#include "ARMBaseInstrInfo.h"
#include "ARM.h"
-#include "ARMAddressingModes.h"
#include "ARMConstantPoolValue.h"
#include "ARMHazardRecognizer.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMRegisterInfo.h"
-#include "ARMGenInstrInfo.inc"
+#include "MCTargetDesc/ARMAddressingModes.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
#include "llvm/GlobalValue.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/MC/MCAsmInfo.h"
+#include "llvm/Support/BranchProbability.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/ADT/STLExtras.h"
+
+#define GET_INSTRINFO_CTOR
+#include "ARMGenInstrInfo.inc"
+
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>
+WidenVMOVS("widen-vmovs", cl::Hidden,
+ cl::desc("Widen ARM vmovs to vmovd when possible"));
+
/// ARM_MLxEntry - Record information about MLA / MLS instructions.
struct ARM_MLxEntry {
unsigned MLxOpc; // MLA / MLS opcode
};
ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
- : TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)),
+ : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
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)
ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
UpdateMI = BuildMI(MF, MI->getDebugLoc(),
- get(isSub ? ARM::SUBrs : ARM::ADDrs), WBReg)
+ get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
.addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
.addImm(Pred).addReg(0).addReg(0);
} else
? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
int BccOpc = !AFI->isThumbFunction()
? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
+ bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
// Shouldn't be a fall through.
assert(TBB && "InsertBranch must not be told to insert a fallthrough");
"ARM branch conditions have two components!");
if (FBB == 0) {
- if (Cond.empty()) // Unconditional branch?
- BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
- else
+ if (Cond.empty()) { // Unconditional branch?
+ if (isThumb)
+ BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
+ else
+ BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
+ } else
BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
.addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
return 1;
// Two-way conditional branch.
BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
.addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
- BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
+ if (isThumb)
+ BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
+ else
+ BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
return 2;
}
const MachineFunction *MF = MBB.getParent();
const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
- // Basic size info comes from the TSFlags field.
const MCInstrDesc &MCID = MI->getDesc();
- uint64_t TSFlags = MCID.TSFlags;
+ if (MCID.getSize())
+ return MCID.getSize();
- unsigned Opc = MI->getOpcode();
- switch ((TSFlags & ARMII::SizeMask) >> ARMII::SizeShift) {
- default: {
// If this machine instr is an inline asm, measure it.
if (MI->getOpcode() == ARM::INLINEASM)
return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
if (MI->isLabel())
return 0;
+ unsigned Opc = MI->getOpcode();
switch (Opc) {
- default:
- llvm_unreachable("Unknown or unset size field for instr!");
case TargetOpcode::IMPLICIT_DEF:
case TargetOpcode::KILL:
case TargetOpcode::PROLOG_LABEL:
case TargetOpcode::EH_LABEL:
case TargetOpcode::DBG_VALUE:
return 0;
- }
- break;
- }
- case ARMII::Size8Bytes: return 8; // ARM instruction x 2.
- case ARMII::Size4Bytes: return 4; // ARM / Thumb2 instruction.
- 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:
// Otherwise, pseudo-instruction sizes are zero.
return 0;
}
- }
- }
return 0; // Not reached
}
bool SPRDest = ARM::SPRRegClass.contains(DestReg);
bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
- unsigned Opc;
- if (SPRDest && SPRSrc)
+ unsigned Opc = 0;
+ if (SPRDest && SPRSrc) {
Opc = ARM::VMOVS;
- else if (GPRDest && SPRSrc)
+
+ // An even S-S copy may be feeding a NEON v2f32 instruction being used for
+ // f32 operations. In that case, it is better to copy the full D-regs with
+ // a VMOVD since that can be converted to a NEON-domain move by
+ // NEONMoveFix.cpp. Check that MI is the original COPY instruction, and
+ // that it really defines the whole D-register.
+ if (WidenVMOVS &&
+ (DestReg - ARM::S0) % 2 == 0 && (SrcReg - ARM::S0) % 2 == 0 &&
+ I != MBB.end() && I->isCopy() &&
+ I->getOperand(0).getReg() == DestReg &&
+ I->getOperand(1).getReg() == SrcReg) {
+ // I is pointing to the ortiginal COPY instruction.
+ // Find the parent D-registers.
