#include "llvm/Constants.h"
#include "llvm/Function.h"
#include "llvm/GlobalValue.h"
-#include "llvm/ADT/STLExtras.h"
#include "llvm/CodeGen/LiveVariables.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
+#include "llvm/ADT/STLExtras.h"
using namespace llvm;
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) {
if (isLoad)
MemMI = BuildMI(MF, MI->getDebugLoc(),
get(MemOpc), MI->getOperand(0).getReg())
- .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
+ .addReg(WBReg).addImm(0).addImm(Pred);
else
MemMI = BuildMI(MF, MI->getDebugLoc(),
get(MemOpc)).addReg(MI->getOperand(1).getReg())
if (isLoad)
MemMI = BuildMI(MF, MI->getDebugLoc(),
get(MemOpc), MI->getOperand(0).getReg())
- .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
+ .addReg(BaseReg).addImm(0).addImm(Pred);
else
MemMI = BuildMI(MF, MI->getDebugLoc(),
get(MemOpc)).addReg(MI->getOperand(1).getReg())
}
/// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
-DISABLE_INLINE
+LLVM_ATTRIBUTE_NOINLINE
static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
unsigned JTI);
static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
switch (RC->getID()) {
case ARM::GPRRegClassID:
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STR))
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
.addReg(SrcReg, getKillRegState(isKill))
- .addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO));
+ .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
break;
case ARM::SPRRegClassID:
AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
.addReg(SrcReg, getKillRegState(isKill))
.addMemOperand(MMO));
} else {
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQ))
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
.addReg(SrcReg, getKillRegState(isKill))
.addFrameIndex(FI)
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
.addMemOperand(MMO));
}
break;
.addMemOperand(MMO));
} else {
MachineInstrBuilder MIB =
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMD))
- .addFrameIndex(FI)
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia)))
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
+ .addFrameIndex(FI))
.addMemOperand(MMO);
MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
break;
case ARM::QQQQPRRegClassID: {
MachineInstrBuilder MIB =
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMD))
- .addFrameIndex(FI)
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia)))
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
+ .addFrameIndex(FI))
.addMemOperand(MMO);
MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
int &FrameIndex) const {
switch (MI->getOpcode()) {
default: break;
- case ARM::STR:
+ case ARM::STRrs:
case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
if (MI->getOperand(1).isFI() &&
MI->getOperand(2).isReg() &&
return MI->getOperand(0).getReg();
}
break;
+ case ARM::STRi12:
case ARM::t2STRi12:
case ARM::tSpill:
case ARM::VSTRD:
return MI->getOperand(2).getReg();
}
break;
- case ARM::VSTMQ:
+ case ARM::VSTMQIA:
if (MI->getOperand(1).isFI() &&
- MI->getOperand(2).isImm() &&
- MI->getOperand(2).getImm() == ARM_AM::getAM4ModeImm(ARM_AM::ia) &&
MI->getOperand(0).getSubReg() == 0) {
FrameIndex = MI->getOperand(1).getIndex();
return MI->getOperand(0).getReg();
switch (RC->getID()) {
case ARM::GPRRegClassID:
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDR), DestReg)
- .addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO));
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
+ .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
break;
case ARM::SPRRegClassID:
AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
.addFrameIndex(FI).addImm(16)
.addMemOperand(MMO));
} else {
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQ), DestReg)
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
.addFrameIndex(FI)
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
.addMemOperand(MMO));
}
break;
.addMemOperand(MMO));
} else {
MachineInstrBuilder MIB =
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMD))
- .addFrameIndex(FI)
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia)))
