#include "llvm/CodeGen/MachineJumpTableInfo.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"
cl::desc("Enable ARM 2-addr to 3-addr conv"));
static cl::opt<bool>
-WidenVMOVS("widen-vmovs", cl::Hidden,
+WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true),
cl::desc("Widen ARM vmovs to vmovd when possible"));
/// ARM_MLxEntry - Record information about MLA / MLS instructions.
unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
const MCInstrDesc &MCID = MI->getDesc();
unsigned NumOps = MCID.getNumOperands();
- bool isLoad = !MCID.mayStore();
+ bool isLoad = !MI->mayStore();
const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
const MachineOperand &Base = MI->getOperand(2);
const MachineOperand &Offset = MI->getOperand(NumOps-3);
unsigned OffImm = MI->getOperand(NumOps-2).getImm();
ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
switch (AddrMode) {
- default:
- assert(false && "Unknown indexed op!");
- return NULL;
+ default: llvm_unreachable("Unknown indexed op!");
case ARMII::AddrMode2: {
bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
unsigned Amt = ARM_AM::getAM2Offset(OffImm);
return false;
}
+bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
+ if (MI->isBundle()) {
+ MachineBasicBlock::const_instr_iterator I = MI;
+ MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
+ while (++I != E && I->isInsideBundle()) {
+ int PIdx = I->findFirstPredOperandIdx();
+ if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
+ return true;
+ }
+ return false;
+ }
+
+ int PIdx = MI->findFirstPredOperandIdx();
+ return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
+}
+
bool ARMBaseInstrInfo::
PredicateInstruction(MachineInstr *MI,
const SmallVectorImpl<MachineOperand> &Pred) const {
std::vector<MachineOperand> &Pred) const {
// FIXME: This confuses implicit_def with optional CPSR def.
const MCInstrDesc &MCID = MI->getDesc();
- if (!MCID.getImplicitDefs() && !MCID.hasOptionalDef())
+ if (!MCID.getImplicitDefs() && !MI->hasOptionalDef())
return false;
bool Found = false;
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.getReg() == ARM::CPSR) {
+ if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
Pred.push_back(MO);
Found = true;
}
/// By default, this returns true for every instruction with a
/// PredicateOperand.
bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
- const MCInstrDesc &MCID = MI->getDesc();
- if (!MCID.isPredicable())
+ if (!MI->isPredicable())
return false;
- if ((MCID.TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
+ if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
ARMFunctionInfo *AFI =
MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
return AFI->isThumb2Function();
if (MCID.getSize())
return MCID.getSize();
- // 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;
+ // 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) {
- case TargetOpcode::IMPLICIT_DEF:
- case TargetOpcode::KILL:
- case TargetOpcode::PROLOG_LABEL:
- case TargetOpcode::EH_LABEL:
- case TargetOpcode::DBG_VALUE:
- return 0;
- 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.
- return MI->getOperand(2).getImm();
- case ARM::Int_eh_sjlj_longjmp:
- return 16;
- case ARM::tInt_eh_sjlj_longjmp:
- return 10;
- case ARM::Int_eh_sjlj_setjmp:
- case ARM::Int_eh_sjlj_setjmp_nofp:
- return 20;
- case ARM::tInt_eh_sjlj_setjmp:
- case ARM::t2Int_eh_sjlj_setjmp:
- case ARM::t2Int_eh_sjlj_setjmp_nofp:
- return 12;
- case ARM::BR_JTr:
- case ARM::BR_JTm:
- case ARM::BR_JTadd:
- case ARM::tBR_JTr:
- case ARM::t2BR_JT:
- 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_JT)
- ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
- unsigned NumOps = MCID.getNumOperands();
- MachineOperand JTOP =
- MI->getOperand(NumOps - (MCID.isPredicable() ? 3 : 2));
- unsigned JTI = JTOP.getIndex();
- const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
- assert(MJTI != 0);
- const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
- assert(JTI < JT.size());
- // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
- // 4 aligned. The assembler / linker may add 2 byte padding just before
- // the JT entries. The size does not include this padding; the
- // constant islands pass does separate bookkeeping for it.
- // FIXME: If we know the size of the function is less than (1 << 16) *2
- // bytes, we can use 16-bit entries instead. Then there won't be an
- // alignment issue.
- unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
- unsigned NumEntries = getNumJTEntries(JT, JTI);
- 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.