+ const TargetRegisterInfo *TRI = &getRegisterInfo();
+ unsigned SrcD = TRI->getMatchingSuperReg(SrcReg, ARM::ssub_0,
+ &ARM::DPRRegClass);
+ unsigned DestD = TRI->getMatchingSuperReg(DestReg, ARM::ssub_0,
+ &ARM::DPRRegClass);
+ // Be careful to not clobber an INSERT_SUBREG that reads and redefines a
+ // D-register. There must be an <imp-def> of destD, and no <imp-use>.
+ if (I->definesRegister(DestD, TRI) && !I->readsRegister(DestD, TRI)) {
+ Opc = ARM::VMOVD;
+ SrcReg = SrcD;
+ DestReg = DestD;
+ if (KillSrc)
+ KillSrc = I->killsRegister(SrcReg, TRI);
+ }
+ }
+ } else if (GPRDest && SPRSrc)
Opc = ARM::VMOVRS;
else if (SPRDest && GPRSrc)
Opc = ARM::VMOVSR;
else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
Opc = ARM::VMOVD;
else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
- Opc = ARM::VMOVQ;
- else if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
- Opc = ARM::VMOVQQ;
- else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
- Opc = ARM::VMOVQQQQ;
- else
- llvm_unreachable("Impossible reg-to-reg copy");
+ Opc = ARM::VORRq;
- MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
- MIB.addReg(SrcReg, getKillRegState(KillSrc));
- if (Opc != ARM::VMOVQQ && Opc != ARM::VMOVQQQQ)
+ if (Opc) {
+ MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
+ MIB.addReg(SrcReg, getKillRegState(KillSrc));
+ if (Opc == ARM::VORRq)
+ MIB.addReg(SrcReg, getKillRegState(KillSrc));
AddDefaultPred(MIB);
+ return;
+ }
+
+ // Generate instructions for VMOVQQ and VMOVQQQQ pseudos in place.
+ if (ARM::QQPRRegClass.contains(DestReg, SrcReg) ||
+ ARM::QQQQPRRegClass.contains(DestReg, SrcReg)) {
+ const TargetRegisterInfo *TRI = &getRegisterInfo();
+ assert(ARM::qsub_0 + 3 == ARM::qsub_3 && "Expected contiguous enum.");
+ unsigned EndSubReg = ARM::QQPRRegClass.contains(DestReg, SrcReg) ?
+ ARM::qsub_1 : ARM::qsub_3;
+ for (unsigned i = ARM::qsub_0, e = EndSubReg + 1; i != e; ++i) {
+ unsigned Dst = TRI->getSubReg(DestReg, i);
+ unsigned Src = TRI->getSubReg(SrcReg, i);
+ MachineInstrBuilder Mov =
+ AddDefaultPred(BuildMI(MBB, I, I->getDebugLoc(), get(ARM::VORRq))
+ .addReg(Dst, RegState::Define)
+ .addReg(Src, getKillRegState(KillSrc))
+ .addReg(Src, getKillRegState(KillSrc)));
+ if (i == EndSubReg) {
+ Mov->addRegisterDefined(DestReg, TRI);
+ if (KillSrc)
+ Mov->addRegisterKilled(SrcReg, TRI);
+ }
+ }
+ return;
+ }
+ llvm_unreachable("Impossible reg-to-reg copy");
}
static const
MFI.getObjectSize(FI),
Align);
- // tGPR is used sometimes in ARM instructions that need to avoid using
- // certain registers. Just treat it as GPR here. Likewise, rGPR.
- if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
- || RC == ARM::rGPRRegisterClass)
- RC = ARM::GPRRegisterClass;
-
- switch (RC->getID()) {
- case ARM::GPRRegClassID:
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
+ switch (RC->getSize()) {
+ case 4:
+ if (ARM::GPRRegClass.hasSubClassEq(RC)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
.addReg(SrcReg, getKillRegState(isKill))
.addFrameIndex(FI).addImm(0).addMemOperand(MMO));
- break;
- case ARM::SPRRegClassID:
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
+ } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
.addReg(SrcReg, getKillRegState(isKill))
.addFrameIndex(FI).addImm(0).addMemOperand(MMO));
- break;
- case ARM::DPRRegClassID:
- case ARM::DPR_VFP2RegClassID:
- case ARM::DPR_8RegClassID:
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
+ } else
+ llvm_unreachable("Unknown reg class!");
+ break;
+ case 8:
+ if (ARM::DPRRegClass.hasSubClassEq(RC)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
.addReg(SrcReg, getKillRegState(isKill))
.addFrameIndex(FI).addImm(0).addMemOperand(MMO));
- break;
- case ARM::QPRRegClassID:
- case ARM::QPR_VFP2RegClassID:
- case ARM::QPR_8RegClassID:
- if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64Pseudo))
+ } else
+ llvm_unreachable("Unknown reg class!");
+ break;
+ case 16:
+ if (ARM::QPRRegClass.hasSubClassEq(RC)) {
+ if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64Pseudo))
.addFrameIndex(FI).addImm(16)
.addReg(SrcReg, getKillRegState(isKill))
.addMemOperand(MMO));
- } else {
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
+ } else {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
.addReg(SrcReg, getKillRegState(isKill))
.addFrameIndex(FI)
.addMemOperand(MMO));
- }
- break;
- case ARM::QQPRRegClassID:
- case ARM::QQPR_VFP2RegClassID:
- if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
- // FIXME: It's possible to only store part of the QQ register if the
- // spilled def has a sub-register index.
