-//===- ARMBaseInstrInfo.cpp - ARM Instruction Information -------*- C++ -*-===//
+//===-- ARMBaseInstrInfo.cpp - ARM Instruction Information ----------------===//
//
// The LLVM Compiler Infrastructure
//
#include "ARMBaseInstrInfo.h"
#include "ARM.h"
+#include "ARMBaseRegisterInfo.h"
#include "ARMConstantPoolValue.h"
#include "ARMHazardRecognizer.h"
#include "ARMMachineFunctionInfo.h"
-#include "ARMRegisterInfo.h"
#include "MCTargetDesc/ARMAddressingModes.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
#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.
struct ARM_MLxEntry {
- unsigned MLxOpc; // MLA / MLS opcode
- unsigned MulOpc; // Expanded multiplication opcode
- unsigned AddSubOpc; // Expanded add / sub opcode
+ uint16_t MLxOpc; // MLA / MLS opcode
+ uint16_t MulOpc; // Expanded multiplication opcode
+ uint16_t 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.
};
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 {
bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
std::vector<MachineOperand> &Pred) const {
- // FIXME: This confuses implicit_def with optional CPSR def.
- const MCInstrDesc &MCID = MI->getDesc();
- if (!MCID.getImplicitDefs() && !MCID.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.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
+ (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;
return;
}
- // Generate instructions for VMOVQQ and VMOVQQQQ pseudos in place.
- if (ARM::QQPRRegClass.contains(DestReg, SrcReg) ||
- ARM::QQQQPRRegClass.contains(DestReg, SrcReg)) {
+ // Handle register classes that require multiple instructions.
+ unsigned BeginIdx = 0;
+ unsigned SubRegs = 0;
+ unsigned Spacing = 1;
+
+ // Use VORRq when possible.
+ if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 2;
+ else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 4;
+ // Fall back to VMOVD.
+ else if (ARM::DPairRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2;
+ else if (ARM::DTripleRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3;
+ else if (ARM::DQuadRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4;
+
+ else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2, Spacing = 2;
+ else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3, Spacing = 2;
+ else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg))
+ Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4, Spacing = 2;
+
+ if (Opc) {
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);
- }
+ MachineInstrBuilder Mov;
+ for (unsigned i = 0; i != SubRegs; ++i) {
+ unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i*Spacing);
+ unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i*Spacing);
+ assert(Dst && Src && "Bad sub-register");
+ Mov = AddDefaultPred(BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst)
+ .addReg(Src));
+ // VORR takes two source operands.
+ if (Opc == ARM::VORRq)
+ Mov.addReg(Src);
}
+ // Add implicit super-register defs and kills to the last instruction.
+ Mov->addRegisterDefined(DestReg, TRI);
+ if (KillSrc)
+ Mov->addRegisterKilled(SrcReg, TRI);
return;
}
+
llvm_unreachable("Impossible reg-to-reg copy");
}
unsigned Align = MFI.getObjectAlignment(FI);
MachineMemOperand *MMO =
- MF.getMachineMemOperand(MachinePointerInfo(
- PseudoSourceValue::getFixedStack(FI)),
+ MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
MachineMemOperand::MOStore,
MFI.getObjectSize(FI),
Align);
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))
+ if (ARM::DPairRegClass.hasSubClassEq(RC)) {
+ // Use aligned spills if the stack can be realigned.
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
.addFrameIndex(FI).addImm(16)
.addReg(SrcReg, getKillRegState(isKill))
.addMemOperand(MMO));
} else
llvm_unreachable("Unknown reg class!");
break;
+ case 24:
+ if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
+ // Use aligned spills if the stack can be realigned.
