#include "VirtRegRewriter.h"
#include "Spiller.h"
#include "llvm/Function.h"
+#include "llvm/CodeGen/CalcSpillWeights.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/LiveStackAnalysis.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
cl::desc("Pre-register allocation live interval splitting"),
cl::init(false), cl::Hidden);
+static cl::opt<bool>
+TrivCoalesceEnds("trivial-coalesce-ends",
+ cl::desc("Attempt trivial coalescing of interval ends"),
+ cl::init(false), cl::Hidden);
+
static RegisterRegAlloc
linearscanRegAlloc("linearscan", "linear scan register allocator",
createLinearScanRegisterAllocator);
namespace {
+ // 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;
const TargetRegisterInfo* tri_;
const TargetInstrInfo* tii_;
BitVector allocatableRegs_;
+ BitVector reservedRegs_;
LiveIntervals* li_;
LiveStacks* ls_;
- const MachineLoopInfo *loopInfo;
+ MachineLoopInfo *loopInfo;
/// handled_ - Intervals are added to the handled_ set in the order of their
/// start value. This is uses for backtracking.
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";
// Make sure PassManager knows which analyses to make available
// to coalescing and which analyses coalescing invalidates.
AU.addRequiredTransitive<RegisterCoalescer>();
+ AU.addRequired<CalculateSpillWeights>();
if (PreSplitIntervals)
AU.addRequiredID(PreAllocSplittingID);
AU.addRequired<LiveStacks>();
/// 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();
SmallVector<LiveInterval*, 8> &SpillIntervals);
/// 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
+ /// try to allocate the definition to the same register as the source,
+ /// if the register is not defined during the life time of the interval.
+ /// This eliminates a copy, and 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.
bool Error = false;
for (unsigned i = 0, e = tri_->getNumRegs(); i != e; ++i) {
if (regUse_[i] != 0) {
- errs() << tri_->getName(i) << " is still in use!\n";
+ dbgs() << tri_->getName(i) << " is still in use!\n";
Error = true;
}
}
SmallVector<unsigned, 256> &inactiveCounts,
bool SkipDGRegs);
- /// assignVirt2StackSlot - assigns this virtual register to a
- /// stack slot. returns the stack slot
- int assignVirt2StackSlot(unsigned virtReg);
+ /// getFirstNonReservedPhysReg - return the first non-reserved physical
+ /// register in the register class.
+ unsigned getFirstNonReservedPhysReg(const TargetRegisterClass *RC) {
+ TargetRegisterClass::iterator aoe = RC->allocation_order_end(*mf_);
+ TargetRegisterClass::iterator i = RC->allocation_order_begin(*mf_);
+ while (i != aoe && reservedRegs_.test(*i))
+ ++i;
+ assert(i != aoe && "All registers reserved?!");
+ return *i;
+ }
void ComputeRelatedRegClasses();
void printIntervals(const char* const str, ItTy i, ItTy e) const {
DEBUG({
if (str)
- errs() << str << " intervals:\n";
+ dbgs() << str << " intervals:\n";
for (; i != e; ++i) {
- errs() << "\t" << *i->first << " -> ";
+ dbgs() << "\t" << *i->first << " -> ";
unsigned reg = i->first->reg;
if (TargetRegisterInfo::isVirtualRegister(reg))
reg = vrm_->getPhys(reg);
- errs() << tri_->getName(reg) << '\n';
+ dbgs() << tri_->getName(reg) << '\n';
}
});
}
char RALinScan::ID = 0;
}
-static RegisterPass<RALinScan>
-X("linearscan-regalloc", "Linear Scan Register Allocator");
+INITIALIZE_PASS(RALinScan, "linearscan-regalloc",
+ "Linear Scan Register Allocator", false, false);
void RALinScan::ComputeRelatedRegClasses() {
// First pass, add all reg classes to the union, and determine at least one
for (TargetRegisterClass::iterator I = (*RCI)->begin(), E = (*RCI)->end();
I != E; ++I) {
HasAliases = HasAliases || *tri_->getAliasSet(*I) != 0;
-
+
const TargetRegisterClass *&PRC = OneClassForEachPhysReg[*I];
if (PRC) {
// Already processed this register. Just make sure we know that
}
}
}
-
+
// Second pass, now that we know conservatively what register classes each reg
// belongs to, add info about aliases. We don't need to do this for targets
// without register aliases.
