#include "ARM.h"
#include "ARMAddressingModes.h"
#include "ARMConstantPoolValue.h"
+#include "ARMHazardRecognizer.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMRegisterInfo.h"
#include "ARMGenInstrInfo.inc"
#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>
-OldARMIfConv("old-arm-if-conversion", cl::Hidden,
- cl::desc("Use old-style ARM if-conversion heuristics"));
+/// ARM_MLxEntry - Record information about MLA / MLS instructions.
+struct ARM_MLxEntry {
+ unsigned MLxOpc; // MLA / MLS opcode
+ unsigned MulOpc; // Expanded multiplication opcode
+ unsigned AddSubOpc; // Expanded add / sub opcode
+ bool NegAcc; // True if the acc is negated before the add / sub.
+ bool HasLane; // True if instruction has an extra "lane" operand.
+};
+
+static const ARM_MLxEntry ARM_MLxTable[] = {
+ // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
+ // fp scalar ops
+ { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
+ { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
+ { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
+ { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
+ { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
+ { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
+ { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
+ { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
+
+ // fp SIMD ops
+ { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
+ { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
+ { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
+ { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
+ { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
+ { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
+ { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
+ { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
+};
ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
: TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)),
Subtarget(STI) {
+ for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
+ if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
+ assert(false && "Duplicated entries?");
+ MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
+ MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
+ }
+}
+
+// Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
+// currently defaults to no prepass hazard recognizer.
+ScheduleHazardRecognizer *ARMBaseInstrInfo::
+CreateTargetHazardRecognizer(const TargetMachine *TM,
+ const ScheduleDAG *DAG) const {
+ if (usePreRAHazardRecognizer()) {
+ const InstrItineraryData *II = TM->getInstrItineraryData();
+ return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
+ }
+ return TargetInstrInfoImpl::CreateTargetHazardRecognizer(TM, DAG);
+}
+
+ScheduleHazardRecognizer *ARMBaseInstrInfo::
+CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
+ const ScheduleDAG *DAG) const {
+ if (Subtarget.isThumb2() || Subtarget.hasVFP2())
+ return (ScheduleHazardRecognizer *)
+ new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget, DAG);
+ return TargetInstrInfoImpl::CreateTargetPostRAHazardRecognizer(II, DAG);
}
MachineInstr *
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())
if (LV) {
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.getReg() &&
- TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
+ if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
unsigned Reg = MO.getReg();
LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
return NewMIs[0];
}
-bool
-ARMBaseInstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MI,
- const std::vector<CalleeSavedInfo> &CSI,
- const TargetRegisterInfo *TRI) const {
- if (CSI.empty())
- return false;
-
- DebugLoc DL;
- if (MI != MBB.end()) DL = MI->getDebugLoc();
-
- for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
- unsigned Reg = CSI[i].getReg();
- bool isKill = true;
-
- // Add the callee-saved register as live-in unless it's LR and
- // @llvm.returnaddress is called. If LR is returned for @llvm.returnaddress
- // then it's already added to the function and entry block live-in sets.
- if (Reg == ARM::LR) {
- MachineFunction &MF = *MBB.getParent();
- if (MF.getFrameInfo()->isReturnAddressTaken() &&
- MF.getRegInfo().isLiveIn(Reg))
- isKill = false;
- }
-
- if (isKill)
- MBB.addLiveIn(Reg);
-
- // Insert the spill to the stack frame. The register is killed at the spill
- //
- const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
- storeRegToStackSlot(MBB, MI, Reg, isKill,
- CSI[i].getFrameIdx(), RC, TRI);
- }
- return true;
-}
-
// Branch analysis.
bool
ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
}
/// 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,
case ARMII::Size2Bytes: return 2; // Thumb1 instruction.
case ARMII::SizeSpecial: {
switch (Opc) {
+ case ARM::MOVi16_ga_pcrel:
+ case ARM::MOVTi16_ga_pcrel:
+ case ARM::t2MOVi16_ga_pcrel:
+ case ARM::t2MOVTi16_ga_pcrel:
+ return 4;
+ case ARM::MOVi32imm:
+ case ARM::t2MOVi32imm:
+ return 8;
case ARM::CONSTPOOL_ENTRY:
// If this machine instr is a constant pool entry, its size is recorded as
// operand #2.
case ARM::BR_JTadd:
case ARM::tBR_JTr:
case ARM::t2BR_JT:
- case ARM::t2TBB:
- case ARM::t2TBH: {
+ case ARM::t2TBB_JT:
+ case ARM::t2TBH_JT: {
// These are jumptable branches, i.e. a branch followed by an inlined
// jumptable. The size is 4 + 4 * number of entries. For TBB, each
// entry is one byte; TBH two byte each.
