//
//===----------------------------------------------------------------------===//
-#include "ARMBaseInstrInfo.h"
#include "ARM.h"
+#include "ARMBaseInstrInfo.h"
#include "ARMBaseRegisterInfo.h"
#include "ARMConstantPoolValue.h"
+#include "ARMFeatures.h"
#include "ARMHazardRecognizer.h"
#include "ARMMachineFunctionInfo.h"
#include "MCTargetDesc/ARMAddressingModes.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
-#define GET_INSTRINFO_CTOR
+#define GET_INSTRINFO_CTOR_DTOR
#include "ARMGenInstrInfo.inc"
using namespace llvm;
// Walk backwards from the end of the basic block until the branch is
// analyzed or we give up.
- while (isPredicated(I) || I->isTerminator()) {
+ while (isPredicated(I) || I->isTerminator() || I->isDebugValue()) {
// Flag to be raised on unanalyzeable instructions. This is useful in cases
// where we want to clean up on the end of the basic block before we bail
if (!MI->isPredicable())
return false;
- if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
- ARMFunctionInfo *AFI =
- MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
- return AFI->isThumb2Function();
+ ARMFunctionInfo *AFI =
+ MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
+
+ if (AFI->isThumb2Function()) {
+ if (getSubtarget().restrictIT())
+ return isV8EligibleForIT(MI);
+ } else { // non-Thumb
+ if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON)
+ return false;
}
+
return true;
}
unsigned DestReg, unsigned SrcReg,
bool KillSrc) const {
bool GPRDest = ARM::GPRRegClass.contains(DestReg);
- bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
+ bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
if (GPRDest && GPRSrc) {
AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
- .addReg(SrcReg, getKillRegState(KillSrc))));
+ .addReg(SrcReg, getKillRegState(KillSrc))));
return;
}
bool SPRDest = ARM::SPRRegClass.contains(DestReg);
- bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
+ bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
unsigned Opc = 0;
if (SPRDest && SPRSrc)
int 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;
+ 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::GPRPairRegClass.contains(DestReg, SrcReg))
- Opc = ARM::MOVr, BeginIdx = ARM::gsub_0, SubRegs = 2;
-
- 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;
+ } 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::GPRPairRegClass.contains(DestReg, SrcReg)) {
+ Opc = Subtarget.isThumb2() ? ARM::tMOVr : ARM::MOVr;
+ BeginIdx = ARM::gsub_0;
+ SubRegs = 2;
+ } 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;
+ }
assert(Opc && "Impossible reg-to-reg copy");
// Copy register tuples backward when the first Dest reg overlaps with SrcReg.
if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) {
- BeginIdx = BeginIdx + ((SubRegs-1)*Spacing);
+ BeginIdx = BeginIdx + ((SubRegs - 1) * Spacing);
Spacing = -Spacing;
}
#ifndef NDEBUG
SmallSet<unsigned, 4> DstRegs;
#endif
for (unsigned i = 0; i != SubRegs; ++i) {
- unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i*Spacing);
- unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i*Spacing);
+ unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i * Spacing);
+ unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i * Spacing);
assert(Dst && Src && "Bad sub-register");
#ifndef NDEBUG
assert(!DstRegs.count(Src) && "destructive vector copy");
DstRegs.insert(Dst);
#endif
- Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst)
- .addReg(Src);
+ Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst).addReg(Src);
// VORR takes two source operands.
if (Opc == ARM::VORRq)
Mov.addReg(Src);
Opcode == ARM::t2LDRpci_pic ||
Opcode == ARM::tLDRpci ||
Opcode == ARM::tLDRpci_pic ||
- Opcode == ARM::MOV_ga_dyn ||
+ Opcode == ARM::LDRLIT_ga_pcrel ||
+ Opcode == ARM::LDRLIT_ga_pcrel_ldr ||
+ Opcode == ARM::tLDRLIT_ga_pcrel ||
Opcode == ARM::MOV_ga_pcrel ||
Opcode == ARM::MOV_ga_pcrel_ldr ||
- Opcode == ARM::t2MOV_ga_dyn ||
Opcode == ARM::t2MOV_ga_pcrel) {
if (MI1->getOpcode() != Opcode)
return false;
if (MO0.getOffset() != MO1.getOffset())
return false;
- if (Opcode == ARM::MOV_ga_dyn ||
+ if (Opcode == ARM::LDRLIT_ga_pcrel ||
+ Opcode == ARM::LDRLIT_ga_pcrel_ldr ||
+ Opcode == ARM::tLDRLIT_ga_pcrel ||
Opcode == ARM::MOV_ga_pcrel ||
Opcode == ARM::MOV_ga_pcrel_ldr ||
- Opcode == ARM::t2MOV_ga_dyn ||
Opcode == ARM::t2MOV_ga_pcrel)
// Ignore the PC labels.
