//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "sched-instrs"
-#include "RegisterPressure.h"
-#include "llvm/Operator.h"
+#define DEBUG_TYPE "misched"
+#include "llvm/CodeGen/ScheduleDAGInstrs.h"
+#include "llvm/ADT/MapVector.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
-#include "llvm/CodeGen/ScheduleDAGInstrs.h"
+#include "llvm/CodeGen/RegisterPressure.h"
+#include "llvm/CodeGen/ScheduleDFS.h"
+#include "llvm/IR/Operator.h"
#include "llvm/MC/MCInstrItineraries.h"
-#include "llvm/Target/TargetMachine.h"
-#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Target/TargetSubtargetInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/Format.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
+#include <queue>
+
using namespace llvm;
static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
const MachineDominatorTree &mdt,
bool IsPostRAFlag,
LiveIntervals *lis)
- : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()),
- InstrItins(mf.getTarget().getInstrItineraryData()), LIS(lis),
- IsPostRA(IsPostRAFlag), UnitLatencies(false), CanHandleTerminators(false),
- LoopRegs(MLI, MDT), FirstDbgValue(0) {
+ : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()), LIS(lis),
+ IsPostRA(IsPostRAFlag), CanHandleTerminators(false), FirstDbgValue(0) {
assert((IsPostRA || LIS) && "PreRA scheduling requires LiveIntervals");
DbgValues.clear();
assert(!(IsPostRA && MRI.getNumVirtRegs()) &&
"Virtual registers must be removed prior to PostRA scheduling");
+
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ SchedModel.init(*ST.getSchedModel(), &ST, TII);
}
/// getUnderlyingObjectFromInt - This is the function that does the work of
// regular getUnderlyingObjectFromInt.
if (U->getOpcode() == Instruction::PtrToInt)
return U->getOperand(0);
- // If we find an add of a constant or a multiplied value, it's
+ // If we find an add of a constant, a multiplied value, or a phi, it's
// likely that the other operand will lead us to the base
// object. We don't have to worry about the case where the
// object address is somehow being computed by the multiply,
// because our callers only care when the result is an
- // identifibale object.
+ // identifiable object.
if (U->getOpcode() != Instruction::Add ||
(!isa<ConstantInt>(U->getOperand(1)) &&
- Operator::getOpcode(U->getOperand(1)) != Instruction::Mul))
+ Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
+ !isa<PHINode>(U->getOperand(1))))
return V;
V = U->getOperand(0);
} else {
} while (1);
}
-/// getUnderlyingObject - This is a wrapper around GetUnderlyingObject
+/// getUnderlyingObjects - This is a wrapper around GetUnderlyingObjects
/// and adds support for basic ptrtoint+arithmetic+inttoptr sequences.
-static const Value *getUnderlyingObject(const Value *V) {
- // First just call Value::getUnderlyingObject to let it do what it does.
+static void getUnderlyingObjects(const Value *V,
+ SmallVectorImpl<Value *> &Objects) {
+ SmallPtrSet<const Value*, 16> Visited;
+ SmallVector<const Value *, 4> Working(1, V);
do {
- V = GetUnderlyingObject(V);
- // If it found an inttoptr, use special code to continue climing.
- if (Operator::getOpcode(V) != Instruction::IntToPtr)
- break;
- const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
- // If that succeeded in finding a pointer, continue the search.
- if (!O->getType()->isPointerTy())
- break;
- V = O;
- } while (1);
- return V;
+ V = Working.pop_back_val();
+
+ SmallVector<Value *, 4> Objs;
+ GetUnderlyingObjects(const_cast<Value *>(V), Objs);
+
+ for (SmallVectorImpl<Value *>::iterator I = Objs.begin(), IE = Objs.end();
+ I != IE; ++I) {
+ V = *I;
+ if (!Visited.insert(V))
+ continue;
+ if (Operator::getOpcode(V) == Instruction::IntToPtr) {
+ const Value *O =
+ getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
+ if (O->getType()->isPointerTy()) {
+ Working.push_back(O);
+ continue;
+ }
+ }
+ Objects.push_back(const_cast<Value *>(V));
+ }
+ } while (!Working.empty());
}
-/// getUnderlyingObjectForInstr - If this machine instr has memory reference
+typedef SmallVector<PointerIntPair<const Value *, 1, bool>, 4>
+UnderlyingObjectsVector;
+
+/// getUnderlyingObjectsForInstr - If this machine instr has memory reference
/// information and it can be tracked to a normal reference to a known
-/// object, return the Value for that object. Otherwise return null.
-static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI,
- const MachineFrameInfo *MFI,
- bool &MayAlias) {
- MayAlias = true;
+/// object, return the Value for that object.
+static void getUnderlyingObjectsForInstr(const MachineInstr *MI,
+ const MachineFrameInfo *MFI,
+ UnderlyingObjectsVector &Objects) {
if (!MI->hasOneMemOperand() ||
!(*MI->memoperands_begin())->getValue() ||
(*MI->memoperands_begin())->isVolatile())
- return 0;
+ return;
const Value *V = (*MI->memoperands_begin())->getValue();
if (!V)
- return 0;
-
- V = getUnderlyingObject(V);
- if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
- // For now, ignore PseudoSourceValues which may alias LLVM IR values
- // because the code that uses this function has no way to cope with
- // such aliases.
- if (PSV->isAliased(MFI))
- return 0;
-
- MayAlias = PSV->mayAlias(MFI);
- return V;
- }
+ return;
+
+ SmallVector<Value *, 4> Objs;
+ getUnderlyingObjects(V, Objs);
+
+ for (SmallVectorImpl<Value *>::iterator I = Objs.begin(), IE = Objs.end();
+ I != IE; ++I) {
+ bool MayAlias = true;
+ V = *I;
+
+ if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
+ // For now, ignore PseudoSourceValues which may alias LLVM IR values
+ // because the code that uses this function has no way to cope with
+ // such aliases.
- if (isIdentifiedObject(V))
- return V;
+ if (PSV->isAliased(MFI)) {
+ Objects.clear();
+ return;
+ }
+
+ MayAlias = PSV->mayAlias(MFI);
+ } else if (!isIdentifiedObject(V)) {
+ Objects.clear();
+ return;
+ }
- return 0;
+ Objects.push_back(UnderlyingObjectsVector::value_type(V, MayAlias));
+ }
}
void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
BB = bb;
- LoopRegs.Deps.clear();
- if (MachineLoop *ML = MLI.getLoopFor(BB))
- if (BB == ML->getLoopLatch())
- LoopRegs.VisitLoop(ML);
}
void ScheduleDAGInstrs::finishBlock() {
BB = 0;
}
-/// Initialize the map with the number of registers.
