//
//===----------------------------------------------------------------------===//
-#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
-#include "ARMGenInstrInfo.inc"
-
using namespace llvm;
+#define DEBUG_TYPE "arm-instrinfo"
+
+#define GET_INSTRINFO_CTOR_DTOR
+#include "ARMGenInstrInfo.inc"
+
static cl::opt<bool>
EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
cl::desc("Enable ARM 2-addr to 3-addr conv"));
CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
const ScheduleDAG *DAG) const {
if (Subtarget.isThumb2() || Subtarget.hasVFP2())
- return (ScheduleHazardRecognizer *)
- new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget, DAG);
+ return (ScheduleHazardRecognizer *)new ARMHazardRecognizer(II, DAG);
return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
}
MachineBasicBlock *&FBB,
SmallVectorImpl<MachineOperand> &Cond,
bool AllowModify) const {
- // If the block has no terminators, it just falls into the block after it.
+ TBB = 0;
+ FBB = 0;
+
MachineBasicBlock::iterator I = MBB.end();
if (I == MBB.begin())
- return false;
+ return false; // Empty blocks are easy.
--I;
- while (I->isDebugValue()) {
- if (I == MBB.begin())
- return false;
- --I;
- }
- if (!isUnpredicatedTerminator(I))
- return false;
- // Get the last instruction in the block.
- MachineInstr *LastInst = I;
+ // Walk backwards from the end of the basic block until the branch is
+ // analyzed or we give up.
+ while (isPredicated(I) || I->isTerminator() || I->isDebugValue()) {
- // If there is only one terminator instruction, process it.
- unsigned LastOpc = LastInst->getOpcode();
- if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
- if (isUncondBranchOpcode(LastOpc)) {
- TBB = LastInst->getOperand(0).getMBB();
- return false;
+ // 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
+ // out.
+ bool CantAnalyze = false;
+
+ // Skip over DEBUG values and predicated nonterminators.
+ while (I->isDebugValue() || !I->isTerminator()) {
+ if (I == MBB.begin())
+ return false;
+ --I;
}
- if (isCondBranchOpcode(LastOpc)) {
- // Block ends with fall-through condbranch.
- TBB = LastInst->getOperand(0).getMBB();
- Cond.push_back(LastInst->getOperand(1));
- Cond.push_back(LastInst->getOperand(2));
- return false;
+
+ if (isIndirectBranchOpcode(I->getOpcode()) ||
+ isJumpTableBranchOpcode(I->getOpcode())) {
+ // Indirect branches and jump tables can't be analyzed, but we still want
+ // to clean up any instructions at the tail of the basic block.
+ CantAnalyze = true;
+ } else if (isUncondBranchOpcode(I->getOpcode())) {
+ TBB = I->getOperand(0).getMBB();
+ } else if (isCondBranchOpcode(I->getOpcode())) {
+ // Bail out if we encounter multiple conditional branches.
+ if (!Cond.empty())
+ return true;
+
+ assert(!FBB && "FBB should have been null.");
+ FBB = TBB;
+ TBB = I->getOperand(0).getMBB();
+ Cond.push_back(I->getOperand(1));
+ Cond.push_back(I->getOperand(2));
+ } else if (I->isReturn()) {
+ // Returns can't be analyzed, but we should run cleanup.
+ CantAnalyze = !isPredicated(I);
+ } else {
+ // We encountered other unrecognized terminator. Bail out immediately.
+ return true;
}
- return true; // Can't handle indirect branch.
- }
- // Get the instruction before it if it is a terminator.
- MachineInstr *SecondLastInst = I;
- unsigned SecondLastOpc = SecondLastInst->getOpcode();
-
- // If AllowModify is true and the block ends with two or more unconditional
- // branches, delete all but the first unconditional branch.
- if (AllowModify && isUncondBranchOpcode(LastOpc)) {
- while (isUncondBranchOpcode(SecondLastOpc)) {
- LastInst->eraseFromParent();
- LastInst = SecondLastInst;
- LastOpc = LastInst->getOpcode();
- if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
- // Return now the only terminator is an unconditional branch.
- TBB = LastInst->getOperand(0).getMBB();
- return false;
- } else {
- SecondLastInst = I;
- SecondLastOpc = SecondLastInst->getOpcode();
+ // Cleanup code - to be run for unpredicated unconditional branches and
+ // returns.
+ if (!isPredicated(I) &&
+ (isUncondBranchOpcode(I->getOpcode()) ||
+ isIndirectBranchOpcode(I->getOpcode()) ||
+ isJumpTableBranchOpcode(I->getOpcode()) ||
+ I->isReturn())) {
+ // Forget any previous condition branch information - it no longer applies.
