//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "sched-instrs"
-#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/Operator.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
-#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/CodeGen/ScheduleDAGInstrs.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/CodeGen/RegisterPressure.h"
+#include "llvm/CodeGen/ScheduleDAGILP.h"
+#include "llvm/CodeGen/ScheduleDAGInstrs.h"
+#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Target/TargetSubtarget.h"
-#include "llvm/Support/Compiler.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 <map>
+#include "llvm/ADT/SmallPtrSet.h"
using namespace llvm;
-namespace {
- class VISIBILITY_HIDDEN LoopDependencies {
- const MachineLoopInfo &MLI;
- const MachineDominatorTree &MDT;
-
- public:
- typedef std::map<unsigned, std::pair<const MachineOperand *, unsigned> >
- LoopDeps;
- LoopDeps Deps;
-
- LoopDependencies(const MachineLoopInfo &mli,
- const MachineDominatorTree &mdt) :
- MLI(mli), MDT(mdt) {}
-
- void VisitLoop(const MachineLoop *Loop) {
- Deps.clear();
- MachineBasicBlock *Header = Loop->getHeader();
- SmallSet<unsigned, 8> LoopLiveIns;
- for (MachineBasicBlock::livein_iterator LI = Header->livein_begin(),
- LE = Header->livein_end(); LI != LE; ++LI)
- LoopLiveIns.insert(*LI);
-
- VisitRegion(MDT.getNode(Header), Loop, LoopLiveIns);
+static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
+ cl::ZeroOrMore, cl::init(false),
+ cl::desc("Enable use of AA during MI GAD construction"));
+
+ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
+ const MachineLoopInfo &mli,
+ const MachineDominatorTree &mdt,
+ bool IsPostRAFlag,
+ LiveIntervals *lis)
+ : 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
+/// looking through basic ptrtoint+arithmetic+inttoptr sequences.
+static const Value *getUnderlyingObjectFromInt(const Value *V) {
+ do {
+ if (const Operator *U = dyn_cast<Operator>(V)) {
+ // If we find a ptrtoint, we can transfer control back to the
+ // 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
+ // 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
+ // identifiable object.
+ if (U->getOpcode() != Instruction::Add ||
+ (!isa<ConstantInt>(U->getOperand(1)) &&
+ Operator::getOpcode(U->getOperand(1)) != Instruction::Mul))
+ return V;
+ V = U->getOperand(0);
+ } else {
+ return V;
+ }
+ assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
+ } while (1);
+}
+
+/// getUnderlyingObject - This is a wrapper around GetUnderlyingObject
+/// 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.
+ 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;
+}
+
+/// getUnderlyingObjectForInstr - 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;
+ if (!MI->hasOneMemOperand() ||
+ !(*MI->memoperands_begin())->getValue() ||
+ (*MI->memoperands_begin())->isVolatile())
+ return 0;
+
+ 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;
+ }
+
+ if (isIdentifiedObject(V))
+ return V;
+
+ return 0;
+}
+
+void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
+ BB = bb;
+}
+
+void ScheduleDAGInstrs::finishBlock() {
+ // Subclasses should no longer refer to the old block.
+ 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
+/// region.
+void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
+ MachineBasicBlock::iterator begin,
+ MachineBasicBlock::iterator end,
+ unsigned endcount) {
+ assert(bb == BB && "startBlock should set BB");
+ RegionBegin = begin;
+ RegionEnd = end;
+ EndIndex = endcount;
+ MISUnitMap.clear();
+
+ ScheduleDAG::clearDAG();
+}
+
+/// Close the current scheduling region. Don't clear any state in case the
+/// driver wants to refer to the previous scheduling region.
+void ScheduleDAGInstrs::exitRegion() {
+ // Nothing to do.
+}
+
+/// addSchedBarrierDeps - Add dependencies from instructions in the current
+/// list of instructions being scheduled to scheduling barrier by adding
+/// the exit SU to the register defs and use list. This is because we want to
+/// make sure instructions which define registers that are either used by
+/// the terminator or are live-out are properly scheduled. This is
+/// especially important when the definition latency of the return value(s)
+/// are too high to be hidden by the branch or when the liveout registers
+/// used by instructions in the fallthrough block.
+void ScheduleDAGInstrs::addSchedBarrierDeps() {
+ MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : 0;
+ ExitSU.setInstr(ExitMI);
+ bool AllDepKnown = ExitMI &&
+ (ExitMI->isCall() || ExitMI->isBarrier());
+ if (ExitMI && AllDepKnown) {
+ // If it's a call or a barrier, add dependencies on the defs and uses of
+ // instruction.
+ for (unsigned i = 0, e = ExitMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = ExitMI->getOperand(i);
+ if (!MO.isReg() || MO.isDef()) continue;
+ unsigned Reg = MO.getReg();
+ if (Reg == 0) continue;
+
+ if (TRI->isPhysicalRegister(Reg))
+ Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
+ else {
+ assert(!IsPostRA && "Virtual register encountered after regalloc.");
+ addVRegUseDeps(&ExitSU, i);
+ }
}
+ } else {
+ // For others, e.g. fallthrough, conditional branch, assume the exit
+ // uses all the registers that are livein to the successor blocks.