-void Reg2SUnitsMap::setRegLimit(unsigned Limit) {
- PhysRegSet.setUniverse(Limit);
- SUnits.resize(Limit);
-}
-
-/// Clear the map without deallocating storage.
-void Reg2SUnitsMap::clear() {
- for (const_iterator I = reg_begin(), E = reg_end(); I != E; ++I) {
- SUnits[*I].clear();
- }
- PhysRegSet.clear();
-}
-
/// Initialize the DAG and common scheduler state for the current scheduling
/// region. This does not actually create the DAG, only clears it. The
/// scheduling driver may call BuildSchedGraph multiple times per scheduling
void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
MachineBasicBlock::iterator begin,
MachineBasicBlock::iterator end,
- unsigned endcount) {
+ unsigned regioninstrs) {
assert(bb == BB && "startBlock should set BB");
RegionBegin = begin;
RegionEnd = end;
- EndIndex = endcount;
- MISUnitMap.clear();
-
- // Check to see if the scheduler cares about latencies.
- UnitLatencies = forceUnitLatencies();
-
- ScheduleDAG::clearDAG();
+ NumRegionInstrs = regioninstrs;
}
/// Close the current scheduling region. Don't clear any state in case the
if (Reg == 0) continue;
if (TRI->isPhysicalRegister(Reg))
- Uses[Reg].push_back(&ExitSU);
+ Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
else {
assert(!IsPostRA && "Virtual register encountered after regalloc.");
- addVRegUseDeps(&ExitSU, i);
+ if (MO.readsReg()) // ignore undef operands
+ addVRegUseDeps(&ExitSU, i);
}
}
} else {
E = (*SI)->livein_end(); I != E; ++I) {
unsigned Reg = *I;
if (!Uses.contains(Reg))
- Uses[Reg].push_back(&ExitSU);
+ Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
}
}
}
/// MO is an operand of SU's instruction that defines a physical register. Add
/// data dependencies from SU to any uses of the physical register.
-void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU,
- const MachineOperand &MO) {
+void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
+ const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
assert(MO.isDef() && "expect physreg def");
// Ask the target if address-backscheduling is desirable, and if so how much.
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
- unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
- unsigned DataLatency = SU->Latency;
- for (const uint16_t *Alias = TRI->getOverlaps(MO.getReg()); *Alias; ++Alias) {
+ for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
+ Alias.isValid(); ++Alias) {
if (!Uses.contains(*Alias))
continue;
- std::vector<SUnit*> &UseList = Uses[*Alias];
- for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
- SUnit *UseSU = UseList[i];
+ for (Reg2SUnitsMap::iterator I = Uses.find(*Alias); I != Uses.end(); ++I) {
+ SUnit *UseSU = I->SU;
if (UseSU == SU)
continue;
- unsigned LDataLatency = DataLatency;
- // Optionally add in a special extra latency for nodes that
- // feed addresses.
- // TODO: Perhaps we should get rid of
- // SpecialAddressLatency and just move this into
- // adjustSchedDependency for the targets that care about it.
- if (SpecialAddressLatency != 0 && !UnitLatencies &&
- UseSU != &ExitSU) {
- MachineInstr *UseMI = UseSU->getInstr();
- const MCInstrDesc &UseMCID = UseMI->getDesc();
- int RegUseIndex = UseMI->findRegisterUseOperandIdx(*Alias);
- assert(RegUseIndex >= 0 && "UseMI doesn't use register!");
- if (RegUseIndex >= 0 &&
- (UseMI->mayLoad() || UseMI->mayStore()) &&
- (unsigned)RegUseIndex < UseMCID.getNumOperands() &&
- UseMCID.OpInfo[RegUseIndex].isLookupPtrRegClass())
- LDataLatency += SpecialAddressLatency;
- }
- // Adjust the dependence latency using operand def/use
- // information (if any), and then allow the target to
- // perform its own adjustments.
- const SDep& dep = SDep(SU, SDep::Data, LDataLatency, *Alias);
- if (!UnitLatencies) {
- computeOperandLatency(SU, UseSU, const_cast<SDep &>(dep));
- ST.adjustSchedDependency(SU, UseSU, const_cast<SDep &>(dep));
+
+ // Adjust the dependence latency using operand def/use information,
+ // then allow the target to perform its own adjustments.
+ int UseOp = I->OpIdx;
+ MachineInstr *RegUse = 0;
+ SDep Dep;
+ if (UseOp < 0)
+ Dep = SDep(SU, SDep::Artificial);
+ else {
+ // Set the hasPhysRegDefs only for physreg defs that have a use within
+ // the scheduling region.
+ SU->hasPhysRegDefs = true;
+ Dep = SDep(SU, SDep::Data, *Alias);
+ RegUse = UseSU->getInstr();
}
- UseSU->addPred(dep);
+ Dep.setLatency(
+ SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
+ UseOp));
+
+ ST.adjustSchedDependency(SU, UseSU, Dep);
+ UseSU->addPred(Dep);
}
}
}
// TODO: Using a latency of 1 here for output dependencies assumes
// there's no cost for reusing registers.
SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
- for (const uint16_t *Alias = TRI->getOverlaps(MO.getReg()); *Alias; ++Alias) {
+ for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
+ Alias.isValid(); ++Alias) {
if (!Defs.contains(*Alias))
continue;
- std::vector<SUnit *> &DefList = Defs[*Alias];
- for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
- SUnit *DefSU = DefList[i];
+ for (Reg2SUnitsMap::iterator I = Defs.find(*Alias); I != Defs.end(); ++I) {
+ SUnit *DefSU = I->SU;
if (DefSU == &ExitSU)
continue;
if (DefSU != SU &&
(Kind != SDep::Output || !MO.isDead() ||
!DefSU->getInstr()->registerDefIsDead(*Alias))) {
if (Kind == SDep::Anti)
- DefSU->addPred(SDep(SU, Kind, 0, /*Reg=*/*Alias));
+ DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
else {
- unsigned AOLat = TII->getOutputLatency(InstrItins, MI, OperIdx,
- DefSU->getInstr());
- DefSU->addPred(SDep(SU, Kind, AOLat, /*Reg=*/*Alias));
+ SDep Dep(SU, Kind, /*Reg=*/*Alias);
+ Dep.setLatency(
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
+ DefSU->addPred(Dep);
}
}
}
}
if (!MO.isDef()) {
+ SU->hasPhysRegUses = true;
// Either insert a new Reg2SUnits entry with an empty SUnits list, or
// retrieve the existing SUnits list for this register's uses.
// Push this SUnit on the use list.
- Uses[MO.getReg()].push_back(SU);
+ Uses.insert(PhysRegSUOper(SU, OperIdx, MO.getReg()));
}
else {
- addPhysRegDataDeps(SU, MO);
+ addPhysRegDataDeps(SU, OperIdx);
+ unsigned Reg = MO.getReg();
- // Either insert a new Reg2SUnits entry with an empty SUnits list, or
- // retrieve the existing SUnits list for this register's defs.