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
+ .addFrameIndex(FI))
.addMemOperand(MMO);
MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
break;
case ARM::QQQQPRRegClassID: {
MachineInstrBuilder MIB =
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMD))
- .addFrameIndex(FI)
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia)))
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
+ .addFrameIndex(FI))
.addMemOperand(MMO);
MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
int &FrameIndex) const {
switch (MI->getOpcode()) {
default: break;
- case ARM::LDR:
+ case ARM::LDRrs:
case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
if (MI->getOperand(1).isFI() &&
MI->getOperand(2).isReg() &&
return MI->getOperand(0).getReg();
}
break;
+ case ARM::LDRi12:
case ARM::t2LDRi12:
case ARM::tRestore:
case ARM::VLDRD:
return MI->getOperand(0).getReg();
}
break;
- case ARM::VLDMQ:
+ case ARM::VLDMQIA:
if (MI->getOperand(1).isFI() &&
- MI->getOperand(2).isImm() &&
- MI->getOperand(2).getImm() == ARM_AM::getAM4ModeImm(ARM_AM::ia) &&
MI->getOperand(0).getSubReg() == 0) {
FrameIndex = MI->getOperand(1).getIndex();
return MI->getOperand(0).getReg();
switch (Load1->getMachineOpcode()) {
default:
return false;
- case ARM::LDR:
- case ARM::LDRB:
+ case ARM::LDRi12:
+ case ARM::LDRBi12:
case ARM::LDRD:
case ARM::LDRH:
case ARM::LDRSB:
switch (Load2->getMachineOpcode()) {
default:
return false;
- case ARM::LDR:
- case ARM::LDRB:
+ case ARM::LDRi12:
+ case ARM::LDRBi12:
case ARM::LDRD:
case ARM::LDRH:
case ARM::LDRSB:
}
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
unsigned NumBits = 0;
unsigned Scale = 1;
switch (AddrMode) {
+ case ARMII::AddrMode_i12: {
+ ImmIdx = FrameRegIdx + 1;
+ InstrOffs = MI.getOperand(ImmIdx).getImm();
+ NumBits = 12;
+ break;
+ }
case ARMII::AddrMode2: {
ImmIdx = FrameRegIdx+2;
InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
if ((unsigned)Offset <= Mask * Scale) {
// Replace the FrameIndex with sp
MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
- if (isSub)
- ImmedOffset |= 1 << NumBits;
+ // FIXME: When addrmode2 goes away, this will simplify (like the
+ // T2 version), as the LDR.i12 versions don't need the encoding
+ // tricks for the offset value.
+ if (isSub) {
+ if (AddrMode == ARMII::AddrMode_i12)
+ ImmedOffset = -ImmedOffset;
+ else
+ ImmedOffset |= 1 << NumBits;
+ }
ImmOp.ChangeToImmediate(ImmedOffset);
Offset = 0;
return true;
// Otherwise, it didn't fit. Pull in what we can to simplify the immed.
ImmedOffset = ImmedOffset & Mask;
- if (isSub)
- ImmedOffset |= 1 << NumBits;
+ if (isSub) {
+ if (AddrMode == ARMII::AddrMode_i12)
+ ImmedOffset = -ImmedOffset;
+ else
+ ImmedOffset |= 1 << NumBits;
+ }
ImmOp.ChangeToImmediate(ImmedOffset);
Offset &= ~(Mask*Scale);
}
}
/// OptimizeCompareInstr - Convert the instruction supplying the argument to the
-/// comparison into one that sets the zero bit in the flags register. Update the
-/// iterator *only* if a transformation took place.
+/// comparison into one that sets the zero bit in the flags register.
bool ARMBaseInstrInfo::
OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
- int CmpValue, MachineBasicBlock::iterator &MII) const {
+ int CmpValue, const MachineRegisterInfo *MRI) const {
if (CmpValue != 0)
return false;
- MachineRegisterInfo &MRI = CmpInstr->getParent()->getParent()->getRegInfo();
- MachineRegisterInfo::def_iterator DI = MRI.def_begin(SrcReg);
- if (llvm::next(DI) != MRI.def_end())
+ MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
+ if (llvm::next(DI) != MRI->def_end())
// Only support one definition.
return false;
if (CmpMask != ~0) {
if (!isSuitableForMask(MI, SrcReg, CmpMask, false)) {
MI = 0;
- for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(SrcReg),
- UE = MRI.use_end(); UI != UE; ++UI) {
+ for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
+ UE = MRI->use_end(); UI != UE; ++UI) {
if (UI->getParent() != CmpInstr->getParent()) continue;
MachineInstr *PotentialAND = &*UI;
if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true))
// want to change.
