//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "regalloc"
+#include "llvm/CodeGen/Passes.h"
#include "AllocationOrder.h"
#include "InterferenceCache.h"
#include "LiveDebugVariables.h"
-#include "LiveRangeEdit.h"
#include "RegAllocBase.h"
-#include "Spiller.h"
#include "SpillPlacement.h"
+#include "Spiller.h"
#include "SplitKit.h"
-#include "VirtRegMap.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/CalcSpillWeights.h"
#include "llvm/CodeGen/EdgeBundles.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/LiveRangeEdit.h"
+#include "llvm/CodeGen/LiveRegMatrix.h"
#include "llvm/CodeGen/LiveStackAnalysis.h"
+#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/RegAllocRegistry.h"
-#include "llvm/Target/TargetOptions.h"
+#include "llvm/CodeGen/RegisterClassInfo.h"
+#include "llvm/CodeGen/VirtRegMap.h"
+#include "llvm/PassAnalysisSupport.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Timer.h"
-
+#include "llvm/Support/raw_ostream.h"
#include <queue>
using namespace llvm;
STATISTIC(NumLocalSplits, "Number of split local live ranges");
STATISTIC(NumEvicted, "Number of interferences evicted");
+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 cl::opt<unsigned>
+LastChanceRecoloringMaxDepth("lcr-max-depth", cl::Hidden,
+ cl::desc("Last chance recoloring max depth"),
+ cl::init(5));
+
+static cl::opt<unsigned> LastChanceRecoloringMaxInterference(
+ "lcr-max-interf", cl::Hidden,
+ cl::desc("Last chance recoloring maximum number of considered"
+ " interference at a time"),
+ cl::init(8));
+
static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
createGreedyRegisterAllocator);
class RAGreedy : public MachineFunctionPass,
public RegAllocBase,
private LiveRangeEdit::Delegate {
+ // Convenient shortcuts.
+ typedef std::priority_queue<std::pair<unsigned, unsigned> > PQueue;
+ typedef SmallPtrSet<LiveInterval *, 4> SmallLISet;
+ typedef SmallSet<unsigned, 16> SmallVirtRegSet;
// context
MachineFunction *MF;
+ // Shortcuts to some useful interface.
+ const TargetInstrInfo *TII;
+ const TargetRegisterInfo *TRI;
+ RegisterClassInfo RCI;
+
// analyses
SlotIndexes *Indexes;
- LiveStacks *LS;
+ MachineBlockFrequencyInfo *MBFI;
MachineDominatorTree *DomTree;
MachineLoopInfo *Loops;
EdgeBundles *Bundles;
LiveDebugVariables *DebugVars;
// state
- std::auto_ptr<Spiller> SpillerInstance;
- std::priority_queue<std::pair<unsigned, unsigned> > Queue;
+ OwningPtr<Spiller> SpillerInstance;
+ PQueue Queue;
unsigned NextCascade;
// Live ranges pass through a number of stages as we try to allocate them.
// 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.
+ /// Newly created live range that has never been queued.
+ RS_New,
+
+ /// Only attempt assignment and eviction. Then requeue as RS_Split.
+ RS_Assign,
+
+ /// Attempt live range splitting if assignment is impossible.
+ RS_Split,
+
+ /// Attempt more aggressive live range splitting that is guaranteed to make
+ /// progress. This is used for split products that may not be making
+ /// progress.
+ RS_Split2,
+
+ /// Live range will be spilled. No more splitting will be attempted.
+ RS_Spill,
+
+ /// There is nothing more we can do to this live range. Abort compilation
+ /// if it can't be assigned.
+ RS_Done
};
+#ifndef NDEBUG
static const char *const StageName[];
+#endif
// RegInfo - Keep additional information about each live range.
struct RegInfo {
void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
ExtraRegInfo.resize(MRI->getNumVirtRegs());
for (;Begin != End; ++Begin) {
- unsigned Reg = (*Begin)->reg;
+ unsigned Reg = *Begin;
if (ExtraRegInfo[Reg].Stage == RS_New)
ExtraRegInfo[Reg].Stage = NewStage;
}
unsigned BrokenHints; ///< Total number of broken hints.
float MaxWeight; ///< Maximum spill weight evicted.
- EvictionCost(unsigned B = 0) : BrokenHints(B), MaxWeight(0) {}
+ EvictionCost(): BrokenHints(0), MaxWeight(0) {}
+
+ bool isMax() const { return BrokenHints == ~0u; }
+
+ void setMax() { BrokenHints = ~0u; }
+
+ void setBrokenHints(unsigned NHints) { BrokenHints = NHints; }
bool operator<(const EvictionCost &O) const {
if (BrokenHints != O.BrokenHints)
};
// splitting state.
- std::auto_ptr<SplitAnalysis> SA;
- std::auto_ptr<SplitEditor> SE;
+ OwningPtr<SplitAnalysis> SA;
+ OwningPtr<SplitEditor> SE;
/// Cached per-block interference maps
InterferenceCache IntfCache;
/// Global live range splitting candidate info.
struct GlobalSplitCandidate {
+ // Register intended for assignment, or 0.
unsigned PhysReg;
+
+ // SplitKit interval index for this candidate.
+ unsigned IntvIdx;
+
+ // Interference for PhysReg.
InterferenceCache::Cursor Intf;
+
+ // Bundles where this candidate should be live.
BitVector LiveBundles;
SmallVector<unsigned, 8> ActiveBlocks;
void reset(InterferenceCache &Cache, unsigned Reg) {
PhysReg = Reg;
+ IntvIdx = 0;
Intf.setPhysReg(Cache, Reg);
LiveBundles.clear();
ActiveBlocks.clear();
}
+
+ // Set B[i] = C for every live bundle where B[i] was NoCand.
+ unsigned getBundles(SmallVectorImpl<unsigned> &B, unsigned C) {
+ unsigned Count = 0;
+ for (int i = LiveBundles.find_first(); i >= 0;
+ i = LiveBundles.find_next(i))
+ if (B[i] == NoCand) {
+ B[i] = C;
+ Count++;
+ }
+ return Count;
+ }
};
- /// Candidate info for for each PhysReg in AllocationOrder.
+ /// Candidate info for each PhysReg in AllocationOrder.
/// This vector never shrinks, but grows to the size of the largest register
/// class.
SmallVector<GlobalSplitCandidate, 32> GlobalCand;
+ enum LLVM_ENUM_INT_TYPE(unsigned) { NoCand = ~0u };
+
+ /// Candidate map. Each edge bundle is assigned to a GlobalCand entry, or to
+ /// NoCand which indicates the stack interval.
+ SmallVector<unsigned, 32> BundleCand;
+
public:
RAGreedy();
virtual void enqueue(LiveInterval *LI);
virtual LiveInterval *dequeue();
virtual unsigned selectOrSplit(LiveInterval&,
- SmallVectorImpl<LiveInterval*>&);
+ SmallVectorImpl<unsigned>&);
/// Perform register allocation.