return MO0.getGlobal() == MO1.getGlobal();
case ARM::VLDRD:
case ARM::VLDRS:
case ARM::t2LDRi8:
+ case ARM::t2LDRBi8:
case ARM::t2LDRDi8:
case ARM::t2LDRSHi8:
case ARM::t2LDRi12:
+ case ARM::t2LDRBi12:
case ARM::t2LDRSHi12:
break;
}
case ARM::VLDRD:
case ARM::VLDRS:
case ARM::t2LDRi8:
+ case ARM::t2LDRBi8:
case ARM::t2LDRSHi8:
case ARM::t2LDRi12:
+ case ARM::t2LDRBi12:
case ARM::t2LDRSHi12:
break;
}
if ((Offset2 - Offset1) / 8 > 64)
return false;
- if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
+ // Check if the machine opcodes are different. If they are different
+ // then we consider them to not be of the same base address,
+ // EXCEPT in the case of Thumb2 byte loads where one is LDRBi8 and the other LDRBi12.
+ // In this case, they are considered to be the same because they are different
+ // encoding forms of the same basic instruction.
+ if ((Load1->getMachineOpcode() != Load2->getMachineOpcode()) &&
+ !((Load1->getMachineOpcode() == ARM::t2LDRBi8 &&
+ Load2->getMachineOpcode() == ARM::t2LDRBi12) ||
+ (Load1->getMachineOpcode() == ARM::t2LDRBi12 &&
+ Load2->getMachineOpcode() == ARM::t2LDRBi8)))
return false; // FIXME: overly conservative?
// Four loads in a row should be sufficient.
bool PreferFalse) const {
assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
"Unknown select instruction");
- const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
+ MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
bool Invert = !DefMI;
if (!DefMI)
if (!DefMI)
return 0;
+ // Find new register class to use.
+ MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
+ unsigned DestReg = MI->getOperand(0).getReg();
+ const TargetRegisterClass *PreviousClass = MRI.getRegClass(FalseReg.getReg());
+ if (!MRI.constrainRegClass(DestReg, PreviousClass))
+ return 0;
+
// Create a new predicated version of DefMI.
// Rfalse is the first use.
MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
- DefMI->getDesc(),
- MI->getOperand(0).getReg());
+ DefMI->getDesc(), DestReg);
// Copy all the DefMI operands, excluding its (null) predicate.
const MCInstrDesc &DefDesc = DefMI->getDesc();
// register operand tied to the first def.
// The tie makes the register allocator ensure the FalseReg is allocated the
// same register as operand 0.
- MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
FalseReg.setImplicit();
NewMI.addOperand(FalseReg);
NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
unsigned DestReg, unsigned BaseReg, int NumBytes,
ARMCC::CondCodes Pred, unsigned PredReg,
const ARMBaseInstrInfo &TII, unsigned MIFlags) {
+ if (NumBytes == 0 && DestReg != BaseReg) {
+ BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), DestReg)
+ .addReg(BaseReg, RegState::Kill)
+ .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
+ .setMIFlags(MIFlags);
+ return;
+ }
+
bool isSub = NumBytes < 0;
if (isSub) NumBytes = -NumBytes;
}
}
+bool llvm::tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget,
+ MachineFunction &MF, MachineInstr *MI,
+ unsigned NumBytes) {
+ // This optimisation potentially adds lots of load and store
+ // micro-operations, it's only really a great benefit to code-size.
+ if (!Subtarget.isMinSize())
+ return false;
+
+ // If only one register is pushed/popped, LLVM can use an LDR/STR
+ // instead. We can't modify those so make sure we're dealing with an
+ // instruction we understand.
+ bool IsPop = isPopOpcode(MI->getOpcode());
+ bool IsPush = isPushOpcode(MI->getOpcode());
+ if (!IsPush && !IsPop)
+ return false;
+
+ bool IsVFPPushPop = MI->getOpcode() == ARM::VSTMDDB_UPD ||
+ MI->getOpcode() == ARM::VLDMDIA_UPD;
+ bool IsT1PushPop = MI->getOpcode() == ARM::tPUSH ||
+ MI->getOpcode() == ARM::tPOP ||
+ MI->getOpcode() == ARM::tPOP_RET;
+
+ assert((IsT1PushPop || (MI->getOperand(0).getReg() == ARM::SP &&
+ MI->getOperand(1).getReg() == ARM::SP)) &&
+ "trying to fold sp update into non-sp-updating push/pop");
+
+ // The VFP push & pop act on D-registers, so we can only fold an adjustment
+ // by a multiple of 8 bytes in correctly. Similarly rN is 4-bytes. Don't try
+ // if this is violated.
+ if (NumBytes % (IsVFPPushPop ? 8 : 4) != 0)
+ return false;
+
+ // ARM and Thumb2 push/pop insts have explicit "sp, sp" operands (+
+ // pred) so the list starts at 4. Thumb1 starts after the predicate.
+ int RegListIdx = IsT1PushPop ? 2 : 4;
+
+ // Calculate the space we'll need in terms of registers.