MachineBasicBlock::const_iterator I = CmpInstr, E = MI,
B = MI->getParent()->begin();
+
+ // Early exit if CmpInstr is at the beginning of the BB.
+ if (I == B) return false;
+
--I;
for (; I != E; --I) {
const MachineInstr &Instr = *I;
for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) {
const MachineOperand &MO = Instr.getOperand(IO);
- if (!MO.isReg() || !MO.isDef()) continue;
+ if (!MO.isReg()) continue;
- // This instruction modifies CPSR before the one we want to change. We
- // can't do this transformation.
+ // This instruction modifies or uses CPSR after the one we want to
+ // change. We can't do this transformation.
if (MO.getReg() == ARM::CPSR)
return false;
}
MI->RemoveOperand(5);
MachineInstrBuilder(MI)
.addReg(ARM::CPSR, RegState::Define | RegState::Implicit);
- MII = llvm::next(MachineBasicBlock::iterator(CmpInstr));
CmpInstr->eraseFromParent();
return true;
}
}
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;
default:
llvm_unreachable("Unexpected multi-uops instruction!");
break;
- case ARM::VLDMQ:
- case ARM::VSTMQ:
+ case ARM::VLDMQIA:
+ case ARM::VLDMQDB:
+ case ARM::VSTMQIA:
+ case ARM::VSTMQDB:
return 2;
// The number of uOps for load / store multiple are determined by the number
// registers.
+ //
// On Cortex-A8, each pair of register loads / stores can be scheduled on the
// same cycle. The scheduling for the first load / store must be done
// separately by assuming the the address is not 64-bit aligned.
+ //
// On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
- // is not 64-bit aligned, then AGU would take an extra cycle.
- // For VFP / NEON load / store multiple, the formula is
- // (#reg / 2) + (#reg % 2) + 1.
- case ARM::VLDMD:
- case ARM::VLDMS:
- case ARM::VLDMD_UPD:
- case ARM::VLDMS_UPD:
- case ARM::VSTMD:
- case ARM::VSTMS:
- case ARM::VSTMD_UPD:
- case ARM::VSTMS_UPD: {
+ // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
+ // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
+ case ARM::VLDMDIA:
+ case ARM::VLDMDDB:
+ case ARM::VLDMDIA_UPD:
+ case ARM::VLDMDDB_UPD:
+ case ARM::VLDMSIA:
+ case ARM::VLDMSDB:
+ case ARM::VLDMSIA_UPD:
+ case ARM::VLDMSDB_UPD:
+ case ARM::VSTMDIA:
+ case ARM::VSTMDDB:
+ case ARM::VSTMDIA_UPD:
+ case ARM::VSTMDDB_UPD:
+ case ARM::VSTMSIA:
+ case ARM::VSTMSDB:
+ case ARM::VSTMSIA_UPD:
+ case ARM::VSTMSDB_UPD: {
unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
return (NumRegs / 2) + (NumRegs % 2) + 1;
}
- case ARM::LDM_RET:
- case ARM::LDM:
- case ARM::LDM_UPD:
- case ARM::STM:
- case ARM::STM_UPD:
- case ARM::tLDM:
- case ARM::tLDM_UPD:
- case ARM::tSTM_UPD:
+
+ case ARM::LDMIA_RET:
+ case ARM::LDMIA:
+ case ARM::LDMDA:
+ case ARM::LDMDB:
+ case ARM::LDMIB:
+ case ARM::LDMIA_UPD:
+ case ARM::LDMDA_UPD:
+ case ARM::LDMDB_UPD:
+ case ARM::LDMIB_UPD:
+ case ARM::STMIA:
+ case ARM::STMDA:
+ case ARM::STMDB:
+ case ARM::STMIB:
+ case ARM::STMIA_UPD:
+ case ARM::STMDA_UPD:
+ case ARM::STMDB_UPD:
+ case ARM::STMIB_UPD:
+ case ARM::tLDMIA:
+ case ARM::tLDMIA_UPD:
+ case ARM::tSTMIA:
+ case ARM::tSTMIA_UPD:
case ARM::tPOP_RET:
case ARM::tPOP:
case ARM::tPUSH:
- case ARM::t2LDM_RET:
- case ARM::t2LDM:
- case ARM::t2LDM_UPD:
- case ARM::t2STM:
- case ARM::t2STM_UPD: {
+ case ARM::t2LDMIA_RET:
+ case ARM::t2LDMIA:
+ case ARM::t2LDMDB:
+ case ARM::t2LDMIA_UPD:
+ case ARM::t2LDMDB_UPD:
+ case ARM::t2STMIA:
+ case ARM::t2STMDB:
+ case ARM::t2STMIA_UPD:
+ case ARM::t2STMDB_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,
} else if (Subtarget.isCortexA9()) {
DefCycle = RegNo;
bool isSLoad = false;
+
switch (DefTID.getOpcode()) {
default: break;
- case ARM::VLDMS:
- case ARM::VLDMS_UPD:
+ case ARM::VLDMSIA:
+ case ARM::VLDMSDB:
+ case ARM::VLDMSIA_UPD:
+ case ARM::VLDMSDB_UPD:
isSLoad = true;
break;
}
+
// If there are odd number of 'S' registers or if it's not 64-bit aligned,
// then it takes an extra cycle.
if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
} else if (Subtarget.isCortexA9()) {
UseCycle = RegNo;
bool isSStore = false;
+
switch (UseTID.getOpcode()) {
default: break;
- case ARM::VSTMS:
- case ARM::VSTMS_UPD:
+ case ARM::VSTMSIA:
+ case ARM::VSTMSDB:
+ case ARM::VSTMSIA_UPD:
+ case ARM::VSTMSDB_UPD:
isSStore = true;
break;
}
+
// If there are odd number of 'S' registers or if it's not 64-bit aligned,
// then it takes an extra cycle.
if ((isSStore && (RegNo % 2)) || UseAlign < 8)
// This may be a def / use of a variable_ops instruction, the operand
// latency might be determinable dynamically. Let the target try to
// figure it out.
- bool LdmBypass = false;
int DefCycle = -1;
+ bool LdmBypass = false;
switch (DefTID.getOpcode()) {
default:
DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
break;
- case ARM::VLDMD:
- case ARM::VLDMS:
- case ARM::VLDMD_UPD:
- case ARM::VLDMS_UPD: {
+
+ case ARM::VLDMDIA:
+ case ARM::VLDMDDB:
+ case ARM::VLDMDIA_UPD:
+ case ARM::VLDMDDB_UPD:
+ case ARM::VLDMSIA:
+ case ARM::VLDMSDB:
+ case ARM::VLDMSIA_UPD:
+ case ARM::VLDMSDB_UPD:
DefCycle = getVLDMDefCycle(ItinData, DefTID, DefClass, DefIdx, DefAlign);
break;
- }
- case ARM::LDM_RET:
- case ARM::LDM:
- case ARM::LDM_UPD:
- case ARM::tLDM:
- case ARM::tLDM_UPD:
+
+ case ARM::LDMIA_RET:
+ case ARM::LDMIA:
+ case ARM::LDMDA:
+ case ARM::LDMDB:
+ case ARM::LDMIB:
+ case ARM::LDMIA_UPD:
+ case ARM::LDMDA_UPD:
+ case ARM::LDMDB_UPD:
+ case ARM::LDMIB_UPD:
+ case ARM::tLDMIA:
+ case ARM::tLDMIA_UPD:
case ARM::tPUSH:
- case ARM::t2LDM_RET:
- case ARM::t2LDM:
- case ARM::t2LDM_UPD: {
+ case ARM::t2LDMIA_RET:
+ case ARM::t2LDMIA:
+ case ARM::t2LDMDB:
+ case ARM::t2LDMIA_UPD:
+ case ARM::t2LDMDB_UPD:
LdmBypass = 1;
DefCycle = getLDMDefCycle(ItinData, DefTID, DefClass, DefIdx, DefAlign);
break;
}
- }
if (DefCycle == -1)
// We can't seem to determine the result latency of the def, assume it's 2.