- std::vector<SUnit *> &DefList = Defs[MO.getReg()];
-
- // If a def is going to wrap back around to the top of the loop,
- // backschedule it.
- if (!UnitLatencies && DefList.empty()) {
- LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(MO.getReg());
- if (I != LoopRegs.Deps.end()) {
- const MachineOperand *UseMO = I->second.first;
- unsigned Count = I->second.second;
- const MachineInstr *UseMI = UseMO->getParent();
- unsigned UseMOIdx = UseMO - &UseMI->getOperand(0);
- const MCInstrDesc &UseMCID = UseMI->getDesc();
- const TargetSubtargetInfo &ST =
- TM.getSubtarget<TargetSubtargetInfo>();
- unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
- // TODO: If we knew the total depth of the region here, we could
- // handle the case where the whole loop is inside the region but
- // is large enough that the isScheduleHigh trick isn't needed.
- if (UseMOIdx < UseMCID.getNumOperands()) {
- // Currently, we only support scheduling regions consisting of
- // single basic blocks. Check to see if the instruction is in
- // the same region by checking to see if it has the same parent.
- if (UseMI->getParent() != MI->getParent()) {
- unsigned Latency = SU->Latency;
- if (UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass())
- Latency += SpecialAddressLatency;
- // This is a wild guess as to the portion of the latency which
- // will be overlapped by work done outside the current
- // scheduling region.
- Latency -= std::min(Latency, Count);
- // Add the artificial edge.
- ExitSU.addPred(SDep(SU, SDep::Order, Latency,
- /*Reg=*/0, /*isNormalMemory=*/false,
- /*isMustAlias=*/false,
- /*isArtificial=*/true));
- } else if (SpecialAddressLatency > 0 &&
- UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass()) {
- // The entire loop body is within the current scheduling region
- // and the latency of this operation is assumed to be greater
- // than the latency of the loop.
- // TODO: Recursively mark data-edge predecessors as
- // isScheduleHigh too.
- SU->isScheduleHigh = true;
- }
- }
- LoopRegs.Deps.erase(I);
+ // clear this register's use list
+ if (Uses.contains(Reg))
+ Uses.eraseAll(Reg);
+
+ if (!MO.isDead()) {
+ Defs.eraseAll(Reg);
+ } else if (SU->isCall) {
+ // Calls will not be reordered because of chain dependencies (see
+ // below). Since call operands are dead, calls may continue to be added
+ // to the DefList making dependence checking quadratic in the size of
+ // the block. Instead, we leave only one call at the back of the
+ // DefList.
+ Reg2SUnitsMap::RangePair P = Defs.equal_range(Reg);
+ Reg2SUnitsMap::iterator B = P.first;
+ Reg2SUnitsMap::iterator I = P.second;
+ for (bool isBegin = I == B; !isBegin; /* empty */) {
+ isBegin = (--I) == B;
+ if (!I->SU->isCall)
+ break;
+ I = Defs.erase(I);
}
}
- // clear this register's use list
- if (Uses.contains(MO.getReg()))
- Uses[MO.getReg()].clear();
-
- if (!MO.isDead())
- DefList.clear();
-
- // Calls will not be reordered because of chain dependencies (see
- // below). Since call operands are dead, calls may continue to be added
- // to the DefList making dependence checking quadratic in the size of
- // the block. Instead, we leave only one call at the back of the
- // DefList.
- if (SU->isCall) {
- while (!DefList.empty() && DefList.back()->isCall)
- DefList.pop_back();
- }
// Defs are pushed in the order they are visited and never reordered.
- DefList.push_back(SU);
+ Defs.insert(PhysRegSUOper(SU, OperIdx, Reg));
}
}
const MachineInstr *MI = SU->getInstr();
unsigned Reg = MI->getOperand(OperIdx).getReg();
- // SSA defs do not have output/anti dependencies.
+ // Singly defined vregs do not have output/anti dependencies.
// The current operand is a def, so we have at least one.
- if (llvm::next(MRI.def_begin(Reg)) == MRI.def_end())
+ // Check here if there are any others...
+ if (MRI.hasOneDef(Reg))
return;
// Add output dependence to the next nearest def of this vreg.
else {
SUnit *DefSU = DefI->SU;
if (DefSU != SU && DefSU != &ExitSU) {
- unsigned OutLatency = TII->getOutputLatency(InstrItins, MI, OperIdx,
- DefSU->getInstr());
- DefSU->addPred(SDep(SU, SDep::Output, OutLatency, Reg));
+ SDep Dep(SU, SDep::Output, Reg);
+ Dep.setLatency(
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
+ DefSU->addPred(Dep);
}
DefI->SU = SU;
}
MachineInstr *MI = SU->getInstr();
unsigned Reg = MI->getOperand(OperIdx).getReg();
+ // Record this local VReg use.
+ VReg2UseMap::iterator UI = VRegUses.find(Reg);
+ for (; UI != VRegUses.end(); ++UI) {
+ if (UI->SU == SU)
+ break;
+ }
+ if (UI == VRegUses.end())
+ VRegUses.insert(VReg2SUnit(Reg, SU));
+
// Lookup this operand's reaching definition.
assert(LIS && "vreg dependencies requires LiveIntervals");
- SlotIndex UseIdx = LIS->getInstructionIndex(MI).getRegSlot();
- LiveInterval *LI = &LIS->getInterval(Reg);
- VNInfo *VNI = LI->getVNInfoBefore(UseIdx);
+ LiveQueryResult LRQ
+ = LIS->getInterval(Reg).Query(LIS->getInstructionIndex(MI));
+ VNInfo *VNI = LRQ.valueIn();
- // Special case: An early-clobber tied operand reads and writes the
- // register one slot early. e.g. InlineAsm.
- //
- // FIXME: Same special case is in shrinkToUses. Hide under an API.
- if (SlotIndex::isSameInstr(VNI->def, UseIdx)) {
- UseIdx = VNI->def;
- VNI = LI->getVNInfoBefore(UseIdx);
- }
// VNI will be valid because MachineOperand::readsReg() is checked by caller.
+ assert(VNI && "No value to read by operand");
MachineInstr *Def = LIS->getInstructionFromIndex(VNI->def);
// Phis and other noninstructions (after coalescing) have a NULL Def.
if (Def) {
if (DefSU) {
// The reaching Def lives within this scheduling region.
// Create a data dependence.
- //
- // TODO: Handle "special" address latencies cleanly.
- const SDep &dep = SDep(DefSU, SDep::Data, DefSU->Latency, Reg);
- if (!UnitLatencies) {
- // Adjust the dependence latency using operand def/use information, then
- // allow the target to perform its own adjustments.
