#define DEBUG_TYPE "regalloc"
#include "AllocationOrder.h"
-#include "LiveIntervalUnion.h"
+#include "InterferenceCache.h"
+#include "LiveDebugVariables.h"
#include "LiveRangeEdit.h"
#include "RegAllocBase.h"
#include "Spiller.h"
#include "SpillPlacement.h"
#include "SplitKit.h"
#include "VirtRegMap.h"
-#include "VirtRegRewriter.h"
+#include "RegisterCoalescer.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Function.h"
#include "llvm/PassAnalysisSupport.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/RegAllocRegistry.h"
-#include "llvm/CodeGen/RegisterCoalescer.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Timer.h"
+#include <queue>
+
using namespace llvm;
+STATISTIC(NumGlobalSplits, "Number of split global live ranges");
+STATISTIC(NumLocalSplits, "Number of split local live ranges");
+STATISTIC(NumEvicted, "Number of interferences evicted");
+
static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
createGreedyRegisterAllocator);
namespace {
-class RAGreedy : public MachineFunctionPass, public RegAllocBase {
+class RAGreedy : public MachineFunctionPass,
+ public RegAllocBase,
+ private LiveRangeEdit::Delegate {
+
// context
MachineFunction *MF;
- BitVector ReservedRegs;
// analyses
SlotIndexes *Indexes;
MachineLoopRanges *LoopRanges;
EdgeBundles *Bundles;
SpillPlacement *SpillPlacer;
+ LiveDebugVariables *DebugVars;
// state
std::auto_ptr<Spiller> SpillerInstance;
- std::auto_ptr<SplitAnalysis> SA;
+ std::priority_queue<std::pair<unsigned, unsigned> > Queue;
+ unsigned NextCascade;
+
+ // Live ranges pass through a number of stages as we try to allocate them.
+ // Some of the stages may also create new live ranges:
+ //
+ // - Region splitting.
+ // - Per-block splitting.
+ // - Local splitting.
+ // - Spilling.
+ //
+ // Ranges produced by one of the stages skip the previous stages when they are
+ // dequeued. This improves performance because we can skip interference checks
+ // that are unlikely to give any results. It also guarantees that the live
+ // range splitting algorithm terminates, something that is otherwise hard to
+ // ensure.
+ enum LiveRangeStage {
+ RS_New, ///< Never seen before.
+ RS_First, ///< First time in the queue.
+ RS_Second, ///< Second time in the queue.
+ RS_Global, ///< Produced by global splitting.
+ RS_Local, ///< Produced by local splitting.
+ RS_Spill ///< Produced by spilling.
+ };
+
+ static const char *const StageName[];
+
+ // RegInfo - Keep additional information about each live range.
+ struct RegInfo {
+ LiveRangeStage Stage;
+
+ // Cascade - Eviction loop prevention. See canEvictInterference().
+ unsigned Cascade;
+
+ RegInfo() : Stage(RS_New), Cascade(0) {}
+ };
+
+ IndexedMap<RegInfo, VirtReg2IndexFunctor> ExtraRegInfo;
+
+ LiveRangeStage getStage(const LiveInterval &VirtReg) const {
+ return ExtraRegInfo[VirtReg.reg].Stage;
+ }
+
+ void setStage(const LiveInterval &VirtReg, LiveRangeStage Stage) {
+ ExtraRegInfo.resize(MRI->getNumVirtRegs());
+ ExtraRegInfo[VirtReg.reg].Stage = Stage;
+ }
+
+ template<typename Iterator>
+ void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
+ ExtraRegInfo.resize(MRI->getNumVirtRegs());
+ for (;Begin != End; ++Begin) {
+ unsigned Reg = (*Begin)->reg;
+ if (ExtraRegInfo[Reg].Stage == RS_New)
+ ExtraRegInfo[Reg].Stage = NewStage;
+ }
+ }
// splitting state.
+ std::auto_ptr<SplitAnalysis> SA;
+ std::auto_ptr<SplitEditor> SE;
+
+ /// Cached per-block interference maps
+ InterferenceCache IntfCache;
+
+ /// All basic blocks where the current register has uses.
+ SmallVector<SpillPlacement::BlockConstraint, 8> SplitConstraints;
- /// All basic blocks where the current register is live.
- SmallVector<SpillPlacement::BlockConstraint, 8> SpillConstraints;
-
- /// Additional information about basic blocks where the current variable is
- /// live. Such a block will look like one of these templates:
- ///
- /// 1. | o---x | Internal to block. Variable is only live in this block.
- /// 2. |---x | Live-in, kill.
- /// 3. | o---| Def, live-out.
- /// 4. |---x o---| Live-in, kill, def, live-out.
- /// 5. |---o---o---| Live-through with uses or defs.
- /// 6. |-----------| Live-through without uses. Transparent.
- ///
- struct BlockInfo {
- MachineBasicBlock *MBB;
- SlotIndex FirstUse; ///< First instr using current reg.
- SlotIndex LastUse; ///< Last instr using current reg.
- SlotIndex Kill; ///< Interval end point inside block.
- SlotIndex Def; ///< Interval start point inside block.
- /// Last possible point for splitting live ranges.
- SlotIndex LastSplitPoint;
- bool Uses; ///< Current reg has uses or defs in block.
- bool LiveThrough; ///< Live in whole block (Templ 5. or 6. above).
- bool LiveIn; ///< Current reg is live in.
- bool LiveOut; ///< Current reg is live out.
-
- // Per-interference pattern scratch data.
- bool OverlapEntry; ///< Interference overlaps entering interval.
- bool OverlapExit; ///< Interference overlaps exiting interval.
+ /// Global live range splitting candidate info.
+ struct GlobalSplitCandidate {
+ unsigned PhysReg;
+ BitVector LiveBundles;
+ SmallVector<unsigned, 8> ActiveBlocks;
+
+ void reset(unsigned Reg) {
+ PhysReg = Reg;
+ LiveBundles.clear();
+ ActiveBlocks.clear();
+ }
};
- /// Basic blocks where var is live. This array is parallel to
- /// SpillConstraints.
- SmallVector<BlockInfo, 8> LiveBlocks;
+ /// Candidate info for for each PhysReg in AllocationOrder.
+ /// This vector never shrinks, but grows to the size of the largest register
+ /// class.
+ SmallVector<GlobalSplitCandidate, 32> GlobalCand;
public:
RAGreedy();
/// RAGreedy analysis usage.
virtual void getAnalysisUsage(AnalysisUsage &AU) const;
-
virtual void releaseMemory();
-
virtual Spiller &spiller() { return *SpillerInstance; }
-
- virtual float getPriority(LiveInterval *LI);
-
+ virtual void enqueue(LiveInterval *LI);
+ virtual LiveInterval *dequeue();
virtual unsigned selectOrSplit(LiveInterval&,
SmallVectorImpl<LiveInterval*>&);
static char ID;
private:
- bool checkUncachedInterference(LiveInterval&, unsigned);
- LiveInterval *getSingleInterference(LiveInterval&, unsigned);
- bool reassignVReg(LiveInterval &InterferingVReg, unsigned OldPhysReg);
- bool reassignInterferences(LiveInterval &VirtReg, unsigned PhysReg);
- float calcInterferenceWeight(LiveInterval&, unsigned);
- void calcLiveBlockInfo(LiveInterval&);
- float calcInterferenceInfo(LiveInterval&, unsigned);
- float calcGlobalSplitCost(const BitVector&);
- void splitAroundRegion(LiveInterval&, unsigned, const BitVector&,
+ void LRE_WillEraseInstruction(MachineInstr*);
+ bool LRE_CanEraseVirtReg(unsigned);
+ void LRE_WillShrinkVirtReg(unsigned);
+ void LRE_DidCloneVirtReg(unsigned, unsigned);
+
+ float calcSpillCost();
+ bool addSplitConstraints(InterferenceCache::Cursor, float&);
+ void addThroughConstraints(InterferenceCache::Cursor, ArrayRef<unsigned>);
+ void growRegion(GlobalSplitCandidate &Cand, InterferenceCache::Cursor);
+ float calcGlobalSplitCost(GlobalSplitCandidate&, InterferenceCache::Cursor);
+ void splitAroundRegion(LiveInterval&, GlobalSplitCandidate&,
SmallVectorImpl<LiveInterval*>&);
-
- unsigned tryReassign(LiveInterval&, AllocationOrder&);
+ void calcGapWeights(unsigned, SmallVectorImpl<float>&);
+ bool canEvict(LiveInterval &A, LiveInterval &B);
+ bool canEvictInterference(LiveInterval&, unsigned, float&);
+
+ unsigned tryAssign(LiveInterval&, AllocationOrder&,
+ SmallVectorImpl<LiveInterval*>&);
+ unsigned tryEvict(LiveInterval&, AllocationOrder&,
+ SmallVectorImpl<LiveInterval*>&, unsigned = ~0u);
unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
SmallVectorImpl<LiveInterval*>&);
+ unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
+ SmallVectorImpl<LiveInterval*>&);
unsigned trySplit(LiveInterval&, AllocationOrder&,
SmallVectorImpl<LiveInterval*>&);
- unsigned trySpillInterferences(LiveInterval&, AllocationOrder&,
- SmallVectorImpl<LiveInterval*>&);
};
} // end anonymous namespace
char RAGreedy::ID = 0;
+#ifndef NDEBUG
+const char *const RAGreedy::StageName[] = {
+ "RS_New",
+ "RS_First",
+ "RS_Second",
+ "RS_Global",
+ "RS_Local",
+ "RS_Spill"
+};
+#endif
+
+// Hysteresis to use when comparing floats.
+// This helps stabilize decisions based on float comparisons.
+const float Hysteresis = 0.98f;
+
+
FunctionPass* llvm::createGreedyRegisterAllocator() {
return new RAGreedy();
}
RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
+ initializeLiveDebugVariablesPass(*PassRegistry::getPassRegistry());
initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
- initializeRegisterCoalescerAnalysisGroup(*PassRegistry::getPassRegistry());
+ initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
initializeLiveStacksPass(*PassRegistry::getPassRegistry());
initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
AU.addRequired<LiveIntervals>();
AU.addRequired<SlotIndexes>();
AU.addPreserved<SlotIndexes>();
+ AU.addRequired<LiveDebugVariables>();
+ AU.addPreserved<LiveDebugVariables>();
if (StrongPHIElim)
AU.addRequiredID(StrongPHIEliminationID);
AU.addRequiredTransitive<RegisterCoalescer>();
MachineFunctionPass::getAnalysisUsage(AU);
}
+
+//===----------------------------------------------------------------------===//
+// 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);
+ return true;
+ }
+ // Unassigned virtreg is probably in the priority queue.