+ unsigned FirstReg = MI->getOperand(RegListIdx).getReg();
+ unsigned RD0Reg, RegsNeeded;
+ if (IsVFPPushPop) {
+ RD0Reg = ARM::D0;
+ RegsNeeded = NumBytes / 8;
+ } else {
+ RD0Reg = ARM::R0;
+ RegsNeeded = NumBytes / 4;
+ }
+
+ // We're going to have to strip all list operands off before
+ // re-adding them since the order matters, so save the existing ones
+ // for later.
+ SmallVector<MachineOperand, 4> RegList;
+ for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i)
+ RegList.push_back(MI->getOperand(i));
+
+ MachineBasicBlock *MBB = MI->getParent();
+ const TargetRegisterInfo *TRI = MF.getRegInfo().getTargetRegisterInfo();
+ const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(&MF);
+
+ // Now try to find enough space in the reglist to allocate NumBytes.
+ for (unsigned CurReg = FirstReg - 1; CurReg >= RD0Reg && RegsNeeded;
+ --CurReg) {
+ if (!IsPop) {
+ // Pushing any register is completely harmless, mark the
+ // register involved as undef since we don't care about it in
+ // the slightest.
+ RegList.push_back(MachineOperand::CreateReg(CurReg, false, false,
+ false, false, true));
+ --RegsNeeded;
+ continue;
+ }
+
+ // However, we can only pop an extra register if it's not live. For
+ // registers live within the function we might clobber a return value
+ // register; the other way a register can be live here is if it's
+ // callee-saved.
+ if (isCalleeSavedRegister(CurReg, CSRegs) ||
+ MBB->computeRegisterLiveness(TRI, CurReg, MI) !=
+ MachineBasicBlock::LQR_Dead) {
+ // VFP pops don't allow holes in the register list, so any skip is fatal
+ // for our transformation. GPR pops do, so we should just keep looking.
+ if (IsVFPPushPop)
+ return false;
+ else
+ continue;
+ }
+
+ // Mark the unimportant registers as <def,dead> in the POP.
+ RegList.push_back(MachineOperand::CreateReg(CurReg, true, false, false,
+ true));
+ --RegsNeeded;
+ }
+
+ if (RegsNeeded > 0)
+ return false;
+
+ // Finally we know we can profitably perform the optimisation so go
+ // ahead: strip all existing registers off and add them back again
+ // in the right order.
+ for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i)
+ MI->RemoveOperand(i);
+
+ // Add the complete list back in.
+ MachineInstrBuilder MIB(MF, &*MI);
+ for (int i = RegList.size() - 1; i >= 0; --i)
+ MIB.addOperand(RegList[i]);
+
+ return true;
+}
+
bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
unsigned FrameReg, int &Offset,
const ARMBaseInstrInfo &TII) {
isSafe = true;
break;
}
- // Condition code is after the operand before CPSR.
- ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
+ // Condition code is after the operand before CPSR except for VSELs.
+ ARMCC::CondCodes CC;
+ bool IsInstrVSel = true;
+ switch (Instr.getOpcode()) {
+ default:
+ IsInstrVSel = false;
+ CC = (ARMCC::CondCodes)Instr.getOperand(IO - 1).getImm();
+ break;
+ case ARM::VSELEQD:
+ case ARM::VSELEQS:
+ CC = ARMCC::EQ;
+ break;
+ case ARM::VSELGTD:
+ case ARM::VSELGTS:
+ CC = ARMCC::GT;
+ break;
+ case ARM::VSELGED:
+ case ARM::VSELGES:
+ CC = ARMCC::GE;
+ break;
+ case ARM::VSELVSS:
+ case ARM::VSELVSD:
+ CC = ARMCC::VS;
+ break;
+ }
+
if (Sub) {
ARMCC::CondCodes NewCC = getSwappedCondition(CC);
if (NewCC == ARMCC::AL)
// 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
+ Sub->getOperand(2).getReg() == SrcReg) {
+ // VSel doesn't support condition code update.
+ if (IsInstrVSel)
+ return false;
+ OperandsToUpdate.push_back(
+ std::make_pair(&((*I).getOperand(IO - 1)), NewCC));
+ }
+ } else
switch (CC) {
default:
// CPSR can be used multiple times, we should continue.
return Latency;
}
+unsigned ARMBaseInstrInfo::getPredicationCost(const MachineInstr *MI) const {
+ if (MI->isCopyLike() || MI->isInsertSubreg() ||
+ MI->isRegSequence() || MI->isImplicitDef())
+ return 0;
+
+ if (MI->isBundle())
+ return 0;
+
+ const MCInstrDesc &MCID = MI->getDesc();
+
+ if (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR)) {
+ // When predicated, CPSR is an additional source operand for CPSR updating
+ // instructions, this apparently increases their latencies.
+ return 1;
+ }
+ return 0;
+}
+
unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
const MachineInstr *MI,
unsigned *PredCost) const {
// FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines
// the full D-register by loading the same value to both lanes. The
// instruction is micro-coded with 2 uops, so don't do this until we can
- // properly schedule micro-coded instuctions. The dispatcher stalls cause
+ // properly schedule micro-coded instructions. The dispatcher stalls cause
// too big regressions.
// Insert the dependency-breaking FCONSTD before MI.