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
+ }
+ } else
+ llvm_unreachable("Unknown reg class!");
+ break;
+ case 32:
+ if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+ // FIXME: It's possible to only store part of the QQ register if the
+ // spilled def has a sub-register index.
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
.addFrameIndex(FI).addImm(16)
.addReg(SrcReg, getKillRegState(isKill))
.addMemOperand(MMO));
- } else {
- MachineInstrBuilder MIB =
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
+ } else {
+ MachineInstrBuilder MIB =
+ 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);
- MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
- AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
- }
- break;
- case ARM::QQQQPRRegClassID: {
- MachineInstrBuilder MIB =
- 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);
- MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
- MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
- MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
- MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
- MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
- AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
- break;
- }
- default:
- llvm_unreachable("Unknown regclass!");
+ .addMemOperand(MMO);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
+ AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
+ }
+ } else
+ llvm_unreachable("Unknown reg class!");
+ break;
+ case 64:
+ if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
+ MachineInstrBuilder MIB =
+ 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);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
+ MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
+ AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
+ } else
+ llvm_unreachable("Unknown reg class!");
+ break;
+ default:
+ llvm_unreachable("Unknown reg class!");
}
}
break;
case ARM::STRi12:
case ARM::t2STRi12:
- case ARM::tSpill:
+ case ARM::tSTRspi:
case ARM::VSTRD:
case ARM::VSTRS:
if (MI->getOperand(1).isFI() &&
return 0;
}
+unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
+ int &FrameIndex) const {
+ const MachineMemOperand *Dummy;
+ return MI->getDesc().mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
+}
+
void ARMBaseInstrInfo::
loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
unsigned DestReg, int FI,
MFI.getObjectSize(FI),
Align);
- // tGPR is used sometimes in ARM instructions that need to avoid using
- // certain registers. Just treat it as GPR here.
- if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
- || RC == ARM::rGPRRegisterClass)
- RC = ARM::GPRRegisterClass;
-
- switch (RC->getID()) {
- case ARM::GPRRegClassID:
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
+ switch (RC->getSize()) {
+ case 4:
+ if (ARM::GPRRegClass.hasSubClassEq(RC)) {
+ 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)
+
+ } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO));
+ } else
+ llvm_unreachable("Unknown reg class!");
break;
- case ARM::DPRRegClassID:
- case ARM::DPR_VFP2RegClassID:
- case ARM::DPR_8RegClassID:
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
+ case 8:
+ if (ARM::DPRRegClass.hasSubClassEq(RC)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO));
+ } else
+ llvm_unreachable("Unknown reg class!");
break;
- case ARM::QPRRegClassID:
- case ARM::QPR_VFP2RegClassID:
- case ARM::QPR_8RegClassID:
- if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64Pseudo), DestReg)
+ case 16:
+ if (ARM::QPRRegClass.hasSubClassEq(RC)) {
+ if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64Pseudo), DestReg)
.addFrameIndex(FI).addImm(16)
.addMemOperand(MMO));
- } else {
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
- .addFrameIndex(FI)
- .addMemOperand(MMO));
- }
+ } else {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
+ .addFrameIndex(FI)
+ .addMemOperand(MMO));
+ }
+ } else
+ llvm_unreachable("Unknown reg class!");
break;
- case ARM::QQPRRegClassID:
- case ARM::QQPR_VFP2RegClassID:
- if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
- AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
+ case 32:
+ if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
.addFrameIndex(FI).addImm(16)
.addMemOperand(MMO));
- } else {
- MachineInstrBuilder MIB =
+ } else {
+ MachineInstrBuilder MIB =