- unsigned EntrySize = (Opc == ARM::t2TBB)
- ? 1 : ((Opc == ARM::t2TBH) ? 2 : 4);
+ unsigned EntrySize = (Opc == ARM::t2TBB_JT)
+ ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
unsigned NumOps = TID.getNumOperands();
MachineOperand JTOP =
MI->getOperand(NumOps - (TID.isPredicable() ? 3 : 2));
// alignment issue.
unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
unsigned NumEntries = getNumJTEntries(JT, JTI);
- if (Opc == ARM::t2TBB && (NumEntries & 1))
+ if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
// Make sure the instruction that follows TBB is 2-byte aligned.
// FIXME: Constant island pass should insert an "ALIGN" instruction
// instead.
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();
ARMConstantPoolValue *ACPV =
static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
- unsigned PCLabelId = AFI->createConstPoolEntryUId();
+ unsigned PCLabelId = AFI->createPICLabelUId();
ARMConstantPoolValue *NewCPV = 0;
// FIXME: The below assumes PIC relocation model and that the function
// is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
}
bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
- const MachineInstr *MI1) const {
+ const MachineInstr *MI1,
+ const MachineRegisterInfo *MRI) const {
int Opcode = MI0->getOpcode();
if (Opcode == ARM::t2LDRpci ||
Opcode == ARM::t2LDRpci_pic ||
Opcode == ARM::tLDRpci ||
- Opcode == ARM::tLDRpci_pic) {
+ Opcode == ARM::tLDRpci_pic ||
+ Opcode == ARM::MOV_ga_dyn ||
+ Opcode == ARM::MOV_ga_pcrel ||
+ Opcode == ARM::MOV_ga_pcrel_ldr ||
+ Opcode == ARM::t2MOV_ga_dyn ||
+ Opcode == ARM::t2MOV_ga_pcrel) {
if (MI1->getOpcode() != Opcode)
return false;
if (MI0->getNumOperands() != MI1->getNumOperands())
if (MO0.getOffset() != MO1.getOffset())
return false;
+ if (Opcode == ARM::MOV_ga_dyn ||
+ Opcode == ARM::MOV_ga_pcrel ||
+ Opcode == ARM::MOV_ga_pcrel_ldr ||
+ Opcode == ARM::t2MOV_ga_dyn ||
+ Opcode == ARM::t2MOV_ga_pcrel)
+ // Ignore the PC labels.
+ return MO0.getGlobal() == MO1.getGlobal();
+
const MachineFunction *MF = MI0->getParent()->getParent();
const MachineConstantPool *MCP = MF->getConstantPool();
int CPI0 = MO0.getIndex();
int CPI1 = MO1.getIndex();
const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
- ARMConstantPoolValue *ACPV0 =
- static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
- ARMConstantPoolValue *ACPV1 =
- static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
- return ACPV0->hasSameValue(ACPV1);
+ bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
+ bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
+ if (isARMCP0 && isARMCP1) {
+ ARMConstantPoolValue *ACPV0 =
+ static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
+ ARMConstantPoolValue *ACPV1 =
+ static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
+ return ACPV0->hasSameValue(ACPV1);
+ } else if (!isARMCP0 && !isARMCP1) {
+ return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
+ }
+ return false;
+ } else if (Opcode == ARM::PICLDR) {
+ if (MI1->getOpcode() != Opcode)
+ return false;
+ if (MI0->getNumOperands() != MI1->getNumOperands())
+ return false;
+
+ unsigned Addr0 = MI0->getOperand(1).getReg();
+ unsigned Addr1 = MI1->getOperand(1).getReg();
+ if (Addr0 != Addr1) {
+ if (!MRI ||
+ !TargetRegisterInfo::isVirtualRegister(Addr0) ||
+ !TargetRegisterInfo::isVirtualRegister(Addr1))
+ return false;
+
+ // This assumes SSA form.
+ MachineInstr *Def0 = MRI->getVRegDef(Addr0);
+ MachineInstr *Def1 = MRI->getVRegDef(Addr1);
+ // Check if the loaded value, e.g. a constantpool of a global address, are
+ // the same.
+ if (!produceSameValue(Def0, Def1, MRI))
+ return false;
+ }
+
+ for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
+ // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
+ const MachineOperand &MO0 = MI0->getOperand(i);
+ const MachineOperand &MO1 = MI1->getOperand(i);
+ if (!MO0.isIdenticalTo(MO1))
+ return false;
+ }
+ return true;
}
return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
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:
}
/// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
-/// determine (in conjuction with areLoadsFromSameBasePtr) if two loads should
+/// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
/// be scheduled togther. On some targets if two loads are loading from
/// addresses in the same cache line, it's better if they are scheduled
/// together. This function takes two integers that represent the load offsets
}
bool ARMBaseInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
- unsigned NumInstrs,
- float Probability) const {
- if (!NumInstrs)
+ unsigned NumCycles,
+ unsigned ExtraPredCycles,
+ float Probability,
+ float Confidence) const {
+ if (!NumCycles)
return false;
-
- // Use old-style heuristics
- if (OldARMIfConv) {
- if (Subtarget.getCPUString() == "generic")
- // Generic (and overly aggressive) if-conversion limits for testing.
