#define DEBUG_TYPE "regalloc"
#include "VirtRegMap.h"
+#include "VirtRegRewriter.h"
#include "Spiller.h"
#include "llvm/Function.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/Debug.h"
-#include "llvm/Support/Compiler.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <set>
#include <queue>
#include <memory>
#include <cmath>
+
using namespace llvm;
STATISTIC(NumIters , "Number of iterations performed");
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;
/// vrm_ - Tracks register assignments.
VirtRegMap* vrm_;
+ std::auto_ptr<VirtRegRewriter> rewriter_;
+
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(unsigned CurPoint);
+ void processActiveIntervals(SlotIndex CurPoint);
/// processInactiveIntervals - expire old intervals and move overlapping
/// ones to the active list.
- void processInactiveIntervals(unsigned 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.
}
void finalizeRegUses() {
+#ifndef NDEBUG
+ // Verify all the registers are "freed".
+ 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";
+ Error = true;
+ }
+ }
+ if (Error)
+ llvm_unreachable(0);
+#endif
regUse_.clear();
regUseBackUp_.clear();
}
/// getFreePhysReg - return a free physical register for this virtual
/// register interval if we have one, otherwise return 0.
unsigned getFreePhysReg(LiveInterval* cur);
- unsigned getFreePhysReg(const TargetRegisterClass *RC,
+ unsigned getFreePhysReg(LiveInterval* cur,
+ const TargetRegisterClass *RC,
unsigned MaxInactiveCount,
SmallVector<unsigned, 256> &inactiveCounts,
bool SkipDGRegs);
template <typename ItTy>
void printIntervals(const char* const str, ItTy i, ItTy e) const {
- if (str) DOUT << str << " intervals:\n";
- for (; i != e; ++i) {
- DOUT << "\t" << *i->first << " -> ";
- unsigned reg = i->first->reg;
- if (TargetRegisterInfo::isVirtualRegister(reg)) {
- reg = vrm_->getPhys(reg);
- }
- DOUT << tri_->getName(reg) << '\n';
- }
+ DEBUG({
+ if (str)
+ errs() << str << " intervals:\n";
+
+ for (; i != e; ++i) {
+ errs() << "\t" << *i->first << " -> ";
+
+ unsigned reg = i->first->reg;
+ if (TargetRegisterInfo::isVirtualRegister(reg))
+ reg = vrm_->getPhys(reg);
+
+ errs() << tri_->getName(reg) << '\n';
+ }
+ });
}
};
char RALinScan::ID = 0;
/// different register classes or because the coalescer was overly
/// conservative.
unsigned RALinScan::attemptTrivialCoalescing(LiveInterval &cur, unsigned Reg) {
- if ((cur.preference && cur.preference == Reg) || !cur.containsOneValue())
+ unsigned Preference = vrm_->getRegAllocPref(cur.reg);
+ if ((Preference && Preference == Reg) || !cur.containsOneValue())
return Reg;
VNInfo *vni = cur.begin()->valno;
- if (!vni->def || vni->def == ~1U || vni->def == ~0U)
+ if ((vni->def == SlotIndex()) ||
+ vni->isUnused() || !vni->isDefAccurate())
return Reg;
MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
- unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
+ unsigned SrcReg, DstReg, SrcSubReg, DstSubReg, PhysReg;
if (!CopyMI ||
!tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubReg, DstSubReg))
return Reg;
+ PhysReg = SrcReg;
if (TargetRegisterInfo::isVirtualRegister(SrcReg)) {
if (!vrm_->isAssignedReg(SrcReg))
return Reg;
- else
- SrcReg = vrm_->getPhys(SrcReg);
+ PhysReg = vrm_->getPhys(SrcReg);
}
- if (Reg == SrcReg)
+ if (Reg == PhysReg)
return Reg;
const TargetRegisterClass *RC = mri_->getRegClass(cur.reg);
- if (!RC->contains(SrcReg))
+ if (!RC->contains(PhysReg))
return Reg;
// Try to coalesce.