+ assert(Uses.empty() && "Uses in set before adding deps?");
+ for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
+ SE = BB->succ_end(); SI != SE; ++SI)
+ for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
+ E = (*SI)->livein_end(); I != E; ++I) {
+ unsigned Reg = *I;
+ if (!Uses.contains(Reg))
+ Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
+ }
+ }
+}
- private:
- void VisitRegion(const MachineDomTreeNode *Node,
- const MachineLoop *Loop,
- const SmallSet<unsigned, 8> &LoopLiveIns) {
- MachineBasicBlock *MBB = Node->getBlock();
- if (!Loop->contains(MBB)) return;
-
- unsigned Count = 0;
- for (MachineBasicBlock::const_iterator I = MBB->begin(), E = MBB->end();
- I != E; ++I, ++Count) {
- const MachineInstr *MI = I;
- for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
- const MachineOperand &MO = MI->getOperand(i);
- if (!MO.isReg() || !MO.isUse())
- continue;
- unsigned MOReg = MO.getReg();
- if (LoopLiveIns.count(MOReg))
- Deps.insert(std::make_pair(MOReg, std::make_pair(&MO, Count)));
+/// 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, 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>();
+
+ for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
+ Alias.isValid(); ++Alias) {
+ if (!Uses.contains(*Alias))
+ continue;
+ std::vector<PhysRegSUOper> &UseList = Uses[*Alias];
+ for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
+ SUnit *UseSU = UseList[i].SU;
+ if (UseSU == SU)
+ continue;
+
+ // Adjust the dependence latency using operand def/use information,
+ // then allow the target to perform its own adjustments.
+ int UseOp = UseList[i].OpIdx;
+ MachineInstr *RegUse = 0;
+ SDep Dep;
+ if (UseOp < 0)
+ Dep = SDep(SU, SDep::Artificial);
+ else {
+ Dep = SDep(SU, SDep::Data, *Alias);
+ RegUse = UseSU->getInstr();
+ Dep.setMinLatency(
+ SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
+ RegUse, UseOp, /*FindMin=*/true));
+ }
+ Dep.setLatency(
+ SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
+ RegUse, UseOp, /*FindMin=*/false));
+
+ ST.adjustSchedDependency(SU, UseSU, Dep);
+ UseSU->addPred(Dep);
+ }
+ }
+}
+
+/// addPhysRegDeps - Add register dependencies (data, anti, and output) from
+/// this SUnit to following instructions in the same scheduling region that
+/// depend the physical register referenced at OperIdx.
+void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
+ const MachineInstr *MI = SU->getInstr();
+ const MachineOperand &MO = MI->getOperand(OperIdx);
+
+ // Optionally add output and anti dependencies. For anti
+ // dependencies we use a latency of 0 because for a multi-issue
+ // target we want to allow the defining instruction to issue
+ // in the same cycle as the using instruction.
+ // 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 (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
+ Alias.isValid(); ++Alias) {
+ if (!Defs.contains(*Alias))
+ continue;
+ std::vector<PhysRegSUOper> &DefList = Defs[*Alias];
+ for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
+ SUnit *DefSU = DefList[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, /*Reg=*/*Alias));
+ else {
+ SDep Dep(SU, Kind, /*Reg=*/*Alias);
+ unsigned OutLatency =
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
+ Dep.setMinLatency(OutLatency);
+ Dep.setLatency(OutLatency);
+ DefSU->addPred(Dep);
}
}
+ }
+ }
+
+ if (!MO.isDef()) {
+ // 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(PhysRegSUOper(SU, OperIdx));
+ }
+ else {
+ addPhysRegDataDeps(SU, OperIdx);
+
+ // Either insert a new Reg2SUnits entry with an empty SUnits list, or
+ // retrieve the existing SUnits list for this register's defs.
+ std::vector<PhysRegSUOper> &DefList = Defs[MO.getReg()];
- const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
- for (unsigned I = 0, E = Children.size(); I != E; ++I)
- VisitRegion(Children[I], Loop, LoopLiveIns);
+ // 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().SU->isCall)
+ DefList.pop_back();
}
- };
+ // Defs are pushed in the order they are visited and never reordered.
+ DefList.push_back(PhysRegSUOper(SU, OperIdx));
+ }
}
-ScheduleDAGInstrs::ScheduleDAGInstrs(MachineBasicBlock *bb,
- const TargetMachine &tm,
- const MachineLoopInfo &mli,
- const MachineDominatorTree &mdt)
- : ScheduleDAG(0, bb, tm), MLI(mli), MDT(mdt) {}
+/// addVRegDefDeps - Add register output and data dependencies from this SUnit
+/// to instructions that occur later in the same scheduling region if they read
+/// from or write to the virtual register defined at OperIdx.