- return NumInstrs <= 10;
- else 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 * NumCycles;
UnpredCost += 1.0; // The branch itself
- UnpredCost += 0.1 * Subtarget.getMispredictionPenalty();
-
- float PredCost = NumInstrs;
-
- return PredCost < UnpredCost;
-
+ UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
+
+ return (float)(NumCycles + ExtraPredCycles) < UnpredCost;
}
-
-bool ARMBaseInstrInfo::
-isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
- MachineBasicBlock &FMBB, unsigned NumF,
- float Probability) const {
- // Use old-style if-conversion heuristics
- if (OldARMIfConv) {
- return NumT && NumF && NumT <= 2 && NumF <= 2;
- }
- if (!NumT || !NumF)
+bool ARMBaseInstrInfo::
+isProfitableToIfCvt(MachineBasicBlock &TMBB,
+ unsigned TCycles, unsigned TExtra,
+ MachineBasicBlock &FMBB,
+ unsigned FCycles, unsigned FExtra,
+ float Probability, float Confidence) const {
+ 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 += 0.1 * Subtarget.getMispredictionPenalty();
-
- float PredCost = NumT + NumF;
-
- return PredCost < UnpredCost;
+ UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
+
+ return (float)(TCycles + FCycles + TExtra + FExtra) < UnpredCost;
}
/// getInstrPredicate - If instruction is predicated, returns its predicate
MachineBasicBlock::iterator &MBBI, DebugLoc dl,
unsigned DestReg, unsigned BaseReg, int NumBytes,
ARMCC::CondCodes Pred, unsigned PredReg,
- const ARMBaseInstrInfo &TII) {
+ const ARMBaseInstrInfo &TII, unsigned MIFlags) {
bool isSub = NumBytes < 0;
if (isSub) NumBytes = -NumBytes;
unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
.addReg(BaseReg, RegState::Kill).addImm(ThisVal)
- .addImm((unsigned)Pred).addReg(PredReg).addReg(0);
+ .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
+ .setMIFlags(MIFlags);
BaseReg = DestReg;
}
}
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);
}
switch (MI->getOpcode()) {
default: break;
case ARM::CMPri:
- case ARM::CMPzri:
case ARM::t2CMPri:
- case ARM::t2CMPzri:
SrcReg = MI->getOperand(0).getReg();
CmpMask = ~0;
CmpValue = MI->getOperand(1).getImm();
case ARM::COPY: {
// Walk down one instruction which is potentially an 'and'.
const MachineInstr &Copy = *MI;
- MachineBasicBlock::iterator AND(next(MachineBasicBlock::iterator(MI)));
+ MachineBasicBlock::iterator AND(
+ llvm::next(MachineBasicBlock::iterator(MI)));
if (AND == MI->getParent()->end()) return false;
MI = AND;
return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
}
/// 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;
}
// Set the "zero" bit in CPSR.
switch (MI->getOpcode()) {
default: break;
+ case ARM::RSBrr:
+ case ARM::RSBri:
+ case ARM::RSCrr:
+ case ARM::RSCri:
+ case ARM::ADDrr:
case ARM::ADDri:
- case ARM::ANDri:
- case ARM::t2ANDri:
+ case ARM::ADCrr:
+ case ARM::ADCri:
+ case ARM::SUBrr:
case ARM::SUBri:
+ case ARM::SBCrr:
+ case ARM::SBCri:
+ case ARM::t2RSBri:
+ case ARM::t2ADDrr:
case ARM::t2ADDri:
+ case ARM::t2ADCrr:
+ case ARM::t2ADCri:
+ case ARM::t2SUBrr:
case ARM::t2SUBri:
- MI->RemoveOperand(5);
- MachineInstrBuilder(MI)
- .addReg(ARM::CPSR, RegState::Define | RegState::Implicit);
- MII = llvm::next(MachineBasicBlock::iterator(CmpInstr));
+ case ARM::t2SBCrr:
+ case ARM::t2SBCri:
+ case ARM::ANDrr:
+ case ARM::ANDri:
+ case ARM::t2ANDrr:
+ case ARM::t2ANDri:
+ case ARM::ORRrr:
+ case ARM::ORRri:
+ case ARM::t2ORRrr:
+ case ARM::t2ORRri:
+ case ARM::EORrr:
+ case ARM::EORri:
+ case ARM::t2EORrr:
+ case ARM::t2EORri: {
+ // Scan forward for the use of CPSR, if it's a conditional code requires
+ // checking of V bit, then this is not safe to do. If we can't find the
+ // CPSR use (i.e. used in another block), then it's not safe to perform
+ // the optimization.
