#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
-#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
cl::init(false), cl::Hidden);
static cl::opt<bool>
-NewSpillFramework("new-spill-framework",
- cl::desc("New spilling framework"),
+TrivCoalesceEnds("trivial-coalesce-ends",
+ cl::desc("Attempt trivial coalescing of interval ends"),
cl::init(false), cl::Hidden);
static RegisterRegAlloc
createLinearScanRegisterAllocator);
namespace {
- struct VISIBILITY_HIDDEN RALinScan : public MachineFunctionPass {
+ // When we allocate a register, add it to a fixed-size queue of
+ // registers to skip in subsequent allocations. This trades a small
+ // amount of register pressure and increased spills for flexibility in
+ // the post-pass scheduler.
+ //
+ // Note that in a the number of registers used for reloading spills
+ // will be one greater than the value of this option.
+ //
+ // One big limitation of this is that it doesn't differentiate between
+ // different register classes. So on x86-64, if there is xmm register
+ // pressure, it can caused fewer GPRs to be held in the queue.
+ static cl::opt<unsigned>
+ NumRecentlyUsedRegs("linearscan-skip-count",
+ cl::desc("Number of registers for linearscan to remember to skip."),
+ cl::init(0),
+ cl::Hidden);
+
+ struct RALinScan : public MachineFunctionPass {
static char ID;
- RALinScan() : MachineFunctionPass(&ID) {}
+ RALinScan() : MachineFunctionPass(&ID) {
+ // Initialize the queue to record recently-used registers.
+ if (NumRecentlyUsedRegs > 0)
+ RecentRegs.resize(NumRecentlyUsedRegs, 0);
+ RecentNext = RecentRegs.begin();
+ }
typedef std::pair<LiveInterval*, LiveInterval::iterator> IntervalPtr;
typedef SmallVector<IntervalPtr, 32> IntervalPtrs;
std::auto_ptr<Spiller> spiller_;
+ // The queue of recently-used registers.
+ SmallVector<unsigned, 4> RecentRegs;
+ SmallVector<unsigned, 4>::iterator RecentNext;
+
+ // Record that we just picked this register.
+ void recordRecentlyUsed(unsigned reg) {
+ assert(reg != 0 && "Recently used register is NOREG!");
+ if (!RecentRegs.empty()) {
+ *RecentNext++ = reg;
+ if (RecentNext == RecentRegs.end())
+ RecentNext = RecentRegs.begin();
+ }
+ }
+
public:
virtual const char* getPassName() const {
return "Linear Scan Register Allocator";
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
AU.addRequired<LiveIntervals>();
+ AU.addPreserved<SlotIndexes>();
if (StrongPHIElim)
AU.addRequiredID(StrongPHIEliminationID);
// Make sure PassManager knows which analyses to make available
/// runOnMachineFunction - register allocate the whole function
bool runOnMachineFunction(MachineFunction&);
+ // Determine if we skip this register due to its being recently used.
+ bool isRecentlyUsed(unsigned reg) const {
+ return std::find(RecentRegs.begin(), RecentRegs.end(), reg) !=
+ RecentRegs.end();
+ }
+
private:
/// linearScan - the linear scan algorithm
void linearScan();
/// processActiveIntervals - expire old intervals and move non-overlapping
/// ones to the inactive list.
- void processActiveIntervals(LiveIndex CurPoint);
+ void processActiveIntervals(SlotIndex CurPoint);
/// processInactiveIntervals - expire old intervals and move overlapping
/// ones to the active list.
- void processInactiveIntervals(LiveIndex CurPoint);
+ void processInactiveIntervals(SlotIndex CurPoint);
/// hasNextReloadInterval - Return the next liveinterval that's being
/// defined by a reload from the same SS as the specified one.
RelatedRegClasses.unionSets(I->second, OneClassForEachPhysReg[*AS]);
}
-/// attemptTrivialCoalescing - If a simple interval is defined by a copy,
-/// try allocate the definition the same register as the source register
-/// if the register is not defined during live time of the interval. This
-/// eliminate a copy. This is used to coalesce copies which were not
-/// coalesced away before allocation either due to dest and src being in
-/// different register classes or because the coalescer was overly
-/// conservative.
+/// attemptTrivialCoalescing - If a simple interval is defined by a copy, try
+/// allocate the definition the same register as the source register if the
+/// register is not defined during live time of the interval. If the interval is
+/// killed by a copy, try to use the destination register. This eliminates a
+/// copy. This is used to coalesce copies which were not coalesced away before
+/// allocation either due to dest and src being in different register classes or
+/// because the coalescer was overly conservative.
unsigned RALinScan::attemptTrivialCoalescing(LiveInterval &cur, unsigned Reg) {
unsigned Preference = vrm_->getRegAllocPref(cur.reg);
if ((Preference && Preference == Reg) || !cur.containsOneValue())
return Reg;
- VNInfo *vni = cur.begin()->valno;
- if ((vni->def == LiveIndex()) ||
- vni->isUnused() || !vni->isDefAccurate())
+ // We cannot handle complicated live ranges. Simple linear stuff only.