+///
+/// TODO: Hoist loop induction variable increments. This has to be
+/// reevaluated. Generally, IV scheduling should be done before coalescing.
+void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
+ const MachineInstr *MI = SU->getInstr();
+ unsigned Reg = MI->getOperand(OperIdx).getReg();
+
+ // Singly defined vregs do not have output/anti dependencies.
+ // The current operand is a def, so we have at least one.
+ // Check here if there are any others...
+ if (MRI.hasOneDef(Reg))
+ return;
+
+ // Add output dependence to the next nearest def of this vreg.
+ //
+ // Unless this definition is dead, the output dependence should be
+ // transitively redundant with antidependencies from this definition's
+ // uses. We're conservative for now until we have a way to guarantee the uses
+ // are not eliminated sometime during scheduling. The output dependence edge
+ // is also useful if output latency exceeds def-use latency.
+ VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
+ if (DefI == VRegDefs.end())
+ VRegDefs.insert(VReg2SUnit(Reg, SU));
+ else {
+ SUnit *DefSU = DefI->SU;
+ if (DefSU != SU && DefSU != &ExitSU) {
+ SDep Dep(SU, SDep::Output, Reg);
+ unsigned OutLatency =
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
+ Dep.setMinLatency(OutLatency);
+ Dep.setLatency(OutLatency);
+ DefSU->addPred(Dep);
+ }
+ DefI->SU = SU;
+ }
+}
+
+/// addVRegUseDeps - Add a register data dependency if the instruction that
+/// defines the virtual register used at OperIdx is mapped to an SUnit. Add a
+/// register antidependency from this SUnit to instructions that occur later in
+/// the same scheduling region if they write the virtual register.
+///
+/// TODO: Handle ExitSU "uses" properly.
+void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
+ MachineInstr *MI = SU->getInstr();
+ unsigned Reg = MI->getOperand(OperIdx).getReg();
+
+ // Lookup this operand's reaching definition.
+ assert(LIS && "vreg dependencies requires LiveIntervals");
+ LiveRangeQuery LRQ(LIS->getInterval(Reg), LIS->getInstructionIndex(MI));
+ VNInfo *VNI = LRQ.valueIn();
+
+ // 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) {
+ SUnit *DefSU = getSUnit(Def);
+ if (DefSU) {
+ // The reaching Def lives within this scheduling region.
+ // Create a data dependence.
+ 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, false));
+ dep.setMinLatency(
+ SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, true));
+
+ 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, 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->hasOrderedMemoryRef() &&
+ (!MI->mayLoad() || !MI->isInvariantLoad(AA))))
+ return true;
+ return false;
+}
+
+// This MI might have either incomplete info, or known to be unsafe
+// to deal with (i.e. volatile object).
+static inline bool isUnsafeMemoryObject(MachineInstr *MI,
+ const MachineFrameInfo *MFI) {
+ if (!MI || MI->memoperands_empty())
+ return true;
+ // We purposefully do no check for hasOneMemOperand() here
+ // in hope to trigger an assert downstream in order to
+ // finish implementation.
+ 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))
+ return true;
+ }
+ // Does this pointer refer to a distinct and identifiable object?
+ if (!isIdentifiedObject(V))
+ return true;
+
+ return false;
+}
+
+/// This returns true if the two MIs need a chain edge betwee them.
+/// If these are not even memory operations, we still may need
+/// chain deps between them. The question really is - could
+/// these two MIs be reordered during scheduling from memory dependency
+/// point of view.
+static bool MIsNeedChainEdge(AliasAnalysis *AA, const MachineFrameInfo *MFI,
+ MachineInstr *MIa,
+ MachineInstr *MIb) {
+ // Cover a trivial case - no edge is need to itself.
+ if (MIa == MIb)
+ return false;
+
+ if (isUnsafeMemoryObject(MIa, MFI) || isUnsafeMemoryObject(MIb, MFI))
+ return true;
+
+ // If we are dealing with two "normal" loads, we do not need an edge
+ // between them - they could be reordered.
+ if (!MIa->mayStore() && !MIb->mayStore())
+ return false;
+
+ // To this point analysis is generic. From here on we do need AA.
+ if (!AA)
+ return true;
+
+ MachineMemOperand *MMOa = *MIa->memoperands_begin();
+ MachineMemOperand *MMOb = *MIb->memoperands_begin();
+
+ // FIXME: Need to handle multiple memory operands to support all targets.
+ if (!MIa->hasOneMemOperand() || !MIb->hasOneMemOperand())
+ llvm_unreachable("Multiple memory operands.");
+
+ // The following interface to AA is fashioned after DAGCombiner::isAlias
+ // and operates with MachineMemOperand offset with some important
+ // assumptions:
+ // - LLVM fundamentally assumes flat address spaces.