+ Cond.clear();
+ FBB = 0;
+
+ // If we can modify the function, delete everything below this
+ // unconditional branch.
+ if (AllowModify) {
+ MachineBasicBlock::iterator DI = std::next(I);
+ while (DI != MBB.end()) {
+ MachineInstr *InstToDelete = DI;
+ ++DI;
+ InstToDelete->eraseFromParent();
+ }
}
}
- }
-
- // If there are three terminators, we don't know what sort of block this is.
- if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
- return true;
- // If the block ends with a B and a Bcc, handle it.
- if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
- TBB = SecondLastInst->getOperand(0).getMBB();
- Cond.push_back(SecondLastInst->getOperand(1));
- Cond.push_back(SecondLastInst->getOperand(2));
- FBB = LastInst->getOperand(0).getMBB();
- return false;
- }
+ if (CantAnalyze)
+ return true;
- // If the block ends with two unconditional branches, handle it. The second
- // one is not executed, so remove it.
- if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
- TBB = SecondLastInst->getOperand(0).getMBB();
- I = LastInst;
- if (AllowModify)
- I->eraseFromParent();
- return false;
- }
+ if (I == MBB.begin())
+ return false;
- // ...likewise if it ends with a branch table followed by an unconditional
- // branch. The branch folder can create these, and we must get rid of them for
- // correctness of Thumb constant islands.
- if ((isJumpTableBranchOpcode(SecondLastOpc) ||
- isIndirectBranchOpcode(SecondLastOpc)) &&
- isUncondBranchOpcode(LastOpc)) {
- I = LastInst;
- if (AllowModify)
- I->eraseFromParent();
- return true;
+ --I;
}
- // Otherwise, can't handle this.
- return true;
+ // We made it past the terminators without bailing out - we must have
+ // analyzed this branch successfully.
+ return false;
}
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;
+}
+
+namespace llvm {
+template <> bool IsCPSRDead<MachineInstr>(MachineInstr *MI) {
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg() || MO.isUndef() || MO.isUse())
+ continue;
+ if (MO.getReg() != ARM::CPSR)
+ continue;
+ if (!MO.isDead())
+ return false;
}
+ // all definitions of CPSR are dead
return true;
}
+}
/// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
LLVM_ATTRIBUTE_NOINLINE
// 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:
+ default:
+ // pseudo-instruction sizes are zero.
return 0;
case TargetOpcode::BUNDLE:
return getInstBundleLength(MI);
++NumEntries;
return NumEntries * EntrySize + InstSize;
}
- default:
- // Otherwise, pseudo-instruction sizes are zero.
- return 0;
}
}
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);
Mov = AddDefaultPred(Mov);
+ // MOVr can set CC.
+ if (Opc == ARM::MOVr)
+ Mov = AddDefaultCC(Mov);
}
// Add implicit super-register defs and kills to the last instruction.
Mov->addRegisterDefined(DestReg, TRI);
return true;
}
-MachineInstr*
-ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
- int FrameIx, uint64_t Offset,
- const MDNode *MDPtr,
- DebugLoc DL) const {
- MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
- .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
- return &*MIB;
-}
-
/// Create a copy of a const pool value. Update CPI to the new index and return
/// the label UID.
static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
unsigned PCLabelId = AFI->createPICLabelUId();
ARMConstantPoolValue *NewCPV = 0;
+
// FIXME: The below assumes PIC relocation model and that the function
// is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
// zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
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.
return false;
// Terminators and labels can't be scheduled around.
- if (MI->isTerminator() || MI->isLabel())
+ if (MI->isTerminator() || MI->isPosition())
return true;
// Treat the start of the IT block as a scheduling boundary, but schedule
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;
}
}
+static bool isAnySubRegLive(unsigned Reg, const TargetRegisterInfo *TRI,
+ MachineInstr *MI) {
+ for (MCSubRegIterator Subreg(Reg, TRI, /* IncludeSelf */ true);
+ Subreg.isValid(); ++Subreg)
+ if (MI->getParent()->computeRegisterLiveness(TRI, *Subreg, MI) !=
+ MachineBasicBlock::LQR_Dead)
+ return true;
+ return false;
+}
+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));
+
+ 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.
+ // TODO: Currently, computeRegisterLiveness() does not report "live" if a
+ // sub reg is live. When computeRegisterLiveness() works for sub reg, it
+ // can replace isAnySubRegLive().