- computeOperandLatency(DefSU, SU, const_cast<SDep &>(dep));
- const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
- ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
- }
+ SDep dep(DefSU, SDep::Data, Reg);
+ // Adjust the dependence latency using operand def/use information, then
+ // allow the target to perform its own adjustments.
+ int DefOp = Def->findRegisterDefOperandIdx(Reg);
+ dep.setLatency(SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx));
+
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
SU->addPred(dep);
}
}
// Add antidependence to the following def of the vreg it uses.
VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
if (DefI != VRegDefs.end() && DefI->SU != SU)
- DefI->SU->addPred(SDep(SU, SDep::Anti, 0, Reg));
+ DefI->SU->addPred(SDep(SU, SDep::Anti, Reg));
}
/// Return true if MI is an instruction we are unable to reason about
/// (like a call or something with unmodeled side effects).
static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
if (MI->isCall() || MI->hasUnmodeledSideEffects() ||
- (MI->hasVolatileMemoryRef() &&
+ (MI->hasOrderedMemoryRef() &&
(!MI->mayLoad() || !MI->isInvariantLoad(AA))))
return true;
return false;
if ((*MI->memoperands_begin())->isVolatile() ||
MI->hasUnmodeledSideEffects())
return true;
-
const Value *V = (*MI->memoperands_begin())->getValue();
if (!V)
return true;
- V = getUnderlyingObject(V);
- if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
- // Similarly to getUnderlyingObjectForInstr:
- // For now, ignore PseudoSourceValues which may alias LLVM IR values
- // because the code that uses this function has no way to cope with
- // such aliases.
- if (PSV->isAliased(MFI))
+ SmallVector<Value *, 4> Objs;
+ getUnderlyingObjects(V, Objs);
+ for (SmallVectorImpl<Value *>::iterator I = Objs.begin(),
+ IE = Objs.end(); I != IE; ++I) {
+ V = *I;
+
+ if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
+ // Similarly to getUnderlyingObjectForInstr:
+ // For now, ignore PseudoSourceValues which may alias LLVM IR values
+ // because the code that uses this function has no way to cope with
+ // such aliases.
+ if (PSV->isAliased(MFI))
+ return true;
+ }
+
+ // Does this pointer refer to a distinct and identifiable object?
+ if (!isIdentifiedObject(V))
return true;
}
- // Does this pointer refer to a distinct and identifiable object?
- if (!isIdentifiedObject(V))
- return true;
return false;
}
// and stop descending.
if (*Depth > 200 ||
MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
- SUb->addPred(SDep(SUa, SDep::Order, /*Latency=*/0, /*Reg=*/0,
- /*isNormalMemory=*/true));
+ SUb->addPred(SDep(SUa, SDep::MayAliasMem));
return *Depth;
}
// Track current depth.
/// checks whether SU can be aliasing any node dominated
/// by it.
static void adjustChainDeps(AliasAnalysis *AA, const MachineFrameInfo *MFI,
- SUnit *SU, SUnit *ExitSU, std::set<SUnit *> &CheckList) {
+ SUnit *SU, SUnit *ExitSU, std::set<SUnit *> &CheckList,
+ unsigned LatencyToLoad) {
if (!SU)
return;
I != IE; ++I) {
if (SU == *I)
continue;
- if (MIsNeedChainEdge(AA, MFI, SU->getInstr(), (*I)->getInstr()))
- (*I)->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0,
- /*isNormalMemory=*/true));
+ if (MIsNeedChainEdge(AA, MFI, SU->getInstr(), (*I)->getInstr())) {
+ SDep Dep(SU, SDep::MayAliasMem);
+ Dep.setLatency(((*I)->getInstr()->mayLoad()) ? LatencyToLoad : 0);
+ (*I)->addPred(Dep);
+ }
// Now go through all the chain successors and iterate from them.
// Keep track of visited nodes.
for (SUnit::const_succ_iterator J = (*I)->Succs.begin(),
bool isNormalMemory = false) {
// If this is a false dependency,
// do not add the edge, but rememeber the rejected node.
- if (!EnableAASchedMI ||
- MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr()))
- SUb->addPred(SDep(SUa, SDep::Order, TrueMemOrderLatency, /*Reg=*/0,
- isNormalMemory));
+ if (!AA || MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
+ SDep Dep(SUa, isNormalMemory ? SDep::MayAliasMem : SDep::Barrier);
+ Dep.setLatency(TrueMemOrderLatency);
+ SUb->addPred(Dep);
+ }
else {
// Duplicate entries should be ignored.
RejectList.insert(SUb);
void ScheduleDAGInstrs::initSUnits() {
// We'll be allocating one SUnit for each real instruction in the region,
// which is contained within a basic block.
- SUnits.reserve(BB->size());
+ SUnits.reserve(NumRegionInstrs);
for (MachineBasicBlock::iterator I = RegionBegin; I != RegionEnd; ++I) {
MachineInstr *MI = I;
SU->isCommutable = MI->isCommutable();
// Assign the Latency field of SU using target-provided information.
- if (UnitLatencies)
- SU->Latency = 1;
- else
- computeLatency(SU);
+ SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
}
}
/// DAG builder is an efficient place to do it because it already visits
/// operands.
void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
- RegPressureTracker *RPTracker) {
+ RegPressureTracker *RPTracker,
+ PressureDiffs *PDiffs) {
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
+ : ST.useAA();
+ AliasAnalysis *AAForDep = UseAA ? AA : 0;
+
+ MISUnitMap.clear();
+ ScheduleDAG::clearDAG();
+
// Create an SUnit for each real instruction.
initSUnits();
+ if (PDiffs)
+ PDiffs->init(SUnits.size());
+
// We build scheduling units by walking a block's instruction list from bottom
// to top.
// so that they can be given more precise dependencies. We track
// separately the known memory locations that may alias and those
// that are known not to alias
- std::map<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs;
- std::map<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
+ MapVector<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs;
+ MapVector<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
std::set<SUnit*> RejectMemNodes;
// Remove any stale debug info; sometimes BuildSchedGraph is called again
assert(Defs.empty() && Uses.empty() &&
"Only BuildGraph should update Defs/Uses");
- Defs.setRegLimit(TRI->getNumRegs());
- Uses.setRegLimit(TRI->getNumRegs());
+ Defs.setUniverse(TRI->getNumRegs());
+ Uses.setUniverse(TRI->getNumRegs());
assert(VRegDefs.empty() && "Only BuildSchedGraph may access VRegDefs");
- // FIXME: Allow SparseSet to reserve space for the creation of virtual
- // registers during scheduling. Don't artificially inflate the Universe
- // because we want to assert that vregs are not created during DAG building.