+ // RegAllocBase will erase it after dequeueing.
+ return false;
+}
+
+void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
+ unsigned PhysReg = VRM->getPhys(VirtReg);
+ if (!PhysReg)
+ return;
+
+ // Register is assigned, put it back on the queue for reassignment.
+ LiveInterval &LI = LIS->getInterval(VirtReg);
+ unassign(LI, PhysReg);
+ enqueue(&LI);
+}
+
+void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
+ // LRE may clone a virtual register because dead code elimination causes it to
+ // be split into connected components. Ensure that the new register gets the
+ // same stage as the parent.
+ ExtraRegInfo.grow(New);
+ ExtraRegInfo[New] = ExtraRegInfo[Old];
+}
+
void RAGreedy::releaseMemory() {
SpillerInstance.reset(0);
+ ExtraRegInfo.clear();
+ GlobalCand.clear();
RegAllocBase::releaseMemory();
}
-float RAGreedy::getPriority(LiveInterval *LI) {
- float Priority = LI->weight;
+void RAGreedy::enqueue(LiveInterval *LI) {
+ // Prioritize live ranges by size, assigning larger ranges first.
+ // The queue holds (size, reg) pairs.
+ const unsigned Size = LI->getSize();
+ const unsigned Reg = LI->reg;
+ assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
+ "Can only enqueue virtual registers");
+ unsigned Prio;
+
+ ExtraRegInfo.grow(Reg);
+ if (ExtraRegInfo[Reg].Stage == RS_New)
+ ExtraRegInfo[Reg].Stage = RS_First;
+
+ if (ExtraRegInfo[Reg].Stage == RS_Second)
+ // 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.
+ Prio = (1u << 31) - 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.
+ Prio = (1u << 31) + Size;
+
+ // Boost ranges that have a physical register hint.
+ if (TargetRegisterInfo::isPhysicalRegister(VRM->getRegAllocPref(Reg)))
+ Prio |= (1u << 30);
+ }
- // Prioritize hinted registers so they are allocated first.
- std::pair<unsigned, unsigned> Hint;
- if (Hint.first || Hint.second) {
- // The hint can be target specific, a virtual register, or a physreg.
- Priority *= 2;
+ Queue.push(std::make_pair(Prio, Reg));
+}
- // Prefer physreg hints above anything else.
- if (Hint.first == 0 && TargetRegisterInfo::isPhysicalRegister(Hint.second))
- Priority *= 2;
- }
- return Priority;
+LiveInterval *RAGreedy::dequeue() {
+ if (Queue.empty())
+ return 0;
+ LiveInterval *LI = &LIS->getInterval(Queue.top().second);
+ Queue.pop();
+ return LI;
}
//===----------------------------------------------------------------------===//
-// Register Reassignment
+// Direct Assignment
//===----------------------------------------------------------------------===//
-// Check interference without using the cache.
-bool RAGreedy::checkUncachedInterference(LiveInterval &VirtReg,
- unsigned PhysReg) {
- for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
- LiveIntervalUnion::Query subQ(&VirtReg, &PhysReg2LiveUnion[*AliasI]);
- if (subQ.checkInterference())
- return true;
- }
- return false;
+/// tryAssign - Try to assign VirtReg to an available register.
+unsigned RAGreedy::tryAssign(LiveInterval &VirtReg,
+ AllocationOrder &Order,
+ SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ Order.rewind();
+ unsigned PhysReg;
+ while ((PhysReg = Order.next()))
+ if (!checkPhysRegInterference(VirtReg, PhysReg))
+ break;
+ if (!PhysReg || Order.isHint(PhysReg))
+ return PhysReg;
+
+ // PhysReg is available. Try to evict interference from a cheaper alternative.
+ unsigned Cost = TRI->getCostPerUse(PhysReg);
+
+ // Most registers have 0 additional cost.
+ if (!Cost)
+ return PhysReg;
+
+ DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " is available at cost " << Cost
+ << '\n');
+ unsigned CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost);
+ return CheapReg ? CheapReg : PhysReg;
}
-/// getSingleInterference - Return the single interfering virtual register
-/// assigned to PhysReg. Return 0 if more than one virtual register is
-/// interfering.
-LiveInterval *RAGreedy::getSingleInterference(LiveInterval &VirtReg,
- unsigned PhysReg) {
- // Check physreg and aliases.
- LiveInterval *Interference = 0;
+
+//===----------------------------------------------------------------------===//
+// Interference eviction
+//===----------------------------------------------------------------------===//
+
+/// canEvict - determine if A can evict the assigned live range B. The eviction
+/// policy defined by this function together with the allocation order defined
+/// by enqueue() decides which registers ultimately end up being split and
+/// spilled.
+///
+/// Cascade numbers are used to prevent infinite loops if this function is a
+/// cyclic relation.
+bool RAGreedy::canEvict(LiveInterval &A, LiveInterval &B) {
+ return A.weight > B.weight;
+}
+
+/// canEvict - Return true if all interferences between VirtReg and PhysReg can
+/// be evicted.
+/// Return false if any interference is heavier than MaxWeight.
+/// On return, set MaxWeight to the maximal spill weight of an interference.
+bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
+ float &MaxWeight) {
+ // Find VirtReg's cascade number. This will be unassigned if VirtReg was never
+ // involved in an eviction before. If a cascade number was assigned, deny
+ // evicting anything with the same or a newer cascade number. This prevents
+ // infinite eviction loops.
+ //
+ // This works out so a register without a cascade number is allowed to evict
+ // anything, and it can be evicted by anything.
+ unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
+ if (!Cascade)
+ Cascade = NextCascade;
+
+ float Weight = 0;
for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
- if (Q.checkInterference()) {
- if (Interference)
- return 0;
- Q.collectInterferingVRegs(1);
- if (!Q.seenAllInterferences())
- return 0;
- Interference = Q.interferingVRegs().front();
+ // If there is 10 or more interferences, chances are one is heavier.
+ if (Q.collectInterferingVRegs(10, MaxWeight) >= 10)
+ return false;
+
+ // Check if any interfering live range is heavier than MaxWeight.
+ for (unsigned i = Q.interferingVRegs().size(); i; --i) {
+ LiveInterval *Intf = Q.interferingVRegs()[i - 1];
+ if (TargetRegisterInfo::isPhysicalRegister(Intf->reg))
+ return false;
+ if (Cascade <= ExtraRegInfo[Intf->reg].Cascade)
+ return false;
+ if (Intf->weight >= MaxWeight)
+ return false;
+ if (!canEvict(VirtReg, *Intf))
+ return false;
+ Weight = std::max(Weight, Intf->weight);
}
}
- return Interference;
+ MaxWeight = Weight;
+ return true;
}
-// Attempt to reassign this virtual register to a different physical register.
-//
-// FIXME: we are not yet caching these "second-level" interferences discovered
-// in the sub-queries. These interferences can change with each call to
-// selectOrSplit. However, we could implement a "may-interfere" cache that
-// could be conservatively dirtied when we reassign or split.
-//
-// FIXME: This may result in a lot of alias queries. We could summarize alias
-// live intervals in their parent register's live union, but it's messy.
-bool RAGreedy::reassignVReg(LiveInterval &InterferingVReg,
- unsigned WantedPhysReg) {
- assert(TargetRegisterInfo::isVirtualRegister(InterferingVReg.reg) &&
- "Can only reassign virtual registers");
- assert(TRI->regsOverlap(WantedPhysReg, VRM->getPhys(InterferingVReg.reg)) &&
- "inconsistent phys reg assigment");
-
- AllocationOrder Order(InterferingVReg.reg, *VRM, ReservedRegs);
+/// tryEvict - Try to evict all interferences for a physreg.
+/// @param VirtReg Currently unassigned virtual register.
+/// @param Order Physregs to try.
+/// @return Physreg to assign VirtReg, or 0.
+unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
+ AllocationOrder &Order,
+ SmallVectorImpl<LiveInterval*> &NewVRegs,
+ unsigned CostPerUseLimit) {
+ NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
+
+ // Keep track of the lightest single interference seen so far.
+ float BestWeight = HUGE_VALF;
+ unsigned BestPhys = 0;
+
+ Order.rewind();
while (unsigned PhysReg = Order.next()) {
- // Don't reassign to a WantedPhysReg alias.
- if (TRI->regsOverlap(PhysReg, WantedPhysReg))
+ if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit)
+ continue;
+ // The first use of a register in a function has cost 1.
+ if (CostPerUseLimit == 1 && !MRI->isPhysRegUsed(PhysReg))
continue;
- if (checkUncachedInterference(InterferingVReg, PhysReg))
+ float Weight = BestWeight;
+ if (!canEvictInterference(VirtReg, PhysReg, Weight))
continue;
- // Reassign the interfering virtual reg to this physical reg.
- unsigned OldAssign = VRM->getPhys(InterferingVReg.reg);
- DEBUG(dbgs() << "reassigning: " << InterferingVReg << " from " <<
- TRI->getName(OldAssign) << " to " << TRI->getName(PhysReg) << '\n');
- PhysReg2LiveUnion[OldAssign].extract(InterferingVReg);
- VRM->clearVirt(InterferingVReg.reg);
- VRM->assignVirt2Phys(InterferingVReg.reg, PhysReg);
- PhysReg2LiveUnion[PhysReg].unify(InterferingVReg);
+ // This is an eviction candidate.
+ DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " interference = "
+ << Weight << '\n');
+ if (BestPhys && Weight >= BestWeight)
+ continue;
- return true;
+ // Best so far.
+ BestPhys = PhysReg;
+ BestWeight = Weight;
+ // Stop if the hint can be used.
+ if (Order.isHint(PhysReg))
+ break;
}
- return false;
-}
-/// reassignInterferences - Reassign all interferences to different physical
-/// registers such that Virtreg can be assigned to PhysReg.
-/// Currently this only works with a single interference.
-/// @param VirtReg Currently unassigned virtual register.
-/// @param PhysReg Physical register to be cleared.
-/// @return True on success, false if nothing was changed.