+ bool isSafe = false;
+ I = CmpInstr;
+ E = MI->getParent()->end();
+ while (!isSafe && ++I != E) {
+ const MachineInstr &Instr = *I;
+ for (unsigned IO = 0, EO = Instr.getNumOperands();
+ !isSafe && IO != EO; ++IO) {
+ const MachineOperand &MO = Instr.getOperand(IO);
+ if (!MO.isReg() || MO.getReg() != ARM::CPSR)
+ continue;
+ if (MO.isDef()) {
+ isSafe = true;
+ break;
+ }
+ // Condition code is after the operand before CPSR.
+ ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
+ switch (CC) {
+ default:
+ isSafe = true;
+ break;
+ case ARMCC::VS:
+ case ARMCC::VC:
+ case ARMCC::GE:
+ case ARMCC::LT:
+ case ARMCC::GT:
+ case ARMCC::LE:
+ return false;
+ }
+ }
+ }
+
+ if (!isSafe)
+ return false;
+
+ // Toggle the optional operand to CPSR.
+ MI->getOperand(5).setReg(ARM::CPSR);
+ MI->getOperand(5).setIsDef(true);
CmpInstr->eraseFromParent();
return true;
}
+ }
return false;
}
+bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
+ MachineInstr *DefMI, unsigned Reg,
+ MachineRegisterInfo *MRI) const {
+ // Fold large immediates into add, sub, or, xor.
+ unsigned DefOpc = DefMI->getOpcode();
+ if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
+ return false;
+ if (!DefMI->getOperand(1).isImm())
+ // Could be t2MOVi32imm <ga:xx>
+ return false;
+
+ if (!MRI->hasOneNonDBGUse(Reg))
+ return false;
+
+ unsigned UseOpc = UseMI->getOpcode();
+ unsigned NewUseOpc = 0;
+ uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
+ uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
+ bool Commute = false;
+ switch (UseOpc) {
+ default: return false;
+ case ARM::SUBrr:
+ case ARM::ADDrr:
+ case ARM::ORRrr:
+ case ARM::EORrr:
+ case ARM::t2SUBrr:
+ case ARM::t2ADDrr:
+ case ARM::t2ORRrr:
+ case ARM::t2EORrr: {
+ Commute = UseMI->getOperand(2).getReg() != Reg;
+ switch (UseOpc) {
+ default: break;
+ case ARM::SUBrr: {
+ if (Commute)
+ return false;
+ ImmVal = -ImmVal;
+ NewUseOpc = ARM::SUBri;
+ // Fallthrough
+ }
+ case ARM::ADDrr:
+ case ARM::ORRrr:
+ case ARM::EORrr: {
+ if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
+ return false;
+ SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
+ SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
+ switch (UseOpc) {
+ default: break;
+ case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
+ case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
+ case ARM::EORrr: NewUseOpc = ARM::EORri; break;
+ }
+ break;
+ }
+ case ARM::t2SUBrr: {
+ if (Commute)
+ return false;
+ ImmVal = -ImmVal;
+ NewUseOpc = ARM::t2SUBri;
+ // Fallthrough
+ }
+ case ARM::t2ADDrr:
+ case ARM::t2ORRrr:
+ case ARM::t2EORrr: {
+ if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
+ return false;
+ SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
+ SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
+ switch (UseOpc) {
+ default: break;
+ case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
+ case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
+ case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
+ }
+ break;
+ }
+ }
+ }
+ }
+
+ unsigned OpIdx = Commute ? 2 : 1;
+ unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
+ bool isKill = UseMI->getOperand(OpIdx).isKill();
+ unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
+ AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
+ *UseMI, UseMI->getDebugLoc(),
+ get(NewUseOpc), NewReg)
+ .addReg(Reg1, getKillRegState(isKill))
+ .addImm(SOImmValV1)));
+ UseMI->setDesc(get(NewUseOpc));
+ UseMI->getOperand(1).setReg(NewReg);
+ UseMI->getOperand(1).setIsKill();
+ UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
+ DefMI->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::VSTMQIA:
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::VLDMDIA_UPD:
+ case ARM::VLDMDDB_UPD:
+ case ARM::VLDMSIA:
+ case ARM::VLDMSIA_UPD:
+ case ARM::VLDMSDB_UPD:
+ case ARM::VSTMDIA:
+ case ARM::VSTMDIA_UPD:
+ case ARM::VSTMDDB_UPD:
+ case ARM::VSTMSIA:
+ 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 {
// Assume the worst.
return NumRegs;
- }
+ }
+ }
+ }
+}
+
+int
+ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
+ const TargetInstrDesc &DefTID,
+ unsigned DefClass,
+ unsigned DefIdx, unsigned DefAlign) const {
+ int RegNo = (int)(DefIdx+1) - DefTID.getNumOperands() + 1;
+ if (RegNo <= 0)
+ // Def is the address writeback.