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
+ .addFrameIndex(FI).addImm(16)
+ .addReg(SrcReg, getKillRegState(isKill))
+ .addMemOperand(MMO));
+ } 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);
+ AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
+ }
+ } else
+ llvm_unreachable("Unknown reg class!");
+ break;
case 32:
- if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
+ if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.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.
return MI->getOperand(0).getReg();
}
break;
- case ARM::VST1q64Pseudo:
+ case ARM::VST1q64:
if (MI->getOperand(0).isFI() &&
MI->getOperand(2).getSubReg() == 0) {
FrameIndex = MI->getOperand(0).getIndex();
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);
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::VLD1q64Pseudo), DestReg)
+ if (ARM::DPairRegClass.hasSubClassEq(RC)) {
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
.addFrameIndex(FI).addImm(16)
.addMemOperand(MMO));
} else {
} else
llvm_unreachable("Unknown reg class!");
break;
- case 32:
- if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
+ case 24:
+ if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
+ if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
+ .addFrameIndex(FI).addImm(16)
+ .addMemOperand(MMO));
+ } else {
+ MachineInstrBuilder MIB =
+ AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
+ .addFrameIndex(FI)
+ .addMemOperand(MMO));
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
+ if (TargetRegisterInfo::isPhysicalRegister(DestReg))
+ MIB.addReg(DestReg, RegState::ImplicitDefine);
+ }
+ } else
+ llvm_unreachable("Unknown reg class!");
+ break;
+ case 32:
+ if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
.addFrameIndex(FI).addImm(16)
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.addReg(DestReg, RegState::Define | RegState::Implicit);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
+ if (TargetRegisterInfo::isPhysicalRegister(DestReg))
+ MIB.addReg(DestReg, RegState::ImplicitDefine);
}
} else
llvm_unreachable("Unknown reg class!");
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);
- MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
- MIB.addReg(DestReg, RegState::Define | RegState::Implicit);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
+ MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
+ if (TargetRegisterInfo::isPhysicalRegister(DestReg))
+ MIB.addReg(DestReg, RegState::ImplicitDefine);
} else
llvm_unreachable("Unknown reg class!");
break;
return MI->getOperand(0).getReg();
}
break;
- case ARM::VLD1q64Pseudo:
+ case ARM::VLD1q64:
if (MI->getOperand(1).isFI() &&
MI->getOperand(0).getSubReg() == 0) {
FrameIndex = MI->getOperand(1).getIndex();
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
// saves compile time, because it doesn't require every single
// stack slot reference to depend on the instruction that does the
// modification.
- if (MI->definesRegister(ARM::SP))
+ // Calls don't actually change the stack pointer, even if they have imp-defs.
+ // No ARM calling conventions change the stack pointer. (X86 calling
+ // conventions sometimes do).
+ if (!MI->isCall() && MI->definesRegister(ARM::SP))
return true;
return false;
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;
}
+/// commuteInstruction - Handle commutable instructions.
+MachineInstr *
+ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
+ switch (MI->getOpcode()) {
+ case ARM::MOVCCr:
+ case ARM::t2MOVCCr: {
+ // MOVCC can be commuted by inverting the condition.
+ unsigned PredReg = 0;
+ ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
+ // MOVCC AL can't be inverted. Shouldn't happen.
+ if (CC == ARMCC::AL || PredReg != ARM::CPSR)
+ return NULL;
+ MI = TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
+ if (!MI)
+ return NULL;
+ // After swapping the MOVCC operands, also invert the condition.
+ MI->getOperand(MI->findFirstPredOperandIdx())
+ .setImm(ARMCC::getOppositeCondition(CC));
+ return MI;
+ }
+ }
+ return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
+}
/// 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
/// This will go away once we can teach tblgen how to set the optional CPSR def
/// operand itself.
struct AddSubFlagsOpcodePair {
- unsigned PseudoOpc;
- unsigned MachineOpc;
+ uint16_t PseudoOpc;
+ uint16_t MachineOpc;
};
-static AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
+static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
{ARM::ADDSri, ARM::ADDri},
{ARM::ADDSrr, ARM::ADDrr},
{ARM::ADDSrsi, ARM::ADDrsi},
{ARM::SUBSrsr, ARM::SUBrsr},
{ARM::RSBSri, ARM::RSBri},
- {ARM::RSBSrr, ARM::RSBrr},
{ARM::RSBSrsi, ARM::RSBrsi},
{ARM::RSBSrsr, ARM::RSBrsr},
};
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;
- }
- }
+ for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
+ if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
+ return AddSubFlagsOpcodeMap[i].MachineOpc;
return 0;
}
}
default:
llvm_unreachable("Unsupported addressing mode!");
- break;
}
Offset += InstrOffs * Scale;
return Offset == 0;
}
+/// analyzeCompare - For a comparison instruction, return the source registers
+/// in SrcReg and SrcReg2 if having two register operands, and the value it
+/// compares against in CmpValue. Return true if the comparison instruction
+/// can be analyzed.