virtual bool runOnMachineFunction(MachineFunction &mf);
static char ID;
private:
- void LRE_WillEraseInstruction(MachineInstr*);
+ unsigned selectOrSplitImpl(LiveInterval &, SmallVectorImpl<unsigned> &,
+ SmallVirtRegSet &, unsigned = 0);
+
bool LRE_CanEraseVirtReg(unsigned);
void LRE_WillShrinkVirtReg(unsigned);
void LRE_DidCloneVirtReg(unsigned, unsigned);
+ void enqueue(PQueue &CurQueue, LiveInterval *LI);
+ LiveInterval *dequeue(PQueue &CurQueue);
- float calcSpillCost();
- bool addSplitConstraints(InterferenceCache::Cursor, float&);
+ BlockFrequency calcSpillCost();
+ bool addSplitConstraints(InterferenceCache::Cursor, BlockFrequency&);
void addThroughConstraints(InterferenceCache::Cursor, ArrayRef<unsigned>);
void growRegion(GlobalSplitCandidate &Cand);
- float calcGlobalSplitCost(GlobalSplitCandidate&);
+ BlockFrequency calcGlobalSplitCost(GlobalSplitCandidate&);
bool calcCompactRegion(GlobalSplitCandidate&);
- void splitAroundRegion(LiveInterval&, GlobalSplitCandidate&,
- SmallVectorImpl<LiveInterval*>&);
+ void splitAroundRegion(LiveRangeEdit&, ArrayRef<unsigned>);
void calcGapWeights(unsigned, SmallVectorImpl<float>&);
+ unsigned canReassign(LiveInterval &VirtReg, unsigned PhysReg);
bool shouldEvict(LiveInterval &A, bool, LiveInterval &B, bool);
bool canEvictInterference(LiveInterval&, unsigned, bool, EvictionCost&);
void evictInterference(LiveInterval&, unsigned,
- SmallVectorImpl<LiveInterval*>&);
+ SmallVectorImpl<unsigned>&);
+ bool mayRecolorAllInterferences(unsigned PhysReg, LiveInterval &VirtReg,
+ SmallLISet &RecoloringCandidates,
+ const SmallVirtRegSet &FixedRegisters);
unsigned tryAssign(LiveInterval&, AllocationOrder&,
- SmallVectorImpl<LiveInterval*>&);
+ SmallVectorImpl<unsigned>&);
unsigned tryEvict(LiveInterval&, AllocationOrder&,
- SmallVectorImpl<LiveInterval*>&, unsigned = ~0u);
+ SmallVectorImpl<unsigned>&, unsigned = ~0u);
unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
- SmallVectorImpl<LiveInterval*>&);
+ SmallVectorImpl<unsigned>&);
+ unsigned tryBlockSplit(LiveInterval&, AllocationOrder&,
+ SmallVectorImpl<unsigned>&);
+ unsigned tryInstructionSplit(LiveInterval&, AllocationOrder&,
+ SmallVectorImpl<unsigned>&);
unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
- SmallVectorImpl<LiveInterval*>&);
+ SmallVectorImpl<unsigned>&);
unsigned trySplit(LiveInterval&, AllocationOrder&,
- SmallVectorImpl<LiveInterval*>&);
+ SmallVectorImpl<unsigned>&);
+ unsigned tryLastChanceRecoloring(LiveInterval &, AllocationOrder &,
+ SmallVectorImpl<unsigned> &,
+ SmallVirtRegSet &, unsigned);
+ bool tryRecoloringCandidates(PQueue &, SmallVectorImpl<unsigned> &,
+ SmallVirtRegSet &, unsigned);
};
} // end anonymous namespace
#ifndef NDEBUG
const char *const RAGreedy::StageName[] = {
- "RS_New",
- "RS_First",
- "RS_Second",
- "RS_Global",
- "RS_Local",
- "RS_Spill"
+ "RS_New",
+ "RS_Assign",
+ "RS_Split",
+ "RS_Split2",
+ "RS_Spill",
+ "RS_Done"
};
#endif
// Hysteresis to use when comparing floats.
// This helps stabilize decisions based on float comparisons.
-const float Hysteresis = 0.98f;
+const float Hysteresis = (2007 / 2048.0f); // 0.97998046875
FunctionPass* llvm::createGreedyRegisterAllocator() {
initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
- initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
- initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
+ initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
initializeLiveStacksPass(*PassRegistry::getPassRegistry());
initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
+ initializeLiveRegMatrixPass(*PassRegistry::getPassRegistry());
initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
}
void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
+ AU.addRequired<MachineBlockFrequencyInfo>();
+ AU.addPreserved<MachineBlockFrequencyInfo>();
AU.addRequired<AliasAnalysis>();
AU.addPreserved<AliasAnalysis>();
AU.addRequired<LiveIntervals>();
+ AU.addPreserved<LiveIntervals>();
AU.addRequired<SlotIndexes>();
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>();
AU.addRequired<MachineDominatorTree>();
AU.addPreserved<MachineLoopInfo>();
AU.addRequired<VirtRegMap>();
AU.addPreserved<VirtRegMap>();
+ AU.addRequired<LiveRegMatrix>();
+ AU.addPreserved<LiveRegMatrix>();
AU.addRequired<EdgeBundles>();
AU.addRequired<SpillPlacement>();
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);
+ if (VRM->hasPhys(VirtReg)) {
+ Matrix->unassign(LIS->getInterval(VirtReg));
return true;
}
// Unassigned virtreg is probably in the priority queue.
}
void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
- unsigned PhysReg = VRM->getPhys(VirtReg);
- if (!PhysReg)
+ if (!VRM->hasPhys(VirtReg))
return;
// Register is assigned, put it back on the queue for reassignment.
LiveInterval &LI = LIS->getInterval(VirtReg);
- unassign(LI, PhysReg);
+ Matrix->unassign(LI);
enqueue(&LI);
}
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. Ensure that the new register gets the
+ // be split into connected components. The new components are much smaller
+ // than the original, so they should get a new chance at being assigned.
// same stage as the parent.
+ ExtraRegInfo[Old].Stage = RS_Assign;
ExtraRegInfo.grow(New);
ExtraRegInfo[New] = ExtraRegInfo[Old];
}
SpillerInstance.reset(0);
ExtraRegInfo.clear();
GlobalCand.clear();
- RegAllocBase::releaseMemory();
}
-void RAGreedy::enqueue(LiveInterval *LI) {
+void RAGreedy::enqueue(LiveInterval *LI) { enqueue(Queue, LI); }
+
+void RAGreedy::enqueue(PQueue &CurQueue, LiveInterval *LI) {
// Prioritize live ranges by size, assigning larger ranges first.
// The queue holds (size, reg) pairs.
const unsigned Size = LI->getSize();
ExtraRegInfo.grow(Reg);
if (ExtraRegInfo[Reg].Stage == RS_New)
- ExtraRegInfo[Reg].Stage = RS_First;
+ ExtraRegInfo[Reg].Stage = RS_Assign;
- if (ExtraRegInfo[Reg].Stage == RS_Second)
+ 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.
- 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;
+ // everything else has been allocated.
+ Prio = Size;
+ } else {
+ if (ExtraRegInfo[Reg].Stage == RS_Assign && !LI->empty() &&
+ LIS->intervalIsInOneMBB(*LI)) {
+ // Allocate original local ranges in linear instruction order. Since they
+ // are singly defined, this produces optimal coloring in the absence of
+ // global interference and other constraints.
+ if (!TRI->reverseLocalAssignment())
+ Prio = LI->beginIndex().getInstrDistance(Indexes->getLastIndex());
+ else {
+ // Allocating bottom up may allow many short LRGs to be assigned first
+ // to one of the cheap registers. This could be much faster for very
+ // large blocks on targets with many physical registers.
+ Prio = Indexes->getZeroIndex().getInstrDistance(LI->beginIndex());
+ }
+ }
+ else {
+ // Allocate global and split ranges in long->short order. Long ranges that
+ // don't fit should be spilled (or split) ASAP so they don't create
+ // interference. Mark a bit to prioritize global above local ranges.
+ Prio = (1u << 29) + Size;
+ }
+ // Mark a higher bit to prioritize global and local above RS_Split.
+ Prio |= (1u << 31);
// Boost ranges that have a physical register hint.
- if (TargetRegisterInfo::isPhysicalRegister(VRM->getRegAllocPref(Reg)))
+ if (VRM->hasKnownPreference(Reg))
Prio |= (1u << 30);
}
-
- Queue.push(std::make_pair(Prio, Reg));
+ // The virtual register number is a tie breaker for same-sized ranges.
+ // Give lower vreg numbers higher priority to assign them first.
