aboutsummaryrefslogtreecommitdiffstats
path: root/lib
diff options
context:
space:
mode:
Diffstat (limited to 'lib')
-rw-r--r--lib/CodeGen/IfConversion.cpp59
-rw-r--r--lib/CodeGen/ScheduleDAGInstrs.cpp8
-rw-r--r--lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp8
-rw-r--r--lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp11
-rw-r--r--lib/Target/ARM/ARMBaseInstrInfo.cpp100
-rw-r--r--lib/Target/ARM/ARMBaseInstrInfo.h22
-rw-r--r--lib/Target/ARM/Thumb2InstrInfo.cpp27
-rw-r--r--lib/Target/ARM/Thumb2InstrInfo.h6
-rw-r--r--lib/Target/TargetInstrInfo.cpp24
9 files changed, 155 insertions, 110 deletions
diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp
index e90985b..c05aa40 100644
--- a/lib/CodeGen/IfConversion.cpp
+++ b/lib/CodeGen/IfConversion.cpp
@@ -93,7 +93,8 @@ namespace {
/// ClobbersPred - True if BB could modify predicates (e.g. has
/// cmp, call, etc.)
/// NonPredSize - Number of non-predicated instructions.
- /// ExtraCost - Extra cost for microcoded instructions.
+ /// ExtraCost - Extra cost for multi-cycle instructions.
+ /// ExtraCost2 - Some instructions are slower when predicated
/// BB - Corresponding MachineBasicBlock.
/// TrueBB / FalseBB- See AnalyzeBranch().
/// BrCond - Conditions for end of block conditional branches.
@@ -110,6 +111,7 @@ namespace {
bool ClobbersPred : 1;
unsigned NonPredSize;
unsigned ExtraCost;
+ unsigned ExtraCost2;
MachineBasicBlock *BB;
MachineBasicBlock *TrueBB;
MachineBasicBlock *FalseBB;
@@ -119,7 +121,7 @@ namespace {
IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false),
HasFallThrough(false), IsUnpredicable(false),
CannotBeCopied(false), ClobbersPred(false), NonPredSize(0),
- ExtraCost(0), BB(0), TrueBB(0), FalseBB(0) {}
+ ExtraCost(0), ExtraCost2(0), BB(0), TrueBB(0), FalseBB(0) {}
};
/// IfcvtToken - Record information about pending if-conversions to attempt:
@@ -203,17 +205,20 @@ namespace {
bool IgnoreBr = false);
void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true);
- bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, unsigned Size,
+ bool MeetIfcvtSizeLimit(MachineBasicBlock &BB,
+ unsigned Cycle, unsigned Extra,
float Prediction, float Confidence) const {
- return Size > 0 && TII->isProfitableToIfCvt(BB, Size,
- Prediction, Confidence);
+ return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra,
+ Prediction, Confidence);
}
- bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, unsigned TSize,
- MachineBasicBlock &FBB, unsigned FSize,
+ bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB,
+ unsigned TCycle, unsigned TExtra,
+ MachineBasicBlock &FBB,
+ unsigned FCycle, unsigned FExtra,
float Prediction, float Confidence) const {
- return TSize > 0 && FSize > 0 &&
- TII->isProfitableToIfCvt(TBB, TSize, FBB, FSize,
+ return TCycle > 0 && FCycle > 0 &&
+ TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra,
Prediction, Confidence);
}
@@ -649,6 +654,7 @@ void IfConverter::ScanInstructions(BBInfo &BBI) {
// Then scan all the instructions.
BBI.NonPredSize = 0;
BBI.ExtraCost = 0;
+ BBI.ExtraCost2 = 0;
BBI.ClobbersPred = false;
for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end();
I != E; ++I) {
@@ -665,9 +671,12 @@ void IfConverter::ScanInstructions(BBInfo &BBI) {
if (!isCondBr) {
if (!isPredicated) {
BBI.NonPredSize++;
- unsigned NumOps = TII->getNumMicroOps(&*I, InstrItins);
- if (NumOps > 1)
- BBI.ExtraCost += NumOps-1;
+ unsigned ExtraPredCost = 0;
+ unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I,
+ &ExtraPredCost);
+ if (NumCycles > 1)
+ BBI.ExtraCost += NumCycles-1;
+ BBI.ExtraCost2 += ExtraPredCost;
} else if (!AlreadyPredicated) {
// FIXME: This instruction is already predicated before the
// if-conversion pass. It's probably something like a conditional move.