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 == SlotIndex()) ||
- 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;
+ if (vni->def != SlotIndex() && vni->isDefAccurate() &&
+ (CopyMI = li_->getInstructionFromIndex(vni->def)) && CopyMI->isCopy())
+ // Defined by a copy, try to extend SrcReg forward
+ CandReg = CopyMI->getOperand(1).getReg();
+ else if (TrivCoalesceEnds &&
+ (CopyMI = li_->getInstructionFromIndex(range.end.getBaseIndex())) &&
+ CopyMI->isCopy() && cur.reg == CopyMI->getOperand(1).getReg())
+ // Only used by a copy, try to extend DstReg backwards
+ CandReg = CopyMI->getOperand(0).getReg();
+ 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_->getInstructionIndex(MI).getDefIndex()))
- O.setIsKill(false);
- }
- }
+ if (li_->conflictsWithPhysReg(cur, *vrm_, CandReg))
+ return Reg;
- ++NumCoalesce;
- return PhysReg;
- }
+ // Try to coalesce.
+ DEBUG(dbgs() << "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) {
tri_ = tm_->getRegisterInfo();
tii_ = tm_->getInstrInfo();
allocatableRegs_ = tri_->getAllocatableSet(fn);
+ reservedRegs_ = tri_->getReservedRegs(fn);
li_ = &getAnalysis<LiveIntervals>();
ls_ = &getAnalysis<LiveStacks>();
loopInfo = &getAnalysis<MachineLoopInfo>();
vrm_ = &getAnalysis<VirtRegMap>();
if (!rewriter_.get()) rewriter_.reset(createVirtRegRewriter());
-
- spiller_.reset(createSpiller(mf_, li_, ls_, loopInfo, vrm_));
-
+
+ spiller_.reset(createSpiller(*this, *mf_, *vrm_));
+
initIntervalSets();
linearScan();
void RALinScan::linearScan() {
// linear scan algorithm
DEBUG({
- errs() << "********** LINEAR SCAN **********\n"
- << "********** Function: "
+ dbgs() << "********** LINEAR SCAN **********\n"
+ << "********** Function: "
<< mf_->getFunction()->getName() << '\n';
printIntervals("fixed", fixed_.begin(), fixed_.end());
});
LiveInterval* cur = unhandled_.top();
unhandled_.pop();
++NumIters;
- DEBUG(errs() << "\n*** CURRENT ***: " << *cur << '\n');
+ DEBUG(dbgs() << "\n*** CURRENT ***: " << *cur << '\n');
assert(!cur->empty() && "Empty interval in unhandled set.");
while (!active_.empty()) {
IntervalPtr &IP = active_.back();
unsigned reg = IP.first->reg;
- DEBUG(errs() << "\tinterval " << *IP.first << " expired\n");
+ DEBUG(dbgs() << "\tinterval " << *IP.first << " expired\n");
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
reg = vrm_->getPhys(reg);
DEBUG({
for (IntervalPtrs::reverse_iterator
i = inactive_.rbegin(); i != inactive_.rend(); ++i)
- errs() << "\tinterval " << *i->first << " expired\n";
+ dbgs() << "\tinterval " << *i->first << " expired\n";
});
inactive_.clear();
}
}
- DEBUG(errs() << *vrm_);
+ DEBUG(dbgs() << *vrm_);
// Look for physical registers that end up not being allocated even though
// register allocator had to spill other registers in its register class.