+ // - MachineOperand offset can *only* result from legalization and
+ // cannot affect queries other than the trivial case of overlap
+ // checking.
+ // - These offsets never wrap and never step outside
+ // of allocated objects.
+ // - There should never be any negative offsets here.
+ //
+ // FIXME: Modify API to hide this math from "user"
+ // FIXME: Even before we go to AA we can reason locally about some
+ // memory objects. It can save compile time, and possibly catch some
+ // corner cases not currently covered.
+
+ assert ((MMOa->getOffset() >= 0) && "Negative MachineMemOperand offset");
+ assert ((MMOb->getOffset() >= 0) && "Negative MachineMemOperand offset");
+
+ int64_t MinOffset = std::min(MMOa->getOffset(), MMOb->getOffset());
+ int64_t Overlapa = MMOa->getSize() + MMOa->getOffset() - MinOffset;
+ int64_t Overlapb = MMOb->getSize() + MMOb->getOffset() - MinOffset;
+
+ AliasAnalysis::AliasResult AAResult = AA->alias(
+ AliasAnalysis::Location(MMOa->getValue(), Overlapa,
+ MMOa->getTBAAInfo()),
+ AliasAnalysis::Location(MMOb->getValue(), Overlapb,
+ MMOb->getTBAAInfo()));
-void ScheduleDAGInstrs::BuildSchedUnits() {
- SUnits.clear();
+ return (AAResult != AliasAnalysis::NoAlias);
+}
+
+/// This recursive function iterates over chain deps of SUb looking for
+/// "latest" node that needs a chain edge to SUa.
+static unsigned
+iterateChainSucc(AliasAnalysis *AA, const MachineFrameInfo *MFI,
+ SUnit *SUa, SUnit *SUb, SUnit *ExitSU, unsigned *Depth,
+ SmallPtrSet<const SUnit*, 16> &Visited) {
+ if (!SUa || !SUb || SUb == ExitSU)
+ return *Depth;
+
+ // Remember visited nodes.
+ if (!Visited.insert(SUb))
+ return *Depth;
+ // If there is _some_ dependency already in place, do not
+ // descend any further.
+ // TODO: Need to make sure that if that dependency got eliminated or ignored
+ // for any reason in the future, we would not violate DAG topology.
+ // Currently it does not happen, but makes an implicit assumption about
+ // future implementation.
+ //
+ // Independently, if we encounter node that is some sort of global
+ // object (like a call) we already have full set of dependencies to it
+ // and we can stop descending.
+ if (SUa->isSucc(SUb) ||
+ isGlobalMemoryObject(AA, SUb->getInstr()))
+ return *Depth;
+
+ // If we do need an edge, or we have exceeded depth budget,
+ // add that edge to the predecessors chain of SUb,
+ // and stop descending.
+ if (*Depth > 200 ||
+ MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
+ SUb->addPred(SDep(SUa, SDep::MayAliasMem));
+ return *Depth;
+ }
+ // Track current depth.
+ (*Depth)++;
+ // Iterate over chain dependencies only.
+ for (SUnit::const_succ_iterator I = SUb->Succs.begin(), E = SUb->Succs.end();
+ I != E; ++I)
+ if (I->isCtrl())
+ iterateChainSucc (AA, MFI, SUa, I->getSUnit(), ExitSU, Depth, Visited);
+ return *Depth;
+}
+
+/// This function assumes that "downward" from SU there exist
+/// tail/leaf of already constructed DAG. It iterates downward and
+/// 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,
+ unsigned LatencyToLoad) {
+ if (!SU)
+ return;
+
+ SmallPtrSet<const SUnit*, 16> Visited;
+ unsigned Depth = 0;
+
+ for (std::set<SUnit *>::iterator I = CheckList.begin(), IE = CheckList.end();
+ I != IE; ++I) {
+ if (SU == *I)
+ continue;
+ 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(),
+ JE = (*I)->Succs.end(); J != JE; ++J)
+ if (J->isCtrl())
+ iterateChainSucc (AA, MFI, SU, J->getSUnit(),
+ ExitSU, &Depth, Visited);
+ }
+}
+
+/// Check whether two objects need a chain edge, if so, add it
+/// otherwise remember the rejected SU.
+static inline
+void addChainDependency (AliasAnalysis *AA, const MachineFrameInfo *MFI,
+ SUnit *SUa, SUnit *SUb,
+ std::set<SUnit *> &RejectList,
+ unsigned TrueMemOrderLatency = 0,
+ 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())) {
+ SDep Dep(SUa, isNormalMemory ? SDep::MayAliasMem : SDep::Barrier);
+ Dep.setLatency(TrueMemOrderLatency);
+ SUb->addPred(Dep);
+ }
+ else {
+ // Duplicate entries should be ignored.