@@ -815,9 +824,9 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) &&
MeetIfcvtSizeLimit(*TrueBBI.BB, (TrueBBI.NonPredSize - (Dups + Dups2) +
- TrueBBI.ExtraCost),
+ TrueBBI.ExtraCost), TrueBBI.ExtraCost2,
*FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) +
- FalseBBI.ExtraCost),
+ FalseBBI.ExtraCost),FalseBBI.ExtraCost2,
Prediction, Confidence) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond) &&
FeasibilityAnalysis(FalseBBI, RevCond)) {
@@ -836,7 +845,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction, Confidence) &&
MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
- Prediction, Confidence) &&
+ TrueBBI.ExtraCost2, Prediction, Confidence) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) {
// Triangle:
// EBB
@@ -851,7 +860,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction, Confidence) &&
MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
- Prediction, Confidence) &&
+ TrueBBI.ExtraCost2, Prediction, Confidence) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) {
Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups));
Enqueued = true;
@@ -859,7 +868,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
if (ValidSimple(TrueBBI, Dups, Prediction, Confidence) &&
MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost,
- Prediction, Confidence) &&
+ TrueBBI.ExtraCost2, Prediction, Confidence) &&
FeasibilityAnalysis(TrueBBI, BBI.BrCond)) {
// Simple (split, no rejoin):
// EBB
@@ -878,7 +887,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
1.0-Prediction, Confidence) &&
MeetIfcvtSizeLimit(*FalseBBI.BB,
FalseBBI.NonPredSize + FalseBBI.ExtraCost,
- 1.0-Prediction, Confidence) &&
+ FalseBBI.ExtraCost2, 1.0-Prediction, Confidence) &&
FeasibilityAnalysis(FalseBBI, RevCond, true)) {
Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups));
Enqueued = true;
@@ -888,7 +897,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
1.0-Prediction, Confidence) &&
MeetIfcvtSizeLimit(*FalseBBI.BB,
FalseBBI.NonPredSize + FalseBBI.ExtraCost,
- 1.0-Prediction, Confidence) &&
+ FalseBBI.ExtraCost2, 1.0-Prediction, Confidence) &&
FeasibilityAnalysis(FalseBBI, RevCond, true, true)) {
Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups));
Enqueued = true;
@@ -897,7 +906,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB,
if (ValidSimple(FalseBBI, Dups, 1.0-Prediction, Confidence) &&
MeetIfcvtSizeLimit(*FalseBBI.BB,
FalseBBI.NonPredSize + FalseBBI.ExtraCost,
- 1.0-Prediction, Confidence) &&
+ FalseBBI.ExtraCost2, 1.0-Prediction, Confidence) &&
FeasibilityAnalysis(FalseBBI, RevCond)) {
Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups));
Enqueued = true;
@@ -1427,9 +1436,11 @@ void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
MachineInstr *MI = MF.CloneMachineInstr(I);
ToBBI.BB->insert(ToBBI.BB->end(), MI);
ToBBI.NonPredSize++;
- unsigned NumOps = TII->getNumMicroOps(MI, InstrItins);
- if (NumOps > 1)
- ToBBI.ExtraCost += NumOps-1;
+ unsigned ExtraPredCost = 0;
+ unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, &ExtraPredCost);
+ if (NumCycles > 1)
+ ToBBI.ExtraCost += NumCycles-1;
+ ToBBI.ExtraCost2 += ExtraPredCost;
if (!TII->isPredicated(I) && !MI->isDebugValue()) {
if (!TII->PredicateInstruction(MI, Cond)) {
@@ -1504,8 +1515,10 @@ void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) {
ToBBI.NonPredSize += FromBBI.NonPredSize;
ToBBI.ExtraCost += FromBBI.ExtraCost;
+ ToBBI.ExtraCost2 += FromBBI.ExtraCost2;
FromBBI.NonPredSize = 0;
FromBBI.ExtraCost = 0;
+ FromBBI.ExtraCost2 = 0;
ToBBI.ClobbersPred |= FromBBI.ClobbersPred;
ToBBI.HasFallThrough = FromBBI.HasFallThrough;
diff --git a/lib/CodeGen/ScheduleDAGInstrs.cpp b/lib/CodeGen/ScheduleDAGInstrs.cpp
index abd68ca..e86a78c 100644
--- a/lib/CodeGen/ScheduleDAGInstrs.cpp
+++ b/lib/CodeGen/ScheduleDAGInstrs.cpp
@@ -238,6 +238,8 @@ void ScheduleDAGInstrs::BuildSchedGraph(AliasAnalysis *AA) {
"Cannot schedule terminators or labels!");
// Create the SUnit for this MI.