+ if (cur.ranges.size() != 1)
return Reg;
- MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
- unsigned SrcReg, DstReg, SrcSubReg, DstSubReg, PhysReg;
- if (!CopyMI ||
- !tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubReg, DstSubReg))
+
+ const LiveRange &range = cur.ranges.front();
+
+ VNInfo *vni = range.valno;
+ if (vni->isUnused())
return Reg;
- PhysReg = SrcReg;
- if (TargetRegisterInfo::isVirtualRegister(SrcReg)) {
- if (!vrm_->isAssignedReg(SrcReg))
+
+ unsigned CandReg;
+ {
+ MachineInstr *CopyMI;
+ unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
+ if (vni->def != SlotIndex() && vni->isDefAccurate() &&
+ (CopyMI = li_->getInstructionFromIndex(vni->def)) &&
+ tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubReg, DstSubReg))
+ // Defined by a copy, try to extend SrcReg forward
+ CandReg = SrcReg;
+ else if (TrivCoalesceEnds &&
+ (CopyMI =
+ li_->getInstructionFromIndex(range.end.getBaseIndex())) &&
+ tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubReg, DstSubReg) &&
+ cur.reg == SrcReg)
+ // Only used by a copy, try to extend DstReg backwards
+ CandReg = DstReg;
+ else
+ return Reg;
+ }
+
+ if (TargetRegisterInfo::isVirtualRegister(CandReg)) {
+ if (!vrm_->isAssignedReg(CandReg))
return Reg;
- PhysReg = vrm_->getPhys(SrcReg);
+ CandReg = vrm_->getPhys(CandReg);
}
- if (Reg == PhysReg)
+ if (Reg == CandReg)
return Reg;
const TargetRegisterClass *RC = mri_->getRegClass(cur.reg);
- if (!RC->contains(PhysReg))
+ if (!RC->contains(CandReg))
return Reg;
- // Try to coalesce.
- if (!li_->conflictsWithPhysRegDef(cur, *vrm_, PhysReg)) {
- DEBUG(errs() << "Coalescing: " << cur << " -> " << tri_->getName(PhysReg)
- << '\n');
- vrm_->clearVirt(cur.reg);
- vrm_->assignVirt2Phys(cur.reg, PhysReg);
-
- // Remove unnecessary kills since a copy does not clobber the register.
- if (li_->hasInterval(SrcReg)) {
- LiveInterval &SrcLI = li_->getInterval(SrcReg);
- for (MachineRegisterInfo::use_iterator I = mri_->use_begin(cur.reg),
- E = mri_->use_end(); I != E; ++I) {
- MachineOperand &O = I.getOperand();
- if (!O.isKill())
- continue;
- MachineInstr *MI = &*I;
- if (SrcLI.liveAt(li_->getDefIndex(li_->getInstructionIndex(MI))))
- O.setIsKill(false);
- }
- }
+ if (li_->conflictsWithPhysReg(cur, *vrm_, CandReg))
+ return Reg;
- ++NumCoalesce;
- return PhysReg;
- }
+ // Try to coalesce.
+ DEBUG(errs() << "Coalescing: " << cur << " -> " << tri_->getName(CandReg)
+ << '\n');
+ vrm_->clearVirt(cur.reg);
+ vrm_->assignVirt2Phys(cur.reg, CandReg);
- return Reg;
+ ++NumCoalesce;
+ return CandReg;
}
bool RALinScan::runOnMachineFunction(MachineFunction &fn) {
vrm_ = &getAnalysis<VirtRegMap>();
if (!rewriter_.get()) rewriter_.reset(createVirtRegRewriter());
- if (NewSpillFramework) {
- spiller_.reset(createSpiller(mf_, li_, ls_, vrm_));
- }
+ spiller_.reset(createSpiller(mf_, li_, loopInfo, vrm_));
initIntervalSets();
for (LiveIntervals::iterator i = li_->begin(), e = li_->end(); i != e; ++i) {
if (TargetRegisterInfo::isPhysicalRegister(i->second->reg)) {
- mri_->setPhysRegUsed(i->second->reg);
- fixed_.push_back(std::make_pair(i->second, i->second->begin()));
- } else
- unhandled_.push(i->second);
+ if (!i->second->empty()) {
+ mri_->setPhysRegUsed(i->second->reg);
+ fixed_.push_back(std::make_pair(i->second, i->second->begin()));
+ }
+ } else {
+ if (i->second->empty()) {
+ assignRegOrStackSlotAtInterval(i->second);
+ }
+ else
+ unhandled_.push(i->second);
+ }
}
}
++NumIters;
DEBUG(errs() << "\n*** CURRENT ***: " << *cur << '\n');
- if (!cur->empty()) {
- processActiveIntervals(cur->beginIndex());
- processInactiveIntervals(cur->beginIndex());
+ assert(!cur->empty() && "Empty interval in unhandled set.");
- assert(TargetRegisterInfo::isVirtualRegister(cur->reg) &&
- "Can only allocate virtual registers!");
- }
+ processActiveIntervals(cur->beginIndex());
+ processInactiveIntervals(cur->beginIndex());
+
+ assert(TargetRegisterInfo::isVirtualRegister(cur->reg) &&
+ "Can only allocate virtual registers!");
// Allocating a virtual register. try to find a free
// physical register or spill an interval (possibly this one) in order to
/// processActiveIntervals - expire old intervals and move non-overlapping ones
/// to the inactive list.