- if (!li_->conflictsWithPhysRegDef(cur, *vrm_, SrcReg)) {
- DOUT << "Coalescing: " << cur << " -> " << tri_->getName(SrcReg)
- << '\n';
+ if (!li_->conflictsWithPhysRegDef(cur, *vrm_, PhysReg)) {
+ DEBUG(errs() << "Coalescing: " << cur << " -> " << tri_->getName(PhysReg)
+ << '\n');
vrm_->clearVirt(cur.reg);
- vrm_->assignVirt2Phys(cur.reg, SrcReg);
+ 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);
+ }
+ }
+
++NumCoalesce;
- return SrcReg;
+ return PhysReg;
}
return Reg;
initRegUses();
vrm_ = &getAnalysis<VirtRegMap>();
- if (!spiller_.get()) spiller_.reset(createSpiller());
-
+ if (!rewriter_.get()) rewriter_.reset(createVirtRegRewriter());
+
+ spiller_.reset(createSpiller(mf_, li_, loopInfo, vrm_));
+
initIntervalSets();
linearScan();
// Rewrite spill code and update the PhysRegsUsed set.
- spiller_->runOnMachineFunction(*mf_, *vrm_, li_);
+ rewriter_->runOnMachineFunction(*mf_, *vrm_, li_);
assert(unhandled_.empty() && "Unhandled live intervals remain!");
NextReloadMap.clear();
DowngradedRegs.clear();
DowngradeMap.clear();
+ spiller_.reset(0);
return true;
}
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);
+ }
}
}
-void RALinScan::linearScan()
-{
+void RALinScan::linearScan() {
// linear scan algorithm
- DOUT << "********** LINEAR SCAN **********\n";
- DOUT << "********** Function: " << mf_->getFunction()->getName() << '\n';
-
- DEBUG(printIntervals("fixed", fixed_.begin(), fixed_.end()));
+ DEBUG({
+ errs() << "********** LINEAR SCAN **********\n"
+ << "********** Function: "
+ << mf_->getFunction()->getName() << '\n';
+ printIntervals("fixed", fixed_.begin(), fixed_.end());
+ });
while (!unhandled_.empty()) {
// pick the interval with the earliest start point
LiveInterval* cur = unhandled_.top();
unhandled_.pop();
++NumIters;
- DOUT << "\n*** CURRENT ***: " << *cur << '\n';
+ DEBUG(errs() << "\n*** CURRENT ***: " << *cur << '\n');
- if (!cur->empty()) {
- processActiveIntervals(cur->beginNumber());
- processInactiveIntervals(cur->beginNumber());
+ 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
// assign it one.
assignRegOrStackSlotAtInterval(cur);
- DEBUG(printIntervals("active", active_.begin(), active_.end()));
- DEBUG(printIntervals("inactive", inactive_.begin(), inactive_.end()));
+ DEBUG({
+ printIntervals("active", active_.begin(), active_.end());
+ printIntervals("inactive", inactive_.begin(), inactive_.end());
+ });
}
// Expire any remaining active intervals
while (!active_.empty()) {
IntervalPtr &IP = active_.back();
unsigned reg = IP.first->reg;
- DOUT << "\tinterval " << *IP.first << " expired\n";
+ DEBUG(errs() << "\tinterval " << *IP.first << " expired\n");
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
reg = vrm_->getPhys(reg);
}
// Expire any remaining inactive intervals
- DEBUG(for (IntervalPtrs::reverse_iterator
- i = inactive_.rbegin(); i != inactive_.rend(); ++i)
- DOUT << "\tinterval " << *i->first << " expired\n");
+ DEBUG({
+ for (IntervalPtrs::reverse_iterator
+ i = inactive_.rbegin(); i != inactive_.rend(); ++i)
+ errs() << "\tinterval " << *i->first << " expired\n";
+ });
inactive_.clear();
// Add live-ins to every BB except for entry. Also perform trivial coalescing.
// Ignore splited live intervals.
if (!isPhys && vrm_->getPreSplitReg(cur.reg))
continue;
+
for (LiveInterval::Ranges::const_iterator I = cur.begin(), E = cur.end();
I != E; ++I) {
const LiveRange &LR = *I;
if (li_->findLiveInMBBs(LR.start, LR.end, LiveInMBBs)) {
for (unsigned i = 0, e = LiveInMBBs.size(); i != e; ++i)
- if (LiveInMBBs[i] != EntryMBB)
+ if (LiveInMBBs[i] != EntryMBB) {
+ assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
+ "Adding a virtual register to livein set?");
LiveInMBBs[i]->addLiveIn(Reg);
+ }
LiveInMBBs.clear();
}
}
}
- DOUT << *vrm_;
+ DEBUG(errs() << *vrm_);
+
+ // Look for physical registers that end up not being allocated even though
+ // register allocator had to spill other registers in its register class.