-bool RAGreedy::reassignInterferences(LiveInterval &VirtReg, unsigned PhysReg) {
- LiveInterval *InterferingVReg = getSingleInterference(VirtReg, PhysReg);
- if (!InterferingVReg)
- return false;
- if (TargetRegisterInfo::isPhysicalRegister(InterferingVReg->reg))
- return false;
- return reassignVReg(*InterferingVReg, PhysReg);
-}
+ if (!BestPhys)
+ return 0;
-/// tryReassign - Try to reassign interferences to different physregs.
-/// @param VirtReg Currently unassigned virtual register.
-/// @param Order Physregs to try.
-/// @return Physreg to assign VirtReg, or 0.
-unsigned RAGreedy::tryReassign(LiveInterval &VirtReg, AllocationOrder &Order) {
- NamedRegionTimer T("Reassign", TimerGroupName, TimePassesIsEnabled);
- Order.rewind();
- while (unsigned PhysReg = Order.next())
- if (reassignInterferences(VirtReg, PhysReg))
- return PhysReg;
- return 0;
+ // We will evict interference. Make sure that VirtReg has a cascade number,
+ // and assign that cascade number to every evicted register. These live
+ // ranges than then only be evicted by a newer cascade, preventing infinite
+ // loops.
+ unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
+ if (!Cascade)
+ Cascade = ExtraRegInfo[VirtReg.reg].Cascade = NextCascade++;
+
+ DEBUG(dbgs() << "evicting " << PrintReg(BestPhys, TRI)
+ << " interference: Cascade " << Cascade << '\n');
+ for (const unsigned *AliasI = TRI->getOverlaps(BestPhys); *AliasI; ++AliasI) {
+ LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
+ assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
+ for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
+ LiveInterval *Intf = Q.interferingVRegs()[i];
+ unassign(*Intf, VRM->getPhys(Intf->reg));
+ assert(ExtraRegInfo[Intf->reg].Cascade < Cascade &&
+ "Cannot decrease cascade number, illegal eviction");
+ ExtraRegInfo[Intf->reg].Cascade = Cascade;
+ ++NumEvicted;
+ NewVRegs.push_back(Intf);
+ }
+ }
+ return BestPhys;
}
// Region Splitting
//===----------------------------------------------------------------------===//
-/// calcLiveBlockInfo - Fill the LiveBlocks array with information about blocks
-/// where VirtReg is live.
-/// The SpillConstraints array is minimally initialized with MBB->getNumber().
-void RAGreedy::calcLiveBlockInfo(LiveInterval &VirtReg) {
- LiveBlocks.clear();
- SpillConstraints.clear();
-
- assert(!VirtReg.empty() && "Cannot allocate an empty interval");
- LiveInterval::const_iterator LVI = VirtReg.begin();
- LiveInterval::const_iterator LVE = VirtReg.end();
-
- SmallVectorImpl<SlotIndex>::const_iterator UseI, UseE;
- UseI = SA->UseSlots.begin();
- UseE = SA->UseSlots.end();
+/// addSplitConstraints - Fill out the SplitConstraints vector based on the
+/// interference pattern in Physreg and its aliases. Add the constraints to
+/// SpillPlacement and return the static cost of this split in Cost, assuming
+/// that all preferences in SplitConstraints are met.
+/// Return false if there are no bundles with positive bias.
+bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf,
+ float &Cost) {
+ ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
- // Loop over basic blocks where VirtReg is live.
- MachineFunction::iterator MFI = Indexes->getMBBFromIndex(LVI->start);
- for (;;) {
- // Block constraints depend on the interference pattern.
- // Just allocate them here, don't compute anything.
- SpillPlacement::BlockConstraint BC;
- BC.Number = MFI->getNumber();
- SpillConstraints.push_back(BC);
-
- BlockInfo BI;
- BI.MBB = MFI;
- SlotIndex Start, Stop;
- tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
+ // Reset interference dependent info.
+ SplitConstraints.resize(UseBlocks.size());
+ float StaticCost = 0;
+ for (unsigned i = 0; i != UseBlocks.size(); ++i) {
+ const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
+ SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
+
+ BC.Number = BI.MBB->getNumber();
+ Intf.moveToBlock(BC.Number);
+ BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
+ BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
+
+ if (!Intf.hasInterference())
+ continue;
- // The last split point is the latest possible insertion point that dominates
- // all successor blocks. If interference reaches LastSplitPoint, it is not
- // possible to insert a split or reload that makes VirtReg live in the
- // outgoing bundle.
- MachineBasicBlock::iterator LSP = LIS->getLastSplitPoint(VirtReg, BI.MBB);
- if (LSP == BI.MBB->end())
- BI.LastSplitPoint = Stop;
- else
- BI.LastSplitPoint = Indexes->getInstructionIndex(LSP);
-
- // LVI is the first live segment overlapping MBB.
- BI.LiveIn = LVI->start <= Start;
- if (!BI.LiveIn)
- BI.Def = LVI->start;
-
- // Find the first and last uses in the block.
- BI.Uses = SA->hasUses(MFI);
- if (BI.Uses && UseI != UseE) {
- BI.FirstUse = *UseI;
- assert(BI.FirstUse >= Start);
- do ++UseI;
- while (UseI != UseE && *UseI < Stop);
- BI.LastUse = UseI[-1];
- assert(BI.LastUse < Stop);
+ // Number of spill code instructions to insert.
+ unsigned Ins = 0;
+
+ // Interference for the live-in value.
+ if (BI.LiveIn) {
+ if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number))
+ BC.Entry = SpillPlacement::MustSpill, ++Ins;
+ else if (Intf.first() < BI.FirstUse)
+ BC.Entry = SpillPlacement::PrefSpill, ++Ins;
+ else if (Intf.first() < BI.LastUse)
+ ++Ins;
}
- // Look for gaps in the live range.
- bool hasGap = false;
- BI.LiveOut = true;
- while (LVI->end < Stop) {
- SlotIndex LastStop = LVI->end;
- if (++LVI == LVE || LVI->start >= Stop) {
- BI.Kill = LastStop;
- BI.LiveOut = false;
- break;
- }
- if (LastStop < LVI->start) {
- hasGap = true;
- BI.Kill = LastStop;
- BI.Def = LVI->start;
- }
+ // Interference for the live-out value.
+ if (BI.LiveOut) {
+ if (Intf.last() >= SA->getLastSplitPoint(BC.Number))
+ BC.Exit = SpillPlacement::MustSpill, ++Ins;
+ else if (Intf.last() > BI.LastUse)
+ BC.Exit = SpillPlacement::PrefSpill, ++Ins;
+ else if (Intf.last() > BI.FirstUse)
+ ++Ins;
}
- // Don't set LiveThrough when the block has a gap.
- BI.LiveThrough = !hasGap && BI.LiveIn && BI.LiveOut;
- LiveBlocks.push_back(BI);
-
- // LVI is now at LVE or LVI->end >= Stop.
- if (LVI == LVE)
- break;
-
- // Live segment ends exactly at Stop. Move to the next segment.
- if (LVI->end == Stop && ++LVI == LVE)
- break;
-
- // Pick the next basic block.
- if (LVI->start < Stop)
- ++MFI;
- else
- MFI = Indexes->getMBBFromIndex(LVI->start);
+ // Accumulate the total frequency of inserted spill code.
+ if (Ins)
+ StaticCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
}
+ Cost = StaticCost;
+
+ // Add constraints for use-blocks. Note that these are the only constraints
+ // that may add a positive bias, it is downhill from here.
+ SpillPlacer->addConstraints(SplitConstraints);
+ return SpillPlacer->scanActiveBundles();
}
-/// calcInterferenceInfo - Compute per-block outgoing and ingoing constraints
-/// when considering interference from PhysReg. Also compute an optimistic local
-/// cost of this interference pattern.
-///
-/// The final cost of a split is the local cost + global cost of preferences
-/// broken by SpillPlacement.
-///
-float RAGreedy::calcInterferenceInfo(LiveInterval &VirtReg, unsigned PhysReg) {
- // Reset interference dependent info.
- for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
- BlockInfo &BI = LiveBlocks[i];
- SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
- BC.Entry = (BI.Uses && BI.LiveIn) ?
- SpillPlacement::PrefReg : SpillPlacement::DontCare;
- BC.Exit = (BI.Uses && BI.LiveOut) ?
- SpillPlacement::PrefReg : SpillPlacement::DontCare;
- BI.OverlapEntry = BI.OverlapExit = false;
- }
- // Add interference info from each PhysReg alias.
- for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
- if (!query(VirtReg, *AI).checkInterference())
- continue;
- LiveIntervalUnion::SegmentIter IntI =
- PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
- if (!IntI.valid())
+/// addThroughConstraints - Add constraints and links to SpillPlacer from the
+/// live-through blocks in Blocks.
+void RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf,
+ ArrayRef<unsigned> Blocks) {
+ const unsigned GroupSize = 8;
+ SpillPlacement::BlockConstraint BCS[GroupSize];
+ unsigned TBS[GroupSize];
+ unsigned B = 0, T = 0;
+
+ for (unsigned i = 0; i != Blocks.size(); ++i) {
+ unsigned Number = Blocks[i];
+ Intf.moveToBlock(Number);
+
+ if (!Intf.hasInterference()) {
+ assert(T < GroupSize && "Array overflow");
+ TBS[T] = Number;
+ if (++T == GroupSize) {
+ SpillPlacer->addLinks(ArrayRef<unsigned>(TBS, T));
+ T = 0;
+ }
continue;
+ }
- for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
- BlockInfo &BI = LiveBlocks[i];
- SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
- SlotIndex Start, Stop;
- tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
-
- // Skip interference-free blocks.
- if (IntI.start() >= Stop)
- continue;
-
- // Handle transparent blocks with interference separately.
- // Transparent blocks never incur any fixed cost.
- if (BI.LiveThrough && !BI.Uses) {
- // Check if interference is live-in - force spill.
- if (BC.Entry != SpillPlacement::MustSpill) {
- BC.Entry = SpillPlacement::PrefSpill;
- IntI.advanceTo(Start);
- if (IntI.valid() && IntI.start() <= Start)
- BC.Entry = SpillPlacement::MustSpill;
- }
-
- // Check if interference is live-out - force spill.
- if (BC.Exit != SpillPlacement::MustSpill) {
- BC.Exit = SpillPlacement::PrefSpill;
- // Any interference overlapping [LastSplitPoint;Stop) forces a spill.
- IntI.advanceTo(BI.LastSplitPoint.getPrevSlot());
- if (IntI.valid() && IntI.start() < Stop)
- BC.Exit = SpillPlacement::MustSpill;
- }
-
- // Nothing more to do for this transparent block.