+ VRegUses.clear();
VRegDefs.setUniverse(MRI.getNumVirtRegs());
+ VRegUses.setUniverse(MRI.getNumVirtRegs());
// Model data dependencies between instructions being scheduled and the
// ExitSU.
addSchedBarrierDeps();
// Walk the list of instructions, from bottom moving up.
- MachineInstr *PrevMI = NULL;
+ MachineInstr *DbgMI = NULL;
for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
MII != MIE; --MII) {
MachineInstr *MI = prior(MII);
- if (MI && PrevMI) {
- DbgValues.push_back(std::make_pair(PrevMI, MI));
- PrevMI = NULL;
+ if (MI && DbgMI) {
+ DbgValues.push_back(std::make_pair(DbgMI, MI));
+ DbgMI = NULL;
}
if (MI->isDebugValue()) {
- PrevMI = MI;
+ DbgMI = MI;
continue;
}
+ SUnit *SU = MISUnitMap[MI];
+ assert(SU && "No SUnit mapped to this MI");
+
if (RPTracker) {
- RPTracker->recede();
+ PressureDiff *PDiff = PDiffs ? &(*PDiffs)[SU->NodeNum] : 0;
+ RPTracker->recede(/*LiveUses=*/0, PDiff);
assert(RPTracker->getPos() == prior(MII) && "RPTracker can't find MI");
}
- assert((!MI->isTerminator() || CanHandleTerminators) && !MI->isLabel() &&
+ assert((CanHandleTerminators || (!MI->isTerminator() && !MI->isLabel())) &&
"Cannot schedule terminators or labels!");
- SUnit *SU = MISUnitMap[MI];
- assert(SU && "No SUnit mapped to this MI");
-
// Add register-based dependencies (data, anti, and output).
+ bool HasVRegDef = false;
for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
const MachineOperand &MO = MI->getOperand(j);
if (!MO.isReg()) continue;
addPhysRegDeps(SU, j);
else {
assert(!IsPostRA && "Virtual register encountered!");
- if (MO.isDef())
+ if (MO.isDef()) {
+ HasVRegDef = true;
addVRegDefDeps(SU, j);
+ }
else if (MO.readsReg()) // ignore undef operands
addVRegUseDeps(SU, j);
}
}
+ // If we haven't seen any uses in this scheduling region, create a
+ // dependence edge to ExitSU to model the live-out latency. This is required
+ // for vreg defs with no in-region use, and prefetches with no vreg def.
+ //
+ // FIXME: NumDataSuccs would be more precise than NumSuccs here. This
+ // check currently relies on being called before adding chain deps.
+ if (SU->NumSuccs == 0 && SU->Latency > 1
+ && (HasVRegDef || MI->mayLoad())) {
+ SDep Dep(SU, SDep::Artificial);
+ Dep.setLatency(SU->Latency - 1);
+ ExitSU.addPred(Dep);
+ }
// Add chain dependencies.
// Chain dependencies used to enforce memory order should have
// after stack slots are lowered to actual addresses.
// TODO: Use an AliasAnalysis and do real alias-analysis queries, and
// produce more precise dependence information.
-#define STORE_LOAD_LATENCY 1
- unsigned TrueMemOrderLatency = 0;
+ unsigned TrueMemOrderLatency = MI->mayStore() ? 1 : 0;
if (isGlobalMemoryObject(AA, MI)) {
// Be conservative with these and add dependencies on all memory
// references, even those that are known to not alias.
- for (std::map<const Value *, SUnit *>::iterator I =
+ for (MapVector<const Value *, SUnit *>::iterator I =
NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
- I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ I->second->addPred(SDep(SU, SDep::Barrier));
}
- for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
+ for (MapVector<const Value *, std::vector<SUnit *> >::iterator I =
NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) {
- for (unsigned i = 0, e = I->second.size(); i != e; ++i)
- I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
+ for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
+ SDep Dep(SU, SDep::Barrier);
+ Dep.setLatency(TrueMemOrderLatency);
+ I->second[i]->addPred(Dep);
+ }
}
// Add SU to the barrier chain.
if (BarrierChain)
- BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ BarrierChain->addPred(SDep(SU, SDep::Barrier));
BarrierChain = SU;
// This is a barrier event that acts as a pivotal node in the DAG,
// so it is safe to clear list of exposed nodes.
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
+ TrueMemOrderLatency);
RejectMemNodes.clear();
NonAliasMemDefs.clear();
NonAliasMemUses.clear();
// fall-through
new_alias_chain:
// Chain all possibly aliasing memory references though SU.
- if (AliasChain)
- addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
+ if (AliasChain) {
+ unsigned ChainLatency = 0;
+ if (AliasChain->getInstr()->mayLoad())
+ ChainLatency = TrueMemOrderLatency;
+ addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes,
+ ChainLatency);
+ }
AliasChain = SU;
for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
- addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
+ addChainDependency(AAForDep, MFI, SU, PendingLoads[k], RejectMemNodes,
TrueMemOrderLatency);
- for (std::map<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(),
+ for (MapVector<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(),
E = AliasMemDefs.end(); I != E; ++I)
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
- for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes);
+ for (MapVector<const Value *, std::vector<SUnit *> >::iterator I =
AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
- addChainDependency(AA, MFI, SU, I->second[i], RejectMemNodes,
+ addChainDependency(AAForDep, MFI, SU, I->second[i], RejectMemNodes,
TrueMemOrderLatency);
}
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
+ TrueMemOrderLatency);
PendingLoads.clear();
AliasMemDefs.clear();
AliasMemUses.clear();
} else if (MI->mayStore()) {
- bool MayAlias = true;
- TrueMemOrderLatency = STORE_LOAD_LATENCY;
- if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
+ UnderlyingObjectsVector Objs;
+ getUnderlyingObjectsForInstr(MI, MFI, Objs);
+
+ if (Objs.empty()) {
+ // Treat all other stores conservatively.
+ goto new_alias_chain;
+ }
+
+ bool MayAlias = false;
+ for (UnderlyingObjectsVector::iterator K = Objs.begin(), KE = Objs.end();
+ K != KE; ++K) {
+ const Value *V = K->getPointer();
+ bool ThisMayAlias = K->getInt();
+ if (ThisMayAlias)
+ MayAlias = true;
+
// A store to a specific PseudoSourceValue. Add precise dependencies.
// Record the def in MemDefs, first adding a dep if there is
// an existing def.
- std::map<const Value *, SUnit *>::iterator I =
- ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
- std::map<const Value *, SUnit *>::iterator IE =
- ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
+ MapVector<const Value *, SUnit *>::iterator I =
+ ((ThisMayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
+ MapVector<const Value *, SUnit *>::iterator IE =
+ ((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
if (I != IE) {
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes,
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes,
0, true);
I->second = SU;
} else {
- if (MayAlias)
+ if (ThisMayAlias)
AliasMemDefs[V] = SU;
else
NonAliasMemDefs[V] = SU;
}
// Handle the uses in MemUses, if there are any.
