cl::desc("Disable cycle-level precision during preRA scheduling"));
// Temporary sched=list-ilp flags until the heuristics are robust.
+// Some options are also available under sched=list-hybrid.
static cl::opt<bool> DisableSchedRegPressure(
"disable-sched-reg-pressure", cl::Hidden, cl::init(false),
cl::desc("Disable regpressure priority in sched=list-ilp"));
static cl::opt<bool> DisableSchedVRegCycle(
"disable-sched-vrcycle", cl::Hidden, cl::init(false),
cl::desc("Disable virtual register cycle interference checks"));
+static cl::opt<bool> DisableSchedPhysRegJoin(
+ "disable-sched-physreg-join", cl::Hidden, cl::init(false),
+ cl::desc("Disable physreg def-use affinity"));
static cl::opt<bool> DisableSchedStalls(
"disable-sched-stalls", cl::Hidden, cl::init(true),
cl::desc("Disable no-stall priority in sched=list-ilp"));
};
} // end anonymous namespace
+/// GetCostForDef - Looks up the register class and cost for a given definition.
+/// Typically this just means looking up the representative register class,
+/// but for untyped values (MVT::untyped) it means inspecting the node's
+/// opcode to determine what register class is being generated.
+static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos,
+ const TargetLowering *TLI,
+ const TargetInstrInfo *TII,
+ const TargetRegisterInfo *TRI,
+ unsigned &RegClass, unsigned &Cost) {
+ EVT VT = RegDefPos.GetValue();
+
+ // Special handling for untyped values. These values can only come from
+ // the expansion of custom DAG-to-DAG patterns.
+ if (VT == MVT::untyped) {
+ const SDNode *Node = RegDefPos.GetNode();
+ unsigned Opcode = Node->getMachineOpcode();
+
+ if (Opcode == TargetOpcode::REG_SEQUENCE) {
+ unsigned DstRCIdx = cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue();
+ const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx);
+ RegClass = RC->getID();
+ Cost = 1;
+ return;
+ }
+
+ unsigned Idx = RegDefPos.GetIdx();
+ const TargetInstrDesc Desc = TII->get(Opcode);
+ const TargetRegisterClass *RC = Desc.getRegClass(Idx, TRI);
+ RegClass = RC->getID();
+ // FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a
+ // better way to determine it.
+ Cost = 1;
+ } else {
+ RegClass = TLI->getRepRegClassFor(VT)->getID();
+ Cost = TLI->getRepRegClassCostFor(VT);
+ }
+}
/// Schedule - Schedule the DAG using list scheduling.
void ScheduleDAGRRList::Schedule() {
for (const unsigned *AliasI = TRI->getOverlaps(Reg); *AliasI; ++AliasI) {
// Check if Ref is live.
- if (!LiveRegDefs[Reg]) continue;
+ if (!LiveRegDefs[*AliasI]) continue;
// Allow multiple uses of the same def.
- if (LiveRegDefs[Reg] == SU) continue;
+ if (LiveRegDefs[*AliasI] == SU) continue;
// Add Reg to the set of interfering live regs.
- if (RegAdded.insert(Reg))
- LRegs.push_back(Reg);
+ if (RegAdded.insert(*AliasI)) {
+ LRegs.push_back(*AliasI);
+ }
}
}
bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
};
+#ifndef NDEBUG
+template<class SF>
+struct reverse_sort : public queue_sort {
+ SF &SortFunc;
+ reverse_sort(SF &sf) : SortFunc(sf) {}
+ reverse_sort(const reverse_sort &RHS) : SortFunc(RHS.SortFunc) {}
+
+ bool operator()(SUnit* left, SUnit* right) const {
+ // reverse left/right rather than simply !SortFunc(left, right)
+ // to expose different paths in the comparison logic.