SUnit *SU = NewSUnit(MI);
+ SU->isCall = TID.isCall();
+ SU->isCommutable = TID.isCommutable();
// Assign the Latency field of SU using target-provided information.
if (UnitLatencies)
@@ -564,9 +566,9 @@ void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) {
// extra time.
if (SU->getInstr()->getDesc().mayLoad())
SU->Latency += 2;
- } else
- SU->Latency =
- InstrItins->getStageLatency(SU->getInstr()->getDesc().getSchedClass());
+ } else {
+ SU->Latency = TII->getInstrLatency(InstrItins, SU->getInstr());
+ }
}
void ScheduleDAGInstrs::ComputeOperandLatency(SUnit *Def, SUnit *Use,
diff --git a/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp b/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
index ea1aaa1..9978d00 100644
--- a/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
+++ b/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
@@ -1589,6 +1589,10 @@ bool src_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
}
bool hybrid_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const{
+ if (left->isCall || right->isCall)
+ // No way to compute latency of calls.
+ return BURRSort(left, right, SPQ);
+
bool LHigh = SPQ->HighRegPressure(left);
bool RHigh = SPQ->HighRegPressure(right);
// Avoid causing spills. If register pressure is high, schedule for
@@ -1648,6 +1652,10 @@ bool hybrid_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const{
bool ilp_ls_rr_sort::operator()(const SUnit *left,
const SUnit *right) const {
+ if (left->isCall || right->isCall)
+ // No way to compute latency of calls.
+ return BURRSort(left, right, SPQ);
+
bool LHigh = SPQ->HighRegPressure(left);
bool RHigh = SPQ->HighRegPressure(right);
// Avoid causing spills. If register pressure is high, schedule for
diff --git a/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp b/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp
index 7d01bd3..429b115 100644
--- a/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp
+++ b/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp
@@ -72,6 +72,7 @@ SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
SUnit *SU = NewSUnit(Old->getNode());
SU->OrigNode = Old->OrigNode;
SU->Latency = Old->Latency;
+ SU->isCall = Old->isCall;
SU->isTwoAddress = Old->isTwoAddress;
SU->isCommutable = Old->isCommutable;
SU->hasPhysRegDefs = Old->hasPhysRegDefs;
@@ -300,6 +301,8 @@ void ScheduleDAGSDNodes::BuildSchedUnits() {
N = N->getOperand(N->getNumOperands()-1).getNode();
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
+ if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
+ NodeSUnit->isCall = true;
}
// Scan down to find any flagged succs.
@@ -316,6 +319,8 @@ void ScheduleDAGSDNodes::BuildSchedUnits() {
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
N = *UI;
+ if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
+ NodeSUnit->isCall = true;
break;
}
if (!HasFlagUse) break;
@@ -438,10 +443,8 @@ void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
// all nodes flagged together into this SUnit.
SU->Latency = 0;
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
- if (N->isMachineOpcode()) {
- SU->Latency += InstrItins->
- getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
- }
+ if (N->isMachineOpcode())
+ SU->Latency += TII->getInstrLatency(InstrItins, N);
}
void ScheduleDAGSDNodes::ComputeOperandLatency(SDNode *Def, SDNode *Use,
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 3257065..b023379 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -40,10 +40,6 @@ static cl::opt<bool>
EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
cl::desc("Enable ARM 2-addr to 3-addr conv"));
-static cl::opt<bool>
-OldARMIfCvt("old-arm-ifcvt", cl::Hidden,
- cl::desc("Use old-style ARM if-conversion heuristics"));
-
ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
: TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)),
Subtarget(STI) {
@@ -1205,53 +1201,36 @@ bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
}
bool ARMBaseInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
- unsigned NumInstrs,
+ unsigned NumCyles,
+ unsigned ExtraPredCycles,
float Probability,
float Confidence) const {
- if (!NumInstrs)
+ if (!NumCyles)
return false;
- // Use old-style heuristics
- if (OldARMIfCvt) {
- if (Subtarget.getCPUString() == "generic")
- // Generic (and overly aggressive) if-conversion limits for testing.
- return NumInstrs <= 10;
- if (Subtarget.hasV7Ops())
- return NumInstrs <= 3;
- return NumInstrs <= 2;
- }
-
// Attempt to estimate the relative costs of predication versus branching.