/// to the inactive list.
void RALinScan::processActiveIntervals(SlotIndex CurPoint)
{
- DEBUG(errs() << "\tprocessing active intervals:\n");
+ DEBUG(dbgs() << "\tprocessing active intervals:\n");
for (unsigned i = 0, e = active_.size(); i != e; ++i) {
LiveInterval *Interval = active_[i].first;
IntervalPos = Interval->advanceTo(IntervalPos, CurPoint);
if (IntervalPos == Interval->end()) { // Remove expired intervals.
- DEBUG(errs() << "\t\tinterval " << *Interval << " expired\n");
+ DEBUG(dbgs() << "\t\tinterval " << *Interval << " expired\n");
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
reg = vrm_->getPhys(reg);
} else if (IntervalPos->start > CurPoint) {
// Move inactive intervals to inactive list.
- DEBUG(errs() << "\t\tinterval " << *Interval << " inactive\n");
+ DEBUG(dbgs() << "\t\tinterval " << *Interval << " inactive\n");
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
reg = vrm_->getPhys(reg);
/// ones to the active list.
void RALinScan::processInactiveIntervals(SlotIndex CurPoint)
{
- DEBUG(errs() << "\tprocessing inactive intervals:\n");
+ DEBUG(dbgs() << "\tprocessing inactive intervals:\n");
for (unsigned i = 0, e = inactive_.size(); i != e; ++i) {
LiveInterval *Interval = inactive_[i].first;
IntervalPos = Interval->advanceTo(IntervalPos, CurPoint);
if (IntervalPos == Interval->end()) { // remove expired intervals.
- DEBUG(errs() << "\t\tinterval " << *Interval << " expired\n");
+ DEBUG(dbgs() << "\t\tinterval " << *Interval << " expired\n");
// Pop off the end of the list.
inactive_[i] = inactive_.back();
--i; --e;
} else if (IntervalPos->start <= CurPoint) {
// move re-activated intervals in active list
- DEBUG(errs() << "\t\tinterval " << *Interval << " active\n");
+ DEBUG(dbgs() << "\t\tinterval " << *Interval << " active\n");
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
reg = vrm_->getPhys(reg);
return IP.end();
}
-static void RevertVectorIteratorsTo(RALinScan::IntervalPtrs &V, SlotIndex 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(),
static
float getConflictWeight(LiveInterval *cur, unsigned Reg, LiveIntervals *li_,
MachineRegisterInfo *mri_,
- const MachineLoopInfo *loopInfo) {
+ MachineLoopInfo *loopInfo) {
float Conflicts = 0;
for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(Reg),
E = mri_->reg_end(); I != E; ++I) {
MachineInstr *MI = &*I;
if (cur->liveAt(li_->getInstructionIndex(MI))) {
unsigned loopDepth = loopInfo->getLoopDepth(MI->getParent());
- Conflicts += powf(10.0f, (float)loopDepth);
+ Conflicts += std::pow(10.0f, (float)loopDepth);
}
}
return Conflicts;
SmallVector<LiveInterval*, 8> SLIs[3];
DEBUG({
- errs() << "\tConsidering " << NumCands << " candidates: ";
+ dbgs() << "\tConsidering " << NumCands << " candidates: ";
for (unsigned i = 0; i != NumCands; ++i)
- errs() << tri_->getName(Candidates[i].first) << " ";
- errs() << "\n";
+ dbgs() << tri_->getName(Candidates[i].first) << " ";
+ dbgs() << "\n";
});
-
+
// Calculate the number of conflicts of each candidate.
for (IntervalPtrs::iterator i = active_.begin(); i != active_.end(); ++i) {
unsigned Reg = i->first->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);
}
};
}
/// assignRegOrStackSlotAtInterval - assign a register if one is available, or
/// spill.
void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) {
- DEBUG(errs() << "\tallocating current interval: ");
+ DEBUG(dbgs() << "\tallocating current interval: ");
// This is an implicitly defined live interval, just assign any register.