- if (!IntI.valid())
- break;
- continue;
- }
+ assert(B < GroupSize && "Array overflow");
+ BCS[B].Number = Number;
- // Now we only have blocks with uses left.
- // Check if the interference overlaps the uses.
- assert(BI.Uses && "Non-transparent block without any uses");
+ // Interference for the live-in value.
+ if (Intf.first() <= Indexes->getMBBStartIdx(Number))
+ BCS[B].Entry = SpillPlacement::MustSpill;
+ else
+ BCS[B].Entry = SpillPlacement::PrefSpill;
- // Check interference on entry.
- if (BI.LiveIn && BC.Entry != SpillPlacement::MustSpill) {
- IntI.advanceTo(Start);
- if (!IntI.valid())
- break;
+ // Interference for the live-out value.
+ if (Intf.last() >= SA->getLastSplitPoint(Number))
+ BCS[B].Exit = SpillPlacement::MustSpill;
+ else
+ BCS[B].Exit = SpillPlacement::PrefSpill;
- // Interference is live-in - force spill.
- if (IntI.start() <= Start)
- BC.Entry = SpillPlacement::MustSpill;
- // Not live in, but before the first use.
- else if (IntI.start() < BI.FirstUse)
- BC.Entry = SpillPlacement::PrefSpill;
- }
+ if (++B == GroupSize) {
+ ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
+ SpillPlacer->addConstraints(Array);
+ B = 0;
+ }
+ }
- // Does interference overlap the uses in the entry segment
- // [FirstUse;Kill)?
- if (BI.LiveIn && !BI.OverlapEntry) {
- IntI.advanceTo(BI.FirstUse);
- if (!IntI.valid())
- break;
- // A live-through interval has no kill.
- // Check [FirstUse;LastUse) instead.
- if (IntI.start() < (BI.LiveThrough ? BI.LastUse : BI.Kill))
- BI.OverlapEntry = true;
- }
+ ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
+ SpillPlacer->addConstraints(Array);
+ SpillPlacer->addLinks(ArrayRef<unsigned>(TBS, T));
+}
- // Does interference overlap the uses in the exit segment [Def;LastUse)?
- if (BI.LiveOut && !BI.LiveThrough && !BI.OverlapExit) {
- IntI.advanceTo(BI.Def);
- if (!IntI.valid())
- break;
- if (IntI.start() < BI.LastUse)
- BI.OverlapExit = true;
- }
+void RAGreedy::growRegion(GlobalSplitCandidate &Cand,
+ InterferenceCache::Cursor Intf) {
+ // Keep track of through blocks that have not been added to SpillPlacer.
+ BitVector Todo = SA->getThroughBlocks();
+ SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks;
+ unsigned AddedTo = 0;
+#ifndef NDEBUG
+ unsigned Visited = 0;
+#endif
- // Check interference on exit.
- if (BI.LiveOut && BC.Exit != SpillPlacement::MustSpill) {
- // Check interference between LastUse and Stop.
- if (BC.Exit != SpillPlacement::PrefSpill) {
- IntI.advanceTo(BI.LastUse);
- if (!IntI.valid())
- break;
- if (IntI.start() < Stop)
- BC.Exit = SpillPlacement::PrefSpill;
- }
- // Is the interference overlapping the last split point?
- IntI.advanceTo(BI.LastSplitPoint.getPrevSlot());
- if (!IntI.valid())
- break;
- if (IntI.start() < Stop)
- BC.Exit = SpillPlacement::MustSpill;
+ for (;;) {
+ ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive();
+ // Find new through blocks in the periphery of PrefRegBundles.
+ for (int i = 0, e = NewBundles.size(); i != e; ++i) {
+ unsigned Bundle = NewBundles[i];
+ // Look at all blocks connected to Bundle in the full graph.
+ ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle);
+ for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end();
+ I != E; ++I) {
+ unsigned Block = *I;
+ if (!Todo.test(Block))
+ continue;
+ Todo.reset(Block);
+ // This is a new through block. Add it to SpillPlacer later.
+ ActiveBlocks.push_back(Block);
+#ifndef NDEBUG
+ ++Visited;
+#endif
}
}
- }
-
- // Accumulate a local cost of this interference pattern.
- float LocalCost = 0;
- for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
- BlockInfo &BI = LiveBlocks[i];
- if (!BI.Uses)
- continue;
- SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
- unsigned Inserts = 0;
-
- // Do we need spill code for the entry segment?
- if (BI.LiveIn)
- Inserts += BI.OverlapEntry || BC.Entry != SpillPlacement::PrefReg;
+ // Any new blocks to add?
+ if (ActiveBlocks.size() == AddedTo)
+ break;
+ addThroughConstraints(Intf,
+ ArrayRef<unsigned>(ActiveBlocks).slice(AddedTo));
+ AddedTo = ActiveBlocks.size();
- // For the exit segment?
- if (BI.LiveOut)
- Inserts += BI.OverlapExit || BC.Exit != SpillPlacement::PrefReg;
+ // Perhaps iterating can enable more bundles?
+ SpillPlacer->iterate();
+ }
+ DEBUG(dbgs() << ", v=" << Visited);
+}
- // The local cost of spill code in this block is the block frequency times
- // the number of spill instructions inserted.
- if (Inserts)
- LocalCost += Inserts * SpillPlacer->getBlockFrequency(BI.MBB);
+/// calcSpillCost - Compute how expensive it would be to split the live range in
+/// 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];
+ unsigned Number = BI.MBB->getNumber();
+ // We normally only need one spill instruction - a load or a store.
+ 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);
+ }
}
- DEBUG(dbgs() << "Local cost of " << PrintReg(PhysReg, TRI) << " = "
- << LocalCost << '\n');
- return LocalCost;
+ return Cost;
}
/// calcGlobalSplitCost - Return the global split cost of following the split
/// pattern in LiveBundles. This cost should be added to the local cost of the
-/// interference pattern in SpillConstraints.
+/// interference pattern in SplitConstraints.
///
-float RAGreedy::calcGlobalSplitCost(const BitVector &LiveBundles) {
+float RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand,
+ InterferenceCache::Cursor Intf) {
float GlobalCost = 0;
- for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
- SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
- unsigned Inserts = 0;
- // Broken entry preference?
- Inserts += LiveBundles[Bundles->getBundle(BC.Number, 0)] !=
- (BC.Entry == SpillPlacement::PrefReg);
- // Broken exit preference?
- Inserts += LiveBundles[Bundles->getBundle(BC.Number, 1)] !=
- (BC.Exit == SpillPlacement::PrefReg);
- if (Inserts)
- GlobalCost += Inserts * SpillPlacer->getBlockFrequency(LiveBlocks[i].MBB);
+ const BitVector &LiveBundles = Cand.LiveBundles;
+ ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
+ for (unsigned i = 0; i != UseBlocks.size(); ++i) {
+ const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
+ SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
+ bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, 0)];
+ bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)];
+ unsigned Ins = 0;
+
+ if (BI.LiveIn)
+ Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
+ if (BI.LiveOut)
+ Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
+ if (Ins)
+ GlobalCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
+ }
+
+ for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) {
+ unsigned Number = Cand.ActiveBlocks[i];
+ bool RegIn = LiveBundles[Bundles->getBundle(Number, 0)];
+ bool RegOut = LiveBundles[Bundles->getBundle(Number, 1)];
+ if (!RegIn && !RegOut)
+ continue;
+ if (RegIn && RegOut) {
+ // We need double spill code if this block has interference.
+ Intf.moveToBlock(Number);
+ if (Intf.hasInterference())
+ GlobalCost += 2*SpillPlacer->getBlockFrequency(Number);
+ continue;
+ }
+ // live-in / stack-out or stack-in live-out.
+ GlobalCost += SpillPlacer->getBlockFrequency(Number);
}
- DEBUG(dbgs() << "Global cost = " << GlobalCost << '\n');
return GlobalCost;
}
/// avoiding interference. The 'stack' interval is the complement constructed by
/// SplitEditor. It will contain the rest.
///
-void RAGreedy::splitAroundRegion(LiveInterval &VirtReg, unsigned PhysReg,
- const BitVector &LiveBundles,
+void RAGreedy::splitAroundRegion(LiveInterval &VirtReg,
+ GlobalSplitCandidate &Cand,
SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ const BitVector &LiveBundles = Cand.LiveBundles;
+
DEBUG({
- dbgs() << "Splitting around region for " << PrintReg(PhysReg, TRI)
+ dbgs() << "Splitting around region for " << PrintReg(Cand.PhysReg, TRI)
<< " with bundles";
for (int i = LiveBundles.find_first(); i>=0; i = LiveBundles.find_next(i))
dbgs() << " EB#" << i;
dbgs() << ".\n";
});
- // First compute interference ranges in the live blocks.
- typedef std::pair<SlotIndex, SlotIndex> IndexPair;
- SmallVector<IndexPair, 8> InterferenceRanges;
- InterferenceRanges.resize(LiveBlocks.size());
- for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
- if (!query(VirtReg, *AI).checkInterference())
- continue;
- LiveIntervalUnion::SegmentIter IntI =
- PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
- if (!IntI.valid())
- continue;
- for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
- const BlockInfo &BI = LiveBlocks[i];
- IndexPair &IP = InterferenceRanges[i];
- SlotIndex Start, Stop;
- tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
- // Skip interference-free blocks.
- if (IntI.start() >= Stop)
- continue;
-
- // First interference in block.
- if (BI.LiveIn) {
- IntI.advanceTo(Start);
- if (!IntI.valid())
- break;
- if (IntI.start() >= Stop)
- continue;
- if (!IP.first.isValid() || IntI.start() < IP.first)
- IP.first = IntI.start();
- }
-
- // Last interference in block.
- if (BI.LiveOut) {
- IntI.advanceTo(Stop);
- if (!IntI.valid() || IntI.start() >= Stop)
- --IntI;
- if (IntI.stop() <= Start)
- continue;
- if (!IP.second.isValid() || IntI.stop() > IP.second)
- IP.second = IntI.stop();
- }
- }
- }
-
- SmallVector<LiveInterval*, 4> SpillRegs;
- LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
- SplitEditor SE(*SA, *LIS, *VRM, *DomTree, LREdit);
+ InterferenceCache::Cursor Intf(IntfCache, Cand.PhysReg);
+ LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+ SE->reset(LREdit);
// Create the main cross-block interval.
- SE.openIntv();
-
- // First add all defs that are live out of a block.
- for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
- BlockInfo &BI = LiveBlocks[i];
- bool RegIn = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
- bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
-
- // Should the register be live out?