+ CurQueue.push(std::make_pair(Prio, ~Reg));
}
-LiveInterval *RAGreedy::dequeue() {
- if (Queue.empty())
+LiveInterval *RAGreedy::dequeue() { return dequeue(Queue); }
+
+LiveInterval *RAGreedy::dequeue(PQueue &CurQueue) {
+ if (CurQueue.empty())
return 0;
- LiveInterval *LI = &LIS->getInterval(Queue.top().second);
- Queue.pop();
+ LiveInterval *LI = &LIS->getInterval(~CurQueue.top().second);
+ CurQueue.pop();
return LI;
}
/// tryAssign - Try to assign VirtReg to an available register.
unsigned RAGreedy::tryAssign(LiveInterval &VirtReg,
AllocationOrder &Order,
- SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ SmallVectorImpl<unsigned> &NewVRegs) {
Order.rewind();
unsigned PhysReg;
while ((PhysReg = Order.next()))
- if (!checkPhysRegInterference(VirtReg, PhysReg))
+ if (!Matrix->checkInterference(VirtReg, PhysReg))
break;
- if (!PhysReg || Order.isHint(PhysReg))
+ if (!PhysReg || Order.isHint())
return PhysReg;
// PhysReg is available, but there may be a better choice.
if (unsigned Hint = MRI->getSimpleHint(VirtReg.reg))
if (Order.isHint(Hint)) {
DEBUG(dbgs() << "missed hint " << PrintReg(Hint, TRI) << '\n');
- EvictionCost MaxCost(1);
+ EvictionCost MaxCost;
+ MaxCost.setBrokenHints(1);
if (canEvictInterference(VirtReg, Hint, true, MaxCost)) {
evictInterference(VirtReg, Hint, NewVRegs);
return Hint;
// Interference eviction
//===----------------------------------------------------------------------===//
+unsigned RAGreedy::canReassign(LiveInterval &VirtReg, unsigned PrevReg) {
+ AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
+ unsigned PhysReg;
+ while ((PhysReg = Order.next())) {
+ if (PhysReg == PrevReg)
+ continue;
+
+ MCRegUnitIterator Units(PhysReg, TRI);
+ for (; Units.isValid(); ++Units) {
+ // Instantiate a "subquery", not to be confused with the Queries array.
+ LiveIntervalUnion::Query subQ(&VirtReg, &Matrix->getLiveUnions()[*Units]);
+ if (subQ.checkInterference())
+ break;
+ }
+ // If no units have interference, break out with the current PhysReg.
+ if (!Units.isValid())
+ break;
+ }
+ if (PhysReg)
+ DEBUG(dbgs() << "can reassign: " << VirtReg << " from "
+ << PrintReg(PrevReg, TRI) << " to " << PrintReg(PhysReg, TRI)
+ << '\n');
+ return PhysReg;
+}
+
/// shouldEvict - determine if A should 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
/// @param BreaksHint True when B is already assigned to its preferred register.
bool RAGreedy::shouldEvict(LiveInterval &A, bool IsHint,
LiveInterval &B, bool BreaksHint) {
- bool CanSplit = getStage(B) <= RS_Second;
+ bool CanSplit = getStage(B) < RS_Spill;
// Be fairly aggressive about following hints as long as the evictee can be
// split.
if (CanSplit && IsHint && !BreaksHint)
return true;
- return A.weight > B.weight;
+ if (A.weight > B.weight) {
+ DEBUG(dbgs() << "should evict: " << B << " w= " << B.weight << '\n');
+ return true;
+ }
+ return false;
}
/// canEvictInterference - Return true if all interferences between VirtReg and
///
/// @param VirtReg Live range that is about to be assigned.
/// @param PhysReg Desired register for assignment.
-/// @prarm IsHint True when PhysReg is VirtReg's preferred register.
+/// @param IsHint True when PhysReg is VirtReg's preferred register.
/// @param MaxCost Only look for cheaper candidates and update with new cost
/// when returning true.
/// @returns True when interference can be evicted cheaper than MaxCost.
bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
bool IsHint, EvictionCost &MaxCost) {
+ // It is only possible to evict virtual register interference.
+ if (Matrix->checkInterference(VirtReg, PhysReg) > LiveRegMatrix::IK_VirtReg)
+ return false;
+
+ bool IsLocal = LIS->intervalIsInOneMBB(VirtReg);
+
// 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
Cascade = NextCascade;
EvictionCost Cost;
- for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
- LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
+ for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
+ LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
// If there is 10 or more interferences, chances are one is heavier.
if (Q.collectInterferingVRegs(10) >= 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;
+ assert(TargetRegisterInfo::isVirtualRegister(Intf->reg) &&
+ "Only expecting virtual register interference from query");
// Never evict spill products. They cannot split or spill.
- if (getStage(*Intf) == RS_Spill)
+ if (getStage(*Intf) == RS_Done)
return false;
// Once a live range becomes small enough, it is urgent that we find a
// register for it. This is indicated by an infinite spill weight. These
// urgent live ranges get to evict almost anything.
- bool Urgent = !VirtReg.isSpillable() && Intf->isSpillable();
+ //
+ // Also allow urgent evictions of unspillable ranges from a strictly
+ // larger allocation order.
+ bool Urgent = !VirtReg.isSpillable() &&
+ (Intf->isSpillable() ||
+ RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(VirtReg.reg)) <
+ RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(Intf->reg)));
// Only evict older cascades or live ranges without a cascade.
unsigned IntfCascade = ExtraRegInfo[Intf->reg].Cascade;
if (Cascade <= IntfCascade) {
// Abort if this would be too expensive.
if (!(Cost < MaxCost))
return false;
- // Finally, apply the eviction policy for non-urgent evictions.
- if (!Urgent && !shouldEvict(VirtReg, IsHint, *Intf, BreaksHint))
+ if (Urgent)
+ continue;
+ // Apply the eviction policy for non-urgent evictions.
+ if (!shouldEvict(VirtReg, IsHint, *Intf, BreaksHint))
+ return false;
+ // If !MaxCost.isMax(), then we're just looking for a cheap register.
+ // Evicting another local live range in this case could lead to suboptimal
+ // coloring.
+ if (!MaxCost.isMax() && IsLocal && LIS->intervalIsInOneMBB(*Intf) &&
+ !canReassign(*Intf, PhysReg)) {
return false;
+ }
}
}
MaxCost = Cost;
/// from being assigned to Physreg. This assumes that canEvictInterference
/// returned true.
void RAGreedy::evictInterference(LiveInterval &VirtReg, unsigned PhysReg,
- SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ SmallVectorImpl<unsigned> &NewVRegs) {
// 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.
DEBUG(dbgs() << "evicting " << PrintReg(PhysReg, TRI)
<< " interference: Cascade " << Cascade << '\n');
- for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
- LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
+
+ // Collect all interfering virtregs first.
+ SmallVector<LiveInterval*, 8> Intfs;
+ for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
+ LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
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 ||
- VirtReg.isSpillable() < Intf->isSpillable()) &&
- "Cannot decrease cascade number, illegal eviction");
- ExtraRegInfo[Intf->reg].Cascade = Cascade;
- ++NumEvicted;
- NewVRegs.push_back(Intf);
- }
+ ArrayRef<LiveInterval*> IVR = Q.interferingVRegs();
+ Intfs.append(IVR.begin(), IVR.end());
+ }
+
+ // Evict them second. This will invalidate the queries.
+ for (unsigned i = 0, e = Intfs.size(); i != e; ++i) {
+ LiveInterval *Intf = Intfs[i];
+ // The same VirtReg may be present in multiple RegUnits. Skip duplicates.