- ++NumEntries;
- return NumEntries * EntrySize + InstSize;
- }
- default:
- // Otherwise, pseudo-instruction sizes are zero.
- return 0;
- }
- return 0; // Not reached
+ switch (Opc) {
+ case TargetOpcode::IMPLICIT_DEF:
+ case TargetOpcode::KILL:
+ case TargetOpcode::PROLOG_LABEL:
+ case TargetOpcode::EH_LABEL:
+ case TargetOpcode::DBG_VALUE:
+ return 0;
+ case TargetOpcode::BUNDLE:
+ return getInstBundleLength(MI);
+ 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.
+ return MI->getOperand(2).getImm();
+ case ARM::Int_eh_sjlj_longjmp:
+ return 16;
+ case ARM::tInt_eh_sjlj_longjmp:
+ return 10;
+ case ARM::Int_eh_sjlj_setjmp:
+ case ARM::Int_eh_sjlj_setjmp_nofp:
+ return 20;
+ case ARM::tInt_eh_sjlj_setjmp:
+ case ARM::t2Int_eh_sjlj_setjmp:
+ case ARM::t2Int_eh_sjlj_setjmp_nofp:
+ return 12;
+ case ARM::BR_JTr:
+ case ARM::BR_JTm:
+ case ARM::BR_JTadd:
+ case ARM::tBR_JTr:
+ case ARM::t2BR_JT:
+ 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_JT)
+ ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
+ unsigned NumOps = MCID.getNumOperands();
+ MachineOperand JTOP =
+ MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
+ unsigned JTI = JTOP.getIndex();
+ const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
+ assert(MJTI != 0);
+ const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
+ assert(JTI < JT.size());
+ // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
+ // 4 aligned. The assembler / linker may add 2 byte padding just before
+ // the JT entries. The size does not include this padding; the
+ // constant islands pass does separate bookkeeping for it.
+ // FIXME: If we know the size of the function is less than (1 << 16) *2
+ // bytes, we can use 16-bit entries instead. Then there won't be an
+ // alignment issue.
+ unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
+ unsigned NumEntries = getNumJTEntries(JT, JTI);
+ 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.
+ ++NumEntries;
+ return NumEntries * EntrySize + InstSize;
+ }
+ default:
+ // Otherwise, pseudo-instruction sizes are zero.
+ return 0;
+ }
+}
+
+unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
+ unsigned Size = 0;
+ MachineBasicBlock::const_instr_iterator I = MI;
+ MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
+ while (++I != E && I->isInsideBundle()) {
+ assert(!I->isBundle() && "No nested bundle!");
+ Size += GetInstSizeInBytes(&*I);
+ }
+ return Size;
}
void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
unsigned Opc = 0;
- if (SPRDest && SPRSrc) {
+ if (SPRDest && SPRSrc)
Opc = ARM::VMOVS;
-
- // 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)
+ else if (GPRDest && SPRSrc)
Opc = ARM::VMOVRS;
else if (SPRDest && GPRSrc)
Opc = ARM::VMOVSR;
unsigned Align = MFI.getObjectAlignment(FI);
MachineMemOperand *MMO =
- MF.getMachineMemOperand(MachinePointerInfo(
- PseudoSourceValue::getFixedStack(FI)),
+ MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
MachineMemOperand::MOStore,
MFI.getObjectSize(FI),
Align);
break;
case 16:
if (ARM::QPRRegClass.hasSubClassEq(RC)) {
- if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
+ // Use aligned spills if the stack can be realigned.
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64Pseudo))
.addFrameIndex(FI).addImm(16)
.addReg(SrcReg, getKillRegState(isKill))
unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
int &FrameIndex) const {
const MachineMemOperand *Dummy;
- return MI->getDesc().mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
+ return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
}
void ARMBaseInstrInfo::
unsigned Align = MFI.getObjectAlignment(FI);
MachineMemOperand *MMO =
MF.getMachineMemOperand(
- MachinePointerInfo(PseudoSourceValue::getFixedStack(FI)),
+ MachinePointerInfo::getFixedStack(FI),
MachineMemOperand::MOLoad,
MFI.getObjectSize(FI),
Align);
break;
case 16:
if (ARM::QPRRegClass.hasSubClassEq(RC)) {
- if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64Pseudo), DestReg)
.addFrameIndex(FI).addImm(16)
.addMemOperand(MMO));
unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
int &FrameIndex) const {
const MachineMemOperand *Dummy;
- return MI->getDesc().mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
+ return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
+}
+
+bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
+ // This hook gets to expand COPY instructions before they become
+ // copyPhysReg() calls. Look for VMOVS instructions that can legally be
+ // widened to VMOVD. We prefer the VMOVD when possible because it may be
+ // changed into a VORR that can go down the NEON pipeline.