bool ARMBaseInstrInfo::
-AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpMask,
- int &CmpValue) const {
+analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
+ int &CmpMask, int &CmpValue) const {
switch (MI->getOpcode()) {
default: break;
case ARM::CMPri:
case ARM::t2CMPri:
SrcReg = MI->getOperand(0).getReg();
+ SrcReg2 = 0;
CmpMask = ~0;
CmpValue = MI->getOperand(1).getImm();
return true;
+ case ARM::CMPrr:
+ case ARM::t2CMPrr:
+ SrcReg = MI->getOperand(0).getReg();
+ SrcReg2 = MI->getOperand(1).getReg();
+ CmpMask = ~0;
+ CmpValue = 0;
+ return true;
case ARM::TSTri:
case ARM::t2TSTri:
SrcReg = MI->getOperand(0).getReg();
+ SrcReg2 = 0;
CmpMask = MI->getOperand(1).getImm();
CmpValue = 0;
return true;
return false;
}
-/// OptimizeCompareInstr - Convert the instruction supplying the argument to the
-/// comparison into one that sets the zero bit in the flags register.
-bool ARMBaseInstrInfo::
-OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
- int CmpValue, const MachineRegisterInfo *MRI) const {
- if (CmpValue != 0)
- return false;
+/// getSwappedCondition - assume the flags are set by MI(a,b), return
+/// the condition code if we modify the instructions such that flags are
+/// set by MI(b,a).
+inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) {
+ switch (CC) {
+ default: return ARMCC::AL;
+ case ARMCC::EQ: return ARMCC::EQ;
+ case ARMCC::NE: return ARMCC::NE;
+ case ARMCC::HS: return ARMCC::LS;
+ case ARMCC::LO: return ARMCC::HI;
+ case ARMCC::HI: return ARMCC::LO;
+ case ARMCC::LS: return ARMCC::HS;
+ case ARMCC::GE: return ARMCC::LE;
+ case ARMCC::LT: return ARMCC::GT;
+ case ARMCC::GT: return ARMCC::LT;
+ case ARMCC::LE: return ARMCC::GE;
+ }
+}
- MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
- if (llvm::next(DI) != MRI->def_end())
- // Only support one definition.
- return false;
+/// isRedundantFlagInstr - check whether the first instruction, whose only
+/// purpose is to update flags, can be made redundant.
+/// CMPrr can be made redundant by SUBrr if the operands are the same.
+/// CMPri can be made redundant by SUBri if the operands are the same.
+/// This function can be extended later on.
+inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg,
+ unsigned SrcReg2, int ImmValue,
+ MachineInstr *OI) {
+ if ((CmpI->getOpcode() == ARM::CMPrr ||
+ CmpI->getOpcode() == ARM::t2CMPrr) &&
+ (OI->getOpcode() == ARM::SUBrr ||
+ OI->getOpcode() == ARM::t2SUBrr) &&
+ ((OI->getOperand(1).getReg() == SrcReg &&
+ OI->getOperand(2).getReg() == SrcReg2) ||
+ (OI->getOperand(1).getReg() == SrcReg2 &&
+ OI->getOperand(2).getReg() == SrcReg)))
+ return true;
+
+ if ((CmpI->getOpcode() == ARM::CMPri ||
+ CmpI->getOpcode() == ARM::t2CMPri) &&
+ (OI->getOpcode() == ARM::SUBri ||
+ OI->getOpcode() == ARM::t2SUBri) &&
+ OI->getOperand(1).getReg() == SrcReg &&
+ OI->getOperand(2).getImm() == ImmValue)
+ return true;
+ return false;
+}
- MachineInstr *MI = &*DI;
+/// optimizeCompareInstr - Convert the instruction supplying the argument to the
+/// comparison into one that sets the zero bit in the flags register;
+/// Remove a redundant Compare instruction if an earlier instruction can set the
+/// flags in the same way as Compare.