+ if (!VRM->hasPhys(Intf->reg))
+ continue;
+ Matrix->unassign(*Intf);
+ assert((ExtraRegInfo[Intf->reg].Cascade < Cascade ||
+ VirtReg.isSpillable() < Intf->isSpillable()) &&
+ "Cannot decrease cascade number, illegal eviction");
+ ExtraRegInfo[Intf->reg].Cascade = Cascade;
+ ++NumEvicted;
+ NewVRegs.push_back(Intf->reg);
}
}
/// @return Physreg to assign VirtReg, or 0.
unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
AllocationOrder &Order,
- SmallVectorImpl<LiveInterval*> &NewVRegs,
+ SmallVectorImpl<unsigned> &NewVRegs,
unsigned CostPerUseLimit) {
NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
// Keep track of the cheapest interference seen so far.
- EvictionCost BestCost(~0u);
+ EvictionCost BestCost;
+ BestCost.setMax();
unsigned BestPhys = 0;
+ unsigned OrderLimit = Order.getOrder().size();
// When we are just looking for a reduced cost per use, don't break any
// hints, and only evict smaller spill weights.
if (CostPerUseLimit < ~0u) {
BestCost.BrokenHints = 0;
BestCost.MaxWeight = VirtReg.weight;
+
+ // Check of any registers in RC are below CostPerUseLimit.
+ const TargetRegisterClass *RC = MRI->getRegClass(VirtReg.reg);
+ unsigned MinCost = RegClassInfo.getMinCost(RC);
+ if (MinCost >= CostPerUseLimit) {
+ DEBUG(dbgs() << RC->getName() << " minimum cost = " << MinCost
+ << ", no cheaper registers to be found.\n");
+ return 0;
+ }
+
+ // It is normal for register classes to have a long tail of registers with
+ // the same cost. We don't need to look at them if they're too expensive.
+ if (TRI->getCostPerUse(Order.getOrder().back()) >= CostPerUseLimit) {
+ OrderLimit = RegClassInfo.getLastCostChange(RC);
+ DEBUG(dbgs() << "Only trying the first " << OrderLimit << " regs.\n");
+ }
}
Order.rewind();
- while (unsigned PhysReg = Order.next()) {
+ while (unsigned PhysReg = Order.next(OrderLimit)) {
if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit)
continue;
// The first use of a callee-saved register in a function has cost 1.
BestPhys = PhysReg;
// Stop if the hint can be used.
- if (Order.isHint(PhysReg))
+ if (Order.isHint())
break;
}
/// 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) {
+ BlockFrequency &Cost) {
ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
// Reset interference dependent info.
SplitConstraints.resize(UseBlocks.size());
- float StaticCost = 0;
+ BlockFrequency StaticCost = 0;
for (unsigned i = 0; i != UseBlocks.size(); ++i) {
const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
Intf.moveToBlock(BC.Number);
BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
+ BC.ChangesValue = BI.FirstDef.isValid();
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;
}
// Accumulate the total frequency of inserted spill code.
- if (Ins)
- StaticCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
+ while (Ins--)
+ StaticCost += SpillPlacer->getBlockFrequency(BC.Number);
}
Cost = StaticCost;
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?
SpillPlacer->prepare(Cand.LiveBundles);
// The static split cost will be zero since Cand.Intf reports no interference.
- float Cost;
+ BlockFrequency Cost;
if (!addSplitConstraints(Cand.Intf, Cost)) {
DEBUG(dbgs() << ", none.\n");
return false;
/// 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();
+BlockFrequency RAGreedy::calcSpillCost() {
+ BlockFrequency Cost = 0;
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;
}
/// pattern in LiveBundles. This cost should be added to the local cost of the
/// interference pattern in SplitConstraints.
///
-float RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand) {
- float GlobalCost = 0;
+BlockFrequency RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand) {
+ BlockFrequency GlobalCost = 0;
const BitVector &LiveBundles = Cand.LiveBundles;
ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
for (unsigned i = 0; i != UseBlocks.size(); ++i) {
Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
if (BI.LiveOut)
Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
- if (Ins)
- GlobalCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
+ while (Ins--)
+ GlobalCost += SpillPlacer->getBlockFrequency(BC.Number);
}
for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) {
if (RegIn && RegOut) {
// We need double spill code if this block has interference.
Cand.Intf.moveToBlock(Number);
- if (Cand.Intf.hasInterference())
- GlobalCost += 2*SpillPlacer->getBlockFrequency(Number);
+ if (Cand.Intf.hasInterference()) {
+ GlobalCost += SpillPlacer->getBlockFrequency(Number);
+ GlobalCost += SpillPlacer->getBlockFrequency(Number);
+ }
continue;
}
// live-in / stack-out or stack-in live-out.
return GlobalCost;
}
-/// splitAroundRegion - Split VirtReg around the region determined by
-/// LiveBundles. Make an effort to avoid interference from PhysReg.
+/// splitAroundRegion - Split the current live range around the regions
+/// determined by BundleCand and GlobalCand.
///
-/// The 'register' interval is going to contain as many uses as possible while
-/// avoiding interference. The 'stack' interval is the complement constructed by
-/// SplitEditor. It will contain the rest.
+/// Before calling this function, GlobalCand and BundleCand must be initialized
+/// so each bundle is assigned to a valid candidate, or NoCand for the
+/// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor
+/// objects must be initialized for the current live range, and intervals
+/// created for the used candidates.
///
-void RAGreedy::splitAroundRegion(LiveInterval &VirtReg,
- GlobalSplitCandidate &Cand,
- SmallVectorImpl<LiveInterval*> &NewVRegs) {
- const BitVector &LiveBundles = Cand.LiveBundles;
-
- DEBUG({
- 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";
- });
-
- InterferenceCache::Cursor &Intf = Cand.Intf;
- LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
- SE->reset(LREdit);
-
- // Create the main cross-block interval.
- const unsigned MainIntv = SE->openIntv();
+/// @param LREdit The LiveRangeEdit object handling the current split.
+/// @param UsedCands List of used GlobalCand entries. Every BundleCand value
+/// must appear in this list.
+void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit,
+ ArrayRef<unsigned> UsedCands) {
+ // These are the intervals created for new global ranges. We may create more
+ // intervals for local ranges.
+ const unsigned NumGlobalIntvs = LREdit.size();
+ 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) {
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)];
+ unsigned Number = BI.MBB->getNumber();
+ unsigned IntvIn = 0, IntvOut = 0;
+ SlotIndex IntfIn, IntfOut;
+ if (BI.LiveIn) {
+ unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
+ if (CandIn != NoCand) {
+ GlobalSplitCandidate &Cand = GlobalCand[CandIn];
+ IntvIn = Cand.IntvIdx;
+ Cand.Intf.moveToBlock(Number);
+ IntfIn = Cand.Intf.first();
+ }
+ }
+ if (BI.LiveOut) {
+ unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
+ if (CandOut != NoCand) {
+ GlobalSplitCandidate &Cand = GlobalCand[CandOut];
+ IntvOut = Cand.IntvIdx;
+ Cand.Intf.moveToBlock(Number);
+ IntfOut = Cand.Intf.last();
+ }
+ }
// Create separate intervals for isolated blocks with multiple uses.
- if (!RegIn && !RegOut) {
+ if (!IntvIn && !IntvOut) {
DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n");
- if (!BI.isOneInstr()) {
+ if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
SE->splitSingleBlock(BI);
- SE->selectIntv(MainIntv);
- }
continue;
}
- Intf.moveToBlock(BI.MBB->getNumber());
-
- if (RegIn && RegOut)
- SE->splitLiveThroughBlock(BI.MBB->getNumber(),
- MainIntv, Intf.first(),
- MainIntv, Intf.last());
- else if (RegIn)
- SE->splitRegInBlock(BI, MainIntv, Intf.first());
+ if (IntvIn && IntvOut)
+ SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
+ else if (IntvIn)
+ SE->splitRegInBlock(BI, IntvIn, IntfIn);
else
- SE->splitRegOutBlock(BI, MainIntv, Intf.last());
+ SE->splitRegOutBlock(BI, IntvOut, IntfOut);
}
- // Handle live-through blocks.