- float UnpredCost = Probability * NumInstrs;
+ float UnpredCost = Probability * NumCyles;
UnpredCost += 1.0; // The branch itself
UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
- float PredCost = NumInstrs;
-
- return PredCost < UnpredCost;
-
+ return (float)(NumCyles + ExtraPredCycles) < UnpredCost;
}
bool ARMBaseInstrInfo::
-isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
- MachineBasicBlock &FMBB, unsigned NumF,
+isProfitableToIfCvt(MachineBasicBlock &TMBB,
+ unsigned TCycles, unsigned TExtra,
+ MachineBasicBlock &FMBB,
+ unsigned FCycles, unsigned FExtra,
float Probability, float Confidence) const {
- // Use old-style if-conversion heuristics
- if (OldARMIfCvt) {
- return NumT && NumF && NumT <= 2 && NumF <= 2;
- }
-
- if (!NumT || !NumF)
+ if (!TCycles || !FCycles)
return false;
// Attempt to estimate the relative costs of predication versus branching.
- float UnpredCost = Probability * NumT + (1.0 - Probability) * NumF;
+ float UnpredCost = Probability * TCycles + (1.0 - Probability) * FCycles;
UnpredCost += 1.0; // The branch itself
UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
- float PredCost = NumT + NumF;
-
- return PredCost < UnpredCost;
+ return (float)(TCycles + FCycles + TExtra + FExtra) < UnpredCost;
}
/// getInstrPredicate - If instruction is predicated, returns its predicate
@@ -1591,8 +1570,8 @@ OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
}
unsigned
-ARMBaseInstrInfo::getNumMicroOps(const MachineInstr *MI,
- const InstrItineraryData *ItinData) const {
+ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
+ const MachineInstr *MI) const {
if (!ItinData || ItinData->isEmpty())
return 1;
@@ -1649,9 +1628,14 @@ ARMBaseInstrInfo::getNumMicroOps(const MachineInstr *MI,
case ARM::t2STM_UPD: {
unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
if (Subtarget.isCortexA8()) {
- // 4 registers would be issued: 1, 2, 1.
- // 5 registers would be issued: 1, 2, 2.
- return 1 + (NumRegs / 2);
+ if (NumRegs < 4)
+ return 2;
+ // 4 registers would be issued: 2, 2.
+ // 5 registers would be issued: 2, 2, 1.
+ UOps = (NumRegs / 2);
+ if (NumRegs % 2)
+ ++UOps;
+ return UOps;
} else if (Subtarget.isCortexA9()) {
UOps = (NumRegs / 2);
// If there are odd number of registers or if it's not 64-bit aligned,
@@ -2025,6 +2009,46 @@ ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
return Latency;
}
+int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *MI,
+ unsigned *PredCost) const {
+ if (MI->isCopyLike() || MI->isInsertSubreg() ||
+ MI->isRegSequence() || MI->isImplicitDef())
+ return 1;
+
+ if (!ItinData || ItinData->isEmpty())
+ return 1;
+
+ const TargetInstrDesc &TID = MI->getDesc();
+ unsigned Class = TID.getSchedClass();
+ unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
+ if (PredCost && TID.hasImplicitDefOfPhysReg(ARM::CPSR))
+ // When predicated, CPSR is an additional source operand for CPSR updating
+ // instructions, this apparently increases their latencies.
+ *PredCost = 1;
+ if (UOps)
+ return ItinData->getStageLatency(Class);
+ return getNumMicroOps(ItinData, MI);
+}
+
+int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
+ SDNode *Node) const {
+ if (!Node->isMachineOpcode())
+ return 1;
+
+ if (!ItinData || ItinData->isEmpty())
+ return 1;
+
+ unsigned Opcode = Node->getMachineOpcode();
+ switch (Opcode) {
+ default:
+ return ItinData->getStageLatency(get(Opcode).getSchedClass());
+ case ARM::VLDMQ:
+ case ARM::VSTMQ:
+ return 2;
+ }
+}
+
bool ARMBaseInstrInfo::
hasHighOperandLatency(const InstrItineraryData *ItinData,
const MachineRegisterInfo *MRI,
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.h b/lib/Target/ARM/ARMBaseInstrInfo.h
index e19bd9c..c11f02c 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.h
+++ b/lib/Target/ARM/ARMBaseInstrInfo.h
@@ -318,18 +318,20 @@ public:
const MachineFunction &MF) const;
virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB,
- unsigned NumInstrs,
+ unsigned NumCyles, unsigned ExtraPredCycles,
float Prob, float Confidence) const;
- virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB,unsigned NumT,
- MachineBasicBlock &FMBB,unsigned NumF,
+ virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB,