+/// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two
+/// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the
+/// condition code of instructions which use the flags.
+bool ARMBaseInstrInfo::
+optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
+ int CmpMask, int CmpValue,
+ const MachineRegisterInfo *MRI) const {
+ // Get the unique definition of SrcReg.
+ MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
+ if (!MI) return false;
// Masked compares sometimes use the same register as the corresponding 'and'.
if (CmpMask != ~0) {
}
}
- // Conservatively refuse to convert an instruction which isn't in the same BB
- // as the comparison.
- if (MI->getParent() != CmpInstr->getParent())
- return false;
-
- // 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();
+ // Get ready to iterate backward from CmpInstr.
+ MachineBasicBlock::iterator I = CmpInstr, E = MI,
+ B = CmpInstr->getParent()->begin();
// Early exit if CmpInstr is at the beginning of the BB.
if (I == B) return false;
+ // There are two possible candidates which can be changed to set CPSR:
+ // One is MI, the other is a SUB instruction.
+ // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
+ // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
+ MachineInstr *Sub = NULL;
+ if (SrcReg2 != 0)
+ // MI is not a candidate for CMPrr.
+ MI = NULL;
+ else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
+ // Conservatively refuse to convert an instruction which isn't in the same
+ // BB as the comparison.
+ // For CMPri, we need to check Sub, thus we can't return here.
+ if (CmpInstr->getOpcode() == ARM::CMPri ||
+ CmpInstr->getOpcode() == ARM::t2CMPri)
+ MI = NULL;
+ else
+ return false;
+ }
+
+ // Check that CPSR isn't set between the comparison instruction and the one we
+ // want to change. At the same time, search for Sub.
+ const TargetRegisterInfo *TRI = &getRegisterInfo();
--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()) continue;
-
+ if (Instr.modifiesRegister(ARM::CPSR, TRI) ||
+ Instr.readsRegister(ARM::CPSR, TRI))
// 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;
+ return false;
+
+ // Check whether CmpInstr can be made redundant by the current instruction.
+ if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
+ Sub = &*I;
+ break;
}
if (I == B)
return false;
}
- // Set the "zero" bit in CPSR.
+ // Return false if no candidates exist.
+ if (!MI && !Sub)
+ return false;
+
+ // The single candidate is called MI.
+ if (!MI) MI = Sub;
+
switch (MI->getOpcode()) {
default: break;
case ARM::RSBrr:
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.
+ // Scan forward for the use of CPSR
+ // When checking against MI: if it's a conditional code requires
+ // checking of V bit, then this is not safe to do.
+ // It is safe to remove CmpInstr if CPSR is redefined or killed.
+ // If we are done with the basic block, we need to check whether CPSR is
+ // live-out.
+ SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
+ OperandsToUpdate;
bool isSafe = false;
I = CmpInstr;
- E = MI->getParent()->end();
+ E = CmpInstr->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.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
+ isSafe = true;
+ break;
+ }
if (!MO.isReg() || MO.getReg() != ARM::CPSR)
continue;
if (MO.isDef()) {
}
// 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 (Sub) {
+ ARMCC::CondCodes NewCC = getSwappedCondition(CC);
+ if (NewCC == ARMCC::AL)
+ return false;
+ // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
+ // on CMP needs to be updated to be based on SUB.
+ // Push the condition code operands to OperandsToUpdate.
+ // If it is safe to remove CmpInstr, the condition code of these
+ // operands will be modified.
+ if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
+ Sub->getOperand(2).getReg() == SrcReg)
+ OperandsToUpdate.push_back(std::make_pair(&((*I).getOperand(IO-1)),
+ NewCC));
}
+ else
+ switch (CC) {
+ default:
+ // CPSR can be used multiple times, we should continue.