+ if (isCalleeSavedRegister(CurReg, CSRegs) ||
+ isAnySubRegLive(CurReg, TRI, MI)) {
+ // 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) {
// Walk down one instruction which is potentially an 'and'.
const MachineInstr &Copy = *MI;
MachineBasicBlock::iterator AND(
- llvm::next(MachineBasicBlock::iterator(MI)));
+ std::next(MachineBasicBlock::iterator(MI)));
if (AND == MI->getParent()->end()) return false;
MI = AND;
return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
if (CmpMask != ~0) {
if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) {
MI = 0;
- for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
- UE = MRI->use_end(); UI != UE; ++UI) {
+ for (MachineRegisterInfo::use_instr_iterator
+ UI = MRI->use_instr_begin(SrcReg), UE = MRI->use_instr_end();
+ UI != UE; ++UI) {
if (UI->getParent() != CmpInstr->getParent()) continue;
MachineInstr *PotentialAND = &*UI;
if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) ||
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.
break;
}
return UOps;
- } else if (Subtarget.isCortexA8()) {
+ } else if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
if (NumRegs < 4)
return 2;
// 4 registers would be issued: 2, 2.
return ItinData->getOperandCycle(DefClass, DefIdx);
int DefCycle;
- if (Subtarget.isCortexA8()) {
+ if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
// (regno / 2) + (regno % 2) + 1
DefCycle = RegNo / 2 + 1;
if (RegNo % 2)
return ItinData->getOperandCycle(DefClass, DefIdx);
int DefCycle;
- if (Subtarget.isCortexA8()) {
+ if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
// 4 registers would be issued: 1, 2, 1.
// 5 registers would be issued: 1, 2, 2.
DefCycle = RegNo / 2;
return ItinData->getOperandCycle(UseClass, UseIdx);
int UseCycle;
- if (Subtarget.isCortexA8()) {
+ if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
// (regno / 2) + (regno % 2) + 1
UseCycle = RegNo / 2 + 1;
if (RegNo % 2)
return ItinData->getOperandCycle(UseClass, UseIdx);
int UseCycle;
- if (Subtarget.isCortexA8()) {
+ if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
UseCycle = RegNo / 2;
if (UseCycle < 2)
UseCycle = 2;
Dist = 0;
MachineBasicBlock::const_iterator I = MI; ++I;
- MachineBasicBlock::const_instr_iterator II =
- llvm::prior(I.getInstrIterator());
+ MachineBasicBlock::const_instr_iterator II = std::prev(I.getInstrIterator());
assert(II->isInsideBundle() && "Empty bundle?");
int Idx = -1;
const MachineInstr *DefMI,
const MCInstrDesc *DefMCID, unsigned DefAlign) {
int Adjust = 0;
- if (Subtarget.isCortexA8() || Subtarget.isLikeA9()) {
+ if (Subtarget.isCortexA8() || Subtarget.isLikeA9() || Subtarget.isCortexA7()) {
// FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
// variants are one cycle cheaper.
switch (DefMCID->getOpcode()) {
UseMCID, UseIdx, UseAlign);
if (Latency > 1 &&
- (Subtarget.isCortexA8() || Subtarget.isLikeA9())) {
+ (Subtarget.isCortexA8() || Subtarget.isLikeA9() ||
+ Subtarget.isCortexA7())) {
// FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
// variants are one cycle cheaper.
switch (DefMCID.getOpcode()) {
case ARM::VLD3d16Pseudo:
case ARM::VLD3d32Pseudo:
case ARM::VLD1d64TPseudo:
+ case ARM::VLD1d64TPseudoWB_fixed:
case ARM::VLD3d8Pseudo_UPD:
case ARM::VLD3d16Pseudo_UPD:
case ARM::VLD3d32Pseudo_UPD:
case ARM::VLD4d16Pseudo:
case ARM::VLD4d32Pseudo:
case ARM::VLD1d64QPseudo:
+ case ARM::VLD1d64QPseudoWB_fixed:
case ARM::VLD4d8Pseudo_UPD:
case ARM::VLD4d16Pseudo_UPD:
case ARM::VLD4d32Pseudo_UPD:
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 {
return true;
// Hoist VFP / NEON instructions with 4 or higher latency.
- int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx,
- /*FindMin=*/false);
+ int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
if (Latency < 0)
Latency = getInstrLatency(ItinData, DefMI);
if (Latency <= 3)
// 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.
}
bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
+ if (MI->getNumOperands() < 4)
+ return true;
unsigned ShOpVal = MI->getOperand(3).getImm();
unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
// Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
projects / oota-llvm.git / blobdiff