegions("compact-regions");
+static cl::opt<SplitEditor::ComplementSpillMode>
+SplitSpillMode("split-spill-mode", cl::Hidden,
+ cl::desc("Spill mode for splitting live ranges"),
+ cl::values(clEnumValN(SplitEditor::SM_Partition, "default", "Default"),
+ clEnumValN(SplitEditor::SM_Size, "size", "Optimize for size"),
+ clEnumValN(SplitEditor::SM_Speed, "speed", "Optimize for speed"),
+ clEnumValEnd),
+ cl::init(SplitEditor::SM_Partition));
static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
createGreedyRegisterAllocator);
}
};
+ // Register mask interference. The current VirtReg is checked for register
+ // mask interference on entry to selectOrSplit(). If there is no
+ // interference, UsableRegs is left empty. If there is interference,
+ // UsableRegs has a bit mask of registers that can be used without register
+ // mask interference.
+ BitVector UsableRegs;
+
+ /// clobberedByRegMask - Returns true if PhysReg is not directly usable
+ /// because of register mask clobbers.
+ bool clobberedByRegMask(unsigned PhysReg) const {
+ return !UsableRegs.empty() && !UsableRegs.test(PhysReg);
+ }
+
// splitting state.
std::auto_ptr<SplitAnalysis> SA;
std::auto_ptr<SplitEditor> SE;
static char ID;
private:
- void LRE_WillEraseInstruction(MachineInstr*);
bool LRE_CanEraseVirtReg(unsigned);
void LRE_WillShrinkVirtReg(unsigned);
void LRE_DidCloneVirtReg(unsigned, unsigned);
SmallVectorImpl<LiveInterval*>&, unsigned = ~0u);
unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
SmallVectorImpl<LiveInterval*>&);
+ unsigned tryBlockSplit(LiveInterval&, AllocationOrder&,
+ SmallVectorImpl<LiveInterval*>&);
unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
SmallVectorImpl<LiveInterval*>&);
unsigned trySplit(LiveInterval&, AllocationOrder&,
initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
- initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
+ initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
initializeLiveStacksPass(*PassRegistry::getPassRegistry());
initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
AU.addPreserved<SlotIndexes>();
AU.addRequired<LiveDebugVariables>();
AU.addPreserved<LiveDebugVariables>();
- if (StrongPHIElim)
- AU.addRequiredID(StrongPHIEliminationID);
- AU.addRequiredTransitive<RegisterCoalescer>();
AU.addRequired<CalculateSpillWeights>();
AU.addRequired<LiveStacks>();
AU.addPreserved<LiveStacks>();
// LiveRangeEdit delegate methods
//===----------------------------------------------------------------------===//
-void RAGreedy::LRE_WillEraseInstruction(MachineInstr *MI) {
- // LRE itself will remove from SlotIndexes and parent basic block.
- VRM->RemoveMachineInstrFromMaps(MI);
-}
-
bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) {
if (unsigned PhysReg = VRM->getPhys(VirtReg)) {
unassign(LIS->getInterval(VirtReg), PhysReg);
}
void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
+ // Cloning a register we haven't even heard about yet? Just ignore it.
+ if (!ExtraRegInfo.inBounds(Old))
+ return;
+
// LRE may clone a virtual register because dead code elimination causes it to
// be split into connected components. The new components are much smaller
// than the original, so they should get a new chance at being assigned.
if (ExtraRegInfo[Reg].Stage == RS_Split) {
// Unsplit ranges that couldn't be allocated immediately are deferred until
- // everything else has been allocated. Long ranges are allocated last so
- // they are split against realistic interference.
- if (CompactRegions)
- Prio = Size;
- else
- Prio = (1u << 31) - Size;
+ // everything else has been allocated.
+ Prio = Size;
} else {
- // Everything else is allocated in long->short order. Long ranges that don't
- // fit should be spilled ASAP so they don't create interference.
+ // Everything is allocated in long->short order. Long ranges that don't fit
+ // should be spilled (or split) ASAP so they don't create interference.
Prio = (1u << 31) + Size;
// Boost ranges that have a physical register hint.