-void RALinScan::processActiveIntervals(LiveIndex CurPoint)
+void RALinScan::processActiveIntervals(SlotIndex CurPoint)
{
DEBUG(errs() << "\tprocessing active intervals:\n");
/// processInactiveIntervals - expire old intervals and move overlapping
/// ones to the active list.
-void RALinScan::processInactiveIntervals(LiveIndex CurPoint)
+void RALinScan::processInactiveIntervals(SlotIndex CurPoint)
{
DEBUG(errs() << "\tprocessing inactive intervals:\n");
return IP.end();
}
-static void RevertVectorIteratorsTo(RALinScan::IntervalPtrs &V, LiveIndex Point){
+static void RevertVectorIteratorsTo(RALinScan::IntervalPtrs &V, SlotIndex Point){
for (unsigned i = 0, e = V.size(); i != e; ++i) {
RALinScan::IntervalPtr &IP = V[i];
LiveInterval::iterator I = std::upper_bound(IP.first->begin(),
if (SI.hasAtLeastOneValue())
VNI = SI.getValNumInfo(0);
else
- VNI = SI.getNextValue(LiveIndex(), 0, false,
+ VNI = SI.getNextValue(SlotIndex(), 0, false,
ls_->getVNInfoAllocator());
LiveInterval &RI = li_->getInterval(cur->reg);
namespace {
struct WeightCompare {
+ private:
+ const RALinScan &Allocator;
+
+ public:
+ WeightCompare(const RALinScan &Alloc) : Allocator(Alloc) {};
+
typedef std::pair<unsigned, float> RegWeightPair;
bool operator()(const RegWeightPair &LHS, const RegWeightPair &RHS) const {
- return LHS.second < RHS.second;
+ return LHS.second < RHS.second && !Allocator.isRecentlyUsed(LHS.first);
}
};
}
backUpRegUses();
std::vector<std::pair<unsigned, float> > SpillWeightsToAdd;
- LiveIndex StartPosition = cur->beginIndex();
+ SlotIndex StartPosition = cur->beginIndex();
const TargetRegisterClass *RCLeader = RelatedRegClasses.getLeaderValue(RC);
// If start of this live interval is defined by a move instruction and its
// one, e.g. X86::mov32to32_. These move instructions are not coalescable.
if (!vrm_->getRegAllocPref(cur->reg) && cur->hasAtLeastOneValue()) {
VNInfo *vni = cur->begin()->valno;
- if ((vni->def != LiveIndex()) && !vni->isUnused() &&
+ if ((vni->def != SlotIndex()) && !vni->isUnused() &&
vni->isDefAccurate()) {
MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
e = RC->allocation_order_end(*mf_); i != e; ++i) {
unsigned reg = *i;
float regWeight = SpillWeights[reg];
- if (minWeight > regWeight)
+ // Skip recently allocated registers.
+ if (minWeight > regWeight && !isRecentlyUsed(reg))
Found = true;
RegsWeights.push_back(std::make_pair(reg, regWeight));
}
}
// Sort all potential spill candidates by weight.