+ if (ls_->getNumIntervals() == 0)
+ return;
+ if (!vrm_->FindUnusedRegisters(li_))
+ return;
}
/// processActiveIntervals - expire old intervals and move non-overlapping ones
/// to the inactive list.
-void RALinScan::processActiveIntervals(unsigned CurPoint)
+void RALinScan::processActiveIntervals(SlotIndex CurPoint)
{
- DOUT << "\tprocessing active intervals:\n";
+ DEBUG(errs() << "\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.
- DOUT << "\t\tinterval " << *Interval << " expired\n";
+ DEBUG(errs() << "\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.
- DOUT << "\t\tinterval " << *Interval << " inactive\n";
+ DEBUG(errs() << "\t\tinterval " << *Interval << " inactive\n");
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
reg = vrm_->getPhys(reg);
/// processInactiveIntervals - expire old intervals and move overlapping
/// ones to the active list.
-void RALinScan::processInactiveIntervals(unsigned CurPoint)
+void RALinScan::processInactiveIntervals(SlotIndex CurPoint)
{
- DOUT << "\tprocessing inactive intervals:\n";
+ DEBUG(errs() << "\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.
- DOUT << "\t\tinterval " << *Interval << " expired\n";
+ DEBUG(errs() << "\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
- DOUT << "\t\tinterval " << *Interval << " active\n";
+ DEBUG(errs() << "\t\tinterval " << *Interval << " active\n");
assert(TargetRegisterInfo::isVirtualRegister(reg) &&
"Can only allocate virtual registers!");
reg = vrm_->getPhys(reg);
// bl should get the same spill weight otherwise it will be choosen
// as a spill candidate since spilling bh doesn't make ebx available.
for (unsigned i = 0, e = Supers.size(); i != e; ++i) {
- for (const unsigned *sr = tri_->getSubRegisters(Supers[i]); *sr; ++sr)
- if (!Processed.count(*sr))
- Weights[*sr] += weight;
+ for (const unsigned *sr = tri_->getSubRegisters(Supers[i]); *sr; ++sr)
+ if (!Processed.count(*sr))
+ Weights[*sr] += weight;
}
}
return IP.end();
}
-static void RevertVectorIteratorsTo(RALinScan::IntervalPtrs &V, unsigned 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(),
/// addStackInterval - Create a LiveInterval for stack if the specified live
/// interval has been spilled.
static void addStackInterval(LiveInterval *cur, LiveStacks *ls_,
- LiveIntervals *li_, float &Weight,
- VirtRegMap &vrm_) {
+ LiveIntervals *li_,
+ MachineRegisterInfo* mri_, VirtRegMap &vrm_) {
int SS = vrm_.getStackSlot(cur->reg);
if (SS == VirtRegMap::NO_STACK_SLOT)
return;
- LiveInterval &SI = ls_->getOrCreateInterval(SS);
- SI.weight += Weight;
+
+ const TargetRegisterClass *RC = mri_->getRegClass(cur->reg);
+ LiveInterval &SI = ls_->getOrCreateInterval(SS, RC);
VNInfo *VNI;
if (SI.hasAtLeastOneValue())
VNI = SI.getValNumInfo(0);
else
- VNI = SI.getNextValue(~0U, 0, ls_->getVNInfoAllocator());
+ VNI = SI.getNextValue(SlotIndex(), 0, false,
+ ls_->getVNInfoAllocator());
LiveInterval &RI = li_->getInterval(cur->reg);
// FIXME: This may be overly conservative.
/// getConflictWeight - Return the number of conflicts between cur
/// live interval and defs and uses of Reg weighted by loop depthes.
-static float getConflictWeight(LiveInterval *cur, unsigned Reg,
- LiveIntervals *li_,
- MachineRegisterInfo *mri_,
- const MachineLoopInfo *loopInfo) {
+static
+float getConflictWeight(LiveInterval *cur, unsigned Reg, LiveIntervals *li_,
+ MachineRegisterInfo *mri_,
+ const MachineLoopInfo *loopInfo) {
float Conflicts = 0;
for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(Reg),
E = mri_->reg_end(); I != E; ++I) {
float Conflicts[3] = { 0.0f, 0.0f, 0.0f };
SmallVector<LiveInterval*, 8> SLIs[3];
- DOUT << "\tConsidering " << NumCands << " candidates: ";
- DEBUG(for (unsigned i = 0; i != NumCands; ++i)
- DOUT << tri_->getName(Candidates[i].first) << " ";
- DOUT << "\n";);
+ DEBUG({
+ errs() << "\tConsidering " << NumCands << " candidates: ";
+ for (unsigned i = 0; i != NumCands; ++i)
+ errs() << tri_->getName(Candidates[i].first) << " ";
+ errs() << "\n";
+ });
// Calculate the number of conflicts of each candidate.