const TargetRegisterClass *RC = mri_->getRegClass(cur->reg);
if (cur->empty()) {
unsigned physReg = vrm_->getRegAllocPref(cur->reg);
if (!physReg)
- physReg = *RC->allocation_order_begin(*mf_);
- DEBUG(errs() << tri_->getName(physReg) << '\n');
+ physReg = getFirstNonReservedPhysReg(RC);
+ DEBUG(dbgs() << tri_->getName(physReg) << '\n');
// Note the register is not really in use.
vrm_->assignVirt2Phys(cur->reg, physReg);
return;
if ((vni->def != SlotIndex()) && !vni->isUnused() &&
vni->isDefAccurate()) {
MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
- unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
- if (CopyMI &&
- tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubReg, DstSubReg)) {
+ if (CopyMI && CopyMI->isCopy()) {
+ unsigned DstSubReg = CopyMI->getOperand(0).getSubReg();
+ unsigned SrcReg = CopyMI->getOperand(1).getReg();
+ unsigned SrcSubReg = CopyMI->getOperand(1).getSubReg();
unsigned Reg = 0;
if (TargetRegisterInfo::isPhysicalRegister(SrcReg))
Reg = SrcReg;
assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
"Can only allocate virtual registers!");
const TargetRegisterClass *RegRC = mri_->getRegClass(Reg);
- // If this is not in a related reg class to the register we're allocating,
+ // If this is not in a related reg class to the register we're allocating,
// don't check it.
if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader &&
cur->overlapsFrom(*i->first, i->second-1)) {
SpillWeightsToAdd.push_back(std::make_pair(Reg, i->first->weight));
}
}
-
+
// Speculatively check to see if we can get a register right now. If not,
// we know we won't be able to by adding more constraints. If so, we can
// check to see if it is valid. Doing an exhaustive search of the fixed_ list
SmallSet<unsigned, 8> RegAliases;
for (const unsigned *AS = tri_->getAliasSet(physReg); *AS; ++AS)
RegAliases.insert(*AS);
-
+
bool ConflictsWithFixed = false;
for (unsigned i = 0, e = fixed_.size(); i != e; ++i) {
IntervalPtr &IP = fixed_[i];
}
}
}
-
+
// Okay, the register picked by our speculative getFreePhysReg call turned
// out to be in use. Actually add all of the conflicting fixed registers to
// regUse_ so we can do an accurate query.
LiveInterval *I = IP.first;
const TargetRegisterClass *RegRC = OneClassForEachPhysReg[I->reg];
- if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader &&
+ if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader &&
I->endIndex() > StartPosition) {
LiveInterval::iterator II = I->advanceTo(IP.second, StartPosition);
IP.second = II;
physReg = getFreePhysReg(cur);
}
}
-
+
// Restore the physical register tracker, removing information about the
// future.
restoreRegUses();
-
+
// If we find a free register, we are done: assign this virtual to
// the free physical register and add this interval to the active
// list.
if (physReg) {
- DEBUG(errs() << tri_->getName(physReg) << '\n');
+ DEBUG(dbgs() << tri_->getName(physReg) << '\n');
vrm_->assignVirt2Phys(cur->reg, physReg);
addRegUse(physReg);
active_.push_back(std::make_pair(cur, cur->begin()));
UpgradeRegister(physReg);
if (LiveInterval *NextReloadLI = hasNextReloadInterval(cur)) {
// "Downgrade" physReg to try to keep physReg from being allocated until
- // the next reload from the same SS is allocated.