+ return ItinData->getOperandCycle(DefClass, DefIdx);
+
+ int DefCycle;
+ if (Subtarget.isCortexA8()) {
+ // (regno / 2) + (regno % 2) + 1
+ DefCycle = RegNo / 2 + 1;
+ if (RegNo % 2)
+ ++DefCycle;
+ } else if (Subtarget.isCortexA9()) {
+ DefCycle = RegNo;
+ bool isSLoad = false;
+
+ switch (DefTID.getOpcode()) {
+ default: break;
+ case ARM::VLDMSIA:
+ 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)
+ ++DefCycle;
+ } else {
+ // Assume the worst.
+ DefCycle = RegNo + 2;
+ }
+
+ return DefCycle;
+}
+
+int
+ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
+ const TargetInstrDesc &DefTID,
+ unsigned DefClass,
+ unsigned DefIdx, unsigned DefAlign) const {
+ int RegNo = (int)(DefIdx+1) - DefTID.getNumOperands() + 1;
+ if (RegNo <= 0)
+ // Def is the address writeback.
+ return ItinData->getOperandCycle(DefClass, DefIdx);
+
+ int DefCycle;
+ if (Subtarget.isCortexA8()) {
+ // 4 registers would be issued: 1, 2, 1.
+ // 5 registers would be issued: 1, 2, 2.
+ DefCycle = RegNo / 2;
+ if (DefCycle < 1)
+ DefCycle = 1;
+ // Result latency is issue cycle + 2: E2.
+ DefCycle += 2;
+ } else if (Subtarget.isCortexA9()) {
+ DefCycle = (RegNo / 2);
+ // If there are odd number of registers or if it's not 64-bit aligned,
+ // then it takes an extra AGU (Address Generation Unit) cycle.
+ if ((RegNo % 2) || DefAlign < 8)
+ ++DefCycle;
+ // Result latency is AGU cycles + 2.
+ DefCycle += 2;
+ } else {
+ // Assume the worst.
+ DefCycle = RegNo + 2;
+ }
+
+ return DefCycle;
+}
+
+int
+ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
+ const TargetInstrDesc &UseTID,
+ unsigned UseClass,
+ unsigned UseIdx, unsigned UseAlign) const {
+ int RegNo = (int)(UseIdx+1) - UseTID.getNumOperands() + 1;
+ if (RegNo <= 0)
+ return ItinData->getOperandCycle(UseClass, UseIdx);
+
+ int UseCycle;
+ if (Subtarget.isCortexA8()) {
+ // (regno / 2) + (regno % 2) + 1
+ UseCycle = RegNo / 2 + 1;
+ if (RegNo % 2)
+ ++UseCycle;
+ } else if (Subtarget.isCortexA9()) {
+ UseCycle = RegNo;
+ bool isSStore = false;
+
+ switch (UseTID.getOpcode()) {
+ default: break;
+ case ARM::VSTMSIA:
+ 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)
+ ++UseCycle;
+ } else {
+ // Assume the worst.
+ UseCycle = RegNo + 2;
+ }
+
+ return UseCycle;
+}
+
+int
+ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
+ const TargetInstrDesc &UseTID,
+ unsigned UseClass,
+ unsigned UseIdx, unsigned UseAlign) const {
+ int RegNo = (int)(UseIdx+1) - UseTID.getNumOperands() + 1;
+ if (RegNo <= 0)
+ return ItinData->getOperandCycle(UseClass, UseIdx);
+
+ int UseCycle;
+ if (Subtarget.isCortexA8()) {
+ UseCycle = RegNo / 2;
+ if (UseCycle < 2)
+ UseCycle = 2;
+ // Read in E3.
+ UseCycle += 2;
+ } else if (Subtarget.isCortexA9()) {
+ UseCycle = (RegNo / 2);
+ // If there are odd number of registers or if it's not 64-bit aligned,
+ // then it takes an extra AGU (Address Generation Unit) cycle.
+ if ((RegNo % 2) || UseAlign < 8)
+ ++UseCycle;
+ } else {
+ // Assume the worst.