- if (!BI.LiveOut || !RegOut)
+ const unsigned MainIntv = SE->openIntv();
+
+ // First handle all the blocks with uses.
+ ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
+ for (unsigned i = 0; i != UseBlocks.size(); ++i) {
+ const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
+ bool RegIn = BI.LiveIn &&
+ LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
+ bool RegOut = BI.LiveOut &&
+ LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
+
+ // Create separate intervals for isolated blocks with multiple uses.
+ //
+ // |---o---o---| Enter and leave on the stack.
+ // ____-----____ Create local interval for uses.
+ //
+ // | o---o---| Defined in block, leave on stack.
+ // -----____ Create local interval for uses.
+ //
+ // |---o---x | Enter on stack, killed in block.
+ // ____----- Create local interval for uses.
+ //
+ if (!RegIn && !RegOut) {
+ DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n");
+ if (!BI.isOneInstr()) {
+ SE->splitSingleBlock(BI);
+ SE->selectIntv(MainIntv);
+ }
continue;
+ }
- IndexPair &IP = InterferenceRanges[i];
SlotIndex Start, Stop;
tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
-
- DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " -> EB#"
- << Bundles->getBundle(BI.MBB->getNumber(), 1)
- << " intf [" << IP.first << ';' << IP.second << ')');
+ Intf.moveToBlock(BI.MBB->getNumber());
+ DEBUG(dbgs() << "EB#" << Bundles->getBundle(BI.MBB->getNumber(), 0)
+ << (BI.LiveIn ? (RegIn ? " => " : " -> ") : " ")
+ << "BB#" << BI.MBB->getNumber()
+ << (BI.LiveOut ? (RegOut ? " => " : " -> ") : " ")
+ << " EB#" << Bundles->getBundle(BI.MBB->getNumber(), 1)
+ << " [" << Start << ';'
+ << SA->getLastSplitPoint(BI.MBB->getNumber()) << '-' << Stop
+ << ") uses [" << BI.FirstUse << ';' << BI.LastUse
+ << ") intf [" << Intf.first() << ';' << Intf.last() << ')');
// The interference interval should either be invalid or overlap MBB.
- assert((!IP.first.isValid() || IP.first < Stop) && "Bad interference");
- assert((!IP.second.isValid() || IP.second > Start) && "Bad interference");
-
- // Check interference leaving the block.
- if (!IP.second.isValid()) {
- // Block is interference-free.
- DEBUG(dbgs() << ", no interference");
- if (!BI.Uses) {
- assert(BI.LiveThrough && "No uses, but not live through block?");
- // Block is live-through without interference.
- DEBUG(dbgs() << ", no uses"
- << (RegIn ? ", live-through.\n" : ", stack in.\n"));
- if (!RegIn)
- SE.enterIntvAtEnd(*BI.MBB);
- continue;
- }
- if (!BI.LiveThrough) {
- DEBUG(dbgs() << ", not live-through.\n");
- SE.useIntv(SE.enterIntvBefore(BI.Def), Stop);
- continue;
- }
- if (!RegIn) {
- // Block is live-through, but entry bundle is on the stack.
- // Reload just before the first use.
- DEBUG(dbgs() << ", not live-in, enter before first use.\n");
- SE.useIntv(SE.enterIntvBefore(BI.FirstUse), Stop);
- continue;
+ assert((!Intf.hasInterference() || Intf.first() < Stop)
+ && "Bad interference");
+ assert((!Intf.hasInterference() || Intf.last() > Start)
+ && "Bad interference");
+
+ // We are now ready to decide where to split in the current block. There
+ // are many variables guiding the decision:
+ //
+ // - RegIn / RegOut: The global splitting algorithm's decisions for our
+ // ingoing and outgoing bundles.
+ //
+ // - BI.BlockIn / BI.BlockOut: Is the live range live-in and/or live-out
+ // from this block.
+ //
+ // - Intf.hasInterference(): Is there interference in this block.
+ //
+ // - Intf.first() / Inft.last(): The range of interference.
+ //
+ // The live range should be split such that MainIntv is live-in when RegIn
+ // is set, and live-out when RegOut is set. MainIntv should never overlap
+ // the interference, and the stack interval should never have more than one
+ // use per block.
+
+ // No splits can be inserted after LastSplitPoint, overlap instead.
+ SlotIndex LastSplitPoint = Stop;
+ if (BI.LiveOut)
+ LastSplitPoint = SA->getLastSplitPoint(BI.MBB->getNumber());
+
+ // At this point, we know that either RegIn or RegOut is set. We dealt with
+ // the all-stack case above.
+
+ // Blocks without interference are relatively easy.
+ if (!Intf.hasInterference()) {
+ DEBUG(dbgs() << ", no interference.\n");
+ SE->selectIntv(MainIntv);
+ // The easiest case has MainIntv live through.
+ //
+ // |---o---o---| Live-in, live-out.
+ // ============= Use MainIntv everywhere.
+ //
+ SlotIndex From = Start, To = Stop;
+
+ // Block entry. Reload before the first use if MainIntv is not live-in.
+ //
+ // |---o-- Enter on stack.
+ // ____=== Reload before first use.
+ //
+ // | o-- Defined in block.
+ // === Use MainIntv from def.
+ //
+ if (!RegIn)
+ From = SE->enterIntvBefore(BI.FirstUse);
+
+ // Block exit. Handle cases where MainIntv is not live-out.
+ if (!BI.LiveOut)
+ //
+ // --x | Killed in block.
+ // === Use MainIntv up to kill.
+ //
+ To = SE->leaveIntvAfter(BI.LastUse);
+ else if (!RegOut) {
+ //
+ // --o---| Live-out on stack.
+ // ===____ Use MainIntv up to last use, switch to stack.
+ //
+ // -----o| Live-out on stack, last use after last split point.
+ // ====== Extend MainIntv to last use, overlapping.
+ // \____ Copy to stack interval before last split point.
+ //
+ if (BI.LastUse < LastSplitPoint)
+ To = SE->leaveIntvAfter(BI.LastUse);
+ else {
+ // The last use is after the last split point, it is probably an
+ // indirect branch.
+ To = SE->leaveIntvBefore(LastSplitPoint);
+ // Run a double interval from the split to the last use. This makes
+ // it possible to spill the complement without affecting the indirect
+ // branch.
+ SE->overlapIntv(To, BI.LastUse);
+ }
}
- DEBUG(dbgs() << ", live-through.\n");
+
+ // Paint in MainIntv liveness for this block.
+ SE->useIntv(From, To);
continue;
}
- // Block has interference.
- DEBUG(dbgs() << ", interference to " << IP.second);
-
- if (!BI.LiveThrough && IP.second <= BI.Def) {
- // The interference doesn't reach the outgoing segment.
- DEBUG(dbgs() << " doesn't affect def from " << BI.Def << '\n');
- SE.useIntv(BI.Def, Stop);
+ // We are now looking at a block with interference, and we know that either
+ // RegIn or RegOut is set.
+ assert(Intf.hasInterference() && (RegIn || RegOut) && "Bad invariant");
+
+ // If the live range is not live through the block, it is possible that the
+ // interference doesn't even overlap. Deal with those cases first. Since
+ // no copy instructions are required, we can tolerate interference starting
+ // or ending at the same instruction that kills or defines our live range.
+
+ // Live-in, killed before interference.
+ //
+ // ~~~ Interference after kill.
+ // |---o---x | Killed in block.
+ // ========= Use MainIntv everywhere.
+ //
+ if (RegIn && !BI.LiveOut && BI.LastUse <= Intf.first()) {
+ DEBUG(dbgs() << ", live-in, killed before interference.\n");
+ SE->selectIntv(MainIntv);
+ SlotIndex To = SE->leaveIntvAfter(BI.LastUse);
+ SE->useIntv(Start, To);
continue;
}
+ // Live-out, defined after interference.
+ //
+ // ~~~ Interference before def.
+ // | o---o---| Defined in block.
+ // ========= Use MainIntv everywhere.
+ //
+ if (RegOut && !BI.LiveIn && BI.FirstUse >= Intf.last()) {
+ DEBUG(dbgs() << ", live-out, defined after interference.\n");
+ SE->selectIntv(MainIntv);
+ SlotIndex From = SE->enterIntvBefore(BI.FirstUse);
+ SE->useIntv(From, Stop);
+ continue;
+ }
- if (!BI.Uses) {
- // No uses in block, avoid interference by reloading as late as possible.
- DEBUG(dbgs() << ", no uses.\n");
- SlotIndex SegStart = SE.enterIntvAtEnd(*BI.MBB);
- assert(SegStart >= IP.second && "Couldn't avoid interference");
+ // The interference is now known to overlap the live range, but it may
+ // still be easy to avoid if all the interference is on one side of the
+ // uses, and we enter or leave on the stack.
+
+ // Live-out on stack, interference after last use.
+ //
+ // ~~~ Interference after last use.
+ // |---o---o---| Live-out on stack.
+ // =========____ Leave MainIntv after last use.
+ //
+ // ~ Interference after last use.
+ // |---o---o--o| Live-out on stack, late last use.
+ // ============ Copy to stack after LSP, overlap MainIntv.
+ // \_____ Stack interval is live-out.
+ //
+ if (!RegOut && Intf.first() > BI.LastUse.getBoundaryIndex()) {
+ assert(RegIn && "Stack-in, stack-out should already be handled");
+ if (BI.LastUse < LastSplitPoint) {
+ DEBUG(dbgs() << ", live-in, stack-out, interference after last use.\n");
+ SE->selectIntv(MainIntv);
+ SlotIndex To = SE->leaveIntvAfter(BI.LastUse);
+ assert(To <= Intf.first() && "Expected to avoid interference");
+ SE->useIntv(Start, To);
+ } else {
+ DEBUG(dbgs() << ", live-in, stack-out, avoid last split point\n");
+ SE->selectIntv(MainIntv);
+ SlotIndex To = SE->leaveIntvBefore(LastSplitPoint);
+ assert(To <= Intf.first() && "Expected to avoid interference");
+ SE->overlapIntv(To, BI.LastUse);
+ SE->useIntv(Start, To);
+ }
continue;
}
- if (IP.second.getBoundaryIndex() < BI.LastUse &&
- IP.second.getBoundaryIndex() <= BI.LastSplitPoint) {
- // There are interference-free uses at the end of the block.
- // Find the first use that can get the live-out register.