+ if (!WidenVMOVS || !MI->isCopy())
+ return false;
+
+ // Look for a copy between even S-registers. That is where we keep floats
+ // when using NEON v2f32 instructions for f32 arithmetic.
+ unsigned DstRegS = MI->getOperand(0).getReg();
+ unsigned SrcRegS = MI->getOperand(1).getReg();
+ if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
+ return false;
+
+ const TargetRegisterInfo *TRI = &getRegisterInfo();
+ unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
+ &ARM::DPRRegClass);
+ unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
+ &ARM::DPRRegClass);
+ if (!DstRegD || !SrcRegD)
+ return false;
+
+ // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
+ // legal if the COPY already defines the full DstRegD, and it isn't a
+ // sub-register insertion.
+ if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
+ return false;
+
+ // A dead copy shouldn't show up here, but reject it just in case.
+ if (MI->getOperand(0).isDead())
+ return false;
+
+ // All clear, widen the COPY.
+ DEBUG(dbgs() << "widening: " << *MI);
+
+ // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
+ // or some other super-register.
+ int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
+ if (ImpDefIdx != -1)
+ MI->RemoveOperand(ImpDefIdx);
+
+ // Change the opcode and operands.
+ MI->setDesc(get(ARM::VMOVD));
+ MI->getOperand(0).setReg(DstRegD);
+ MI->getOperand(1).setReg(SrcRegD);
+ AddDefaultPred(MachineInstrBuilder(MI));
+
+ // We are now reading SrcRegD instead of SrcRegS. This may upset the
+ // register scavenger and machine verifier, so we need to indicate that we
+ // are reading an undefined value from SrcRegD, but a proper value from
+ // SrcRegS.
+ MI->getOperand(1).setIsUndef();
+ MachineInstrBuilder(MI).addReg(SrcRegS, RegState::Implicit);
+
+ // SrcRegD may actually contain an unrelated value in the ssub_1
+ // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
+ if (MI->getOperand(1).isKill()) {
+ MI->getOperand(1).setIsKill(false);
+ MI->addRegisterKilled(SrcRegS, TRI, true);
+ }
+
+ DEBUG(dbgs() << "replaced by: " << *MI);
+ return true;
}
MachineInstr*
// instructions, so that's probably OK, but is PIC always correct when
// we get here?
if (ACPV->isGlobalValue())
- NewCPV = new ARMConstantPoolValue(ACPV->getGV(), PCLabelId,
- ARMCP::CPValue, 4);
+ NewCPV = ARMConstantPoolConstant::
+ Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
+ ARMCP::CPValue, 4);
else if (ACPV->isExtSymbol())
- NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(),
- ACPV->getSymbol(), PCLabelId, 4);
+ NewCPV = ARMConstantPoolSymbol::
+ Create(MF.getFunction()->getContext(),
+ cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
else if (ACPV->isBlockAddress())
- NewCPV = new ARMConstantPoolValue(ACPV->getBlockAddress(), PCLabelId,
- ARMCP::CPBlockAddress, 4);
+ NewCPV = ARMConstantPoolConstant::
+ Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
+ ARMCP::CPBlockAddress, 4);
else if (ACPV->isLSDA())
- NewCPV = new ARMConstantPoolValue(MF.getFunction(), PCLabelId,
- ARMCP::CPLSDA, 4);
+ NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
+ ARMCP::CPLSDA, 4);
+ else if (ACPV->isMachineBasicBlock())
+ NewCPV = ARMConstantPoolMBB::
+ Create(MF.getFunction()->getContext(),
+ cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
else
llvm_unreachable("Unexpected ARM constantpool value type!!");
CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
return false;
// Terminators and labels can't be scheduled around.