for (IntervalPtrs::iterator i = active_.begin(); i != active_.end(); ++i) {
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);
}
};
}
namespace {
struct LISorter {
bool operator()(LiveInterval* A, LiveInterval* B) {
- return A->beginNumber() < B->beginNumber();
+ return A->beginIndex() < B->beginIndex();
}
};
}
/// assignRegOrStackSlotAtInterval - assign a register if one is available, or
/// spill.
-void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur)
-{
- DOUT << "\tallocating current interval: ";
+void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) {
+ DEBUG(errs() << "\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 = cur->preference;
+ unsigned physReg = vrm_->getRegAllocPref(cur->reg);
if (!physReg)
physReg = *RC->allocation_order_begin(*mf_);
- DOUT << tri_->getName(physReg) << '\n';
+ DEBUG(errs() << tri_->getName(physReg) << '\n');
// Note the register is not really in use.
vrm_->assignVirt2Phys(cur->reg, physReg);
return;
backUpRegUses();
std::vector<std::pair<unsigned, float> > SpillWeightsToAdd;
- unsigned StartPosition = cur->beginNumber();
+ SlotIndex StartPosition = cur->beginIndex();
const TargetRegisterClass *RCLeader = RelatedRegClasses.getLeaderValue(RC);
// If start of this live interval is defined by a move instruction and its
// register class, then we should try to assign it the same register.
// This can happen when the move is from a larger register class to a smaller
// one, e.g. X86::mov32to32_. These move instructions are not coalescable.
- if (!cur->preference && cur->hasAtLeastOneValue()) {
+ if (!vrm_->getRegAllocPref(cur->reg) && cur->hasAtLeastOneValue()) {
VNInfo *vni = cur->begin()->valno;
- if (vni->def && vni->def != ~1U && vni->def != ~0U) {
+ if ((vni->def != SlotIndex()) && !vni->isUnused() &&
+ vni->isDefAccurate()) {
MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
unsigned SrcReg, DstReg, SrcSubReg, DstSubReg;
if (CopyMI &&
if (DstSubReg)
Reg = tri_->getMatchingSuperReg(Reg, DstSubReg, RC);
if (Reg && allocatableRegs_[Reg] && RC->contains(Reg))
- cur->preference = Reg;
+ mri_->setRegAllocationHint(cur->reg, 0, Reg);
}
}
}
// Okay, this reg is on the fixed list. Check to see if we actually
// conflict.
LiveInterval *I = IP.first;
- if (I->endNumber() > StartPosition) {
+ if (I->endIndex() > StartPosition) {
LiveInterval::iterator II = I->advanceTo(IP.second, StartPosition);
IP.second = II;
if (II != I->begin() && II->start > StartPosition)
const TargetRegisterClass *RegRC = OneClassForEachPhysReg[I->reg];
if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader &&
- I->endNumber() > StartPosition) {
+ I->endIndex() > StartPosition) {
LiveInterval::iterator II = I->advanceTo(IP.second, StartPosition);
IP.second = II;
if (II != I->begin() && II->start > StartPosition)
// the free physical register and add this interval to the active
// list.
if (physReg) {
- DOUT << tri_->getName(physReg) << '\n';
+ DEBUG(errs() << tri_->getName(physReg) << '\n');
vrm_->assignVirt2Phys(cur->reg, physReg);
addRegUse(physReg);
active_.push_back(std::make_pair(cur, cur->begin()));
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.
- NextReloadLI->preference = physReg;
+ mri_->setRegAllocationHint(NextReloadLI->reg, 0, physReg);
DowngradeRegister(cur, physReg);
}
return;
}
- DOUT << "no free registers\n";
+ DEBUG(errs() << "no free registers\n");
// Compile the spill weights into an array that is better for scanning.
std::vector<float> SpillWeights(tri_->getNumRegs(), 0.0f);
updateSpillWeights(SpillWeights, reg, i->first->weight, RC);
}
- DOUT << "\tassigning stack slot at interval "<< *cur << ":\n";
+ DEBUG(errs() << "\tassigning stack slot at interval "<< *cur << ":\n");
// Find a register to spill.
float minWeight = HUGE_VALF;
- unsigned minReg = 0; /*cur->preference*/; // Try the pref register first.