- SmallVectorImpl<SlotIndex>::const_iterator UI =
- std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
- IP.second.getBoundaryIndex());
- assert(UI != SA->UseSlots.end() && "Couldn't find last use");
- SlotIndex Use = *UI;
- DEBUG(dbgs() << ", free use at " << Use << ".\n");
- assert(Use <= BI.LastUse && "Couldn't find last use");
- SlotIndex SegStart = SE.enterIntvBefore(Use);
- assert(SegStart >= IP.second && "Couldn't avoid interference");
- assert(SegStart < BI.LastSplitPoint && "Impossible split point");
- SE.useIntv(SegStart, Stop);
+ // Live-in on stack, interference before first use.
+ //
+ // ~~~ Interference before first use.
+ // |---o---o---| Live-in on stack.
+ // ____========= Enter MainIntv before first use.
+ //
+ if (!RegIn && Intf.last() < BI.FirstUse.getBaseIndex()) {
+ assert(RegOut && "Stack-in, stack-out should already be handled");
+ DEBUG(dbgs() << ", stack-in, interference before first use.\n");
+ SE->selectIntv(MainIntv);
+ SlotIndex From = SE->enterIntvBefore(BI.FirstUse);
+ assert(From >= Intf.last() && "Expected to avoid interference");
+ SE->useIntv(From, Stop);
continue;
}
- // Interference is after the last use.
- DEBUG(dbgs() << " after last use.\n");
- SlotIndex SegStart = SE.enterIntvAtEnd(*BI.MBB);
- assert(SegStart >= IP.second && "Couldn't avoid interference");
+ // The interference is overlapping somewhere we wanted to use MainIntv. That
+ // means we need to create a local interval that can be allocated a
+ // different register.
+ unsigned LocalIntv = SE->openIntv();
+ DEBUG(dbgs() << ", creating local interval " << LocalIntv << ".\n");
+
+ // We may be creating copies directly between MainIntv and LocalIntv,
+ // bypassing the stack interval. When we do that, we should never use the
+ // leaveIntv* methods as they define values in the stack interval. By
+ // starting from the end of the block and working our way backwards, we can
+ // get by with only enterIntv* methods.
+ //
+ // When selecting split points, we generally try to maximize the stack
+ // interval as long at it contains no uses, maximize the main interval as
+ // long as it doesn't overlap interference, and minimize the local interval
+ // that we don't know how to allocate yet.
+
+ // Handle the block exit, set Pos to the first handled slot.
+ SlotIndex Pos = BI.LastUse;
+ if (RegOut) {
+ assert(Intf.last() < LastSplitPoint && "Cannot be live-out in register");
+ // Create a snippet of MainIntv that is live-out.
+ //
+ // ~~~ Interference overlapping uses.
+ // --o---| Live-out in MainIntv.
+ // ----=== Switch from LocalIntv to MainIntv after interference.
+ //
+ SE->selectIntv(MainIntv);
+ Pos = SE->enterIntvAfter(Intf.last());
+ assert(Pos >= Intf.last() && "Expected to avoid interference");
+ SE->useIntv(Pos, Stop);
+ SE->selectIntv(LocalIntv);
+ } else if (BI.LiveOut) {
+ if (BI.LastUse < LastSplitPoint) {
+ // Live-out on the stack.
+ //
+ // ~~~ Interference overlapping uses.
+ // --o---| Live-out on stack.
+ // ---____ Switch from LocalIntv to stack after last use.
+ //
+ Pos = SE->leaveIntvAfter(BI.LastUse);
+ } else {
+ // Live-out on the stack, last use after last split point.
+ //
+ // ~~~ Interference overlapping uses.
+ // --o--o| Live-out on stack, late use.
+ // ------ Copy to stack before LSP, overlap LocalIntv.
+ // \__
+ //
+ Pos = SE->leaveIntvBefore(LastSplitPoint);
+ // We need to overlap LocalIntv so it can reach LastUse.
+ SE->overlapIntv(Pos, BI.LastUse);
+ }
+ }
+
+ // When not live-out, leave Pos at LastUse. We have handled everything from
+ // Pos to Stop. Find the starting point for LocalIntv.
+ assert(SE->currentIntv() == LocalIntv && "Expecting local interval");
+
+ if (RegIn) {
+ assert(Start < Intf.first() && "Cannot be live-in with interference");
+ // Live-in in MainIntv, only use LocalIntv for interference.
+ //
+ // ~~~ Interference overlapping uses.
+ // |---o-- Live-in in MainIntv.
+ // ====--- Switch to LocalIntv before interference.
+ //
+ SlotIndex Switch = SE->enterIntvBefore(std::min(Pos, Intf.first()));
+ assert(Switch <= Intf.first() && "Expected to avoid interference");
+ SE->useIntv(Switch, Pos);
+ SE->selectIntv(MainIntv);
+ SE->useIntv(Start, Switch);
+ } else {
+ // Live-in on stack, enter LocalIntv before first use.
+ //
+ // ~~~ Interference overlapping uses.
+ // |---o-- Live-in in MainIntv.
+ // ____--- Reload to LocalIntv before interference.
+ //
+ // Defined in block.
+ //
+ // ~~~ Interference overlapping uses.
+ // | o-- Defined in block.
+ // --- Begin LocalIntv at first use.
+ //
+ SlotIndex Switch = SE->enterIntvBefore(std::min(Pos, BI.FirstUse));
+ SE->useIntv(Switch, Pos);
+ }
}
- // Now all defs leading to live bundles are handled, do everything else.
- for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
- BlockInfo &BI = LiveBlocks[i];
- bool RegIn = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
- bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
+ // Handle live-through blocks.
+ SE->selectIntv(MainIntv);
+ for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) {
+ unsigned Number = Cand.ActiveBlocks[i];
+ bool RegIn = LiveBundles[Bundles->getBundle(Number, 0)];
+ bool RegOut = LiveBundles[Bundles->getBundle(Number, 1)];
+ DEBUG(dbgs() << "Live through BB#" << Number << '\n');
+ if (RegIn && RegOut) {
+ Intf.moveToBlock(Number);
+ if (!Intf.hasInterference()) {
+ SE->useIntv(Indexes->getMBBStartIdx(Number),
+ Indexes->getMBBEndIdx(Number));
+ continue;
+ }
+ }
+ MachineBasicBlock *MBB = MF->getBlockNumbered(Number);
+ if (RegIn)
+ SE->leaveIntvAtTop(*MBB);
+ if (RegOut)
+ SE->enterIntvAtEnd(*MBB);
+ }
+
+ ++NumGlobalSplits;
+
+ SmallVector<unsigned, 8> IntvMap;
+ SE->finish(&IntvMap);
+ DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
+
+ ExtraRegInfo.resize(MRI->getNumVirtRegs());
+ unsigned OrigBlocks = SA->getNumLiveBlocks();
+
+ // Sort out the new intervals created by splitting. We get four kinds:
+ // - Remainder intervals should not be split again.
+ // - Candidate intervals can be assigned to Cand.PhysReg.
+ // - Block-local splits are candidates for local splitting.
+ // - DCE leftovers should go back on the queue.
+ for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
+ LiveInterval &Reg = *LREdit.get(i);
- // Is the register live-in?
- if (!BI.LiveIn || !RegIn)
+ // Ignore old intervals from DCE.
+ if (getStage(Reg) != RS_New)
continue;
- // We have an incoming register. Check for interference.
- IndexPair &IP = InterferenceRanges[i];
- SlotIndex Start, Stop;
- tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
+ // Remainder interval. Don't try splitting again, spill if it doesn't
+ // allocate.
+ if (IntvMap[i] == 0) {
+ setStage(Reg, RS_Global);
+ continue;
+ }
- DEBUG(dbgs() << "EB#" << Bundles->getBundle(BI.MBB->getNumber(), 0)
- << " -> BB#" << BI.MBB->getNumber());
-
- // Check interference entering the block.
- if (!IP.first.isValid()) {
- // Block is interference-free.
- DEBUG(dbgs() << ", no interference");
- if (!BI.Uses) {
- assert(BI.LiveThrough && "No uses, but not live through block?");
- // Block is live-through without interference.
- if (RegOut) {
- DEBUG(dbgs() << ", no uses, live-through.\n");
- SE.useIntv(Start, Stop);
- } else {
- DEBUG(dbgs() << ", no uses, stack-out.\n");
- SE.leaveIntvAtTop(*BI.MBB);
- }
- continue;
- }
- if (!BI.LiveThrough) {
- DEBUG(dbgs() << ", killed in block.\n");
- SE.useIntv(Start, SE.leaveIntvAfter(BI.Kill));
- continue;
+ // Main interval. Allow repeated splitting as long as the number of live
+ // blocks is strictly decreasing.
+ if (IntvMap[i] == MainIntv) {
+ if (SA->countLiveBlocks(&Reg) >= OrigBlocks) {
+ DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks
+ << " blocks as original.\n");
+ // Don't allow repeated splitting as a safe guard against looping.
+ setStage(Reg, RS_Global);
}
- if (!RegOut) {
- // Block is live-through, but exit bundle is on the stack.
- // Spill immediately after the last use.
- if (BI.LastUse < BI.LastSplitPoint) {
- DEBUG(dbgs() << ", uses, stack-out.\n");
- SE.useIntv(Start, SE.leaveIntvAfter(BI.LastUse));
- continue;
- }
- // The last use is after the last split point, it is probably an
- // indirect jump.
- DEBUG(dbgs() << ", uses at " << BI.LastUse << " after split point "
- << BI.LastSplitPoint << ", stack-out.\n");
- SlotIndex SegEnd;
- if (BI.LastSplitPoint == Start)
- SegEnd = SE.leaveIntvAtTop(*BI.MBB);
- else {
- MachineBasicBlock::iterator I =
- LIS->getInstructionFromIndex(BI.LastSplitPoint);
- do assert(I != BI.MBB->begin() && "Expected instruction");
- while ((--I)->isDebugValue());
- SegEnd = SE.leaveIntvAfter(LIS->getInstructionIndex(I));
- }
- SE.useIntv(Start, SegEnd);
- // Run a double interval from the split to the last use.
- // This makes it possible to spill the complement without affecting the
- // indirect branch.
- SE.overlapIntv(SegEnd, BI.LastUse);
- continue;
- }
- // Register is live-through.
- DEBUG(dbgs() << ", uses, live-through.\n");
- SE.useIntv(Start, Stop);
continue;
}
- // Block has interference.