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);
- MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
- AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
- }
+ .addMemOperand(MMO);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
+ MIB.addReg(DestReg, RegState::Define | RegState::Implicit);
+ }
+ } else
+ llvm_unreachable("Unknown reg class!");
break;
- case ARM::QQQQPRRegClassID: {
- MachineInstrBuilder MIB =
+ case 64:
+ if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
+ MachineInstrBuilder MIB =
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);
- MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
- MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
- MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI);
- MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI);
- MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI);
- AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
+ .addMemOperand(MMO);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
+ MIB.addReg(DestReg, RegState::Define | RegState::Implicit);
+ } else
+ llvm_unreachable("Unknown reg class!");
break;
- }
default:
llvm_unreachable("Unknown regclass!");
}
break;
case ARM::LDRi12:
case ARM::t2LDRi12:
- case ARM::tRestore:
+ case ARM::tLDRspi:
case ARM::VLDRD:
case ARM::VLDRS:
if (MI->getOperand(1).isFI() &&
return 0;
}
+unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
+ int &FrameIndex) const {
+ const MachineMemOperand *Dummy;
+ return MI->getDesc().mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
+}
+
MachineInstr*
ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
int FrameIx, uint64_t Offset,
else if (ACPV->isLSDA())
NewCPV = new ARMConstantPoolValue(MF.getFunction(), PCLabelId,
ARMCP::CPLSDA, 4);
+ else if (ACPV->isMachineBasicBlock())
+ NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(),
+ ACPV->getMBB(), PCLabelId,
+ ARMCP::CPMachineBasicBlock, 4);
else
llvm_unreachable("Unexpected ARM constantpool value type!!");
CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
return false;
}
-bool ARMBaseInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
- unsigned NumCycles,
- unsigned ExtraPredCycles,
- float Probability,
- float Confidence) const {
+bool ARMBaseInstrInfo::
+isProfitableToIfCvt(MachineBasicBlock &MBB,
+ unsigned NumCycles, unsigned ExtraPredCycles,
+ const BranchProbability &Probability) const {
if (!NumCycles)
return false;
// Attempt to estimate the relative costs of predication versus branching.
- float UnpredCost = Probability * NumCycles;
- UnpredCost += 1.0; // The branch itself
- UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
+ unsigned UnpredCost = Probability.getNumerator() * NumCycles;
+ UnpredCost /= Probability.getDenominator();
+ UnpredCost += 1; // The branch itself
+ UnpredCost += Subtarget.getMispredictionPenalty() / 10;
- return (float)(NumCycles + ExtraPredCycles) < UnpredCost;
+ return (NumCycles + ExtraPredCycles) <= UnpredCost;
}
bool ARMBaseInstrInfo::
unsigned TCycles, unsigned TExtra,
MachineBasicBlock &FMBB,
unsigned FCycles, unsigned FExtra,
- float Probability, float Confidence) const {
+ const BranchProbability &Probability) const {
if (!TCycles || !FCycles)
return false;
// Attempt to estimate the relative costs of predication versus branching.
- float UnpredCost = Probability * TCycles + (1.0 - Probability) * FCycles;
- UnpredCost += 1.0; // The branch itself
- UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
+ unsigned TUnpredCost = Probability.getNumerator() * TCycles;
+ TUnpredCost /= Probability.getDenominator();
+
+ uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
+ unsigned FUnpredCost = Comp * FCycles;
+ FUnpredCost /= Probability.getDenominator();
- return (float)(TCycles + FCycles + TExtra + FExtra) < UnpredCost;
+ unsigned UnpredCost = TUnpredCost + FUnpredCost;
+ UnpredCost += 1; // The branch itself
+ UnpredCost += Subtarget.getMispredictionPenalty() / 10;
+
+ return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
}
/// getInstrPredicate - If instruction is predicated, returns its predicate
}
+/// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
+/// instruction is encoded with an 'S' bit is determined by the optional CPSR
+/// def operand.
+///
+/// This will go away once we can teach tblgen how to set the optional CPSR def
+/// operand itself.