- 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;
- Intf.moveToBlock(Number);
- SE->splitLiveThroughBlock(Number, RegIn ? MainIntv : 0, Intf.first(),
- RegOut ? MainIntv : 0, Intf.last());
+ // Handle live-through blocks. The relevant live-through blocks are stored in
+ // the ActiveBlocks list with each candidate. We need to filter out
+ // duplicates.
+ BitVector Todo = SA->getThroughBlocks();
+ for (unsigned c = 0; c != UsedCands.size(); ++c) {
+ ArrayRef<unsigned> Blocks = GlobalCand[UsedCands[c]].ActiveBlocks;
+ for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
+ unsigned Number = Blocks[i];
+ if (!Todo.test(Number))
+ continue;
+ Todo.reset(Number);
+
+ unsigned IntvIn = 0, IntvOut = 0;
+ SlotIndex IntfIn, IntfOut;
+
+ unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
+ if (CandIn != NoCand) {
+ GlobalSplitCandidate &Cand = GlobalCand[CandIn];
+ IntvIn = Cand.IntvIdx;
+ Cand.Intf.moveToBlock(Number);
+ IntfIn = Cand.Intf.first();
+ }
+
+ unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
+ if (CandOut != NoCand) {
+ GlobalSplitCandidate &Cand = GlobalCand[CandOut];
+ IntvOut = Cand.IntvIdx;
+ Cand.Intf.moveToBlock(Number);
+ IntfOut = Cand.Intf.last();
+ }
+ if (!IntvIn && !IntvOut)
+ continue;
+ SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
+ }
}
++NumGlobalSplits;
SmallVector<unsigned, 8> IntvMap;
SE->finish(&IntvMap);
- DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
+ DebugVars->splitRegister(Reg, LREdit.regs(), *LIS);
ExtraRegInfo.resize(MRI->getNumVirtRegs());
unsigned OrigBlocks = SA->getNumLiveBlocks();
// - 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);
+ LiveInterval &Reg = LIS->getInterval(LREdit.get(i));
// Ignore old intervals from DCE.
if (getStage(Reg) != RS_New)
// Remainder interval. Don't try splitting again, spill if it doesn't
// allocate.
if (IntvMap[i] == 0) {
- setStage(Reg, RS_Global);
+ setStage(Reg, RS_Spill);
continue;
}
- // Main interval. Allow repeated splitting as long as the number of live
+ // Global intervals. Allow repeated splitting as long as the number of live
// blocks is strictly decreasing.
- if (IntvMap[i] == MainIntv) {
+ if (IntvMap[i] < NumGlobalIntvs) {
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);
+ setStage(Reg, RS_Split2);
}
continue;
}
}
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;
+ SmallVectorImpl<unsigned> &NewVRegs) {
unsigned NumCands = 0;
+ unsigned BestCand = NoCand;
+ BlockFrequency BestCost;
+ SmallVector<unsigned, 8> UsedCands;
+
+ // Check if we can split this live range around a compact region.
+ bool HasCompact = calcCompactRegion(GlobalCand.front());
+ if (HasCompact) {
+ // Yes, keep GlobalCand[0] as the compact region candidate.
+ NumCands = 1;
+ BestCost = BlockFrequency::getMaxFrequency();
+ } else {
+ // No benefit from the compact region, our fallback will be per-block
+ // splitting. Make sure we find a solution that is cheaper than spilling.
+ BestCost = calcSpillCost();
+ DEBUG(dbgs() << "Cost of isolating all blocks = ";
+ MBFI->printBlockFreq(dbgs(), BestCost) << '\n');
+ }
Order.rewind();
while (unsigned PhysReg = Order.next()) {
unsigned WorstCount = ~0u;
unsigned Worst = 0;
for (unsigned i = 0; i != NumCands; ++i) {
- if (i == BestCand)
+ if (i == BestCand || !GlobalCand[i].PhysReg)
continue;
unsigned Count = GlobalCand[i].LiveBundles.count();
if (Count < WorstCount)
}
--NumCands;
GlobalCand[Worst] = GlobalCand[NumCands];
+ if (BestCand == NumCands)
+ BestCand = Worst;
}
if (GlobalCand.size() <= NumCands)
Cand.reset(IntfCache, PhysReg);
SpillPlacer->prepare(Cand.LiveBundles);
- float Cost;
+ BlockFrequency Cost;
if (!addSplitConstraints(Cand.Intf, Cost)) {
DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tno positive bundles\n");
continue;
}
- DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = " << Cost);
+ DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = ";
+ MBFI->printBlockFreq(dbgs(), Cost));
if (Cost >= BestCost) {
DEBUG({
if (BestCand == NoCand)
Cost += calcGlobalSplitCost(Cand);
DEBUG({
- dbgs() << ", total = " << Cost << " with bundles";
+ dbgs() << ", total = "; MBFI->printBlockFreq(dbgs(), Cost)
+ << " with bundles";
for (int i = Cand.LiveBundles.find_first(); i>=0;
i = Cand.LiveBundles.find_next(i))
dbgs() << " EB#" << i;
});
if (Cost < BestCost) {
BestCand = NumCands;
- BestCost = Hysteresis * Cost; // Prevent rounding effects.
+ BestCost = Cost;
}
++NumCands;
}
- if (BestCand == NoCand)
+ // No solutions found, fall back to single block splitting.
+ if (!HasCompact && BestCand == NoCand)
+ return 0;
+
+ // Prepare split editor.
+ LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
+ SE->reset(LREdit, SplitSpillMode);
+
+ // Assign all edge bundles to the preferred candidate, or NoCand.
+ BundleCand.assign(Bundles->getNumBundles(), NoCand);
+
+ // Assign bundles for the best candidate region.
+ if (BestCand != NoCand) {
+ GlobalSplitCandidate &Cand = GlobalCand[BestCand];
+ if (unsigned B = Cand.getBundles(BundleCand, BestCand)) {
+ UsedCands.push_back(BestCand);
+ Cand.IntvIdx = SE->openIntv();
+ DEBUG(dbgs() << "Split for " << PrintReg(Cand.PhysReg, TRI) << " in "
+ << B << " bundles, intv " << Cand.IntvIdx << ".\n");
+ (void)B;
+ }
+ }
+
+ // Assign bundles for the compact region.
+ if (HasCompact) {
+ GlobalSplitCandidate &Cand = GlobalCand.front();
+ assert(!Cand.PhysReg && "Compact region has no physreg");
+ if (unsigned B = Cand.getBundles(BundleCand, 0)) {
+ UsedCands.push_back(0);
+ Cand.IntvIdx = SE->openIntv();
+ DEBUG(dbgs() << "Split for compact region in " << B << " bundles, intv "
+ << Cand.IntvIdx << ".\n");
+ (void)B;
+ }
+ }
+
+ splitAroundRegion(LREdit, UsedCands);
+ return 0;
+}
+
+
+//===----------------------------------------------------------------------===//
+// 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<unsigned> &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, *MF, *LIS, VRM, 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(), *LIS);
+
+ 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 = LIS->getInterval(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;
+}
+
+
+//===----------------------------------------------------------------------===//
+// Per-Instruction Splitting
+//===----------------------------------------------------------------------===//
+
+/// Get the number of allocatable registers that match the constraints of \p Reg
+/// on \p MI and that are also in \p SuperRC.