SmallVectorImpl<LiveInterval*> &NewVRegs) {
Order.rewind();
unsigned PhysReg;
- while ((PhysReg = Order.next()))
+ while ((PhysReg = Order.next())) {
+ if (clobberedByRegMask(PhysReg))
+ continue;
if (!checkPhysRegInterference(VirtReg, PhysReg))
break;
+ }
if (!PhysReg || Order.isHint(PhysReg))
return PhysReg;
// If we missed a simple hint, try to cheaply evict interference from the
// preferred register.
if (unsigned Hint = MRI->getSimpleHint(VirtReg.reg))
- if (Order.isHint(Hint)) {
+ if (Order.isHint(Hint) && !clobberedByRegMask(Hint)) {
DEBUG(dbgs() << "missed hint " << PrintReg(Hint, TRI) << '\n');
EvictionCost MaxCost(1);
if (canEvictInterference(VirtReg, Hint, true, MaxCost)) {
Order.rewind();
while (unsigned PhysReg = Order.next()) {
+ if (clobberedByRegMask(PhysReg))
+ continue;
if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit)
continue;
// The first use of a callee-saved register in a function has cost 1.
Intf.moveToBlock(BC.Number);
BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
+ BC.ChangesValue = BI.FirstDef;
if (!Intf.hasInterference())
continue;
if (BI.LiveIn) {
if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number))
BC.Entry = SpillPlacement::MustSpill, ++Ins;
- else if (Intf.first() < BI.FirstUse)
+ else if (Intf.first() < BI.FirstInstr)
BC.Entry = SpillPlacement::PrefSpill, ++Ins;
- else if (Intf.first() < BI.LastUse)
+ else if (Intf.first() < BI.LastInstr)
++Ins;
}
if (BI.LiveOut) {
if (Intf.last() >= SA->getLastSplitPoint(BC.Number))
BC.Exit = SpillPlacement::MustSpill, ++Ins;
- else if (Intf.last() > BI.LastUse)
+ else if (Intf.last() > BI.LastInstr)
BC.Exit = SpillPlacement::PrefSpill, ++Ins;
- else if (Intf.last() > BI.FirstUse)
+ else if (Intf.last() > BI.FirstInstr)
++Ins;
}
if (Cand.PhysReg)
addThroughConstraints(Cand.Intf, NewBlocks);
else
- SpillPlacer->addPrefSpill(NewBlocks);
+ // Provide a strong negative bias on through blocks to prevent unwanted
+ // liveness on loop backedges.
+ SpillPlacer->addPrefSpill(NewBlocks, /* Strong= */ true);
AddedTo = ActiveBlocks.size();
// Perhaps iterating can enable more bundles?
/// SA around all use blocks instead of forming bundle regions.
float RAGreedy::calcSpillCost() {
float Cost = 0;
- const LiveInterval &LI = SA->getParent();
ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
for (unsigned i = 0; i != UseBlocks.size(); ++i) {
const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
Cost += SpillPlacer->getBlockFrequency(Number);
// Unless the value is redefined in the block.
- if (BI.LiveIn && BI.LiveOut) {
- SlotIndex Start, Stop;
- tie(Start, Stop) = Indexes->getMBBRange(Number);
- LiveInterval::const_iterator I = LI.find(Start);
- assert(I != LI.end() && "Expected live-in value");
- // Is there a different live-out value? If so, we need an extra spill
- // instruction.
- if (I->end < Stop)
- Cost += SpillPlacer->getBlockFrequency(Number);
- }
+ if (BI.LiveIn && BI.LiveOut && BI.FirstDef)
+ Cost += SpillPlacer->getBlockFrequency(Number);
}
return Cost;
}
DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs << " globals.\n");
assert(NumGlobalIntvs && "No global intervals configured");
+ // Isolate even single instructions when dealing with a proper sub-class.
+ // That guarantees register class inflation for the stack interval because it
+ // is all copies.
+ unsigned Reg = SA->getParent().reg;
+ bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
+
// First handle all the blocks with uses.
ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
for (unsigned i = 0; i != UseBlocks.size(); ++i) {
// Create separate intervals for isolated blocks with multiple uses.
if (!IntvIn && !IntvOut) {
DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n");
- if (!BI.isOneInstr())
+ if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
SE->splitSingleBlock(BI);
continue;
}
SmallVector<unsigned, 8> IntvMap;
SE->finish(&IntvMap);
- DebugVars->splitRegister(SA->getParent().reg, LREdit.regs());
+ DebugVars->splitRegister(Reg, LREdit.regs());
ExtraRegInfo.resize(MRI->getNumVirtRegs());
unsigned OrigBlocks = SA->getNumLiveBlocks();
SmallVector<unsigned, 8> UsedCands;
// Check if we can split this live range around a compact region.
- bool HasCompact = CompactRegions && calcCompactRegion(GlobalCand.front());
+ bool HasCompact = calcCompactRegion(GlobalCand.front());
if (HasCompact) {
// Yes, keep GlobalCand[0] as the compact region candidate.
NumCands = 1;
}
--NumCands;
GlobalCand[Worst] = GlobalCand[NumCands];
+ if (BestCand == NumCands)
+ BestCand = Worst;
}
if (GlobalCand.size() <= NumCands)
// Prepare split editor.
LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
- SE->reset(LREdit);
+ SE->reset(LREdit, SplitSpillMode);
// Assign all edge bundles to the preferred candidate, or NoCand.
BundleCand.assign(Bundles->getNumBundles(), NoCand);
Cand.IntvIdx = SE->openIntv();
DEBUG(dbgs() << "Split for " << PrintReg(Cand.PhysReg, TRI) << " in "
<< B << " bundles, intv " << Cand.IntvIdx << ".\n");
+ (void)B;
}
}
Cand.IntvIdx = SE->openIntv();
DEBUG(dbgs() << "Split for compact region in " << B << " bundles, intv "
<< Cand.IntvIdx << ".\n");
+ (void)B;
}
}
}
+//===----------------------------------------------------------------------===//
+// Per-Block Splitting
+//===----------------------------------------------------------------------===//
+
+/// tryBlockSplit - Split a global live range around every block with uses. This
+/// creates a lot of local live ranges, that will be split by tryLocalSplit if
+/// they don't allocate.
+unsigned RAGreedy::tryBlockSplit(LiveInterval &VirtReg, AllocationOrder &Order,
+ SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed");
+ unsigned Reg = VirtReg.reg;
+ bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
+ LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+ SE->reset(LREdit, SplitSpillMode);
+ ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
+ for (unsigned i = 0; i != UseBlocks.size(); ++i) {
+ const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
+ if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
+ SE->splitSingleBlock(BI);
+ }
+ // No blocks were split.
+ if (LREdit.empty())
+ return 0;
+
+ // We did split for some blocks.
+ SmallVector<unsigned, 8> IntvMap;
+ SE->finish(&IntvMap);
+
+ // Tell LiveDebugVariables about the new ranges.
+ DebugVars->splitRegister(Reg, LREdit.regs());
+
+ ExtraRegInfo.resize(MRI->getNumVirtRegs());
+
+ // Sort out the new intervals created by splitting. The remainder interval
+ // goes straight to spilling, the new local ranges get to stay RS_New.
+ for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
+ LiveInterval &LI = *LREdit.get(i);
+ if (getStage(LI) == RS_New && IntvMap[i] == 0)
+ setStage(LI, RS_Spill);
+ }
+
+ if (VerifyEnabled)
+ MF->verify(this, "After splitting live range around basic blocks");
+ return 0;
+}
+
//===----------------------------------------------------------------------===//
// Local Splitting
//===----------------------------------------------------------------------===//
SmallVectorImpl<float> &GapWeight) {
assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
- const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
+ ArrayRef<SlotIndex> Uses = SA->getUseSlots();
const unsigned NumGaps = Uses.size()-1;
// Start and end points for the interference check.