+ unsigned minReg = 0;
bool Found = false;
std::vector<std::pair<unsigned,float> > RegsWeights;
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 {
- cerr << "Ran out of registers during register allocation!\n";
- exit(1);
+ assert(false && "Ran out of registers during register allocation!");
+ llvm_report_error("Ran out of registers during register allocation!");
}
return;
}
--LastCandidate;
}
- DOUT << "\t\tregister(s) with min weight(s): ";
- DEBUG(for (unsigned i = 0; i != LastCandidate; ++i)
- DOUT << tri_->getName(RegsWeights[i].first)
- << " (" << RegsWeights[i].second << ")\n");
+ DEBUG({
+ errs() << "\t\tregister(s) with min weight(s): ";
+
+ for (unsigned i = 0; i != LastCandidate; ++i)
+ errs() << tri_->getName(RegsWeights[i].first)
+ << " (" << RegsWeights[i].second << ")\n";
+ });
// If the current has the minimum weight, we need to spill it and
// add any added intervals back to unhandled, and restart
// linearscan.
if (cur->weight != HUGE_VALF && cur->weight <= minWeight) {
- DOUT << "\t\t\tspilling(c): " << *cur << '\n';
- float SSWeight;
+ DEBUG(errs() << "\t\t\tspilling(c): " << *cur << '\n');
SmallVector<LiveInterval*, 8> spillIs;
- std::vector<LiveInterval*> added =
- li_->addIntervalsForSpills(*cur, spillIs, loopInfo, *vrm_, SSWeight);
+ std::vector<LiveInterval*> added;
+
+ added = spiller_->spill(cur, spillIs);
+
std::sort(added.begin(), added.end(), LISorter());
- addStackInterval(cur, ls_, li_, SSWeight, *vrm_);
+ addStackInterval(cur, ls_, li_, mri_, *vrm_);
if (added.empty())
return; // Early exit if all spills were folded.
LiveInterval *ReloadLi = added[i];
if (ReloadLi->weight == HUGE_VALF &&
li_->getApproximateInstructionCount(*ReloadLi) == 0) {
- unsigned ReloadIdx = ReloadLi->beginNumber();
+ SlotIndex ReloadIdx = ReloadLi->beginIndex();
MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx);
int ReloadSS = vrm_->getStackSlot(ReloadLi->reg);
if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) {
// Last reload of same SS is in the same MBB. We want to try to
// allocate both reloads the same register and make sure the reg
// isn't clobbered in between if at all possible.
- assert(LastReload->beginNumber() < ReloadIdx);
+ assert(LastReload->beginIndex() < ReloadIdx);
NextReloadMap.insert(std::make_pair(LastReload->reg, ReloadLi->reg));
}
LastReloadMBB = ReloadMBB;
// The earliest start of a Spilled interval indicates up to where
// in handled we need to roll back
- unsigned earliestStart = cur->beginNumber();
+
+ 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
while (!spillIs.empty()) {
LiveInterval *sli = spillIs.back();
spillIs.pop_back();
- DOUT << "\t\t\tspilling(a): " << *sli << '\n';
- earliestStart = std::min(earliestStart, sli->beginNumber());
- float SSWeight;
- std::vector<LiveInterval*> newIs =
- li_->addIntervalsForSpills(*sli, spillIs, loopInfo, *vrm_, SSWeight);
- addStackInterval(sli, ls_, li_, SSWeight, *vrm_);
+ DEBUG(errs() << "\t\t\tspilling(a): " << *sli << '\n');
+ earliestStartInterval =
+ (earliestStartInterval->beginIndex() < sli->beginIndex()) ?