+static unsigned getNumAllocatableRegsForConstraints(
+ const MachineInstr *MI, unsigned Reg, const TargetRegisterClass *SuperRC,
+ const TargetInstrInfo *TII, const TargetRegisterInfo *TRI,
+ const RegisterClassInfo &RCI) {
+ assert(SuperRC && "Invalid register class");
+
+ const TargetRegisterClass *ConstrainedRC =
+ MI->getRegClassConstraintEffectForVReg(Reg, SuperRC, TII, TRI,
+ /* ExploreBundle */ true);
+ if (!ConstrainedRC)
+ return 0;
+ return RCI.getNumAllocatableRegs(ConstrainedRC);
+}
+
+/// tryInstructionSplit - Split a live range around individual instructions.
+/// This is normally not worthwhile since the spiller is doing essentially the
+/// same thing. However, when the live range is in a constrained register
+/// class, it may help to insert copies such that parts of the live range can
+/// be moved to a larger register class.
+///
+/// This is similar to spilling to a larger register class.
+unsigned
+RAGreedy::tryInstructionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
+ SmallVectorImpl<unsigned> &NewVRegs) {
+ const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg);
+ // There is no point to this if there are no larger sub-classes.
+ if (!RegClassInfo.isProperSubClass(CurRC))
return 0;
- splitAroundRegion(VirtReg, GlobalCand[BestCand], NewVRegs);
+ // Always enable split spill mode, since we're effectively spilling to a
+ // register.
+ LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
+ SE->reset(LREdit, SplitEditor::SM_Size);
+
+ ArrayRef<SlotIndex> Uses = SA->getUseSlots();
+ if (Uses.size() <= 1)
+ return 0;
+
+ DEBUG(dbgs() << "Split around " << Uses.size() << " individual instrs.\n");
+
+ const TargetRegisterClass *SuperRC = TRI->getLargestLegalSuperClass(CurRC);
+ unsigned SuperRCNumAllocatableRegs = RCI.getNumAllocatableRegs(SuperRC);
+ // Split around every non-copy instruction if this split will relax
+ // the constraints on the virtual register.
+ // Otherwise, splitting just inserts uncoalescable copies that do not help
+ // the allocation.
+ for (unsigned i = 0; i != Uses.size(); ++i) {
+ if (const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]))
+ if (MI->isFullCopy() ||
+ SuperRCNumAllocatableRegs ==
+ getNumAllocatableRegsForConstraints(MI, VirtReg.reg, SuperRC, TII,
+ TRI, RCI)) {
+ DEBUG(dbgs() << " skip:\t" << Uses[i] << '\t' << *MI);
+ continue;
+ }
+ SE->openIntv();
+ SlotIndex SegStart = SE->enterIntvBefore(Uses[i]);
+ SlotIndex SegStop = SE->leaveIntvAfter(Uses[i]);
+ SE->useIntv(SegStart, SegStop);
+ }
+
+ if (LREdit.empty()) {
+ DEBUG(dbgs() << "All uses were copies.\n");
+ return 0;
+ }
+
+ SmallVector<unsigned, 8> IntvMap;
+ SE->finish(&IntvMap);
+ DebugVars->splitRegister(VirtReg.reg, LREdit.regs(), *LIS);
+ ExtraRegInfo.resize(MRI->getNumVirtRegs());
+
+ // Assign all new registers to RS_Spill. This was the last chance.
+ setStage(LREdit.begin(), LREdit.end(), RS_Spill);
return 0;
}
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);
// Add interference from each overlapping register.
- for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
- if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
- .checkInterference())
+ for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
+ if (!Matrix->query(const_cast<LiveInterval&>(SA->getParent()), *Units)
+ .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 =
+ Matrix->getLiveUnions()[*Units] .find(StartIdx);
for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
// Skip the gaps before IntI.
while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
break;
}
}
+
+ // Add fixed interference.
+ for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
+ const LiveRange &LR = LIS->getRegUnit(*Units);
+ LiveRange::const_iterator I = LR.find(StartIdx);
+ LiveRange::const_iterator E = LR.end();
+
+ // Same loop as above. Mark any overlapped gaps as HUGE_VALF.
+ for (unsigned Gap = 0; I != E && I->start < StopIdx; ++I) {
+ while (Uses[Gap+1].getBoundaryIndex() < I->start)
+ if (++Gap == NumGaps)
+ break;
+ if (Gap == NumGaps)
+ break;
+
+ for (; Gap != NumGaps; ++Gap) {
+ GapWeight[Gap] = llvm::huge_valf;
+ if (Uses[Gap+1].getBaseIndex() >= I->end)
+ break;
+ }
+ if (Gap == NumGaps)
+ break;
+ }
+ }
}
/// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
/// basic block.
///
unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
- SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ SmallVectorImpl<unsigned> &NewVRegs) {
assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
// 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';
});
+ // If VirtReg is live across any register mask operands, compute a list of
+ // gaps with register masks.
+ SmallVector<unsigned, 8> RegMaskGaps;
+ if (Matrix->checkRegMaskInterference(VirtReg)) {
+ // Get regmask slots for the whole block.
+ ArrayRef<SlotIndex> RMS = LIS->getRegMaskSlotsInBlock(BI.MBB->getNumber());
+ DEBUG(dbgs() << RMS.size() << " regmasks in block:");
+ // Constrain to VirtReg's live range.
+ unsigned ri = std::lower_bound(RMS.begin(), RMS.end(),
+ Uses.front().getRegSlot()) - RMS.begin();
+ unsigned re = RMS.size();
+ for (unsigned i = 0; i != NumGaps && ri != re; ++i) {
+ // Look for Uses[i] <= RMS <= Uses[i+1].
+ assert(!SlotIndex::isEarlierInstr(RMS[ri], Uses[i]));
+ if (SlotIndex::isEarlierInstr(Uses[i+1], RMS[ri]))
+ continue;
+ // Skip a regmask on the same instruction as the last use. It doesn't
+ // overlap the live range.
+ if (SlotIndex::isSameInstr(Uses[i+1], RMS[ri]) && i+1 == NumGaps)
+ break;
+ DEBUG(dbgs() << ' ' << RMS[ri] << ':' << Uses[i] << '-' << Uses[i+1]);
+ RegMaskGaps.push_back(i);
+ // Advance ri to the next gap. A regmask on one of the uses counts in
+ // both gaps.
+ while (ri != re && SlotIndex::isEarlierInstr(RMS[ri], Uses[i+1]))
+ ++ri;
+ }
+ DEBUG(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
//
// Instead we use these rules:
//
- // 1. Allow any split for ranges with getStage() < RS_Local. (Except for the
+ // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the
// noop split, of course).
- // 2. Require progress be made for ranges with getStage() >= RS_Local. All
+ // 2. Require progress be made for ranges with getStage() == RS_Split2. 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,
+ // 3. New ranges with the same number of instructions are marked RS_Split2,
// 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;
+ bool ProgressRequired = getStage(VirtReg) >= RS_Split2;
// Best split candidate.
unsigned BestBefore = NumGaps;
unsigned BestAfter = 0;
float BestDiff = 0;
- const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB->getNumber());
+ const float blockFreq =
+ SpillPlacer->getBlockFrequency(BI.MBB->getNumber()).getFrequency() *
+ (1.0f / MBFI->getEntryFreq());
SmallVector<float, 8> GapWeight;
Order.rewind();
// order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
calcGapWeights(PhysReg, GapWeight);
+ // Remove any gaps with regmask clobbers.
+ if (Matrix->checkRegMaskInterference(VirtReg, PhysReg))
+ for (unsigned i = 0, e = RegMaskGaps.size(); i != e; ++i)
+ GapWeight[RegMaskGaps[i]] = llvm::huge_valf;
+
// Try to find the best sequence of gaps to close.
// The new spill weight must be larger than any gap interference.
// Legally, without causing looping?
bool Legal = !ProgressRequired || NewGaps < NumGaps;
- if (Legal && MaxGap < HUGE_VALF) {
+ if (Legal && MaxGap < llvm::huge_valf) {
// Estimate the new spill weight. Each instruction reads or writes the
// register. Conservatively assume there are no read-modify-write
// instructions.