- std::map<const Value *, std::vector<SUnit *> >::iterator J =
- ((MayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
- std::map<const Value *, std::vector<SUnit *> >::iterator JE =
- ((MayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
+ MapVector<const Value *, std::vector<SUnit *> >::iterator J =
+ ((ThisMayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
+ MapVector<const Value *, std::vector<SUnit *> >::iterator JE =
+ ((ThisMayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
if (J != JE) {
for (unsigned i = 0, e = J->second.size(); i != e; ++i)
- addChainDependency(AA, MFI, SU, J->second[i], RejectMemNodes,
+ addChainDependency(AAForDep, MFI, SU, J->second[i], RejectMemNodes,
TrueMemOrderLatency, true);
J->second.clear();
}
- if (MayAlias) {
- // Add dependencies from all the PendingLoads, i.e. loads
- // with no underlying object.
- for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
- addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
- TrueMemOrderLatency);
- // Add dependence on alias chain, if needed.
- if (AliasChain)
- addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
- // But we also should check dependent instructions for the
- // SU in question.
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
- }
- // Add dependence on barrier chain, if needed.
- // There is no point to check aliasing on barrier event. Even if
- // SU and barrier _could_ be reordered, they should not. In addition,
- // we have lost all RejectMemNodes below barrier.
- if (BarrierChain)
- BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
- } else {
- // Treat all other stores conservatively.
- goto new_alias_chain;
}
+ if (MayAlias) {
+ // Add dependencies from all the PendingLoads, i.e. loads
+ // with no underlying object.
+ for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
+ addChainDependency(AAForDep, MFI, SU, PendingLoads[k], RejectMemNodes,
+ TrueMemOrderLatency);
+ // Add dependence on alias chain, if needed.
+ if (AliasChain)
+ addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes);
+ // But we also should check dependent instructions for the
+ // SU in question.
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
+ TrueMemOrderLatency);
+ }
+ // Add dependence on barrier chain, if needed.
+ // There is no point to check aliasing on barrier event. Even if
+ // SU and barrier _could_ be reordered, they should not. In addition,
+ // we have lost all RejectMemNodes below barrier.
+ if (BarrierChain)
+ BarrierChain->addPred(SDep(SU, SDep::Barrier));
if (!ExitSU.isPred(SU))
// Push store's up a bit to avoid them getting in between cmp
// and branches.
- ExitSU.addPred(SDep(SU, SDep::Order, 0,
- /*Reg=*/0, /*isNormalMemory=*/false,
- /*isMustAlias=*/false,
- /*isArtificial=*/true));
+ ExitSU.addPred(SDep(SU, SDep::Artificial));
} else if (MI->mayLoad()) {
bool MayAlias = true;
- TrueMemOrderLatency = 0;
if (MI->isInvariantLoad(AA)) {
// Invariant load, no chain dependencies needed!
} else {
- if (const Value *V =
- getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
- // A load from a specific PseudoSourceValue. Add precise dependencies.
- std::map<const Value *, SUnit *>::iterator I =
- ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
- std::map<const Value *, SUnit *>::iterator IE =
- ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
- if (I != IE)
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes, 0, true);
- if (MayAlias)
- AliasMemUses[V].push_back(SU);
- else
- NonAliasMemUses[V].push_back(SU);
- } else {
+ UnderlyingObjectsVector Objs;
+ getUnderlyingObjectsForInstr(MI, MFI, Objs);
+
+ if (Objs.empty()) {
// A load with no underlying object. Depend on all
// potentially aliasing stores.
- for (std::map<const Value *, SUnit *>::iterator I =
+ for (MapVector<const Value *, SUnit *>::iterator I =
AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes);
PendingLoads.push_back(SU);
MayAlias = true;
+ } else {
+ MayAlias = false;
+ }
+
+ for (UnderlyingObjectsVector::iterator
+ J = Objs.begin(), JE = Objs.end(); J != JE; ++J) {
+ const Value *V = J->getPointer();
+ bool ThisMayAlias = J->getInt();
+
+ if (ThisMayAlias)
+ MayAlias = true;
+
+ // A load from a specific PseudoSourceValue. Add precise dependencies.
+ MapVector<const Value *, SUnit *>::iterator I =
+ ((ThisMayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
+ MapVector<const Value *, SUnit *>::iterator IE =
+ ((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
+ if (I != IE)
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes,
+ 0, true);
+ if (ThisMayAlias)
+ AliasMemUses[V].push_back(SU);
+ else
+ NonAliasMemUses[V].push_back(SU);
}
if (MayAlias)
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes, /*Latency=*/0);
// Add dependencies on alias and barrier chains, if needed.
if (MayAlias && AliasChain)
- addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
+ addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes);
if (BarrierChain)
- BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ BarrierChain->addPred(SDep(SU, SDep::Barrier));
}
}
}
- if (PrevMI)
- FirstDbgValue = PrevMI;
+ if (DbgMI)
+ FirstDbgValue = DbgMI;
Defs.clear();
Uses.clear();
PendingLoads.clear();
}
-void ScheduleDAGInstrs::computeLatency(SUnit *SU) {
- // Compute the latency for the node.
- if (!InstrItins || InstrItins->isEmpty()) {
- SU->Latency = 1;
-
- // Simplistic target-independent heuristic: assume that loads take
- // extra time.
- if (SU->getInstr()->mayLoad())
- SU->Latency += 2;
- } else {
- SU->Latency = TII->getInstrLatency(InstrItins, SU->getInstr());
- }
-}
-
-void ScheduleDAGInstrs::computeOperandLatency(SUnit *Def, SUnit *Use,
- SDep& dep) const {
- if (!InstrItins || InstrItins->isEmpty())
- return;
-
- // For a data dependency with a known register...
- if ((dep.getKind() != SDep::Data) || (dep.getReg() == 0))
- return;
-
- const unsigned Reg = dep.getReg();
-
- // ... find the definition of the register in the defining
- // instruction
- MachineInstr *DefMI = Def->getInstr();
- int DefIdx = DefMI->findRegisterDefOperandIdx(Reg);
- if (DefIdx != -1) {
- const MachineOperand &MO = DefMI->getOperand(DefIdx);
- if (MO.isReg() && MO.isImplicit() &&
- DefIdx >= (int)DefMI->getDesc().getNumOperands()) {
- // This is an implicit def, getOperandLatency() won't return the correct
- // latency. e.g.
- // %D6<def>, %D7<def> = VLD1q16 %R2<kill>, 0, ..., %Q3<imp-def>
- // %Q1<def> = VMULv8i16 %Q1<kill>, %Q3<kill>, ...
- // What we want is to compute latency between def of %D6/%D7 and use of
- // %Q3 instead.