+ RejectList.insert(SUb);
+ DEBUG(dbgs() << "\tReject chain dep between SU("
+ << SUa->NodeNum << ") and SU("
+ << SUb->NodeNum << ")\n");
+ }
+}
+
+/// Create an SUnit for each real instruction, numbered in top-down toplological
+/// order. The instruction order A < B, implies that no edge exists from B to A.
+///
+/// Map each real instruction to its SUnit.
+///
+/// After initSUnits, the SUnits vector cannot be resized and the scheduler may
+/// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
+/// instead of pointers.
+///
+/// MachineScheduler relies on initSUnits numbering the nodes by their order in
+/// the original instruction list.
+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());
+ for (MachineBasicBlock::iterator I = RegionBegin; I != RegionEnd; ++I) {
+ MachineInstr *MI = I;
+ if (MI->isDebugValue())
+ continue;
+
+ SUnit *SU = newSUnit(MI);
+ MISUnitMap[MI] = SU;
+
+ SU->isCall = MI->isCall();
+ SU->isCommutable = MI->isCommutable();
+
+ // Assign the Latency field of SU using target-provided information.
+ SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
+ }
+}
+
+/// If RegPressure is non null, compute register pressure as a side effect. The
+/// DAG builder is an efficient place to do it because it already visits
+/// operands.
+void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
+ RegPressureTracker *RPTracker) {
+ // Create an SUnit for each real instruction.
+ initSUnits();
+
// We build scheduling units by walking a block's instruction list from bottom
// to top.
- // Remember where defs and uses of each physical register are as we procede.
- std::vector<SUnit *> Defs[TargetRegisterInfo::FirstVirtualRegister] = {};
- std::vector<SUnit *> Uses[TargetRegisterInfo::FirstVirtualRegister] = {};
-
- // Remember where unknown loads are after the most recent unknown store
- // as we procede.
- std::vector<SUnit *> PendingLoads;
-
- // Remember where a generic side-effecting instruction is as we procede. If
- // ChainMMO is null, this is assumed to have arbitrary side-effects. If
- // ChainMMO is non-null, then Chain makes only a single memory reference.
- SUnit *Chain = 0;
- MachineMemOperand *ChainMMO = 0;
-
- // Memory references to specific known memory locations are tracked so that
- // they can be given more precise dependencies.
- std::map<const Value *, SUnit *> MemDefs;
- std::map<const Value *, std::vector<SUnit *> > MemUses;
-
- // Terminators can perform control transfers, we we need to make sure that
- // all the work of the block is done before the terminator.
- SUnit *Terminator = 0;
-
- LoopDependencies LoopRegs(MLI, MDT);
-
- // Track which regs are live into a loop, to help guide back-edge-aware
- // scheduling.
- SmallSet<unsigned, 8> LoopLiveInRegs;
- if (MachineLoop *ML = MLI.getLoopFor(BB))
- if (BB == ML->getLoopLatch()) {
- MachineBasicBlock *Header = ML->getHeader();
- for (MachineBasicBlock::livein_iterator I = Header->livein_begin(),
- E = Header->livein_end(); I != E; ++I)
- LoopLiveInRegs.insert(*I);
- LoopRegs.VisitLoop(ML);
- }
+ // Remember where a generic side-effecting instruction is as we procede.
+ SUnit *BarrierChain = 0, *AliasChain = 0;
- // Check to see if the scheduler cares about latencies.
- bool UnitLatencies = ForceUnitLatencies();
+ // Memory references to specific known memory locations are tracked
+ // 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;
+ std::set<SUnit*> RejectMemNodes;
- // Ask the target if address-backscheduling is desirable, and if so how much.
- unsigned SpecialAddressLatency =
- TM.getSubtarget<TargetSubtarget>().getSpecialAddressLatency();
+ // Remove any stale debug info; sometimes BuildSchedGraph is called again
+ // without emitting the info from the previous call.
+ DbgValues.clear();
+ FirstDbgValue = NULL;
+
+ assert(Defs.empty() && Uses.empty() &&
+ "Only BuildGraph should update Defs/Uses");
+ Defs.setRegLimit(TRI->getNumRegs());
+ Uses.setRegLimit(TRI->getNumRegs());
- for (MachineBasicBlock::iterator MII = BB->end(), MIE = BB->begin();
+ 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.
+ VRegDefs.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;
+ for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
MII != MIE; --MII) {
MachineInstr *MI = prior(MII);
- const TargetInstrDesc &TID = MI->getDesc();
- SUnit *SU = NewSUnit(MI);
+ if (MI && PrevMI) {
+ DbgValues.push_back(std::make_pair(PrevMI, MI));
+ PrevMI = NULL;
+ }
- // Assign the Latency field of SU using target-provided information.