+ // the next reload from the same SS is allocated.
mri_->setRegAllocationHint(NextReloadLI->reg, 0, physReg);
DowngradeRegister(cur, physReg);
}
return;
}
- DEBUG(errs() << "no free registers\n");
+ DEBUG(dbgs() << "no free registers\n");
// Compile the spill weights into an array that is better for scanning.
std::vector<float> SpillWeights(tri_->getNumRegs(), 0.0f);
for (std::vector<std::pair<unsigned, float> >::iterator
I = SpillWeightsToAdd.begin(), E = SpillWeightsToAdd.end(); I != E; ++I)
updateSpillWeights(SpillWeights, I->first, I->second, RC);
-
+
// for each interval in active, update spill weights.
for (IntervalPtrs::const_iterator i = active_.begin(), e = active_.end();
i != e; ++i) {
reg = vrm_->getPhys(reg);
updateSpillWeights(SpillWeights, reg, i->first->weight, RC);
}
-
- DEBUG(errs() << "\tassigning stack slot at interval "<< *cur << ":\n");
+
+ DEBUG(dbgs() << "\tassigning stack slot at interval "<< *cur << ":\n");
// Find a register to spill.
float minWeight = HUGE_VALF;
e = RC->allocation_order_end(*mf_); i != e; ++i) {
unsigned reg = *i;
float regWeight = SpillWeights[reg];
- if (minWeight > regWeight)
+ // Don't even consider reserved regs.
+ if (reservedRegs_.test(reg))
+ continue;
+ // Skip recently allocated registers and reserved registers.
+ if (minWeight > regWeight && !isRecentlyUsed(reg))
Found = true;
RegsWeights.push_back(std::make_pair(reg, regWeight));
}
-
+
// If we didn't find a register that is spillable, try aliases?
if (!Found) {
for (TargetRegisterClass::iterator i = RC->allocation_order_begin(*mf_),
e = RC->allocation_order_end(*mf_); i != e; ++i) {
unsigned reg = *i;
+ if (reservedRegs_.test(reg))
+ continue;
// No need to worry about if the alias register size < regsize of RC.
// We are going to spill all registers that alias it anyway.
for (const unsigned* as = tri_->getAliasSet(reg); *as; ++as)
}
// 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) {
// All registers must have inf weight. Just grab one!
- minReg = BestPhysReg ? BestPhysReg : *RC->allocation_order_begin(*mf_);
+ minReg = BestPhysReg ? BestPhysReg : getFirstNonReservedPhysReg(RC);
if (cur->weight == HUGE_VALF ||
li_->getApproximateInstructionCount(*cur) == 0) {
// Spill a physical register around defs and uses.
assignRegOrStackSlotAtInterval(cur);
} else {
assert(false && "Ran out of registers during register allocation!");
- llvm_report_error("Ran out of registers during register allocation!");
+ report_fatal_error("Ran out of registers during register allocation!");
}
return;
}
}
DEBUG({
- errs() << "\t\tregister(s) with min weight(s): ";
+ dbgs() << "\t\tregister(s) with min weight(s): ";
for (unsigned i = 0; i != LastCandidate; ++i)
- errs() << tri_->getName(RegsWeights[i].first)
+ dbgs() << tri_->getName(RegsWeights[i].first)
<< " (" << RegsWeights[i].second << ")\n";
});
// add any added intervals back to unhandled, and restart
// linearscan.
if (cur->weight != HUGE_VALF && cur->weight <= minWeight) {
- DEBUG(errs() << "\t\t\tspilling(c): " << *cur << '\n');
- SmallVector<LiveInterval*, 8> spillIs;
- std::vector<LiveInterval*> added;
-
- added = spiller_->spill(cur, spillIs);
+ DEBUG(dbgs() << "\t\t\tspilling(c): " << *cur << '\n');
+ SmallVector<LiveInterval*, 8> spillIs, added;
+ spiller_->spill(cur, added, spillIs);
std::sort(added.begin(), added.end(), LISorter());
addStackInterval(cur, ls_, li_, mri_, *vrm_);
// The earliest start of a Spilled interval indicates up to where
// in handled we need to roll back
-
- LiveInterval *earliestStartInterval = cur;
+ assert(!spillIs.empty() && "No spill intervals?");
+ SlotIndex earliestStart = spillIs[0]->beginIndex();
// 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
// track of the earliest start of all spilled live intervals since this will
// mark our rollback point.