- unsigned Op2 = DefMI->findRegisterDefOperandIdx(Reg, false, true, TRI);
- if (DefMI->getOperand(Op2).isReg())
- DefIdx = Op2;
- }
- MachineInstr *UseMI = Use->getInstr();
- // For all uses of the register, calculate the maxmimum latency
- int Latency = -1;
- if (UseMI) {
- for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
- const MachineOperand &MO = UseMI->getOperand(i);
- if (!MO.isReg() || !MO.isUse())
- continue;
- unsigned MOReg = MO.getReg();
- if (MOReg != Reg)
- continue;
-
- int UseCycle = TII->getOperandLatency(InstrItins, DefMI, DefIdx,
- UseMI, i);
- Latency = std::max(Latency, UseCycle);
- }
- } else {
- // UseMI is null, then it must be a scheduling barrier.
- if (!InstrItins || InstrItins->isEmpty())
- return;
- unsigned DefClass = DefMI->getDesc().getSchedClass();
- Latency = InstrItins->getOperandCycle(DefClass, DefIdx);
- }
-
- // If we found a latency, then replace the existing dependence latency.
- if (Latency >= 0)
- dep.setLatency(Latency);
- }
-}
-
void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
SU->getInstr()->dump();
+#endif
}
std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
else if (SU == &ExitSU)
oss << "<exit>";
else
- SU->getInstr()->print(oss);
+ SU->getInstr()->print(oss, &TM, /*SkipOpers=*/true);
return oss.str();
}
std::string ScheduleDAGInstrs::getDAGName() const {
return "dag." + BB->getFullName();
}
+
+//===----------------------------------------------------------------------===//
+// SchedDFSResult Implementation
+//===----------------------------------------------------------------------===//
+
+namespace llvm {
+/// \brief Internal state used to compute SchedDFSResult.
+class SchedDFSImpl {
+ SchedDFSResult &R;
+
+ /// Join DAG nodes into equivalence classes by their subtree.
+ IntEqClasses SubtreeClasses;
+ /// List PredSU, SuccSU pairs that represent data edges between subtrees.
+ std::vector<std::pair<const SUnit*, const SUnit*> > ConnectionPairs;
+
+ struct RootData {
+ unsigned NodeID;
+ unsigned ParentNodeID; // Parent node (member of the parent subtree).
+ unsigned SubInstrCount; // Instr count in this tree only, not children.
+
+ RootData(unsigned id): NodeID(id),
+ ParentNodeID(SchedDFSResult::InvalidSubtreeID),
+ SubInstrCount(0) {}
+
+ unsigned getSparseSetIndex() const { return NodeID; }
+ };
+
+ SparseSet<RootData> RootSet;
+
+public:
+ SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
+ RootSet.setUniverse(R.DFSNodeData.size());
+ }
+
+ /// Return true if this node been visited by the DFS traversal.
+ ///
+ /// During visitPostorderNode the Node's SubtreeID is assigned to the Node
+ /// ID. Later, SubtreeID is updated but remains valid.
+ bool isVisited(const SUnit *SU) const {
+ return R.DFSNodeData[SU->NodeNum].SubtreeID
+ != SchedDFSResult::InvalidSubtreeID;
+ }
+
+ /// Initialize this node's instruction count. We don't need to flag the node
+ /// visited until visitPostorder because the DAG cannot have cycles.
+ void visitPreorder(const SUnit *SU) {
+ R.DFSNodeData[SU->NodeNum].InstrCount =
+ SU->getInstr()->isTransient() ? 0 : 1;
+ }
+
+ /// Called once for each node after all predecessors are visited. Revisit this
+ /// node's predecessors and potentially join them now that we know the ILP of
+ /// the other predecessors.
+ void visitPostorderNode(const SUnit *SU) {
+ // Mark this node as the root of a subtree. It may be joined with its
+ // successors later.
+ R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
+ RootData RData(SU->NodeNum);
+ RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
+
+ // If any predecessors are still in their own subtree, they either cannot be
+ // joined or are large enough to remain separate. If this parent node's
+ // total instruction count is not greater than a child subtree by at least
+ // the subtree limit, then try to join it now since splitting subtrees is
+ // only useful if multiple high-pressure paths are possible.
+ unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
+ for (SUnit::const_pred_iterator
+ PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
+ if (PI->getKind() != SDep::Data)
+ continue;
+ unsigned PredNum = PI->getSUnit()->NodeNum;
+ if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
+ joinPredSubtree(*PI, SU, /*CheckLimit=*/false);
+
+ // Either link or merge the TreeData entry from the child to the parent.
+ if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
+ // If the predecessor's parent is invalid, this is a tree edge and the
+ // current node is the parent.
+ if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
+ RootSet[PredNum].ParentNodeID = SU->NodeNum;
+ }
+ else if (RootSet.count(PredNum)) {
+ // The predecessor is not a root, but is still in the root set. This
+ // must be the new parent that it was just joined to. Note that
+ // RootSet[PredNum].ParentNodeID may either be invalid or may still be
+ // set to the original parent.
+ RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
+ RootSet.erase(PredNum);
+ }
+ }
+ RootSet[SU->NodeNum] = RData;
+ }
+
+ /// Called once for each tree edge after calling visitPostOrderNode on the
+ /// predecessor. Increment the parent node's instruction count and
+ /// preemptively join this subtree to its parent's if it is small enough.
+ void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
+ R.DFSNodeData[Succ->NodeNum].InstrCount
+ += R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
+ joinPredSubtree(PredDep, Succ);
+ }
+
+ /// Add a connection for cross edges.
+ void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
+ ConnectionPairs.push_back(std::make_pair(PredDep.getSUnit(), Succ));
+ }
+
+ /// Set each node's subtree ID to the representative ID and record connections
+ /// between trees.
+ void finalize() {
+ SubtreeClasses.compress();
+ R.DFSTreeData.resize(SubtreeClasses.getNumClasses());
+ assert(SubtreeClasses.getNumClasses() == RootSet.size()
+ && "number of roots should match trees");
+ for (SparseSet<RootData>::const_iterator
+ RI = RootSet.begin(), RE = RootSet.end(); RI != RE; ++RI) {
+ unsigned TreeID = SubtreeClasses[RI->NodeID];
+ if (RI->ParentNodeID != SchedDFSResult::InvalidSubtreeID)
+ R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[RI->ParentNodeID];
+ R.DFSTreeData[TreeID].SubInstrCount = RI->SubInstrCount;
+ // Note that SubInstrCount may be greater than InstrCount if we joined
+ // subtrees across a cross edge. InstrCount will be attributed to the
+ // original parent, while SubInstrCount will be attributed to the joined
+ // parent.