- if (UnitLatencies)
- SU->Latency = 1;
- else
- ComputeLatency(SU);
+ if (MI->isDebugValue()) {
+ PrevMI = MI;
+ continue;
+ }
+ if (RPTracker) {
+ RPTracker->recede();
+ assert(RPTracker->getPos() == prior(MII) && "RPTracker can't find MI");
+ }
+
+ assert((!MI->isTerminator() || CanHandleTerminators) && !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).
for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
- assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!");
- std::vector<SUnit *> &UseList = Uses[Reg];
- std::vector<SUnit *> &DefList = Defs[Reg];
- // Optionally add output and anti dependencies.
- // TODO: Using a latency of 1 here assumes there's no cost for
- // reusing registers.
- SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
- for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
- SUnit *DefSU = DefList[i];
- if (DefSU != SU &&
- (Kind != SDep::Output || !MO.isDead() ||
- !DefSU->getInstr()->registerDefIsDead(Reg)))
- DefSU->addPred(SDep(SU, Kind, /*Latency=*/1, /*Reg=*/Reg));
- }
- for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
- std::vector<SUnit *> &DefList = Defs[*Alias];
- for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
- SUnit *DefSU = DefList[i];
- if (DefSU != SU &&
- (Kind != SDep::Output || !MO.isDead() ||
- !DefSU->getInstr()->registerDefIsDead(Reg)))
- DefSU->addPred(SDep(SU, Kind, /*Latency=*/1, /*Reg=*/ *Alias));
- }
- }
-
- if (MO.isDef()) {
- // Add any data dependencies.
- unsigned DataLatency = SU->Latency;
- for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
- SUnit *UseSU = UseList[i];
- if (UseSU != SU) {
- unsigned LDataLatency = DataLatency;
- // Optionally add in a special extra latency for nodes that
- // feed addresses.
- // TODO: Do this for register aliases too.
- if (SpecialAddressLatency != 0 && !UnitLatencies) {
- MachineInstr *UseMI = UseSU->getInstr();
- const TargetInstrDesc &UseTID = UseMI->getDesc();
- int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg);
- assert(RegUseIndex >= 0 && "UseMI doesn's use register!");
- if ((UseTID.mayLoad() || UseTID.mayStore()) &&
- (unsigned)RegUseIndex < UseTID.getNumOperands() &&
- UseTID.OpInfo[RegUseIndex].isLookupPtrRegClass())
- LDataLatency += SpecialAddressLatency;
- }
- UseSU->addPred(SDep(SU, SDep::Data, LDataLatency, Reg));
- }
- }
- for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
- std::vector<SUnit *> &UseList = Uses[*Alias];
- for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
- SUnit *UseSU = UseList[i];
- if (UseSU != SU)
- UseSU->addPred(SDep(SU, SDep::Data, DataLatency, *Alias));
- }
- }
-
- // If a def is going to wrap back around to the top of the loop,
- // backschedule it.
- // TODO: Blocks in loops without terminators can benefit too.
- if (!UnitLatencies && Terminator && DefList.empty()) {
- LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(Reg);
- 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 TargetInstrDesc &UseTID = UseMI->getDesc();
- // 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 < UseTID.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 (UseTID.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 artifical edge.
- Terminator->addPred(SDep(SU, SDep::Order, Latency,
- /*Reg=*/0, /*isNormalMemory=*/false,
- /*isMustAlias=*/false,
- /*isArtificial=*/true));
- } else if (SpecialAddressLatency > 0 &&
- UseTID.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);
- }
- }
-
- UseList.clear();
- if (!MO.isDead())
- DefList.clear();
- DefList.push_back(SU);
- } else {
- UseList.push_back(SU);
+ if (TRI->isPhysicalRegister(Reg))
+ addPhysRegDeps(SU, j);
+ else {
+ assert(!IsPostRA && "Virtual register encountered!");
+ if (MO.isDef())
+ addVRegDefDeps(SU, j);
+ else if (MO.readsReg()) // ignore undef operands
+ addVRegUseDeps(SU, j);
}
}
// Add chain dependencies.
+ // Chain dependencies used to enforce memory order should have
+ // latency of 0 (except for true dependency of Store followed by
+ // aliased Load... we estimate that with a single cycle of latency
+ // assuming the hardware will bypass)
// Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
// after stack slots are lowered to actual addresses.
// TODO: Use an AliasAnalysis and do real alias-analysis queries, and
// produce more precise dependence information.
- if (TID.isCall() || TID.isTerminator() || TID.hasUnmodeledSideEffects()) {
- new_chain:
- // This is the conservative case. Add dependencies on all memory
- // references.
- if (Chain)
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- Chain = SU;
- for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
- PendingLoads[k]->addPred(SDep(SU, SDep::Order, SU->Latency));
- PendingLoads.clear();
- for (std::map<const Value *, SUnit *>::iterator I = MemDefs.begin(),
- E = MemDefs.end(); I != E; ++I) {
- I->second->addPred(SDep(SU, SDep::Order, SU->Latency));
- I->second = SU;
+ 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 =
+ NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
+ I->second->addPred(SDep(SU, SDep::Barrier));
+ }
+ for (std::map<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) {
+ 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::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,
+ TrueMemOrderLatency);
+ RejectMemNodes.clear();
+ NonAliasMemDefs.clear();
+ NonAliasMemUses.clear();
+
+ // fall-through
+ new_alias_chain:
+ // Chain all possibly aliasing memory references though SU.