- std::vector<LiveInterval*> added;
+ SmallVector<LiveInterval*, 8> added;
while (!spillIs.empty()) {
LiveInterval *sli = spillIs.back();
spillIs.pop_back();
- DEBUG(errs() << "\t\t\tspilling(a): " << *sli << '\n');
- earliestStartInterval =
- (earliestStartInterval->beginIndex() < sli->beginIndex()) ?
- earliestStartInterval : sli;
-
- std::vector<LiveInterval*> newIs;
- newIs = spiller_->spill(sli, spillIs);
+ DEBUG(dbgs() << "\t\t\tspilling(a): " << *sli << '\n');
+ if (sli->beginIndex() < earliestStart)
+ earliestStart = sli->beginIndex();
+ spiller_->spill(sli, added, spillIs);
addStackInterval(sli, ls_, li_, mri_, *vrm_);
- std::copy(newIs.begin(), newIs.end(), std::back_inserter(added));
spilled.insert(sli->reg);
}
- SlotIndex earliestStart = earliestStartInterval->beginIndex();
+ // Include any added intervals in earliestStart.
+ for (unsigned i = 0, e = added.size(); i != e; ++i) {
+ SlotIndex SI = added[i]->beginIndex();
+ if (SI < earliestStart)
+ earliestStart = SI;
+ }
- DEBUG(errs() << "\t\trolling back to: " << earliestStart << '\n');
+ DEBUG(dbgs() << "\t\trolling back to: " << earliestStart << '\n');
// Scan handled in reverse order up to the earliest start of a
// spilled live interval and undo each one, restoring the state of
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');
+ DEBUG(dbgs() << "\t\t\tundo changes for: " << *i << '\n');
handled_.pop_back();
// When undoing a live interval allocation we must know if it is active or
LiveInterval *HI = handled_[i];
if (!HI->expiredAt(earliestStart) &&
HI->expiredAt(cur->beginIndex())) {
- DEBUG(errs() << "\t\t\tundo changes for: " << *HI << '\n');
+ DEBUG(dbgs() << "\t\t\tundo changes for: " << *HI << '\n');
active_.push_back(std::make_pair(HI, HI->begin()));
assert(!TargetRegisterInfo::isPhysicalRegister(HI->reg));
addRegUse(vrm_->getPhys(HI->reg));
// Ignore "downgraded" registers.
if (SkipDGRegs && DowngradedRegs.count(Reg))
continue;
- if (isRegAvail(Reg)) {
+ // Skip reserved registers.
+ if (reservedRegs_.test(Reg))
+ continue;
+ // 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
// Ignore "downgraded" registers.
if (SkipDGRegs && DowngradedRegs.count(Reg))
continue;
+ // Skip reserved registers.
+ if (reservedRegs_.test(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;
}
unsigned RALinScan::getFreePhysReg(LiveInterval *cur) {
SmallVector<unsigned, 256> inactiveCounts;
unsigned MaxInactiveCount = 0;
-
+
const TargetRegisterClass *RC = mri_->getRegClass(cur->reg);
const TargetRegisterClass *RCLeader = RelatedRegClasses.getLeaderValue(RC);
-
+
for (IntervalPtrs::iterator i = inactive_.begin(), e = inactive_.end();
i != e; ++i) {
unsigned reg = i->first->reg;
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
- // If this is not in a related reg class to the register we're allocating,
+ // If this is not in a related reg class to the register we're allocating,
// don't check it.
const TargetRegisterClass *RegRC = mri_->getRegClass(reg);
if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader) {
// available first.
unsigned Preference = vrm_->getRegAllocPref(cur->reg);
if (Preference) {
- DEBUG(errs() << "(preferred: " << tri_->getName(Preference) << ") ");
- if (isRegAvail(Preference) &&
+ DEBUG(dbgs() << "(preferred: " << tri_->getName(Preference) << ") ");
+ if (isRegAvail(Preference) &&
RC->contains(Preference))
return Preference;
}