- SlotIndex StartIdx = BI.LiveIn ? BI.FirstUse.getBaseIndex() : BI.FirstUse;
- SlotIndex StopIdx = BI.LiveOut ? BI.LastUse.getBoundaryIndex() : BI.LastUse;
+ SlotIndex StartIdx =
+ BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr;
+ SlotIndex StopIdx =
+ BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr;
GapWeight.assign(NumGaps, 0.0f);
.checkInterference())
continue;
- // We know that VirtReg is a continuous interval from FirstUse to LastUse,
- // so we don't need InterferenceQuery.
+ // We know that VirtReg is a continuous interval from FirstInstr to
+ // LastInstr, so we don't need InterferenceQuery.
//
// Interference that overlaps an instruction is counted in both gaps
// surrounding the instruction. The exception is interference before
// StartIdx and after StopIdx.
//
- LiveIntervalUnion::SegmentIter IntI = PhysReg2LiveUnion[*AI].find(StartIdx);
+ LiveIntervalUnion::SegmentIter IntI = getLiveUnion(*AI).find(StartIdx);
for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
// Skip the gaps before IntI.
while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
// while only covering a single block - A phi-def can use undef values from
// predecessors, and the block could be a single-block loop.
// We don't bother doing anything clever about such a case, we simply assume
- // that the interval is continuous from FirstUse to LastUse. We should make
- // sure that we don't do anything illegal to such an interval, though.
+ // that the interval is continuous from FirstInstr to LastInstr. We should
+ // make sure that we don't do anything illegal to such an interval, though.
- const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
+ ArrayRef<SlotIndex> Uses = SA->getUseSlots();
if (Uses.size() <= 2)
return 0;
const unsigned NumGaps = Uses.size()-1;
DEBUG({
dbgs() << "tryLocalSplit: ";
for (unsigned i = 0, e = Uses.size(); i != e; ++i)
- dbgs() << ' ' << SA->UseSlots[i];
+ dbgs() << ' ' << Uses[i];
dbgs() << '\n';
});
/// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
SmallVectorImpl<LiveInterval*>&NewVRegs) {
+ // Ranges must be Split2 or less.
+ if (getStage(VirtReg) >= RS_Spill)
+ return 0;
+
// Local intervals are handled separately.
if (LIS->intervalIsInOneMBB(VirtReg)) {
NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
- // Ranges must be Split2 or less.
- if (getStage(VirtReg) >= RS_Spill)
- return 0;
-
SA->analyze(&VirtReg);
// FIXME: SplitAnalysis may repair broken live ranges coming from the
return PhysReg;
}
- // Then isolate blocks with multiple uses.
- SplitAnalysis::BlockPtrSet Blocks;
- if (SA->getMultiUseBlocks(Blocks)) {
- LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
- SE->reset(LREdit);
- SE->splitSingleBlocks(Blocks);
- setStage(NewVRegs.begin(), NewVRegs.end(), RS_Spill);
- if (VerifyEnabled)
- MF->verify(this, "After splitting live range around basic blocks");
- }
-
- // Don't assign any physregs.
- return 0;
+ // Then isolate blocks.
+ return tryBlockSplit(VirtReg, Order, NewVRegs);
}
unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ // Check if VirtReg is live across any calls.
+ UsableRegs.clear();
+ if (LIS->checkRegMaskInterference(VirtReg, UsableRegs))
+ DEBUG(dbgs() << "Live across regmasks.\n");
+
// First try assigning a free register.
AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
ExtraRegInfo.clear();
ExtraRegInfo.resize(MRI->getNumVirtRegs());
NextCascade = 1;
- IntfCache.init(MF, &PhysReg2LiveUnion[0], Indexes, TRI);
+ IntfCache.init(MF, &getLiveUnion(0), Indexes, TRI);
GlobalCand.resize(32); // This will grow as needed.
allocatePhysRegs();
}
// Write out new DBG_VALUE instructions.
- DebugVars->emitDebugValues(VRM);
+ {
+ NamedRegionTimer T("Emit Debug Info", TimerGroupName, TimePassesIsEnabled);
+ DebugVars->emitDebugValues(VRM);
+ }
// The pass output is in VirtRegMap. Release all the transient data.
releaseMemory();
projects / oota-llvm.git / blobdiff