+ UseCycle = 1;
+ }
+ return UseCycle;
+}
+
+int
+ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
+ const TargetInstrDesc &DefTID,
+ unsigned DefIdx, unsigned DefAlign,
+ const TargetInstrDesc &UseTID,
+ unsigned UseIdx, unsigned UseAlign) const {
+ unsigned DefClass = DefTID.getSchedClass();
+ unsigned UseClass = UseTID.getSchedClass();
+
+ if (DefIdx < DefTID.getNumDefs() && UseIdx < UseTID.getNumOperands())
+ return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
+
+ // 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.
+ int DefCycle = -1;
+ bool LdmBypass = false;
+ switch (DefTID.getOpcode()) {
+ default:
+ DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
+ break;
+
+ case ARM::VLDMDIA:
+ case ARM::VLDMDIA_UPD:
+ case ARM::VLDMDDB_UPD:
+ case ARM::VLDMSIA:
+ case ARM::VLDMSIA_UPD:
+ case ARM::VLDMSDB_UPD:
+ DefCycle = getVLDMDefCycle(ItinData, DefTID, DefClass, DefIdx, DefAlign);
+ break;
+
+ 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::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.
+ DefCycle = 2;
+
+ int UseCycle = -1;
+ switch (UseTID.getOpcode()) {
+ default:
+ UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
+ break;
+
+ case ARM::VSTMDIA:
+ case ARM::VSTMDIA_UPD:
+ case ARM::VSTMDDB_UPD:
+ case ARM::VSTMSIA:
+ case ARM::VSTMSIA_UPD:
+ case ARM::VSTMSDB_UPD:
+ UseCycle = getVSTMUseCycle(ItinData, UseTID, UseClass, UseIdx, UseAlign);
+ break;
+
+ 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::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.
+ UseCycle = 1;
+
+ UseCycle = DefCycle - UseCycle + 1;
+ if (UseCycle > 0) {
+ if (LdmBypass) {
+ // It's a variable_ops instruction so we can't use DefIdx here. Just use
+ // first def operand.
+ if (ItinData->hasPipelineForwarding(DefClass, DefTID.getNumOperands()-1,
+ UseClass, UseIdx))
+ --UseCycle;
+ } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
+ UseClass, UseIdx)) {
+ --UseCycle;
+ }
+ }
+
+ return UseCycle;
+}
+
+int
+ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *DefMI, unsigned DefIdx,
+ const MachineInstr *UseMI, unsigned UseIdx) const {
+ if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
+ DefMI->isRegSequence() || DefMI->isImplicitDef())
+ return 1;
+
+ const TargetInstrDesc &DefTID = DefMI->getDesc();
+ 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;
+ 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;
+ }
+ }
+ }
+
+ if (DefAlign < 8 && Subtarget.isCortexA9())
+ switch (DefTID.getOpcode()) {
+ default: break;
+ case ARM::VLD1q8:
+ case ARM::VLD1q16:
+ case ARM::VLD1q32:
+ case ARM::VLD1q64:
+ case ARM::VLD1q8_UPD:
+ case ARM::VLD1q16_UPD:
+ case ARM::VLD1q32_UPD:
+ case ARM::VLD1q64_UPD:
+ case ARM::VLD2d8:
+ case ARM::VLD2d16:
+ case ARM::VLD2d32:
+ case ARM::VLD2q8:
+ case ARM::VLD2q16:
+ case ARM::VLD2q32:
+ case ARM::VLD2d8_UPD:
+ case ARM::VLD2d16_UPD:
+ case ARM::VLD2d32_UPD:
+ case ARM::VLD2q8_UPD:
+ case ARM::VLD2q16_UPD:
+ case ARM::VLD2q32_UPD:
+ case ARM::VLD3d8:
+ case ARM::VLD3d16:
+ case ARM::VLD3d32:
+ case ARM::VLD1d64T:
+ case ARM::VLD3d8_UPD:
+ case ARM::VLD3d16_UPD:
+ case ARM::VLD3d32_UPD:
+ case ARM::VLD1d64T_UPD:
+ case ARM::VLD3q8_UPD:
+ case ARM::VLD3q16_UPD:
+ case ARM::VLD3q32_UPD:
+ case ARM::VLD4d8:
+ case ARM::VLD4d16:
+ case ARM::VLD4d32:
+ case ARM::VLD1d64Q:
+ case ARM::VLD4d8_UPD:
+ case ARM::VLD4d16_UPD:
+ case ARM::VLD4d32_UPD:
+ case ARM::VLD1d64Q_UPD:
+ case ARM::VLD4q8_UPD:
+ case ARM::VLD4q16_UPD:
+ case ARM::VLD4q32_UPD:
+ case ARM::VLD1DUPq8:
+ case ARM::VLD1DUPq16:
+ case ARM::VLD1DUPq32:
+ case ARM::VLD1DUPq8_UPD:
+ case ARM::VLD1DUPq16_UPD:
+ case ARM::VLD1DUPq32_UPD:
+ case ARM::VLD2DUPd8:
+ case ARM::VLD2DUPd16:
+ case ARM::VLD2DUPd32:
+ case ARM::VLD2DUPd8_UPD:
+ case ARM::VLD2DUPd16_UPD:
+ case ARM::VLD2DUPd32_UPD:
+ case ARM::VLD4DUPd8:
+ case ARM::VLD4DUPd16:
+ case ARM::VLD4DUPd32:
+ case ARM::VLD4DUPd8_UPD:
+ case ARM::VLD4DUPd16_UPD:
+ case ARM::VLD4DUPd32_UPD:
+ case ARM::VLD1LNd8:
+ case ARM::VLD1LNd16:
+ case ARM::VLD1LNd32:
+ case ARM::VLD1LNd8_UPD:
+ case ARM::VLD1LNd16_UPD:
+ case ARM::VLD1LNd32_UPD:
+ case ARM::VLD2LNd8:
+ case ARM::VLD2LNd16:
+ case ARM::VLD2LNd32:
+ case ARM::VLD2LNq16:
+ case ARM::VLD2LNq32:
+ case ARM::VLD2LNd8_UPD:
+ case ARM::VLD2LNd16_UPD:
+ case ARM::VLD2LNd32_UPD:
+ case ARM::VLD2LNq16_UPD:
+ case ARM::VLD2LNq32_UPD:
+ case ARM::VLD4LNd8:
+ case ARM::VLD4LNd16:
+ case ARM::VLD4LNd32:
+ case ARM::VLD4LNq16:
+ case ARM::VLD4LNq32:
+ case ARM::VLD4LNd8_UPD:
+ case ARM::VLD4LNd16_UPD:
+ case ARM::VLD4LNd32_UPD:
+ case ARM::VLD4LNq16_UPD:
+ case ARM::VLD4LNq32_UPD:
+ // If the address is not 64-bit aligned, the latencies of these
+ // instructions increases by one.
+ ++Latency;
+ break;
+ }
+
+ return Latency;
+}
+
+int
+ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
+ SDNode *DefNode, unsigned DefIdx,
+ SDNode *UseNode, unsigned UseIdx) const {
+ if (!DefNode->isMachineOpcode())
+ return 1;
+
+ const TargetInstrDesc &DefTID = get(DefNode->getMachineOpcode());
+
+ if (isZeroCost(DefTID.Opcode))
+ return 0;
+
+ if (!ItinData || ItinData->isEmpty())
+ return DefTID.mayLoad() ? 3 : 1;
+
+ 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);
+ unsigned DefAlign = !DefMN->memoperands_empty()
+ ? (*DefMN->memoperands_begin())->getAlignment() : 0;
+ const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
+ unsigned UseAlign = !UseMN->memoperands_empty()
+ ? (*UseMN->memoperands_begin())->getAlignment() : 0;
+ 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;
+ }
+ }
}
+
+ if (DefAlign < 8 && Subtarget.isCortexA9())
+ switch (DefTID.getOpcode()) {
+ default: break;
+ case ARM::VLD1q8Pseudo:
+ case ARM::VLD1q16Pseudo:
+ case ARM::VLD1q32Pseudo:
+ case ARM::VLD1q64Pseudo:
+ case ARM::VLD1q8Pseudo_UPD:
+ case ARM::VLD1q16Pseudo_UPD:
+ case ARM::VLD1q32Pseudo_UPD:
+ case ARM::VLD1q64Pseudo_UPD:
+ case ARM::VLD2d8Pseudo:
+ case ARM::VLD2d16Pseudo:
+ case ARM::VLD2d32Pseudo:
+ case ARM::VLD2q8Pseudo:
+ case ARM::VLD2q16Pseudo:
+ case ARM::VLD2q32Pseudo:
+ case ARM::VLD2d8Pseudo_UPD:
+ case ARM::VLD2d16Pseudo_UPD:
+ case ARM::VLD2d32Pseudo_UPD:
+ case ARM::VLD2q8Pseudo_UPD:
+ case ARM::VLD2q16Pseudo_UPD:
+ case ARM::VLD2q32Pseudo_UPD:
+ case ARM::VLD3d8Pseudo:
+ case ARM::VLD3d16Pseudo:
+ case ARM::VLD3d32Pseudo:
+ case ARM::VLD1d64TPseudo:
+ case ARM::VLD3d8Pseudo_UPD:
+ case ARM::VLD3d16Pseudo_UPD:
+ case ARM::VLD3d32Pseudo_UPD:
+ case ARM::VLD1d64TPseudo_UPD:
+ case ARM::VLD3q8Pseudo_UPD:
+ case ARM::VLD3q16Pseudo_UPD:
+ case ARM::VLD3q32Pseudo_UPD:
+ case ARM::VLD3q8oddPseudo:
+ case ARM::VLD3q16oddPseudo:
+ case ARM::VLD3q32oddPseudo:
+ case ARM::VLD3q8oddPseudo_UPD:
+ case ARM::VLD3q16oddPseudo_UPD:
+ case ARM::VLD3q32oddPseudo_UPD:
+ case ARM::VLD4d8Pseudo:
+ case ARM::VLD4d16Pseudo:
+ case ARM::VLD4d32Pseudo:
+ case ARM::VLD1d64QPseudo:
+ case ARM::VLD4d8Pseudo_UPD:
+ case ARM::VLD4d16Pseudo_UPD:
+ case ARM::VLD4d32Pseudo_UPD:
+ case ARM::VLD1d64QPseudo_UPD:
+ case ARM::VLD4q8Pseudo_UPD:
+ case ARM::VLD4q16Pseudo_UPD:
+ case ARM::VLD4q32Pseudo_UPD:
+ case ARM::VLD4q8oddPseudo:
+ case ARM::VLD4q16oddPseudo:
+ case ARM::VLD4q32oddPseudo:
+ case ARM::VLD4q8oddPseudo_UPD:
+ case ARM::VLD4q16oddPseudo_UPD:
+ case ARM::VLD4q32oddPseudo_UPD:
+ case ARM::VLD1DUPq8Pseudo:
+ case ARM::VLD1DUPq16Pseudo:
+ case ARM::VLD1DUPq32Pseudo:
+ case ARM::VLD1DUPq8Pseudo_UPD:
+ case ARM::VLD1DUPq16Pseudo_UPD:
+ case ARM::VLD1DUPq32Pseudo_UPD:
+ case ARM::VLD2DUPd8Pseudo:
+ case ARM::VLD2DUPd16Pseudo:
+ case ARM::VLD2DUPd32Pseudo:
+ case ARM::VLD2DUPd8Pseudo_UPD:
+ case ARM::VLD2DUPd16Pseudo_UPD:
+ case ARM::VLD2DUPd32Pseudo_UPD:
+ case ARM::VLD4DUPd8Pseudo:
+ case ARM::VLD4DUPd16Pseudo:
+ case ARM::VLD4DUPd32Pseudo:
+ case ARM::VLD4DUPd8Pseudo_UPD:
+ case ARM::VLD4DUPd16Pseudo_UPD:
+ case ARM::VLD4DUPd32Pseudo_UPD:
+ case ARM::VLD1LNq8Pseudo:
+ case ARM::VLD1LNq16Pseudo:
+ case ARM::VLD1LNq32Pseudo:
+ case ARM::VLD1LNq8Pseudo_UPD:
+ case ARM::VLD1LNq16Pseudo_UPD:
+ case ARM::VLD1LNq32Pseudo_UPD:
+ case ARM::VLD2LNd8Pseudo:
+ case ARM::VLD2LNd16Pseudo:
+ case ARM::VLD2LNd32Pseudo:
+ case ARM::VLD2LNq16Pseudo:
+ case ARM::VLD2LNq32Pseudo:
+ case ARM::VLD2LNd8Pseudo_UPD:
+ case ARM::VLD2LNd16Pseudo_UPD:
+ case ARM::VLD2LNd32Pseudo_UPD:
+ case ARM::VLD2LNq16Pseudo_UPD:
+ case ARM::VLD2LNq32Pseudo_UPD:
+ case ARM::VLD4LNd8Pseudo:
+ case ARM::VLD4LNd16Pseudo:
+ case ARM::VLD4LNd32Pseudo:
+ case ARM::VLD4LNq16Pseudo:
+ case ARM::VLD4LNq32Pseudo:
+ case ARM::VLD4LNd8Pseudo_UPD:
+ case ARM::VLD4LNd16Pseudo_UPD:
+ case ARM::VLD4LNd32Pseudo_UPD:
+ case ARM::VLD4LNq16Pseudo_UPD:
+ case ARM::VLD4LNq32Pseudo_UPD:
+ // If the address is not 64-bit aligned, the latencies of these
+ // instructions increases by one.
+ ++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::VSTMQIA:
+ 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;
+}
+
+bool
+ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
+ unsigned &AddSubOpc,
+ bool &NegAcc, bool &HasLane) const {
+ DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
+ if (I == MLxEntryMap.end())
+ return false;
+
+ const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
+ MulOpc = Entry.MulOpc;
+ AddSubOpc = Entry.AddSubOpc;
+ NegAcc = Entry.NegAcc;
+ HasLane = Entry.HasLane;
+ return true;
}