+ 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;
+ // If CPSR is not killed nor re-defined, we should check whether it is
+ // live-out. If it is live-out, do not optimize.
+ if (!isSafe) {
+ MachineBasicBlock *MBB = CmpInstr->getParent();
+ for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
+ SE = MBB->succ_end(); SI != SE; ++SI)
+ if ((*SI)->isLiveIn(ARM::CPSR))
+ return false;
+ }
// Toggle the optional operand to CPSR.
MI->getOperand(5).setReg(ARM::CPSR);
MI->getOperand(5).setIsDef(true);
CmpInstr->eraseFromParent();
+
+ // Modify the condition code of operands in OperandsToUpdate.
+ // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
+ // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
+ for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++)
+ OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second);
return true;
}
}
if (!MRI->hasOneNonDBGUse(Reg))
return false;
+ const MCInstrDesc &DefMCID = DefMI->getDesc();
+ if (DefMCID.hasOptionalDef()) {
+ unsigned NumOps = DefMCID.getNumOperands();
+ const MachineOperand &MO = DefMI->getOperand(NumOps-1);
+ if (MO.getReg() == ARM::CPSR && !MO.isDead())
+ // If DefMI defines CPSR and it is not dead, it's obviously not safe
+ // to delete DefMI.
+ return false;
+ }
+
+ const MCInstrDesc &UseMCID = UseMI->getDesc();
+ if (UseMCID.hasOptionalDef()) {
+ unsigned NumOps = UseMCID.getNumOperands();
+ if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
+ // If the instruction sets the flag, do not attempt this optimization
+ // since it may change the semantics of the code.
+ return false;
+ }
+
unsigned UseOpc = UseMI->getOpcode();
unsigned NewUseOpc = 0;
uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
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)));
const MCInstrDesc &Desc = MI->getDesc();
unsigned Class = Desc.getSchedClass();
- unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
- if (UOps)
- return UOps;
+ int ItinUOps = ItinData->getNumMicroOps(Class);
+ if (ItinUOps >= 0)
+ return ItinUOps;
unsigned Opc = MI->getOpcode();
switch (Opc) {
default:
llvm_unreachable("Unexpected multi-uops instruction!");
- break;
case ARM::VLDMQIA:
case ARM::VSTMQIA:
return 2;
//
// 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.
+ // separately by assuming 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
return 2;
// 4 registers would be issued: 2, 2.
// 5 registers would be issued: 2, 2, 1.
- UOps = (NumRegs / 2);
+ int A8UOps = (NumRegs / 2);
if (NumRegs % 2)
- ++UOps;
- return UOps;
+ ++A8UOps;
+ return A8UOps;
} else if (Subtarget.isCortexA9()) {
- UOps = (NumRegs / 2);
+ int A9UOps = (NumRegs / 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 ((NumRegs % 2) ||
!MI->hasOneMemOperand() ||
(*MI->memoperands_begin())->getAlignment() < 8)
- ++UOps;
- return UOps;
+ ++A9UOps;
+ return A9UOps;
} else {
// Assume the worst.
return NumRegs;
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;
+static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
+ const MachineInstr *MI, unsigned Reg,
+ unsigned &DefIdx, unsigned &Dist) {
+ Dist = 0;
- const MCInstrDesc &DefMCID = DefMI->getDesc();
- if (!ItinData || ItinData->isEmpty())
- return DefMCID.mayLoad() ? 3 : 1;
+ MachineBasicBlock::const_iterator I = MI; ++I;
+ MachineBasicBlock::const_instr_iterator II =
+ llvm::prior(I.getInstrIterator());
+ assert(II->isInsideBundle() && "Empty bundle?");
- const MCInstrDesc &UseMCID = 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;
- }
+ int Idx = -1;
+ while (II->isInsideBundle()) {
+ Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
+ if (Idx != -1)
+ break;
+ --II;
+ ++Dist;
+ }
- // CPSR set and branch can be paired in the same cycle.