- if (MI->getDesc().isTerminator() || MI->isLabel())
+ if (MI->isTerminator() || MI->isLabel())
return true;
// Treat the start of the IT block as a scheduling boundary, but schedule
int llvm::getMatchingCondBranchOpcode(int Opc) {
if (Opc == ARM::B)
return ARM::Bcc;
- else if (Opc == ARM::tB)
+ if (Opc == ARM::tB)
return ARM::tBcc;
- else if (Opc == ARM::t2B)
- return ARM::t2Bcc;
+ if (Opc == ARM::t2B)
+ return ARM::t2Bcc;
llvm_unreachable("Unknown unconditional branch opcode!");
- return 0;
}
{ARM::SUBSrsr, ARM::SUBrsr},
{ARM::RSBSri, ARM::RSBri},
- {ARM::RSBSrr, ARM::RSBrr},
{ARM::RSBSrsi, ARM::RSBrsi},
{ARM::RSBSrsr, ARM::RSBrsr},
}
default:
llvm_unreachable("Unsupported addressing mode!");
- break;
}
Offset += InstrOffs * Scale;
// Check that CPSR isn't set between the comparison instruction and the one we
// want to change.
- MachineBasicBlock::const_iterator I = CmpInstr, E = MI,
- B = MI->getParent()->begin();
+ MachineBasicBlock::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;
bool isKill = UseMI->getOperand(OpIdx).isKill();
unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
- *UseMI, UseMI->getDebugLoc(),
+ UseMI, UseMI->getDebugLoc(),
get(NewUseOpc), NewReg)
.addReg(Reg1, getKillRegState(isKill))
.addImm(SOImmValV1)));
switch (Opc) {
default:
llvm_unreachable("Unexpected multi-uops instruction!");
- break;
case ARM::VLDMQIA:
case ARM::VSTMQIA:
return 2;
return UseCycle;
}
+static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
+ const MachineInstr *MI, unsigned Reg,
+ unsigned &DefIdx, unsigned &Dist) {
+ Dist = 0;
+
+ MachineBasicBlock::const_iterator I = MI; ++I;
+ MachineBasicBlock::const_instr_iterator II =
+ llvm::prior(I.getInstrIterator());
+ assert(II->isInsideBundle() && "Empty bundle?");
+
+ int Idx = -1;
+ while (II->isInsideBundle()) {
+ Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
+ if (Idx != -1)
+ break;
+ --II;
+ ++Dist;
+ }
+
+ assert(Idx != -1 && "Cannot find bundled definition!");
+ DefIdx = Idx;
+ return II;
+}
+
+static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
+ const MachineInstr *MI, unsigned Reg,
+ unsigned &UseIdx, unsigned &Dist) {
+ Dist = 0;
+
+ MachineBasicBlock::const_instr_iterator II = MI; ++II;
+ assert(II->isInsideBundle() && "Empty bundle?");
+ MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
+
+ // FIXME: This doesn't properly handle multiple uses.
+ int Idx = -1;
+ while (II != E && II->isInsideBundle()) {
+ Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
+ if (Idx != -1)
+ break;
+ if (II->getOpcode() != ARM::t2IT)
+ ++Dist;
+ ++II;
+ }
+
+ if (Idx == -1) {
+ Dist = 0;
+ return 0;
+ }
+
+ UseIdx = Idx;
+ return II;
+}
+
int
ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
const MachineInstr *DefMI, unsigned DefIdx,
DefMI->isRegSequence() || DefMI->isImplicitDef())
return 1;
- const MCInstrDesc &DefMCID = DefMI->getDesc();
if (!ItinData || ItinData->isEmpty())
- return DefMCID.mayLoad() ? 3 : 1;
+ return DefMI->mayLoad() ? 3 : 1;
- const MCInstrDesc &UseMCID = UseMI->getDesc();
+ const MCInstrDesc *DefMCID = &DefMI->getDesc();
+ const MCInstrDesc *UseMCID = &UseMI->getDesc();
const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
- if (DefMO.getReg() == ARM::CPSR) {
+ unsigned Reg = DefMO.getReg();
+ if (Reg == 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 (UseMCID.isBranch())
+ if (UseMI->isBranch())
return 0;
+
+ // Otherwise it takes the instruction latency (generally one).
+ int Latency = getInstrLatency(ItinData, DefMI);
+
+ // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
+ // its uses. Instructions which are otherwise scheduled between them may
+ // incur a code size penalty (not able to use the CPSR setting 16-bit
+ // instructions).