- std::sort(RegsWeights.begin(), RegsWeights.end(), WeightCompare());
+ std::sort(RegsWeights.begin(), RegsWeights.end(), WeightCompare(*this));
minReg = RegsWeights[0].first;
minWeight = RegsWeights[0].second;
if (minWeight == HUGE_VALF) {
DowngradedRegs.clear();
assignRegOrStackSlotAtInterval(cur);
} else {
+ assert(false && "Ran out of registers during register allocation!");
llvm_report_error("Ran out of registers during register allocation!");
}
return;
SmallVector<LiveInterval*, 8> spillIs;
std::vector<LiveInterval*> added;
- if (!NewSpillFramework) {
- added = li_->addIntervalsForSpills(*cur, spillIs, loopInfo, *vrm_);
- } else {
- added = spiller_->spill(cur);
- }
+ added = spiller_->spill(cur, spillIs);
std::sort(added.begin(), added.end(), LISorter());
addStackInterval(cur, ls_, li_, mri_, *vrm_);
LiveInterval *ReloadLi = added[i];
if (ReloadLi->weight == HUGE_VALF &&
li_->getApproximateInstructionCount(*ReloadLi) == 0) {
- LiveIndex ReloadIdx = ReloadLi->beginIndex();
+ SlotIndex ReloadIdx = ReloadLi->beginIndex();
MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx);
int ReloadSS = vrm_->getStackSlot(ReloadLi->reg);
if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) {
// The earliest start of a Spilled interval indicates up to where
// in handled we need to roll back
+ assert(!spillIs.empty() && "No spill intervals?");
+ SlotIndex earliestStart = spillIs[0]->beginIndex();
- LiveInterval *earliestStartInterval = cur;
-
// Spill live intervals of virtual regs mapped to the physical register we
// want to clear (and its aliases). We only spill those that overlap with the
// current interval as the rest do not affect its allocation. we also keep
LiveInterval *sli = spillIs.back();
spillIs.pop_back();
DEBUG(errs() << "\t\t\tspilling(a): " << *sli << '\n');
- earliestStartInterval =
- (earliestStartInterval->beginIndex() < sli->beginIndex()) ?
- earliestStartInterval : sli;
+ if (sli->beginIndex() < earliestStart)
+ earliestStart = sli->beginIndex();
std::vector<LiveInterval*> newIs;
- if (!NewSpillFramework) {
- newIs = li_->addIntervalsForSpills(*sli, spillIs, loopInfo, *vrm_);
- } else {
- newIs = spiller_->spill(sli);
- }
+ newIs = spiller_->spill(sli, spillIs, &earliestStart);
addStackInterval(sli, ls_, li_, mri_, *vrm_);
std::copy(newIs.begin(), newIs.end(), std::back_inserter(added));
spilled.insert(sli->reg);
}
- LiveIndex earliestStart = earliestStartInterval->beginIndex();
-
DEBUG(errs() << "\t\trolling back to: " << earliestStart << '\n');
// Scan handled in reverse order up to the earliest start of a
while (!handled_.empty()) {
LiveInterval* i = handled_.back();
// If this interval starts before t we are done.
- if (i->beginIndex() < earliestStart)
+ if (!i->empty() && i->beginIndex() < earliestStart)
break;
DEBUG(errs() << "\t\t\tundo changes for: " << *i << '\n');
handled_.pop_back();
LiveInterval *ReloadLi = added[i];
if (ReloadLi->weight == HUGE_VALF &&
li_->getApproximateInstructionCount(*ReloadLi) == 0) {
- LiveIndex ReloadIdx = ReloadLi->beginIndex();
+ SlotIndex ReloadIdx = ReloadLi->beginIndex();
MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx);
int ReloadSS = vrm_->getStackSlot(ReloadLi->reg);
if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) {
// Ignore "downgraded" registers.
if (SkipDGRegs && DowngradedRegs.count(Reg))
continue;
- if (isRegAvail(Reg)) {
+ // Skip recently allocated registers.
+ if (isRegAvail(Reg) && !isRecentlyUsed(Reg)) {
FreeReg = Reg;
if (FreeReg < inactiveCounts.size())
FreeRegInactiveCount = inactiveCounts[FreeReg];
// If there are no free regs, or if this reg has the max inactive count,
// return this register.
- if (FreeReg == 0 || FreeRegInactiveCount == MaxInactiveCount)
+ if (FreeReg == 0 || FreeRegInactiveCount == MaxInactiveCount) {
+ // Remember what register we picked so we can skip it next time.
+ if (FreeReg != 0) recordRecentlyUsed(FreeReg);
return FreeReg;
-
+ }
+
// Continue scanning the registers, looking for the one with the highest
// inactive count. Alkis found that this reduced register pressure very
// slightly on X86 (in rev 1.94 of this file), though this should probably be
if (SkipDGRegs && DowngradedRegs.count(Reg))
continue;
if (isRegAvail(Reg) && Reg < inactiveCounts.size() &&
- FreeRegInactiveCount < inactiveCounts[Reg]) {
+ FreeRegInactiveCount < inactiveCounts[Reg] && !isRecentlyUsed(Reg)) {
FreeReg = Reg;
FreeRegInactiveCount = inactiveCounts[Reg];
if (FreeRegInactiveCount == MaxInactiveCount)
}
}
+ // Remember what register we picked so we can skip it next time.
+ recordRecentlyUsed(FreeReg);
+
return FreeReg;
}