- if (UseMCID.isBranch())
- return 0;
+ 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;
}
- 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);
+ if (Idx == -1) {
+ Dist = 0;
+ return 0;
+ }
- if (Latency > 1 &&
- (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
+ UseIdx = Idx;
+ return II;
+}
+
+/// Return the number of cycles to add to (or subtract from) the static
+/// itinerary based on the def opcode and alignment. The caller will ensure that
+/// adjusted latency is at least one cycle.
+static int adjustDefLatency(const ARMSubtarget &Subtarget,
+ const MachineInstr *DefMI,
+ const MCInstrDesc *DefMCID, unsigned DefAlign) {
+ int Adjust = 0;
+ if (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: {
unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
if (ShImm == 0 ||
(ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
- --Latency;
+ --Adjust;
break;
}
case ARM::t2LDRs:
// Thumb2 mode: lsl only.
unsigned ShAmt = DefMI->getOperand(3).getImm();
if (ShAmt == 0 || ShAmt == 2)
- --Latency;
+ --Adjust;
break;
}
}
}
- if (DefAlign < 8 && Subtarget.isCortexA9())
- switch (DefMCID.getOpcode()) {
+ if (DefAlign < 8 && Subtarget.isCortexA9()) {
+ 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::VLD4LNq32_UPD:
// If the address is not 64-bit aligned, the latencies of these
// instructions increases by one.
- ++Latency;
+ ++Adjust;
break;
}
+ }
+ return Adjust;
+}
+
+
+
+int
+ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
+ const MachineInstr *DefMI, unsigned DefIdx,
+ const MachineInstr *UseMI,
+ unsigned UseIdx) const {
+ // No operand latency. The caller may fall back to getInstrLatency.
+ if (!ItinData || ItinData->isEmpty())
+ return -1;
+
+ const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
+ unsigned Reg = DefMO.getReg();
+ const MCInstrDesc *DefMCID = &DefMI->getDesc();
+ const MCInstrDesc *UseMCID = &UseMI->getDesc();
+ unsigned DefAdj = 0;
+ if (DefMI->isBundle()) {
+ DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
+ DefMCID = &DefMI->getDesc();
+ }
+ if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
+ DefMI->isRegSequence() || DefMI->isImplicitDef()) {
+ return 1;
+ }
+
+ unsigned UseAdj = 0;
+ if (UseMI->isBundle()) {
+ unsigned NewUseIdx;
+ const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
+ Reg, NewUseIdx, UseAdj);
+ if (!NewUseMI)
+ return -1;
+
+ UseMI = NewUseMI;
+ UseIdx = NewUseIdx;
+ UseMCID = &UseMI->getDesc();
+ }
+
+ 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 (UseMI->isBranch())
+ return 0;
+
+ // Otherwise it takes the instruction latency (generally one).
+ unsigned 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;
+ }
+
+ if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
+ return -1;
+
+ unsigned DefAlign = DefMI->hasOneMemOperand()
+ ? (*DefMI->memoperands_begin())->getAlignment() : 0;
+ unsigned UseAlign = UseMI->hasOneMemOperand()
+ ? (*UseMI->memoperands_begin())->getAlignment() : 0;
+
+ // Get the itinerary's latency if possible, and handle variable_ops.
+ int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
+ *UseMCID, UseIdx, UseAlign);
+ // Unable to find operand latency. The caller may resort to getInstrLatency.
+ if (Latency < 0)
+ return Latency;
+
+ // Adjust for IT block position.
+ int Adj = DefAdj + UseAdj;
+
+ // Adjust for dynamic def-side opcode variants not captured by the itinerary.
+ Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
+ if (Adj >= 0 || (int)Latency > -Adj) {
+ return Latency + Adj;
+ }
+ // Return the itinerary latency, which may be zero but not less than zero.