+struct AddSubFlagsOpcodePair {
+ unsigned PseudoOpc;
+ unsigned MachineOpc;
+};
+
+static AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
+ {ARM::ADDSri, ARM::ADDri},
+ {ARM::ADDSrr, ARM::ADDrr},
+ {ARM::ADDSrsi, ARM::ADDrsi},
+ {ARM::ADDSrsr, ARM::ADDrsr},
+
+ {ARM::SUBSri, ARM::SUBri},
+ {ARM::SUBSrr, ARM::SUBrr},
+ {ARM::SUBSrsi, ARM::SUBrsi},
+ {ARM::SUBSrsr, ARM::SUBrsr},
+
+ {ARM::RSBSri, ARM::RSBri},
+ {ARM::RSBSrr, ARM::RSBrr},
+ {ARM::RSBSrsi, ARM::RSBrsi},
+ {ARM::RSBSrsr, ARM::RSBrsr},
+
+ {ARM::t2ADDSri, ARM::t2ADDri},
+ {ARM::t2ADDSrr, ARM::t2ADDrr},
+ {ARM::t2ADDSrs, ARM::t2ADDrs},
+
+ {ARM::t2SUBSri, ARM::t2SUBri},
+ {ARM::t2SUBSrr, ARM::t2SUBrr},
+ {ARM::t2SUBSrs, ARM::t2SUBrs},
+
+ {ARM::t2RSBSri, ARM::t2RSBri},
+ {ARM::t2RSBSrs, ARM::t2RSBrs},
+};
+
+unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
+ static const int NPairs =
+ sizeof(AddSubFlagsOpcodeMap) / sizeof(AddSubFlagsOpcodePair);
+ for (AddSubFlagsOpcodePair *OpcPair = &AddSubFlagsOpcodeMap[0],
+ *End = &AddSubFlagsOpcodeMap[NPairs]; OpcPair != End; ++OpcPair) {
+ if (OldOpc == OpcPair->PseudoOpc) {
+ return OpcPair->MachineOpc;
+ }
+ }
+ return 0;
+}
+
void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI, DebugLoc dl,
unsigned DestReg, unsigned BaseReg, int NumBytes,
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::STMDA_UPD:
case ARM::STMDB_UPD:
case ARM::STMIB_UPD:
- case ARM::tSTMIA:
case ARM::tSTMIA_UPD:
case ARM::tPOP_RET:
case ARM::tPOP:
return false;
}
+bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
+ StringRef &ErrInfo) const {
+ if (convertAddSubFlagsOpcode(MI->getOpcode())) {
+ ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
+ return false;
+ }
+ return true;
+}
+
bool
ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
unsigned &AddSubOpc,
HasLane = Entry.HasLane;
return true;
}
+
+//===----------------------------------------------------------------------===//
+// Execution domains.
+//===----------------------------------------------------------------------===//
+//
+// Some instructions go down the NEON pipeline, some go down the VFP pipeline,
+// and some can go down both. The vmov instructions go down the VFP pipeline,
+// but they can be changed to vorr equivalents that are executed by the NEON
+// pipeline.
+//
+// We use the following execution domain numbering:
+//
+enum ARMExeDomain {
+ ExeGeneric = 0,
+ ExeVFP = 1,
+ ExeNEON = 2
+};
+//
+// Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
+//
+std::pair<uint16_t, uint16_t>
+ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
+ // VMOVD is a VFP instruction, but can be changed to NEON if it isn't
+ // predicated.
+ if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
+ return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
+
+ // No other instructions can be swizzled, so just determine their domain.
+ unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
+
+ if (Domain & ARMII::DomainNEON)
+ return std::make_pair(ExeNEON, 0);
+
+ // Certain instructions can go either way on Cortex-A8.
+ // Treat them as NEON instructions.
+ if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
+ return std::make_pair(ExeNEON, 0);
+
+ if (Domain & ARMII::DomainVFP)
+ return std::make_pair(ExeVFP, 0);
+
+ return std::make_pair(ExeGeneric, 0);
+}
+
+void
+ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
+ // We only know how to change VMOVD into VORR.
+ assert(MI->getOpcode() == ARM::VMOVD && "Can only swizzle VMOVD");
+ if (Domain != ExeNEON)
+ return;
+
+ // Zap the predicate operands.
+ assert(!isPredicated(MI) && "Cannot predicate a VORRd");
+ MI->RemoveOperand(3);
+ MI->RemoveOperand(2);
+
+ // Change to a VORRd which requires two identical use operands.
+ MI->setDesc(get(ARM::VORRd));
+
+ // Add the extra source operand and new predicates.
+ // This will go before any implicit ops.
+ AddDefaultPred(MachineInstrBuilder(MI).addReg(MI->getOperand(1).getReg()));
+}