<< '-' << Uses[BestAfter] << ", " << BestDiff
<< ", " << (BestAfter - BestBefore + 1) << " instrs\n");
- LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+ LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
SE->reset(LREdit);
SE->openIntv();
SE->useIntv(SegStart, SegStop);
SmallVector<unsigned, 8> IntvMap;
SE->finish(&IntvMap);
- DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
+ DebugVars->splitRegister(VirtReg.reg, LREdit.regs(), *LIS);
// 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,
+ // RS_Split2 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;
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));
+ setStage(LIS->getInterval(LREdit.get(i)), RS_Split2);
+ DEBUG(dbgs() << PrintReg(LREdit.get(i)));
}
DEBUG(dbgs() << '\n');
}
/// assignable.
/// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
- SmallVectorImpl<LiveInterval*>&NewVRegs) {
+ SmallVectorImpl<unsigned>&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);
SA->analyze(&VirtReg);
- return tryLocalSplit(VirtReg, Order, NewVRegs);
+ unsigned PhysReg = tryLocalSplit(VirtReg, Order, NewVRegs);
+ if (PhysReg || !NewVRegs.empty())
+ return PhysReg;
+ return tryInstructionSplit(VirtReg, Order, NewVRegs);
}
NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
- // 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
// an assertion when the coalescer is fixed.
if (SA->didRepairRange()) {
// VirtReg has changed, so all cached queries are invalid.
- invalidateVirtRegs();
+ Matrix->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())
- return PhysReg;
+ // First try to split around a region spanning multiple blocks. RS_Split2
+ // ranges already made dubious progress with region splitting, so they go
+ // straight to single block splitting.
+ if (getStage(VirtReg) < RS_Split2) {
+ unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
+ if (PhysReg || !NewVRegs.empty())
+ return PhysReg;
+ }
+
+ // Then isolate blocks.
+ return tryBlockSplit(VirtReg, Order, NewVRegs);
+}
+
+//===----------------------------------------------------------------------===//
+// Last Chance Recoloring
+//===----------------------------------------------------------------------===//
- // 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_Global);
- if (VerifyEnabled)
- MF->verify(this, "After splitting live range around basic blocks");
+/// mayRecolorAllInterferences - Check if the virtual registers that
+/// interfere with \p VirtReg on \p PhysReg (or one of its aliases) may be
+/// recolored to free \p PhysReg.
+/// When true is returned, \p RecoloringCandidates has been augmented with all
+/// the live intervals that need to be recolored in order to free \p PhysReg
+/// for \p VirtReg.
+/// \p FixedRegisters contains all the virtual registers that cannot be
+/// recolored.
+bool
+RAGreedy::mayRecolorAllInterferences(unsigned PhysReg, LiveInterval &VirtReg,
+ SmallLISet &RecoloringCandidates,
+ const SmallVirtRegSet &FixedRegisters) {
+ const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg);
+
+ for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
+ LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
+ // If there is LastChanceRecoloringMaxInterference or more interferences,
+ // chances are one would not be recolorable.
+ if (Q.collectInterferingVRegs(LastChanceRecoloringMaxInterference) >=
+ LastChanceRecoloringMaxInterference) {
+ DEBUG(dbgs() << "Early abort: too many interferences.\n");
+ return false;
+ }
+ for (unsigned i = Q.interferingVRegs().size(); i; --i) {
+ LiveInterval *Intf = Q.interferingVRegs()[i - 1];
+ // If Intf is done and sit on the same register class as VirtReg,
+ // it would not be recolorable as it is in the same state as VirtReg.
+ if ((getStage(*Intf) == RS_Done &&
+ MRI->getRegClass(Intf->reg) == CurRC) ||
+ FixedRegisters.count(Intf->reg)) {
+ DEBUG(dbgs() << "Early abort: the inteference is not recolorable.\n");
+ return false;
+ }
+ RecoloringCandidates.insert(Intf);
+ }
}
+ return true;
+}
- // Don't assign any physregs.
- return 0;
+/// tryLastChanceRecoloring - Try to assign a color to \p VirtReg by recoloring
+/// its interferences.
+/// Last chance recoloring chooses a color for \p VirtReg and recolors every
+/// virtual register that was using it. The recoloring process may recursively
+/// use the last chance recoloring. Therefore, when a virtual register has been
+/// assigned a color by this mechanism, it is marked as Fixed, i.e., it cannot
+/// be last-chance-recolored again during this recoloring "session".
+/// E.g.,
+/// Let
+/// vA can use {R1, R2 }
+/// vB can use { R2, R3}
+/// vC can use {R1 }
+/// Where vA, vB, and vC cannot be split anymore (they are reloads for
+/// instance) and they all interfere.
+///
+/// vA is assigned R1
+/// vB is assigned R2
+/// vC tries to evict vA but vA is already done.
+/// Regular register allocation fails.
+///
+/// Last chance recoloring kicks in:
+/// vC does as if vA was evicted => vC uses R1.
+/// vC is marked as fixed.
+/// vA needs to find a color.
+/// None are available.
+/// vA cannot evict vC: vC is a fixed virtual register now.
+/// vA does as if vB was evicted => vA uses R2.
+/// vB needs to find a color.
+/// R3 is available.
+/// Recoloring => vC = R1, vA = R2, vB = R3
+///
+/// \p Order defines the prefered allocation order for \p VirtReg.
+/// \p NewRegs will contain any new virtual register that have been created
+/// (split, spill) during the process and that must be assigned.
+/// \p FixedRegisters contains all the virtual registers that cannot be
+/// recolored.
+/// \p Depth gives the current depth of the last chance recoloring.
+/// \return a physical register that can be used for VirtReg or ~0u if none
+/// exists.
+unsigned RAGreedy::tryLastChanceRecoloring(LiveInterval &VirtReg,
+ AllocationOrder &Order,
+ SmallVectorImpl<unsigned> &NewVRegs,
+ SmallVirtRegSet &FixedRegisters,
+ unsigned Depth) {
+ DEBUG(dbgs() << "Try last chance recoloring for " << VirtReg << '\n');
+ // Ranges must be Done.
+ assert(getStage(VirtReg) >= RS_Done &&
+ "Last chance recoloring should really be last chance");
+ // Set the max depth to LastChanceRecoloringMaxDepth.
+ // We may want to reconsider that if we end up with a too large search space
+ // for target with hundreds of registers.
+ // Indeed, in that case we may want to cut the search space earlier.
+ if (Depth >= LastChanceRecoloringMaxDepth) {
+ DEBUG(dbgs() << "Abort because max depth has been reached.\n");
+ return ~0u;
+ }
+
+ // Set of Live intervals that will need to be recolored.
+ SmallLISet RecoloringCandidates;
+ // Record the original mapping virtual register to physical register in case
+ // the recoloring fails.
+ DenseMap<unsigned, unsigned> VirtRegToPhysReg;
+ // Mark VirtReg as fixed, i.e., it will not be recolored pass this point in
+ // this recoloring "session".
+ FixedRegisters.insert(VirtReg.reg);
+
+ Order.rewind();
+ while (unsigned PhysReg = Order.next()) {
+ DEBUG(dbgs() << "Try to assign: " << VirtReg << " to "
+ << PrintReg(PhysReg, TRI) << '\n');
+ RecoloringCandidates.clear();
+ VirtRegToPhysReg.clear();
+
+ // It is only possible to recolor virtual register interference.
+ if (Matrix->checkInterference(VirtReg, PhysReg) >
+ LiveRegMatrix::IK_VirtReg) {
+ DEBUG(dbgs() << "Some inteferences are not with virtual registers.\n");
+
+ continue;
+ }
+
+ // Early give up on this PhysReg if it is obvious we cannot recolor all
+ // the interferences.