+ earliestStartInterval : sli;
+
+ std::vector<LiveInterval*> newIs;
+ newIs = spiller_->spill(sli, spillIs);
+ addStackInterval(sli, ls_, li_, mri_, *vrm_);
std::copy(newIs.begin(), newIs.end(), std::back_inserter(added));
spilled.insert(sli->reg);
}
- DOUT << "\t\trolling back to: " << earliestStart << '\n';
+ SlotIndex earliestStart = earliestStartInterval->beginIndex();
+
+ DEBUG(errs() << "\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->beginNumber() < earliestStart)
+ if (i->beginIndex() < earliestStart)
break;
- DOUT << "\t\t\tundo changes for: " << *i << '\n';
+ DEBUG(errs() << "\t\t\tundo changes for: " << *i << '\n');
handled_.pop_back();
// When undoing a live interval allocation we must know if it is active or
// It interval has a preference, it must be defined by a copy. Clear the
// preference now since the source interval allocation may have been
// undone as well.
- i->preference = 0;
+ mri_->setRegAllocationHint(i->reg, 0, 0);
else {
UpgradeRegister(ii->second);
}
for (unsigned i = 0, e = handled_.size(); i != e; ++i) {
LiveInterval *HI = handled_[i];
if (!HI->expiredAt(earliestStart) &&
- HI->expiredAt(cur->beginNumber())) {
- DOUT << "\t\t\tundo changes for: " << *HI << '\n';
+ HI->expiredAt(cur->beginIndex())) {
+ DEBUG(errs() << "\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));
LiveInterval *ReloadLi = added[i];
if (ReloadLi->weight == HUGE_VALF &&
li_->getApproximateInstructionCount(*ReloadLi) == 0) {
- unsigned ReloadIdx = ReloadLi->beginNumber();
+ SlotIndex ReloadIdx = ReloadLi->beginIndex();
MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx);
int ReloadSS = vrm_->getStackSlot(ReloadLi->reg);
if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) {
// Last reload of same SS is in the same MBB. We want to try to
// allocate both reloads the same register and make sure the reg
// isn't clobbered in between if at all possible.
- assert(LastReload->beginNumber() < ReloadIdx);
+ assert(LastReload->beginIndex() < ReloadIdx);
NextReloadMap.insert(std::make_pair(LastReload->reg, ReloadLi->reg));
}
LastReloadMBB = ReloadMBB;
}
}
-unsigned RALinScan::getFreePhysReg(const TargetRegisterClass *RC,
+unsigned RALinScan::getFreePhysReg(LiveInterval* cur,
+ const TargetRegisterClass *RC,
unsigned MaxInactiveCount,
SmallVector<unsigned, 256> &inactiveCounts,
bool SkipDGRegs) {
unsigned FreeReg = 0;
unsigned FreeRegInactiveCount = 0;
- TargetRegisterClass::iterator I = RC->allocation_order_begin(*mf_);
- TargetRegisterClass::iterator E = RC->allocation_order_end(*mf_);
+ std::pair<unsigned, unsigned> Hint = mri_->getRegAllocationHint(cur->reg);
+ // Resolve second part of the hint (if possible) given the current allocation.
+ unsigned physReg = Hint.second;
+ if (physReg &&
+ TargetRegisterInfo::isVirtualRegister(physReg) && vrm_->hasPhys(physReg))
+ physReg = vrm_->getPhys(physReg);
+
+ TargetRegisterClass::iterator I, E;
+ tie(I, E) = tri_->getAllocationOrder(RC, Hint.first, physReg, *mf_);
assert(I != E && "No allocatable register in this register class!");
// Scan for the first available register.
// 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;
}
// If copy coalescer has assigned a "preferred" register, check if it's
// available first.
- if (cur->preference) {
- DOUT << "(preferred: " << tri_->getName(cur->preference) << ") ";
- if (isRegAvail(cur->preference) &&
- RC->contains(cur->preference))
- return cur->preference;
+ unsigned Preference = vrm_->getRegAllocPref(cur->reg);
+ if (Preference) {
+ DEBUG(errs() << "(preferred: " << tri_->getName(Preference) << ") ");
+ if (isRegAvail(Preference) &&
+ RC->contains(Preference))
+ return Preference;
}
if (!DowngradedRegs.empty()) {
- unsigned FreeReg = getFreePhysReg(RC, MaxInactiveCount, inactiveCounts,
+ unsigned FreeReg = getFreePhysReg(cur, RC, MaxInactiveCount, inactiveCounts,
true);
if (FreeReg)
return FreeReg;
}
- return getFreePhysReg(RC, MaxInactiveCount, inactiveCounts, false);
+ return getFreePhysReg(cur, RC, MaxInactiveCount, inactiveCounts, false);
}
FunctionPass* llvm::createLinearScanRegisterAllocator() {