+ if (AliasChain) {
+ unsigned ChainLatency = 0;
+ if (AliasChain->getInstr()->mayLoad())
+ ChainLatency = TrueMemOrderLatency;
+ addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes,
+ ChainLatency);
+ }
+ AliasChain = SU;
+ for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
+ addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
+ TrueMemOrderLatency);
+ for (std::map<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 =
- MemUses.begin(), E = MemUses.end(); I != E; ++I) {
+ AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
- I->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency));
- I->second.clear();
+ addChainDependency(AA, MFI, SU, I->second[i], RejectMemNodes,
+ TrueMemOrderLatency);
}
- // See if it is known to just have a single memory reference.
- MachineInstr *ChainMI = Chain->getInstr();
- const TargetInstrDesc &ChainTID = ChainMI->getDesc();
- if (!ChainTID.isCall() && !ChainTID.isTerminator() &&
- !ChainTID.hasUnmodeledSideEffects() &&
- ChainMI->hasOneMemOperand() &&
- !ChainMI->memoperands_begin()->isVolatile() &&
- ChainMI->memoperands_begin()->getValue())
- // We know that the Chain accesses one specific memory location.
- ChainMMO = &*ChainMI->memoperands_begin();
- else
- // Unknown memory accesses. Assume the worst.
- ChainMMO = 0;
- } else if (TID.mayStore()) {
- if (MI->hasOneMemOperand() &&
- MI->memoperands_begin()->getValue() &&
- !MI->memoperands_begin()->isVolatile() &&
- isa<PseudoSourceValue>(MI->memoperands_begin()->getValue())) {
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
+ TrueMemOrderLatency);
+ PendingLoads.clear();
+ AliasMemDefs.clear();
+ AliasMemUses.clear();
+ } else if (MI->mayStore()) {
+ bool MayAlias = true;
+ if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
// A store to a specific PseudoSourceValue. Add precise dependencies.
- const Value *V = MI->memoperands_begin()->getValue();
- // Handle the def in MemDefs, if there is one.
- std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V);
- if (I != MemDefs.end()) {
- I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
- /*isNormalMemory=*/true));
+ // 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());
+ if (I != IE) {
+ addChainDependency(AA, MFI, SU, I->second, RejectMemNodes,
+ 0, true);
I->second = SU;
} else {
- MemDefs[V] = SU;
+ if (MayAlias)
+ 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 =
- MemUses.find(V);
- if (J != MemUses.end()) {
+ ((MayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
+ std::map<const Value *, std::vector<SUnit *> >::iterator JE =
+ ((MayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
+ if (J != JE) {
for (unsigned i = 0, e = J->second.size(); i != e; ++i)
- J->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
- /*isNormalMemory=*/true));
+ addChainDependency(AA, MFI, SU, J->second[i], RejectMemNodes,
+ TrueMemOrderLatency, true);
J->second.clear();
}
- // Add a general dependence too, if needed.
- if (Chain)
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- } else
+ 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,
+ 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));
+ } else {
// Treat all other stores conservatively.
- goto new_chain;
- } else if (TID.mayLoad()) {
- if (TII->isInvariantLoad(MI)) {
+ goto new_alias_chain;
+ }
+
+ if (!ExitSU.isPred(SU))
+ // Push store's up a bit to avoid them getting in between cmp
+ // and branches.
+ ExitSU.addPred(SDep(SU, SDep::Artificial));
+ } else if (MI->mayLoad()) {
+ bool MayAlias = true;
+ if (MI->isInvariantLoad(AA)) {
// Invariant load, no chain dependencies needed!
- } else if (MI->hasOneMemOperand() &&
- MI->memoperands_begin()->getValue() &&
- !MI->memoperands_begin()->isVolatile() &&
- isa<PseudoSourceValue>(MI->memoperands_begin()->getValue())) {
- // A load from a specific PseudoSourceValue. Add precise dependencies.
- const Value *V = MI->memoperands_begin()->getValue();
- std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V);
- if (I != MemDefs.end())
- I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
- /*isNormalMemory=*/true));
- MemUses[V].push_back(SU);
-
- // Add a general dependence too, if needed.
- if (Chain && (!ChainMMO ||
- (ChainMMO->isStore() || ChainMMO->isVolatile())))
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- } else if (MI->hasVolatileMemoryRef()) {
- // Treat volatile loads conservatively. Note that this includes
- // cases where memoperand information is unavailable.
- goto new_chain;
} else {
- // A normal load. Just depend on the general chain.
- if (Chain)
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- PendingLoads.push_back(SU);
+ 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 {
+ // A load with no underlying object. Depend on all
+ // potentially aliasing stores.