+ if (Latency > 0 && Subtarget.isThumb2()) {
+ const MachineFunction *MF = DefMI->getParent()->getParent();
+ if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize))
+ --Latency;
+ }
+ return Latency;
}
unsigned DefAlign = DefMI->hasOneMemOperand()
? (*DefMI->memoperands_begin())->getAlignment() : 0;
unsigned UseAlign = UseMI->hasOneMemOperand()
? (*UseMI->memoperands_begin())->getAlignment() : 0;
- int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
- UseMCID, UseIdx, UseAlign);
+
+ unsigned DefAdj = 0;
+ if (DefMI->isBundle()) {
+ DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
+ if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
+ DefMI->isRegSequence() || DefMI->isImplicitDef())
+ return 1;
+ DefMCID = &DefMI->getDesc();
+ }
+ unsigned UseAdj = 0;
+ if (UseMI->isBundle()) {
+ unsigned NewUseIdx;
+ const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
+ Reg, NewUseIdx, UseAdj);
+ if (NewUseMI) {
+ UseMI = NewUseMI;
+ UseIdx = NewUseIdx;
+ UseMCID = &UseMI->getDesc();
+ }
+ }
+
+ int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
+ *UseMCID, UseIdx, UseAlign);
+ int Adj = DefAdj + UseAdj;
+ if (Adj) {
+ Latency -= (int)(DefAdj + UseAdj);
+ if (Latency < 1)
+ return 1;
+ }
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 (DefMCID.getOpcode()) {
+ switch (DefMCID->getOpcode()) {
default: break;
case ARM::LDRrs:
case ARM::LDRBrs: {
}
if (DefAlign < 8 && Subtarget.isCortexA9())
- switch (DefMCID.getOpcode()) {
+ switch (DefMCID->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::VLD1q8wb_fixed:
+ case ARM::VLD1q16wb_fixed:
+ case ARM::VLD1q32wb_fixed:
+ case ARM::VLD1q64wb_fixed:
+ case ARM::VLD1q8wb_register:
+ case ARM::VLD1q16wb_register:
+ case ARM::VLD1q32wb_register:
+ case ARM::VLD1q64wb_register:
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::VLD2d8wb_fixed:
+ case ARM::VLD2d16wb_fixed:
+ case ARM::VLD2d32wb_fixed:
+ case ARM::VLD2q8wb_fixed:
+ case ARM::VLD2q16wb_fixed:
+ case ARM::VLD2q32wb_fixed:
+ case ARM::VLD2d8wb_register:
+ case ARM::VLD2d16wb_register:
+ case ARM::VLD2d32wb_register:
+ case ARM::VLD2q8wb_register:
+ case ARM::VLD2q16wb_register:
+ case ARM::VLD2q32wb_register:
case ARM::VLD3d8:
case ARM::VLD3d16:
case ARM::VLD3d32:
case ARM::VLD3d8_UPD:
case ARM::VLD3d16_UPD:
case ARM::VLD3d32_UPD:
- case ARM::VLD1d64T_UPD:
+ case ARM::VLD1d64Twb_fixed:
+ case ARM::VLD1d64Twb_register:
case ARM::VLD3q8_UPD:
case ARM::VLD3q16_UPD:
case ARM::VLD3q32_UPD:
case ARM::VLD4d8_UPD:
case ARM::VLD4d16_UPD:
case ARM::VLD4d32_UPD:
- case ARM::VLD1d64Q_UPD:
+ case ARM::VLD1d64Qwb_fixed:
+ case ARM::VLD1d64Qwb_register:
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::VLD1DUPq8wb_fixed:
+ case ARM::VLD1DUPq16wb_fixed:
+ case ARM::VLD1DUPq32wb_fixed:
+ case ARM::VLD1DUPq8wb_register:
+ case ARM::VLD1DUPq16wb_register:
+ case ARM::VLD1DUPq32wb_register:
case ARM::VLD2DUPd8:
case ARM::VLD2DUPd16:
case ARM::VLD2DUPd32:
- case ARM::VLD2DUPd8_UPD:
- case ARM::VLD2DUPd16_UPD:
- case ARM::VLD2DUPd32_UPD:
+ case ARM::VLD2DUPd8wb_fixed:
+ case ARM::VLD2DUPd16wb_fixed:
+ case ARM::VLD2DUPd32wb_fixed:
+ case ARM::VLD2DUPd8wb_register:
+ case ARM::VLD2DUPd16wb_register:
+ case ARM::VLD2DUPd32wb_register:
case ARM::VLD4DUPd8:
case ARM::VLD4DUPd16:
case ARM::VLD4DUPd32:
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::VLD1q8PseudoWB_register:
+ case ARM::VLD1q16PseudoWB_register:
+ case ARM::VLD1q32PseudoWB_register:
+ case ARM::VLD1q64PseudoWB_register:
+ case ARM::VLD1q8PseudoWB_fixed:
+ case ARM::VLD1q16PseudoWB_fixed:
+ case ARM::VLD1q32PseudoWB_fixed:
+ case ARM::VLD1q64PseudoWB_fixed:
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::VLD2d8PseudoWB_fixed:
+ case ARM::VLD2d16PseudoWB_fixed:
+ case ARM::VLD2d32PseudoWB_fixed:
+ case ARM::VLD2q8PseudoWB_fixed:
+ case ARM::VLD2q16PseudoWB_fixed:
+ case ARM::VLD2q32PseudoWB_fixed:
+ case ARM::VLD2d8PseudoWB_register:
+ case ARM::VLD2d16PseudoWB_register:
+ case ARM::VLD2d32PseudoWB_register:
+ case ARM::VLD2q8PseudoWB_register:
+ case ARM::VLD2q16PseudoWB_register:
+ case ARM::VLD2q32PseudoWB_register:
case ARM::VLD3d8Pseudo:
case ARM::VLD3d16Pseudo:
case ARM::VLD3d32Pseudo:
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::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::VLD1DUPq8Pseudo:
case ARM::VLD1DUPq16Pseudo:
case ARM::VLD1DUPq32Pseudo:
- case ARM::VLD1DUPq8Pseudo_UPD:
- case ARM::VLD1DUPq16Pseudo_UPD:
- case ARM::VLD1DUPq32Pseudo_UPD:
+ case ARM::VLD1DUPq8PseudoWB_fixed:
+ case ARM::VLD1DUPq16PseudoWB_fixed:
+ case ARM::VLD1DUPq32PseudoWB_fixed:
+ case ARM::VLD1DUPq8PseudoWB_register:
+ case ARM::VLD1DUPq16PseudoWB_register:
+ case ARM::VLD1DUPq32PseudoWB_register:
case ARM::VLD2DUPd8Pseudo:
case ARM::VLD2DUPd16Pseudo:
case ARM::VLD2DUPd32Pseudo:
- case ARM::VLD2DUPd8Pseudo_UPD:
- case ARM::VLD2DUPd16Pseudo_UPD:
- case ARM::VLD2DUPd32Pseudo_UPD:
+ case ARM::VLD2DUPd8PseudoWB_fixed:
+ case ARM::VLD2DUPd16PseudoWB_fixed:
+ case ARM::VLD2DUPd32PseudoWB_fixed:
+ case ARM::VLD2DUPd8PseudoWB_register:
+ case ARM::VLD2DUPd16PseudoWB_register:
+ case ARM::VLD2DUPd32PseudoWB_register:
case ARM::VLD4DUPd8Pseudo:
case ARM::VLD4DUPd16Pseudo:
case ARM::VLD4DUPd32Pseudo:
return Latency;
}
+unsigned
+ARMBaseInstrInfo::getOutputLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *DefMI, unsigned DefIdx,
+ const MachineInstr *DepMI) const {
+ unsigned Reg = DefMI->getOperand(DefIdx).getReg();
+ if (DepMI->readsRegister(Reg, &getRegisterInfo()) || !isPredicated(DepMI))
+ return 1;
+
+ // If the second MI is predicated, then there is an implicit use dependency.
+ return getOperandLatency(ItinData, DefMI, DefIdx, DepMI,
+ DepMI->getNumOperands());
+}
+
int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
const MachineInstr *MI,
unsigned *PredCost) const {
if (!ItinData || ItinData->isEmpty())
return 1;
+ if (MI->isBundle()) {
+ int Latency = 0;
+ MachineBasicBlock::const_instr_iterator I = MI;
+ MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
+ while (++I != E && I->isInsideBundle()) {
+ if (I->getOpcode() != ARM::t2IT)
+ Latency += getInstrLatency(ItinData, I, PredCost);
+ }
+ return Latency;
+ }
+
const MCInstrDesc &MCID = MI->getDesc();
unsigned Class = MCID.getSchedClass();
unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
- if (PredCost && MCID.hasImplicitDefOfPhysReg(ARM::CPSR))
+ if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR)))
// When predicated, CPSR is an additional source operand for CPSR updating
// instructions, this apparently increases their latencies.
*PredCost = 1;
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).addOperand(MI->getOperand(1)));
+}
projects / oota-llvm.git / blobdiff