+ return SortFunc(right, left);
+ }
+};
+#endif // NDEBUG
+
/// bu_ls_rr_sort - Priority function for bottom up register pressure
// reduction scheduler.
struct bu_ls_rr_sort : public queue_sort {
};
template<class SF>
-class RegReductionPriorityQueue : public RegReductionPQBase {
- static SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker) {
- std::vector<SUnit *>::iterator Best = Q.begin();
- for (std::vector<SUnit *>::iterator I = llvm::next(Q.begin()),
- E = Q.end(); I != E; ++I)
- if (Picker(*Best, *I))
- Best = I;
- SUnit *V = *Best;
- if (Best != prior(Q.end()))
- std::swap(*Best, Q.back());
- Q.pop_back();
- return V;
+static SUnit *popFromQueueImpl(std::vector<SUnit*> &Q, SF &Picker) {
+ std::vector<SUnit *>::iterator Best = Q.begin();
+ for (std::vector<SUnit *>::iterator I = llvm::next(Q.begin()),
+ E = Q.end(); I != E; ++I)
+ if (Picker(*Best, *I))
+ Best = I;
+ SUnit *V = *Best;
+ if (Best != prior(Q.end()))
+ std::swap(*Best, Q.back());
+ Q.pop_back();
+ return V;
+}
+
+template<class SF>
+SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker, ScheduleDAG *DAG) {
+#ifndef NDEBUG
+ if (DAG->StressSched) {
+ reverse_sort<SF> RPicker(Picker);
+ return popFromQueueImpl(Q, RPicker);
}
+#endif
+ (void)DAG;
+ return popFromQueueImpl(Q, Picker);
+}
+template<class SF>
+class RegReductionPriorityQueue : public RegReductionPQBase {
SF Picker;
public:
SUnit *pop() {
if (Queue.empty()) return NULL;
- SUnit *V = popFromQueue(Queue, Picker);
+ SUnit *V = popFromQueue(Queue, Picker, scheduleDAG);
V->NodeQueueId = 0;
return V;
}
std::vector<SUnit*> DumpQueue = Queue;
SF DumpPicker = Picker;
while (!DumpQueue.empty()) {
- SUnit *SU = popFromQueue(DumpQueue, DumpPicker);
+ SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG);
if (isBottomUp())
dbgs() << "Height " << SU->getHeight() << ": ";
else
// Static Node Priority for Register Pressure Reduction
//===----------------------------------------------------------------------===//
+// Check for special nodes that bypass scheduling heuristics.
+// Currently this pushes TokenFactor nodes down, but may be used for other
+// pseudo-ops as well.
+//
+// Return -1 to schedule right above left, 1 for left above right.
+// Return 0 if no bias exists.
+static int checkSpecialNodes(const SUnit *left, const SUnit *right) {
+ bool LSchedLow = left->isScheduleLow;
+ bool RSchedLow = right->isScheduleLow;
+ if (LSchedLow != RSchedLow)
+ return LSchedLow < RSchedLow ? 1 : -1;
+ return 0;
+}
+
/// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
/// Smaller number is the higher priority.
static unsigned
// If SU does not have a register def, schedule it close to its uses
// because it does not lengthen any live ranges.
return 0;
+#if 1
return SethiUllmanNumbers[SU->NodeNum];
+#else
+ unsigned Priority = SethiUllmanNumbers[SU->NodeNum];
+ if (SU->isCallOp) {
+ // FIXME: This assumes all of the defs are used as call operands.
+ int NP = (int)Priority - SU->getNode()->getNumValues();
+ return (NP > 0) ? NP : 0;
+ }
+ return Priority;
+#endif
}
//===----------------------------------------------------------------------===//
}
for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
RegDefPos.IsValid(); RegDefPos.Advance()) {
- EVT VT = RegDefPos.GetValue();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- unsigned Cost = TLI->getRepRegClassCostFor(VT);
+ unsigned RCId, Cost;
+ GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost);
+
if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
return true;
}
RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
if (SkipRegDefs)
continue;
- EVT VT = RegDefPos.GetValue();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+
+ unsigned RCId, Cost;
+ GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost);
+ RegPressure[RCId] += Cost;
break;
}
}
RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
if (SkipRegDefs > 0)
continue;
- EVT VT = RegDefPos.GetValue();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT)) {
+ unsigned RCId, Cost;
+ GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost);
+ if (RegPressure[RCId] < Cost) {
// Register pressure tracking is imprecise. This can happen. But we try
// hard not to let it happen because it likely results in poor scheduling.
DEBUG(dbgs() << " SU(" << SU->NodeNum << ") has too many regdefs\n");
RegPressure[RCId] = 0;
}
else {
- RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
+ RegPressure[RCId] -= Cost;
}
}
dumpRegPressure();
}
static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
+ // Schedule physical register definitions close to their use. This is
+ // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as
+ // long as shortening physreg live ranges is generally good, we can defer
+ // creating a subtarget hook.