+ }
+ R.SubtreeConnections.resize(SubtreeClasses.getNumClasses());
+ R.SubtreeConnectLevels.resize(SubtreeClasses.getNumClasses());
+ DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
+ for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
+ R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
+ DEBUG(dbgs() << " SU(" << Idx << ") in tree "
+ << R.DFSNodeData[Idx].SubtreeID << '\n');
+ }
+ for (std::vector<std::pair<const SUnit*, const SUnit*> >::const_iterator
+ I = ConnectionPairs.begin(), E = ConnectionPairs.end();
+ I != E; ++I) {
+ unsigned PredTree = SubtreeClasses[I->first->NodeNum];
+ unsigned SuccTree = SubtreeClasses[I->second->NodeNum];
+ if (PredTree == SuccTree)
+ continue;
+ unsigned Depth = I->first->getDepth();
+ addConnection(PredTree, SuccTree, Depth);
+ addConnection(SuccTree, PredTree, Depth);
+ }
+ }
+
+protected:
+ /// Join the predecessor subtree with the successor that is its DFS
+ /// parent. Apply some heuristics before joining.
+ bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
+ bool CheckLimit = true) {
+ assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
+
+ // Check if the predecessor is already joined.
+ const SUnit *PredSU = PredDep.getSUnit();
+ unsigned PredNum = PredSU->NodeNum;
+ if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
+ return false;
+
+ // Four is the magic number of successors before a node is considered a
+ // pinch point.
+ unsigned NumDataSucs = 0;
+ for (SUnit::const_succ_iterator SI = PredSU->Succs.begin(),
+ SE = PredSU->Succs.end(); SI != SE; ++SI) {
+ if (SI->getKind() == SDep::Data) {
+ if (++NumDataSucs >= 4)
+ return false;
+ }
+ }
+ if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
+ return false;
+ R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
+ SubtreeClasses.join(Succ->NodeNum, PredNum);
+ return true;
+ }
+
+ /// Called by finalize() to record a connection between trees.
+ void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
+ if (!Depth)
+ return;
+
+ do {
+ SmallVectorImpl<SchedDFSResult::Connection> &Connections =
+ R.SubtreeConnections[FromTree];
+ for (SmallVectorImpl<SchedDFSResult::Connection>::iterator
+ I = Connections.begin(), E = Connections.end(); I != E; ++I) {
+ if (I->TreeID == ToTree) {
+ I->Level = std::max(I->Level, Depth);
+ return;
+ }
+ }
+ Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
+ FromTree = R.DFSTreeData[FromTree].ParentTreeID;
+ } while (FromTree != SchedDFSResult::InvalidSubtreeID);
+ }
+};
+} // namespace llvm
+
+namespace {
+/// \brief Manage the stack used by a reverse depth-first search over the DAG.
+class SchedDAGReverseDFS {
+ std::vector<std::pair<const SUnit*, SUnit::const_pred_iterator> > DFSStack;
+public:
+ bool isComplete() const { return DFSStack.empty(); }
+
+ void follow(const SUnit *SU) {
+ DFSStack.push_back(std::make_pair(SU, SU->Preds.begin()));
+ }
+ void advance() { ++DFSStack.back().second; }
+
+ const SDep *backtrack() {
+ DFSStack.pop_back();
+ return DFSStack.empty() ? 0 : llvm::prior(DFSStack.back().second);
+ }
+
+ const SUnit *getCurr() const { return DFSStack.back().first; }
+
+ SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
+
+ SUnit::const_pred_iterator getPredEnd() const {
+ return getCurr()->Preds.end();
+ }
+};
+} // anonymous
+
+static bool hasDataSucc(const SUnit *SU) {
+ for (SUnit::const_succ_iterator
+ SI = SU->Succs.begin(), SE = SU->Succs.end(); SI != SE; ++SI) {
+ if (SI->getKind() == SDep::Data && !SI->getSUnit()->isBoundaryNode())
+ return true;
+ }
+ return false;
+}
+
+/// Compute an ILP metric for all nodes in the subDAG reachable via depth-first
+/// search from this root.
+void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
+ if (!IsBottomUp)
+ llvm_unreachable("Top-down ILP metric is unimplemnted");
+
+ SchedDFSImpl Impl(*this);
+ for (ArrayRef<SUnit>::const_iterator
+ SI = SUnits.begin(), SE = SUnits.end(); SI != SE; ++SI) {
+ const SUnit *SU = &*SI;
+ if (Impl.isVisited(SU) || hasDataSucc(SU))
+ continue;
+
+ SchedDAGReverseDFS DFS;
+ Impl.visitPreorder(SU);
+ DFS.follow(SU);
+ for (;;) {
+ // Traverse the leftmost path as far as possible.
+ while (DFS.getPred() != DFS.getPredEnd()) {
+ const SDep &PredDep = *DFS.getPred();
+ DFS.advance();
+ // Ignore non-data edges.
+ if (PredDep.getKind() != SDep::Data
+ || PredDep.getSUnit()->isBoundaryNode()) {
+ continue;
+ }
+ // An already visited edge is a cross edge, assuming an acyclic DAG.
+ if (Impl.isVisited(PredDep.getSUnit())) {
+ Impl.visitCrossEdge(PredDep, DFS.getCurr());
+ continue;
+ }
+ Impl.visitPreorder(PredDep.getSUnit());
+ DFS.follow(PredDep.getSUnit());
+ }
+ // Visit the top of the stack in postorder and backtrack.
+ const SUnit *Child = DFS.getCurr();
+ const SDep *PredDep = DFS.backtrack();
+ Impl.visitPostorderNode(Child);
+ if (PredDep)
+ Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
+ if (DFS.isComplete())
+ break;
+ }
+ }
+ Impl.finalize();
+}
+
+/// The root of the given SubtreeID was just scheduled. For all subtrees
+/// connected to this tree, record the depth of the connection so that the
+/// nearest connected subtrees can be prioritized.
+void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
+ for (SmallVectorImpl<Connection>::const_iterator
+ I = SubtreeConnections[SubtreeID].begin(),
+ E = SubtreeConnections[SubtreeID].end(); I != E; ++I) {
+ SubtreeConnectLevels[I->TreeID] =
+ std::max(SubtreeConnectLevels[I->TreeID], I->Level);
+ DEBUG(dbgs() << " Tree: " << I->TreeID
+ << " @" << SubtreeConnectLevels[I->TreeID] << '\n');
+ }
+}
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+void ILPValue::print(raw_ostream &OS) const {
+ OS << InstrCount << " / " << Length << " = ";
+ if (!Length)
+ OS << "BADILP";
+ else
+ OS << format("%g", ((double)InstrCount / Length));
+}
+
+void ILPValue::dump() const {
+ dbgs() << *this << '\n';
+}
+
+namespace llvm {
+
+raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {
+ Val.print(OS);
+ return OS;
+}
+
+} // namespace llvm
+#endif // !NDEBUG || LLVM_ENABLE_DUMP