default:
UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
break;
- case ARM::VSTMD:
- case ARM::VSTMS:
- case ARM::VSTMD_UPD:
- case ARM::VSTMS_UPD: {
+
+ case ARM::VSTMDIA:
+ case ARM::VSTMDDB:
+ case ARM::VSTMDIA_UPD:
+ case ARM::VSTMDDB_UPD:
+ case ARM::VSTMSIA:
+ case ARM::VSTMSDB:
+ case ARM::VSTMSIA_UPD:
+ case ARM::VSTMSDB_UPD:
UseCycle = getVSTMUseCycle(ItinData, UseTID, UseClass, UseIdx, UseAlign);
break;
- }
- case ARM::STM:
- case ARM::STM_UPD:
- case ARM::tSTM_UPD:
+
+ case ARM::STMIA:
+ case ARM::STMDA:
+ case ARM::STMDB:
+ case ARM::STMIB:
+ case ARM::STMIA_UPD:
+ case ARM::STMDA_UPD:
+ case ARM::STMDB_UPD:
+ case ARM::STMIB_UPD:
+ case ARM::tSTMIA:
+ case ARM::tSTMIA_UPD:
case ARM::tPOP_RET:
case ARM::tPOP:
- case ARM::t2STM:
- case ARM::t2STM_UPD: {
+ case ARM::t2STMIA:
+ case ARM::t2STMDB:
+ case ARM::t2STMIA_UPD:
+ case ARM::t2STMDB_UPD:
UseCycle = getSTMUseCycle(ItinData, UseTID, UseClass, UseIdx, UseAlign);
break;
}
- }
if (UseCycle == -1)
// Assume it's read in the first stage.
UseClass, UseIdx))
--UseCycle;
} else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
- UseClass, UseIdx))
+ UseClass, UseIdx)) {
--UseCycle;
+ }
}
return UseCycle;
if (!ItinData || ItinData->isEmpty())
return DefTID.mayLoad() ? 3 : 1;
+
const TargetInstrDesc &UseTID = UseMI->getDesc();
+ const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
+ if (DefMO.getReg() == ARM::CPSR) {
+ if (DefMI->getOpcode() == ARM::FMSTAT) {
+ // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
+ return Subtarget.isCortexA9() ? 1 : 20;
+ }
+
+ // CPSR set and branch can be paired in the same cycle.
+ if (UseTID.isBranch())
+ return 0;
+ }
+
unsigned DefAlign = DefMI->hasOneMemOperand()
? (*DefMI->memoperands_begin())->getAlignment() : 0;
unsigned UseAlign = UseMI->hasOneMemOperand()
? (*UseMI->memoperands_begin())->getAlignment() : 0;
- return getOperandLatency(ItinData, DefTID, DefIdx, DefAlign,
- UseTID, UseIdx, UseAlign);
+ int Latency = getOperandLatency(ItinData, DefTID, DefIdx, DefAlign,
+ UseTID, UseIdx, UseAlign);
+
+ if (Latency > 1 &&
+ (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
+ // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
+ // variants are one cycle cheaper.