+ if (!mayRecolorAllInterferences(PhysReg, VirtReg, RecoloringCandidates,
+ FixedRegisters)) {
+ DEBUG(dbgs() << "Some inteferences cannot be recolored.\n");
+ continue;
+ }
+
+ // RecoloringCandidates contains all the virtual registers that interfer
+ // with VirtReg on PhysReg (or one of its aliases).
+ // Enqueue them for recoloring and perform the actual recoloring.
+ PQueue RecoloringQueue;
+ for (SmallLISet::iterator It = RecoloringCandidates.begin(),
+ EndIt = RecoloringCandidates.end();
+ It != EndIt; ++It) {
+ unsigned ItVirtReg = (*It)->reg;
+ enqueue(RecoloringQueue, *It);
+ assert(VRM->hasPhys(ItVirtReg) &&
+ "Interferences are supposed to be with allocated vairables");
+
+ // Record the current allocation.
+ VirtRegToPhysReg[ItVirtReg] = VRM->getPhys(ItVirtReg);
+ // unset the related struct.
+ Matrix->unassign(**It);
+ }
+
+ // Do as if VirtReg was assigned to PhysReg so that the underlying
+ // recoloring has the right information about the interferes and
+ // available colors.
+ Matrix->assign(VirtReg, PhysReg);
+
+ // Save the current recoloring state.
+ // If we cannot recolor all the interferences, we will have to start again
+ // at this point for the next physical register.
+ SmallVirtRegSet SaveFixedRegisters(FixedRegisters);
+ if (tryRecoloringCandidates(RecoloringQueue, NewVRegs, FixedRegisters,
+ Depth)) {
+ // Do not mess up with the global assignment process.
+ // I.e., VirtReg must be unassigned.
+ Matrix->unassign(VirtReg);
+ return PhysReg;
+ }
+
+ DEBUG(dbgs() << "Fail to assign: " << VirtReg << " to "
+ << PrintReg(PhysReg, TRI) << '\n');
+
+ // The recoloring attempt failed, undo the changes.
+ FixedRegisters = SaveFixedRegisters;
+ Matrix->unassign(VirtReg);
+
+ for (SmallLISet::iterator It = RecoloringCandidates.begin(),
+ EndIt = RecoloringCandidates.end();
+ It != EndIt; ++It) {
+ unsigned ItVirtReg = (*It)->reg;
+ if (VRM->hasPhys(ItVirtReg))
+ Matrix->unassign(**It);
+ Matrix->assign(**It, VirtRegToPhysReg[ItVirtReg]);
+ }
+ }
+
+ // Last chance recoloring did not worked either, give up.
+ return ~0u;
}
+/// tryRecoloringCandidates - Try to assign a new color to every register
+/// in \RecoloringQueue.
+/// \p NewRegs will contain any new virtual register created during the
+/// recoloring process.
+/// \p FixedRegisters[in/out] contains all the registers that have been
+/// recolored.
+/// \return true if all virtual registers in RecoloringQueue were successfully
+/// recolored, false otherwise.
+bool RAGreedy::tryRecoloringCandidates(PQueue &RecoloringQueue,
+ SmallVectorImpl<unsigned> &NewVRegs,
+ SmallVirtRegSet &FixedRegisters,
+ unsigned Depth) {
+ while (!RecoloringQueue.empty()) {
+ LiveInterval *LI = dequeue(RecoloringQueue);
+ DEBUG(dbgs() << "Try to recolor: " << *LI << '\n');
+ unsigned PhysReg;
+ PhysReg = selectOrSplitImpl(*LI, NewVRegs, FixedRegisters, Depth + 1);
+ if (PhysReg == ~0u || !PhysReg)
+ return false;
+ DEBUG(dbgs() << "Recoloring of " << *LI
+ << " succeeded with: " << PrintReg(PhysReg, TRI) << '\n');
+ Matrix->assign(*LI, PhysReg);
+ FixedRegisters.insert(LI->reg);
+ }
+ return true;
+}
//===----------------------------------------------------------------------===//
// Main Entry Point
//===----------------------------------------------------------------------===//
unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
- SmallVectorImpl<LiveInterval*> &NewVRegs) {
+ SmallVectorImpl<unsigned> &NewVRegs) {
+ SmallVirtRegSet FixedRegisters;
+ return selectOrSplitImpl(VirtReg, NewVRegs, FixedRegisters);
+}
+
+unsigned RAGreedy::selectOrSplitImpl(LiveInterval &VirtReg,
+ SmallVectorImpl<unsigned> &NewVRegs,
+ SmallVirtRegSet &FixedRegisters,
+ unsigned Depth) {
// First try assigning a free register.
AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
<< " 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
+ // queue. The RS_Split ranges already failed to do this, and they should not
// get a second chance until they have been split.
- if (Stage != RS_Second)
+ if (Stage != RS_Split)
if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
return PhysReg;
// 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);
+ if (Stage < RS_Split) {
+ setStage(VirtReg, RS_Split);
DEBUG(dbgs() << "wait for second round\n");
- NewVRegs.push_back(&VirtReg);
+ NewVRegs.push_back(VirtReg.reg);
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 || !VirtReg.isSpillable())
- return ~0u;
+ if (Stage >= RS_Done || !VirtReg.isSpillable())
+ return tryLastChanceRecoloring(VirtReg, Order, NewVRegs, FixedRegisters,
+ Depth);
// Try splitting VirtReg or interferences.
unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
// Finally spill VirtReg itself.
NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
- LiveRangeEdit LRE(VirtReg, NewVRegs, this);
+ LiveRangeEdit LRE(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
spiller().spill(LRE);
- setStage(NewVRegs.begin(), NewVRegs.end(), RS_Spill);
+ setStage(NewVRegs.begin(), NewVRegs.end(), RS_Done);
if (VerifyEnabled)
MF->verify(this, "After spilling");
bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
- << "********** Function: "
- << ((Value*)mf.getFunction())->getName() << '\n');
+ << "********** Function: " << mf.getName() << '\n');
MF = &mf;
+ TRI = MF->getTarget().getRegisterInfo();
+ TII = MF->getTarget().getInstrInfo();
+ RCI.runOnMachineFunction(mf);
if (VerifyEnabled)
MF->verify(this, "Before greedy register allocator");
- RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
+ RegAllocBase::init(getAnalysis<VirtRegMap>(),
+ getAnalysis<LiveIntervals>(),
+ getAnalysis<LiveRegMatrix>());
Indexes = &getAnalysis<SlotIndexes>();
+ MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
DomTree = &getAnalysis<MachineDominatorTree>();
SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
Loops = &getAnalysis<MachineLoopInfo>();
SpillPlacer = &getAnalysis<SpillPlacement>();
DebugVars = &getAnalysis<LiveDebugVariables>();
+ calculateSpillWeightsAndHints(*LIS, mf, *Loops, *MBFI);
+
+ DEBUG(LIS->dump());
+
SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
- SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
+ SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree, *MBFI));
ExtraRegInfo.clear();
ExtraRegInfo.resize(MRI->getNumVirtRegs());
NextCascade = 1;
- IntfCache.init(MF, &PhysReg2LiveUnion[0], Indexes, TRI);
+ IntfCache.init(MF, Matrix->getLiveUnions(), Indexes, LIS, TRI);
+ GlobalCand.resize(32); // This will grow as needed.
allocatePhysRegs();
- addMBBLiveIns(MF);
- LIS->addKillFlags();
-
- // Run rewriter
- {
- NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
- VRM->rewrite(Indexes);
- }
-
- // Write out new DBG_VALUE instructions.
- DebugVars->emitDebugValues(VRM);
-
- // The pass output is in VirtRegMap. Release all the transient data.
releaseMemory();
-
return true;
}
projects / oota-llvm.git / blobdiff