+ if (!DisableSchedPhysRegJoin) {
+ bool LHasPhysReg = left->hasPhysRegDefs;
+ bool RHasPhysReg = right->hasPhysRegDefs;
+ if (LHasPhysReg != RHasPhysReg) {
+ DEBUG(++FactorCount[FactRegUses]);
+ #ifndef NDEBUG
+ const char *PhysRegMsg[] = {" has no physreg", " defines a physreg"};
+ #endif
+ DEBUG(dbgs() << " SU (" << left->NodeNum << ") "
+ << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") "
+ << PhysRegMsg[RHasPhysReg] << "\n");
+ return LHasPhysReg < RHasPhysReg;
+ }
+ }
+
+ // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down.
unsigned LPriority = SPQ->getNodePriority(left);
unsigned RPriority = SPQ->getNodePriority(right);
+
+ // Be really careful about hoisting call operands above previous calls.
+ // Only allows it if it would reduce register pressure.
+ if (left->isCall && right->isCallOp) {
+ unsigned RNumVals = right->getNode()->getNumValues();
+ RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0;
+ }
+ if (right->isCall && left->isCallOp) {
+ unsigned LNumVals = left->getNode()->getNumValues();
+ LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0;
+ }
+
if (LPriority != RPriority) {
DEBUG(++FactorCount[FactStatic]);
return LPriority > RPriority;
}
- else if(LPriority == 0) {
- // Schedule zero-latency TokenFactor below any other special
- // nodes. The alternative may be to avoid artificially boosting the
- // TokenFactor's height when it is scheduled, but we currently rely on an
- // instruction's final height to equal the cycle in which it is scheduled,
- // so heights are monotonically increasing.
- unsigned LOpc = left->getNode() ? left->getNode()->getOpcode() : 0;
- unsigned ROpc = right->getNode() ? right->getNode()->getOpcode() : 0;
- if (LOpc == ISD::TokenFactor)
- return false;
- if (ROpc == ISD::TokenFactor)
- return true;
+
+ // One or both of the nodes are calls and their sethi-ullman numbers are the
+ // same, then keep source order.
+ if (left->isCall || right->isCall) {
+ unsigned LOrder = SPQ->getNodeOrdering(left);
+ unsigned ROrder = SPQ->getNodeOrdering(right);
+
+ // Prefer an ordering where the lower the non-zero order number, the higher
+ // the preference.
+ if ((LOrder || ROrder) && LOrder != ROrder)
+ return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
}
// Try schedule def + use closer when Sethi-Ullman numbers are the same.
return LScratch > RScratch;
}
- if (!DisableSchedCycles) {
+ // Comparing latency against a call makes little sense unless the node
+ // is register pressure-neutral.
+ if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0))
+ return (left->NodeQueueId > right->NodeQueueId);
+
+ // Do not compare latencies when one or both of the nodes are calls.
+ if (!DisableSchedCycles &&
+ !(left->isCall || right->isCall)) {
int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
if (result != 0)
return result > 0;
// Bottom up
bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ if (int res = checkSpecialNodes(left, right))
+ return res > 0;
+
return BURRSort(left, right, SPQ);
}
// Source order, otherwise bottom up.
bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ if (int res = checkSpecialNodes(left, right))
+ return res > 0;
+
unsigned LOrder = SPQ->getNodeOrdering(left);
unsigned ROrder = SPQ->getNodeOrdering(right);
// Return true if right should be scheduled with higher priority than left.
bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ if (int res = checkSpecialNodes(left, right))
+ return res > 0;
+
if (left->isCall || right->isCall)
// No way to compute latency of calls.
return BURRSort(left, right, SPQ);
// list-ilp is currently an experimental scheduler that allows various
// heuristics to be enabled prior to the normal register reduction logic.
bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ if (int res = checkSpecialNodes(left, right))
+ return res > 0;
+
if (left->isCall || right->isCall)
// No way to compute latency of calls.
return BURRSort(left, right, SPQ);
// Top down
bool td_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
+ if (int res = checkSpecialNodes(left, right))
+ return res < 0;
+
unsigned LPriority = SPQ->getNodePriority(left);
unsigned RPriority = SPQ->getNodePriority(right);
bool LIsTarget = left->getNode() && left->getNode()->isMachineOpcode();
projects / oota-llvm.git / blobdiff