+ unsigned NumT, unsigned ExtraT,
+ MachineBasicBlock &FMBB,
+ unsigned NumF, unsigned ExtraF,
float Probability, float Confidence) const;
virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
- unsigned NumInstrs,
+ unsigned NumCyles,
float Probability,
float Confidence) const {
- return NumInstrs == 1;
+ return NumCyles == 1;
}
/// AnalyzeCompare - For a comparison instruction, return the source register
@@ -345,8 +347,8 @@ public:
const MachineRegisterInfo *MRI,
MachineBasicBlock::iterator &MII) const;
- virtual unsigned getNumMicroOps(const MachineInstr *MI,
- const InstrItineraryData *ItinData) const;
+ virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
+ const MachineInstr *MI) const;
virtual
int getOperandLatency(const InstrItineraryData *ItinData,
@@ -379,6 +381,12 @@ private:
const TargetInstrDesc &UseTID,
unsigned UseIdx, unsigned UseAlign) const;
+ int getInstrLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *MI, unsigned *PredCost = 0) const;
+
+ int getInstrLatency(const InstrItineraryData *ItinData,
+ SDNode *Node) const;
+
bool hasHighOperandLatency(const InstrItineraryData *ItinData,
const MachineRegisterInfo *MRI,
const MachineInstr *DefMI, unsigned DefIdx,
diff --git a/lib/Target/ARM/Thumb2InstrInfo.cpp b/lib/Target/ARM/Thumb2InstrInfo.cpp
index 0a0f314..719b140 100644
--- a/lib/Target/ARM/Thumb2InstrInfo.cpp
+++ b/lib/Target/ARM/Thumb2InstrInfo.cpp
@@ -42,33 +42,6 @@ unsigned Thumb2InstrInfo::getUnindexedOpcode(unsigned Opc) const {
return 0;
}
-bool Thumb2InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
- unsigned NumInstrs,
- float Prediction,
- float Confidence) const {
- if (!OldT2IfCvt)
- return ARMBaseInstrInfo::isProfitableToIfCvt(MBB, NumInstrs,
- Prediction, Confidence);
- return NumInstrs && NumInstrs <= 3;
-}
-
-bool Thumb2InstrInfo::
-isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
- MachineBasicBlock &FMBB, unsigned NumF,
- float Prediction, float Confidence) const {
- if (!OldT2IfCvt)
- return ARMBaseInstrInfo::isProfitableToIfCvt(TMBB, NumT,
- FMBB, NumF,
- Prediction, Confidence);
-
- // FIXME: Catch optimization such as:
- // r0 = movne
- // r0 = moveq
- return NumT && NumF &&
- NumT <= 3 && NumF <= 3;
-}
-
-
void
Thumb2InstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
MachineBasicBlock *NewDest) const {
diff --git a/lib/Target/ARM/Thumb2InstrInfo.h b/lib/Target/ARM/Thumb2InstrInfo.h
index b348ad0..9ed7eea 100644
--- a/lib/Target/ARM/Thumb2InstrInfo.h
+++ b/lib/Target/ARM/Thumb2InstrInfo.h
@@ -38,12 +38,6 @@ public:
bool isLegalToSplitMBBAt(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI) const;
- bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumInstrs,
- float Prediction, float Confidence) const;
- bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumTInstrs,
- MachineBasicBlock &FMBB, unsigned NumFInstrs,
- float Prediction, float Confidence) const;
-
void copyPhysReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I, DebugLoc DL,
unsigned DestReg, unsigned SrcReg,
diff --git a/lib/Target/TargetInstrInfo.cpp b/lib/Target/TargetInstrInfo.cpp
index 62818d0..eca97ab 100644
--- a/lib/Target/TargetInstrInfo.cpp
+++ b/lib/Target/TargetInstrInfo.cpp
@@ -50,8 +50,8 @@ TargetInstrInfo::~TargetInstrInfo() {
}
unsigned
-TargetInstrInfo::getNumMicroOps(const MachineInstr *MI,
- const InstrItineraryData *ItinData) const {
+TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
+ const MachineInstr *MI) const {
if (!ItinData || ItinData->isEmpty())
return 1;
@@ -94,6 +94,26 @@ TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
}
+int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *MI,
+ unsigned *PredCost) const {
+ if (!ItinData || ItinData->isEmpty())
+ return 1;
+
+ return ItinData->getStageLatency(MI->getDesc().getSchedClass());
+}
+
+int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
+ SDNode *N) const {
+ if (!ItinData || ItinData->isEmpty())
+ return 1;
+
+ if (!N->isMachineOpcode())
+ return 1;
+
+ return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
+}
+
bool TargetInstrInfo::hasLowDefLatency(const InstrItineraryData *ItinData,
const MachineInstr *DefMI,
unsigned DefIdx) const {