#define DEBUG_TYPE "arm-ldst-opt"
#include "ARM.h"
-#include "ARMAddressingModes.h"
#include "ARMBaseInstrInfo.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMRegisterInfo.h"
+#include "MCTargetDesc/ARMAddressingModes.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/RegisterScavenging.h"
+#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
namespace {
struct ARMLoadStoreOpt : public MachineFunctionPass {
static char ID;
- ARMLoadStoreOpt() : MachineFunctionPass(&ID) {}
+ ARMLoadStoreOpt() : MachineFunctionPass(ID) {}
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
private:
struct MemOpQueueEntry {
int Offset;
+ unsigned Reg;
+ bool isKill;
unsigned Position;
MachineBasicBlock::iterator MBBI;
bool Merged;
- MemOpQueueEntry(int o, int p, MachineBasicBlock::iterator i)
- : Offset(o), Position(p), MBBI(i), Merged(false) {}
+ MemOpQueueEntry(int o, unsigned r, bool k, unsigned p,
+ MachineBasicBlock::iterator i)
+ : Offset(o), Reg(r), isKill(k), Position(p), MBBI(i), Merged(false) {}
};
typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
typedef MemOpQueue::iterator MemOpQueueIter;
int Offset, unsigned Base, bool BaseKill, int Opcode,
ARMCC::CondCodes Pred, unsigned PredReg, unsigned Scratch,
DebugLoc dl, SmallVector<std::pair<unsigned, bool>, 8> &Regs);
+ void MergeOpsUpdate(MachineBasicBlock &MBB,
+ MemOpQueue &MemOps,
+ unsigned memOpsBegin,
+ unsigned memOpsEnd,
+ unsigned insertAfter,
+ int Offset,
+ unsigned Base,
+ bool BaseKill,
+ int Opcode,
+ ARMCC::CondCodes Pred,
+ unsigned PredReg,
+ unsigned Scratch,
+ DebugLoc dl,
+ SmallVector<MachineBasicBlock::iterator, 4> &Merges);
void MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex, unsigned Base,
int Opcode, unsigned Size,
ARMCC::CondCodes Pred, unsigned PredReg,
char ARMLoadStoreOpt::ID = 0;
}
-static int getLoadStoreMultipleOpcode(int Opcode) {
+static int getLoadStoreMultipleOpcode(int Opcode, ARM_AM::AMSubMode Mode) {
switch (Opcode) {
- case ARM::LDR:
- NumLDMGened++;
- return ARM::LDM;
- case ARM::STR:
- NumSTMGened++;
- return ARM::STM;
+ default: llvm_unreachable("Unhandled opcode!");
+ case ARM::LDRi12:
+ ++NumLDMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::LDMIA;
+ case ARM_AM::da: return ARM::LDMDA;
+ case ARM_AM::db: return ARM::LDMDB;
+ case ARM_AM::ib: return ARM::LDMIB;
+ }
+ break;
+ case ARM::STRi12:
+ ++NumSTMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::STMIA;
+ case ARM_AM::da: return ARM::STMDA;
+ case ARM_AM::db: return ARM::STMDB;
+ case ARM_AM::ib: return ARM::STMIB;
+ }
+ break;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
- NumLDMGened++;
- return ARM::t2LDM;
+ ++NumLDMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::t2LDMIA;
+ case ARM_AM::db: return ARM::t2LDMDB;
+ }
+ break;
case ARM::t2STRi8:
case ARM::t2STRi12:
- NumSTMGened++;
- return ARM::t2STM;
+ ++NumSTMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::t2STMIA;
+ case ARM_AM::db: return ARM::t2STMDB;
+ }
+ break;
case ARM::VLDRS:
- NumVLDMGened++;
- return ARM::VLDMS;
+ ++NumVLDMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VLDMSIA;
+ case ARM_AM::db: return 0; // Only VLDMSDB_UPD exists.
+ }
+ break;
case ARM::VSTRS:
- NumVSTMGened++;
- return ARM::VSTMS;
+ ++NumVSTMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VSTMSIA;
+ case ARM_AM::db: return 0; // Only VSTMSDB_UPD exists.
+ }
+ break;
case ARM::VLDRD:
- NumVLDMGened++;
- return ARM::VLDMD;
+ ++NumVLDMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VLDMDIA;
+ case ARM_AM::db: return 0; // Only VLDMDDB_UPD exists.
+ }
+ break;
case ARM::VSTRD:
- NumVSTMGened++;
- return ARM::VSTMD;
- default: llvm_unreachable("Unhandled opcode!");
+ ++NumVSTMGened;
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VSTMDIA;
+ case ARM_AM::db: return 0; // Only VSTMDDB_UPD exists.