+ for (std::map<const Value *, SUnit *>::iterator I =
+ AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
+ addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
+
+ PendingLoads.push_back(SU);
+ MayAlias = true;
+ }
+ if (MayAlias)
+ 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);
+ if (BarrierChain)
+ BarrierChain->addPred(SDep(SU, SDep::Barrier));
}
}
-
- // Add chain edges from the terminator to ensure that all the work of the
- // block is completed before any control transfers.
- if (Terminator && SU->Succs.empty())
- Terminator->addPred(SDep(SU, SDep::Order, SU->Latency));
- if (TID.isTerminator() || MI->isLabel())
- Terminator = SU;
}
-}
-
-void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) {
- const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
-
- // Compute the latency for the node. We use the sum of the latencies for
- // all nodes flagged together into this SUnit.
- SU->Latency =
- InstrItins.getLatency(SU->getInstr()->getDesc().getSchedClass());
+ if (PrevMI)
+ FirstDbgValue = PrevMI;
- // Simplistic target-independent heuristic: assume that loads take
- // extra time.
- if (InstrItins.isEmpty())
- if (SU->getInstr()->getDesc().mayLoad())
- SU->Latency += 2;
+ Defs.clear();
+ Uses.clear();
+ VRegDefs.clear();
+ PendingLoads.clear();
}
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 {
std::string s;
raw_string_ostream oss(s);
- SU->getInstr()->print(oss);
+ if (SU == &EntrySU)
+ oss << "<entry>";
+ else if (SU == &ExitSU)
+ oss << "<exit>";
+ else
+ SU->getInstr()->print(oss);
return oss.str();
}
-// EmitSchedule - Emit the machine code in scheduled order.
-MachineBasicBlock *ScheduleDAGInstrs::EmitSchedule() {
- // For MachineInstr-based scheduling, we're rescheduling the instructions in
- // the block, so start by removing them from the block.
- while (!BB->empty())
- BB->remove(BB->begin());
+/// Return the basic block label. It is not necessarilly unique because a block
+/// contains multiple scheduling regions. But it is fine for visualization.
+std::string ScheduleDAGInstrs::getDAGName() const {
+ return "dag." + BB->getFullName();
+}
- for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
- SUnit *SU = Sequence[i];
- if (!SU) {
- // Null SUnit* is a noop.
- EmitNoop();
- continue;
- }
+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; }
+
+ void backtrack() { DFSStack.pop_back(); }
+
+ const SUnit *getCurr() const { return DFSStack.back().first; }
+
+ SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
- BB->push_back(SU->getInstr());
+ SUnit::const_pred_iterator getPredEnd() const {
+ return getCurr()->Preds.end();
}
+};
+} // anonymous
- return BB;
+void ScheduleDAGILP::resize(unsigned NumSUnits) {
+ ILPValues.resize(NumSUnits);
}
+
+ILPValue ScheduleDAGILP::getILP(const SUnit *SU) {
+ return ILPValues[SU->NodeNum];
+}
+
+// A leaf node has an ILP of 1/1.
+static ILPValue initILP(const SUnit *SU) {
+ unsigned Cnt = SU->getInstr()->isTransient() ? 0 : 1;
+ return ILPValue(Cnt, 1 + SU->getDepth());
+}
+
+/// Compute an ILP metric for all nodes in the subDAG reachable via depth-first
+/// search from this root.
+void ScheduleDAGILP::computeILP(const SUnit *Root) {
+ if (!IsBottomUp)
+ llvm_unreachable("Top-down ILP metric is unimplemnted");
+
+ SchedDAGReverseDFS DFS;
+ // Mark a node visited by validating it.
+ ILPValues[Root->NodeNum] = initILP(Root);
+ DFS.follow(Root);
+ for (;;) {
+ // Traverse the leftmost path as far as possible.
+ while (DFS.getPred() != DFS.getPredEnd()) {
+ const SUnit *PredSU = DFS.getPred()->getSUnit();
+ DFS.advance();
+ // If the pred is already valid, skip it.
+ if (ILPValues[PredSU->NodeNum].isValid())
+ continue;
+ ILPValues[PredSU->NodeNum] = initILP(PredSU);
+ DFS.follow(PredSU);
+ }
+ // Visit the top of the stack in postorder and backtrack.
+ unsigned PredCount = ILPValues[DFS.getCurr()->NodeNum].InstrCount;
+ DFS.backtrack();
+ if (DFS.isComplete())
+ break;
+ // Add the recently finished predecessor's bottom-up descendent count.
+ ILPValues[DFS.getCurr()->NodeNum].InstrCount += PredCount;
+ }
+}
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+void ILPValue::print(raw_ostream &OS) const {
+ if (!isValid())
+ OS << "BADILP";
+ OS << InstrCount << " / " << Cycles << " = "
+ << format("%g", ((double)InstrCount / Cycles));
+}
+
+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
projects / oota-llvm.git / blobdiff