- DEBUG(dbgs() << ", interference from " << IP.first);
+ // Other intervals are treated as new. This includes local intervals created
+ // for blocks with multiple uses, and anything created by DCE.
+ }
+
+ if (VerifyEnabled)
+ MF->verify(this, "After splitting live range around region");
+}
+
+unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
+ SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ float BestCost = Hysteresis * calcSpillCost();
+ DEBUG(dbgs() << "Cost of isolating all blocks = " << BestCost << '\n');
+ const unsigned NoCand = ~0u;
+ unsigned BestCand = NoCand;
- if (!BI.LiveThrough && IP.first >= BI.Kill) {
- // The interference doesn't reach the outgoing segment.
- DEBUG(dbgs() << " doesn't affect kill at " << BI.Kill << '\n');
- SE.useIntv(Start, BI.Kill);
+ Order.rewind();
+ for (unsigned Cand = 0; unsigned PhysReg = Order.next(); ++Cand) {
+ if (GlobalCand.size() <= Cand)
+ GlobalCand.resize(Cand+1);
+ GlobalCand[Cand].reset(PhysReg);
+
+ SpillPlacer->prepare(GlobalCand[Cand].LiveBundles);
+ float Cost;
+ InterferenceCache::Cursor Intf(IntfCache, PhysReg);
+ if (!addSplitConstraints(Intf, Cost)) {
+ DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tno positive bundles\n");
continue;
}
-
- if (!BI.Uses) {
- // No uses in block, avoid interference by spilling as soon as possible.
- DEBUG(dbgs() << ", no uses.\n");
- SlotIndex SegEnd = SE.leaveIntvAtTop(*BI.MBB);
- assert(SegEnd <= IP.first && "Couldn't avoid interference");
+ DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = " << Cost);
+ if (Cost >= BestCost) {
+ DEBUG({
+ if (BestCand == NoCand)
+ dbgs() << " worse than no bundles\n";
+ else
+ dbgs() << " worse than "
+ << PrintReg(GlobalCand[BestCand].PhysReg, TRI) << '\n';
+ });
continue;
}
- if (IP.first.getBaseIndex() > BI.FirstUse) {
- // There are interference-free uses at the beginning of the block.
- // Find the last use that can get the register.
- SmallVectorImpl<SlotIndex>::const_iterator UI =
- std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
- IP.first.getBaseIndex());
- assert(UI != SA->UseSlots.begin() && "Couldn't find first use");
- SlotIndex Use = (--UI)->getBoundaryIndex();
- DEBUG(dbgs() << ", free use at " << *UI << ".\n");
- SlotIndex SegEnd = SE.leaveIntvAfter(Use);
- assert(SegEnd <= IP.first && "Couldn't avoid interference");
- SE.useIntv(Start, SegEnd);
+ growRegion(GlobalCand[Cand], Intf);
+
+ SpillPlacer->finish();
+
+ // No live bundles, defer to splitSingleBlocks().
+ if (!GlobalCand[Cand].LiveBundles.any()) {
+ DEBUG(dbgs() << " no bundles.\n");
continue;
}
- // Interference is before the first use.
- DEBUG(dbgs() << " before first use.\n");
- SlotIndex SegEnd = SE.leaveIntvAtTop(*BI.MBB);
- assert(SegEnd <= IP.first && "Couldn't avoid interference");
+ Cost += calcGlobalSplitCost(GlobalCand[Cand], Intf);
+ DEBUG({
+ dbgs() << ", total = " << Cost << " with bundles";
+ for (int i = GlobalCand[Cand].LiveBundles.find_first(); i>=0;
+ i = GlobalCand[Cand].LiveBundles.find_next(i))
+ dbgs() << " EB#" << i;
+ dbgs() << ".\n";
+ });
+ if (Cost < BestCost) {
+ BestCand = Cand;
+ BestCost = Hysteresis * Cost; // Prevent rounding effects.
+ }
}
- SE.closeIntv();
+ if (BestCand == NoCand)
+ return 0;
- // FIXME: Should we be more aggressive about splitting the stack region into
- // per-block segments? The current approach allows the stack region to
- // separate into connected components. Some components may be allocatable.
- SE.finish();
+ splitAroundRegion(VirtReg, GlobalCand[BestCand], NewVRegs);
+ return 0;
+}
- if (VerifyEnabled) {
- MF->verify(this, "After splitting live range around region");
-#ifndef NDEBUG
- // Make sure that at least one of the new intervals can allocate to PhysReg.
- // That was the whole point of splitting the live range.
- bool found = false;
- for (LiveRangeEdit::iterator I = LREdit.begin(), E = LREdit.end(); I != E;
- ++I)
- if (!checkUncachedInterference(**I, PhysReg)) {
- found = true;
+//===----------------------------------------------------------------------===//
+// Local Splitting
+//===----------------------------------------------------------------------===//
+
+
+/// calcGapWeights - Compute the maximum spill weight that needs to be evicted
+/// in order to use PhysReg between two entries in SA->UseSlots.
+///
+/// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
+///
+void RAGreedy::calcGapWeights(unsigned PhysReg,
+ 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;
+ 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;
+
+ GapWeight.assign(NumGaps, 0.0f);
+
+ // Add interference from each overlapping register.
+ for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
+ if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
+ .checkInterference())
+ continue;
+
+ // We know that VirtReg is a continuous interval from FirstUse to LastUse,
+ // 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);
+ for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
+ // Skip the gaps before IntI.
+ while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
+ if (++Gap == NumGaps)
+ break;
+ if (Gap == NumGaps)
break;
+
+ // Update the gaps covered by IntI.
+ const float weight = IntI.value()->weight;
+ for (; Gap != NumGaps; ++Gap) {
+ GapWeight[Gap] = std::max(GapWeight[Gap], weight);
+ if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
+ break;
}
- assert(found && "No allocatable intervals after pointless splitting");
-#endif
+ if (Gap == NumGaps)
+ break;
+ }
}
}
-unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
- SmallVectorImpl<LiveInterval*> &NewVRegs) {
- calcLiveBlockInfo(VirtReg);
- BitVector LiveBundles, BestBundles;
- float BestCost = 0;
- unsigned BestReg = 0;
+/// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
+/// basic block.
+///
+unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
+ SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
+ const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
+
+ // Note that it is possible to have an interval that is live-in or live-out
+ // 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.
+
+ const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
+ 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() << '\n';
+ });
+
+ // Since we allow local split results to be split again, there is a risk of
+ // creating infinite loops. It is tempting to require that the new live
+ // ranges have less instructions than the original. That would guarantee
+ // convergence, but it is too strict. A live range with 3 instructions can be
+ // split 2+3 (including the COPY), and we want to allow that.
+ //
+ // Instead we use these rules:
+ //
+ // 1. Allow any split for ranges with getStage() < RS_Local. (Except for the
+ // noop split, of course).
+ // 2. Require progress be made for ranges with getStage() >= RS_Local. All
+ // the new ranges must have fewer instructions than before the split.
+ // 3. New ranges with the same number of instructions are marked RS_Local,
+ // smaller ranges are marked RS_New.
+ //
+ // These rules allow a 3 -> 2+3 split once, which we need. They also prevent
+ // excessive splitting and infinite loops.
+ //
+ bool ProgressRequired = getStage(VirtReg) >= RS_Local;
+
+ // Best split candidate.
+ unsigned BestBefore = NumGaps;
+ unsigned BestAfter = 0;
+ float BestDiff = 0;
+
+ const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB->getNumber());
+ SmallVector<float, 8> GapWeight;
+
Order.rewind();
while (unsigned PhysReg = Order.next()) {
- float Cost = calcInterferenceInfo(VirtReg, PhysReg);
- if (BestReg && Cost >= BestCost)
- continue;
+ // Keep track of the largest spill weight that would need to be evicted in
+ // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
+ calcGapWeights(PhysReg, GapWeight);
- SpillPlacer->placeSpills(SpillConstraints, LiveBundles);
- // No live bundles, defer to splitSingleBlocks().
- if (!LiveBundles.any())
- continue;
+ // Try to find the best sequence of gaps to close.
+ // The new spill weight must be larger than any gap interference.
+
+ // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
+ unsigned SplitBefore = 0, SplitAfter = 1;
+
+ // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
+ // It is the spill weight that needs to be evicted.
+ float MaxGap = GapWeight[0];
+
+ for (;;) {
+ // Live before/after split?
+ const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
+ const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
- Cost += calcGlobalSplitCost(LiveBundles);
- if (!BestReg || Cost < BestCost) {
- BestReg = PhysReg;
- BestCost = Cost;
- BestBundles.swap(LiveBundles);
+ DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
+ << Uses[SplitBefore] << '-' << Uses[SplitAfter]
+ << " i=" << MaxGap);
+
+ // Stop before the interval gets so big we wouldn't be making progress.
+ if (!LiveBefore && !LiveAfter) {
+ DEBUG(dbgs() << " all\n");
+ break;
+ }
+ // Should the interval be extended or shrunk?
+ bool Shrink = true;
+
+ // How many gaps would the new range have?
+ unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter;
+
+ // Legally, without causing looping?
+ bool Legal = !ProgressRequired || NewGaps < NumGaps;
+
+ if (Legal && MaxGap < HUGE_VALF) {
+ // Estimate the new spill weight. Each instruction reads or writes the
+ // register. Conservatively assume there are no read-modify-write
+ // instructions.
+ //
+ // Try to guess the size of the new interval.
+ const float EstWeight = normalizeSpillWeight(blockFreq * (NewGaps + 1),
+ Uses[SplitBefore].distance(Uses[SplitAfter]) +
+ (LiveBefore + LiveAfter)*SlotIndex::InstrDist);
+ // Would this split be possible to allocate?
+ // Never allocate all gaps, we wouldn't be making progress.
+ DEBUG(dbgs() << " w=" << EstWeight);
+ if (EstWeight * Hysteresis >= MaxGap) {
+ Shrink = false;
+ float Diff = EstWeight - MaxGap;
+ if (Diff > BestDiff) {
+ DEBUG(dbgs() << " (best)");
+ BestDiff = Hysteresis * Diff;
+ BestBefore = SplitBefore;
+ BestAfter = SplitAfter;
+ }
+ }
+ }
+
+ // Try to shrink.
+ if (Shrink) {
+ if (++SplitBefore < SplitAfter) {
+ DEBUG(dbgs() << " shrink\n");
+ // Recompute the max when necessary.