return Latency;
}
if (DefAlign < 8 && Subtarget.isCortexA9())
switch (DefMCID.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::VLD1q8:
+ case ARM::VLD1q16:
+ case ARM::VLD1q32:
+ case ARM::VLD1q64:
+ case ARM::VLD1q8wb_register:
+ case ARM::VLD1q16wb_register:
+ case ARM::VLD1q32wb_register:
+ case ARM::VLD1q64wb_register:
+ case ARM::VLD1q8wb_fixed:
+ case ARM::VLD1q16wb_fixed:
+ case ARM::VLD1q32wb_fixed:
+ case ARM::VLD1q64wb_fixed:
+ case ARM::VLD2d8:
+ case ARM::VLD2d16:
+ case ARM::VLD2d32:
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::VLD2d8wb_fixed:
+ case ARM::VLD2d16wb_fixed:
+ case ARM::VLD2d32wb_fixed:
+ case ARM::VLD2q8PseudoWB_fixed:
+ case ARM::VLD2q16PseudoWB_fixed:
+ case ARM::VLD2q32PseudoWB_fixed:
+ case ARM::VLD2d8wb_register:
+ case ARM::VLD2d16wb_register:
+ case ARM::VLD2d32wb_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::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::VLD1DUPq8:
+ case ARM::VLD1DUPq16:
+ case ARM::VLD1DUPq32:
+ 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::VLD2DUPd8wb_fixed:
+ case ARM::VLD2DUPd16wb_fixed:
+ case ARM::VLD2DUPd32wb_fixed:
+ case ARM::VLD2DUPd8wb_register:
+ case ARM::VLD2DUPd16wb_register:
+ case ARM::VLD2DUPd32wb_register:
case ARM::VLD4DUPd8Pseudo:
case ARM::VLD4DUPd16Pseudo:
case ARM::VLD4DUPd32Pseudo:
return Latency;
}
-int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
- const MachineInstr *MI,
- unsigned *PredCost) const {
+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 getInstrLatency(ItinData, DefMI);
+}
+
+unsigned 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;
+ // An instruction scheduler typically runs on unbundled instructions, however
+ // other passes may query the latency of a bundled instruction.
+ if (MI->isBundle()) {
+ unsigned 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;
- if (UOps)
- return ItinData->getStageLatency(Class);
- return getNumMicroOps(ItinData, MI);
+ }
+ // Be sure to call getStageLatency for an empty itinerary in case it has a
+ // valid MinLatency property.
+ if (!ItinData)
+ return MI->mayLoad() ? 3 : 1;
+
+ unsigned Class = MCID.getSchedClass();
+
+ // For instructions with variable uops, use uops as latency.
+ if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0)
+ return getNumMicroOps(ItinData, MI);
+
+ // For the common case, fall back on the itinerary's latency.
+ unsigned Latency = ItinData->getStageLatency(Class);
+
+ // Adjust for dynamic def-side opcode variants not captured by the itinerary.
+ unsigned DefAlign = MI->hasOneMemOperand()
+ ? (*MI->memoperands_begin())->getAlignment() : 0;
+ int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
+ if (Adj >= 0 || (int)Latency > -Adj) {
+ return Latency + Adj;
+ }
+ return Latency;
}
int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
return true;
// Hoist VFP / NEON instructions with 4 or higher latency.
- int Latency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
+ int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx,
+ /*FindMin=*/false);
+ if (Latency < 0)
+ Latency = getInstrLatency(ItinData, DefMI);
if (Latency <= 3)
return false;
return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
//
// We use the following execution domain numbering:
//
-// 0: Generic
-// 1: VFP
-// 2: NEON
+enum ARMExeDomain {
+ ExeGeneric = 0,
+ ExeVFP = 1,
+ ExeNEON = 2
+};
//
// Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
//
// 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(1, 3);
+ 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(2, 0);
+ 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(2, 0);
+ return std::make_pair(ExeNEON, 0);
if (Domain & ARMII::DomainVFP)
- return std::make_pair(1, 0);
+ return std::make_pair(ExeVFP, 0);
- return std::make_pair(0, 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 != 2)
+ 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));
- MachineInstrBuilder(MI).addReg(MI->getOperand(1).getReg());
+
+ // Add the extra source operand and new predicates.
+ // This will go before any implicit ops.
+ AddDefaultPred(MachineInstrBuilder(MI).addOperand(MI->getOperand(1)));
+}
+
+bool ARMBaseInstrInfo::hasNOP() const {
+ return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
}
projects / oota-llvm.git / blobdiff