+ switch (DefTID.getOpcode()) {
+ default: break;
+ case ARM::LDRrs:
+ case ARM::LDRBrs: {
+ unsigned ShOpVal = DefMI->getOperand(3).getImm();
+ unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
+ if (ShImm == 0 ||
+ (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
+ --Latency;
+ break;
+ }
+ case ARM::t2LDRs:
+ case ARM::t2LDRBs:
+ case ARM::t2LDRHs:
+ case ARM::t2LDRSHs: {
+ // Thumb2 mode: lsl only.
+ unsigned ShAmt = DefMI->getOperand(3).getImm();
+ if (ShAmt == 0 || ShAmt == 2)
+ --Latency;
+ break;
+ }
+ }
+ }
+
+ return Latency;
}
int
if (!ItinData || ItinData->isEmpty())
return DefTID.mayLoad() ? 3 : 1;
- if (!UseNode->isMachineOpcode())
- return ItinData->getOperandCycle(DefTID.getSchedClass(), DefIdx);
+ if (!UseNode->isMachineOpcode()) {
+ int Latency = ItinData->getOperandCycle(DefTID.getSchedClass(), DefIdx);
+ if (Subtarget.isCortexA9())
+ return Latency <= 2 ? 1 : Latency - 1;
+ else
+ return Latency <= 3 ? 1 : Latency - 2;
+ }
const TargetInstrDesc &UseTID = get(UseNode->getMachineOpcode());
const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
unsigned UseAlign = !UseMN->memoperands_empty()
? (*UseMN->memoperands_begin())->getAlignment() : 0;
- return getOperandLatency(ItinData, DefTID, DefIdx, DefAlign,
- UseTID, UseIdx, UseAlign);
+ int Latency = getOperandLatency(ItinData, DefTID, DefIdx, DefAlign,
+ UseTID, UseIdx, UseAlign);
+
+ if (Latency > 1 &&
+ (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
+ // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
+ // variants are one cycle cheaper.
+ switch (DefTID.getOpcode()) {
+ default: break;
+ case ARM::LDRrs:
+ case ARM::LDRBrs: {
+ unsigned ShOpVal =
+ cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
+ unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
+ if (ShImm == 0 ||
+ (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
+ --Latency;
+ break;
+ }
+ case ARM::t2LDRs:
+ case ARM::t2LDRBs:
+ case ARM::t2LDRHs:
+ case ARM::t2LDRSHs: {
+ // Thumb2 mode: lsl only.
+ unsigned ShAmt =
+ cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
+ if (ShAmt == 0 || ShAmt == 2)
+ --Latency;
+ break;
+ }
+ }
+ }
+
+ 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::VLDMQIA:
+ case ARM::VLDMQDB:
+ case ARM::VSTMQIA:
+ case ARM::VSTMQDB:
+ return 2;
+ }
+}
+
+bool ARMBaseInstrInfo::
+hasHighOperandLatency(const InstrItineraryData *ItinData,
+ const MachineRegisterInfo *MRI,
+ const MachineInstr *DefMI, unsigned DefIdx,
+ const MachineInstr *UseMI, unsigned UseIdx) const {
+ unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
+ unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
+ if (Subtarget.isCortexA8() &&
+ (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
+ // CortexA8 VFP instructions are not pipelined.
+ return true;
+
+ // Hoist VFP / NEON instructions with 4 or higher latency.
+ int Latency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
+ if (Latency <= 3)
+ return false;
+ return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
+ UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
+}
+
+bool ARMBaseInstrInfo::
+hasLowDefLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *DefMI, unsigned DefIdx) const {
+ if (!ItinData || ItinData->isEmpty())
+ return false;
+
+ unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
+ if (DDomain == ARMII::DomainGeneral) {
+ unsigned DefClass = DefMI->getDesc().getSchedClass();
+ int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
+ return (DefCycle != -1 && DefCycle <= 2);
+ }
+ return false;
}
projects / oota-llvm.git / blobdiff