+ if (GapWeight[SplitBefore - 1] >= MaxGap) {
+ MaxGap = GapWeight[SplitBefore];
+ for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
+ MaxGap = std::max(MaxGap, GapWeight[i]);
+ }
+ continue;
+ }
+ MaxGap = 0;
+ }
+
+ // Try to extend the interval.
+ if (SplitAfter >= NumGaps) {
+ DEBUG(dbgs() << " end\n");
+ break;
+ }
+
+ DEBUG(dbgs() << " extend\n");
+ MaxGap = std::max(MaxGap, GapWeight[SplitAfter++]);
}
}
- if (!BestReg)
+ // Didn't find any candidates?
+ if (BestBefore == NumGaps)
return 0;
- splitAroundRegion(VirtReg, BestReg, BestBundles, NewVRegs);
+ DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
+ << '-' << Uses[BestAfter] << ", " << BestDiff
+ << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
+
+ LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+ SE->reset(LREdit);
+
+ SE->openIntv();
+ SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
+ SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]);
+ SE->useIntv(SegStart, SegStop);
+ SmallVector<unsigned, 8> IntvMap;
+ SE->finish(&IntvMap);
+ DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
+
+ // If the new range has the same number of instructions as before, mark it as
+ // RS_Local so the next split will be forced to make progress. Otherwise,
+ // leave the new intervals as RS_New so they can compete.
+ bool LiveBefore = BestBefore != 0 || BI.LiveIn;
+ bool LiveAfter = BestAfter != NumGaps || BI.LiveOut;
+ unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter;
+ if (NewGaps >= NumGaps) {
+ DEBUG(dbgs() << "Tagging non-progress ranges: ");
+ assert(!ProgressRequired && "Didn't make progress when it was required.");
+ for (unsigned i = 0, e = IntvMap.size(); i != e; ++i)
+ if (IntvMap[i] == 1) {
+ setStage(*LREdit.get(i), RS_Local);
+ DEBUG(dbgs() << PrintReg(LREdit.get(i)->reg));
+ }
+ DEBUG(dbgs() << '\n');
+ }
+ ++NumLocalSplits;
+
return 0;
}
-
//===----------------------------------------------------------------------===//
// Live Range Splitting
//===----------------------------------------------------------------------===//
/// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
SmallVectorImpl<LiveInterval*>&NewVRegs) {
- NamedRegionTimer T("Splitter", TimerGroupName, TimePassesIsEnabled);
- SA->analyze(&VirtReg);
+ // Local intervals are handled separately.
+ if (LIS->intervalIsInOneMBB(VirtReg)) {
+ NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
+ SA->analyze(&VirtReg);
+ return tryLocalSplit(VirtReg, Order, NewVRegs);
+ }
+
+ NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
- // Don't attempt splitting on local intervals for now. TBD.
- if (LIS->intervalIsInOneMBB(VirtReg))
+ // Don't iterate global splitting.
+ // Move straight to spilling if this range was produced by a global split.
+ if (getStage(VirtReg) >= RS_Global)
return 0;
+ SA->analyze(&VirtReg);
+
+ // FIXME: SplitAnalysis may repair broken live ranges coming from the
+ // coalescer. That may cause the range to become allocatable which means that
+ // tryRegionSplit won't be making progress. This check should be replaced with
+ // an assertion when the coalescer is fixed.
+ if (SA->didRepairRange()) {
+ // VirtReg has changed, so all cached queries are invalid.
+ invalidateVirtRegs();
+ if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
+ return PhysReg;
+ }
+
// First try to split around a region spanning multiple blocks.
unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
if (PhysReg || !NewVRegs.empty())
// Then isolate blocks with multiple uses.
SplitAnalysis::BlockPtrSet Blocks;
if (SA->getMultiUseBlocks(Blocks)) {
- SmallVector<LiveInterval*, 4> SpillRegs;
- LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
- SplitEditor(*SA, *LIS, *VRM, *DomTree, LREdit).splitSingleBlocks(Blocks);
+ LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+ SE->reset(LREdit);
+ SE->splitSingleBlocks(Blocks);
+ setStage(NewVRegs.begin(), NewVRegs.end(), RS_Global);
if (VerifyEnabled)
MF->verify(this, "After splitting live range around basic blocks");
}
}
-//===----------------------------------------------------------------------===//
-// Spilling
-//===----------------------------------------------------------------------===//
-
-/// calcInterferenceWeight - Calculate the combined spill weight of
-/// interferences when assigning VirtReg to PhysReg.
-float RAGreedy::calcInterferenceWeight(LiveInterval &VirtReg, unsigned PhysReg){
- float Sum = 0;
- for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
- LiveIntervalUnion::Query &Q = query(VirtReg, *AI);
- Q.collectInterferingVRegs();
- if (Q.seenUnspillableVReg())
- return HUGE_VALF;
- for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i)
- Sum += Q.interferingVRegs()[i]->weight;
- }
- return Sum;
-}
-
-/// trySpillInterferences - Try to spill interfering registers instead of the
-/// current one. Only do it if the accumulated spill weight is smaller than the
-/// current spill weight.
-unsigned RAGreedy::trySpillInterferences(LiveInterval &VirtReg,
- AllocationOrder &Order,
- SmallVectorImpl<LiveInterval*> &NewVRegs) {
- NamedRegionTimer T("Spill Interference", TimerGroupName, TimePassesIsEnabled);
- unsigned BestPhys = 0;
- float BestWeight = 0;
-
- Order.rewind();
- while (unsigned PhysReg = Order.next()) {
- float Weight = calcInterferenceWeight(VirtReg, PhysReg);
- if (Weight == HUGE_VALF || Weight >= VirtReg.weight)
- continue;
- if (!BestPhys || Weight < BestWeight)
- BestPhys = PhysReg, BestWeight = Weight;
- }
-
- // No candidates found.
- if (!BestPhys)
- return 0;
-
- // Collect all interfering registers.
- SmallVector<LiveInterval*, 8> Spills;
- for (const unsigned *AI = TRI->getOverlaps(BestPhys); *AI; ++AI) {
- LiveIntervalUnion::Query &Q = query(VirtReg, *AI);
- Spills.append(Q.interferingVRegs().begin(), Q.interferingVRegs().end());
- for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
- LiveInterval *VReg = Q.interferingVRegs()[i];
- PhysReg2LiveUnion[*AI].extract(*VReg);
- VRM->clearVirt(VReg->reg);
- }
- }
-
- // Spill them all.
- DEBUG(dbgs() << "spilling " << Spills.size() << " interferences with weight "
- << BestWeight << '\n');
- for (unsigned i = 0, e = Spills.size(); i != e; ++i)
- spiller().spill(Spills[i], NewVRegs, Spills);
- return BestPhys;
-}
-
-
//===----------------------------------------------------------------------===//
// Main Entry Point
//===----------------------------------------------------------------------===//
unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
SmallVectorImpl<LiveInterval*> &NewVRegs) {
// First try assigning a free register.
- AllocationOrder Order(VirtReg.reg, *VRM, ReservedRegs);
- while (unsigned PhysReg = Order.next()) {
- if (!checkPhysRegInterference(VirtReg, PhysReg))
- return PhysReg;
- }
-
- // Try to reassign interferences.
- if (unsigned PhysReg = tryReassign(VirtReg, Order))
+ AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
+ if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
return PhysReg;
+ LiveRangeStage Stage = getStage(VirtReg);
+ DEBUG(dbgs() << StageName[Stage]
+ << " Cascade " << ExtraRegInfo[VirtReg.reg].Cascade << '\n');
+
+ // Try to evict a less worthy live range, but only for ranges from the primary
+ // queue. The RS_Second ranges already failed to do this, and they should not
+ // get a second chance until they have been split.
+ if (Stage != RS_Second)
+ if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
+ return PhysReg;
+
assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
+ // The first time we see a live range, don't try to split or spill.
+ // Wait until the second time, when all smaller ranges have been allocated.
+ // This gives a better picture of the interference to split around.
+ if (Stage == RS_First) {
+ setStage(VirtReg, RS_Second);
+ DEBUG(dbgs() << "wait for second round\n");
+ NewVRegs.push_back(&VirtReg);
+ return 0;
+ }
+
+ // If we couldn't allocate a register from spilling, there is probably some
+ // invalid inline assembly. The base class wil report it.
+ if (Stage >= RS_Spill)
+ return ~0u;
+
// Try splitting VirtReg or interferences.
unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
if (PhysReg || !NewVRegs.empty())
return PhysReg;
- // Try to spill another interfering reg with less spill weight.
- PhysReg = trySpillInterferences(VirtReg, Order, NewVRegs);
- if (PhysReg)
- return PhysReg;
-
// Finally spill VirtReg itself.
NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
- SmallVector<LiveInterval*, 1> pendingSpills;
- spiller().spill(&VirtReg, NewVRegs, pendingSpills);
+ LiveRangeEdit LRE(VirtReg, NewVRegs, this);
+ spiller().spill(LRE);
+ setStage(NewVRegs.begin(), NewVRegs.end(), RS_Spill);
+
+ if (VerifyEnabled)
+ MF->verify(this, "After spilling");
// The live virtual register requesting allocation was spilled, so tell
// the caller not to allocate anything during this round.
RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
Indexes = &getAnalysis<SlotIndexes>();
DomTree = &getAnalysis<MachineDominatorTree>();
- ReservedRegs = TRI->getReservedRegs(*MF);
SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
Loops = &getAnalysis<MachineLoopInfo>();
LoopRanges = &getAnalysis<MachineLoopRanges>();
Bundles = &getAnalysis<EdgeBundles>();
SpillPlacer = &getAnalysis<SpillPlacement>();
+ DebugVars = &getAnalysis<LiveDebugVariables>();
- SA.reset(new SplitAnalysis(*MF, *LIS, *Loops));
+ SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
+ SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
+ ExtraRegInfo.clear();
+ ExtraRegInfo.resize(MRI->getNumVirtRegs());
+ NextCascade = 1;
+ IntfCache.init(MF, &PhysReg2LiveUnion[0], Indexes, TRI);
allocatePhysRegs();
addMBBLiveIns(MF);
+ LIS->addKillFlags();
// Run rewriter
{
NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
- std::auto_ptr<VirtRegRewriter> rewriter(createVirtRegRewriter());
- rewriter->runOnMachineFunction(*MF, *VRM, LIS);
+ VRM->rewrite(Indexes);
}
+ // Write out new DBG_VALUE instructions.
+ DebugVars->emitDebugValues(VRM);
+
// The pass output is in VirtRegMap. Release all the transient data.
releaseMemory();