+ }
+ break;
}
+
return 0;
}
+namespace llvm {
+ namespace ARM_AM {
+
+AMSubMode getLoadStoreMultipleSubMode(int Opcode) {
+ switch (Opcode) {
+ default: llvm_unreachable("Unhandled opcode!");
+ case ARM::LDMIA_RET:
+ case ARM::LDMIA:
+ case ARM::LDMIA_UPD:
+ case ARM::STMIA:
+ case ARM::STMIA_UPD:
+ case ARM::t2LDMIA_RET:
+ case ARM::t2LDMIA:
+ case ARM::t2LDMIA_UPD:
+ case ARM::t2STMIA:
+ case ARM::t2STMIA_UPD:
+ case ARM::VLDMSIA:
+ case ARM::VLDMSIA_UPD:
+ case ARM::VSTMSIA:
+ case ARM::VSTMSIA_UPD:
+ case ARM::VLDMDIA:
+ case ARM::VLDMDIA_UPD:
+ case ARM::VSTMDIA:
+ case ARM::VSTMDIA_UPD:
+ return ARM_AM::ia;
+
+ case ARM::LDMDA:
+ case ARM::LDMDA_UPD:
+ case ARM::STMDA:
+ case ARM::STMDA_UPD:
+ return ARM_AM::da;
+
+ case ARM::LDMDB:
+ case ARM::LDMDB_UPD:
+ case ARM::STMDB:
+ case ARM::STMDB_UPD:
+ case ARM::t2LDMDB:
+ case ARM::t2LDMDB_UPD:
+ case ARM::t2STMDB:
+ case ARM::t2STMDB_UPD:
+ case ARM::VLDMSDB_UPD:
+ case ARM::VSTMSDB_UPD:
+ case ARM::VLDMDDB_UPD:
+ case ARM::VSTMDDB_UPD:
+ return ARM_AM::db;
+
+ case ARM::LDMIB:
+ case ARM::LDMIB_UPD:
+ case ARM::STMIB:
+ case ARM::STMIB_UPD:
+ return ARM_AM::ib;
+ }
+
+ return ARM_AM::bad_am_submode;
+}
+
+ } // end namespace ARM_AM
+} // end namespace llvm
+
static bool isT2i32Load(unsigned Opc) {
return Opc == ARM::t2LDRi12 || Opc == ARM::t2LDRi8;
}
static bool isi32Load(unsigned Opc) {
- return Opc == ARM::LDR || isT2i32Load(Opc);
+ return Opc == ARM::LDRi12 || isT2i32Load(Opc);
}
static bool isT2i32Store(unsigned Opc) {
}
static bool isi32Store(unsigned Opc) {
- return Opc == ARM::STR || isT2i32Store(Opc);
+ return Opc == ARM::STRi12 || isT2i32Store(Opc);
}
/// MergeOps - Create and insert a LDM or STM with Base as base register and
return false;
ARM_AM::AMSubMode Mode = ARM_AM::ia;
- bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
- if (isAM4 && Offset == 4) {
- if (isThumb2)
- // Thumb2 does not support ldmib / stmib.
- return false;
+ // VFP and Thumb2 do not support IB or DA modes.
+ bool isNotVFP = isi32Load(Opcode) || isi32Store(Opcode);
+ bool haveIBAndDA = isNotVFP && !isThumb2;
+ if (Offset == 4 && haveIBAndDA)
Mode = ARM_AM::ib;
- } else if (isAM4 && Offset == -4 * (int)NumRegs + 4) {
- if (isThumb2)
- // Thumb2 does not support ldmda / stmda.
- return false;
+ else if (Offset == -4 * (int)NumRegs + 4 && haveIBAndDA)
Mode = ARM_AM::da;
- } else if (isAM4 && Offset == -4 * (int)NumRegs) {
+ else if (Offset == -4 * (int)NumRegs && isNotVFP)
+ // VLDM/VSTM do not support DB mode without also updating the base reg.
Mode = ARM_AM::db;
- } else if (Offset != 0) {
+ else if (Offset != 0) {
+ // Check if this is a supported opcode before we insert instructions to
+ // calculate a new base register.
+ if (!getLoadStoreMultipleOpcode(Opcode, Mode)) return false;
+
// If starting offset isn't zero, insert a MI to materialize a new base.
// But only do so if it is cost effective, i.e. merging more than two
// loads / stores.
if (NewBase == 0)
return false;
}
- int BaseOpc = !isThumb2
- ? ARM::ADDri
- : ((Base == ARM::SP) ? ARM::t2ADDrSPi : ARM::t2ADDri);
+ int BaseOpc = !isThumb2 ? ARM::ADDri : ARM::t2ADDri;
if (Offset < 0) {
- BaseOpc = !isThumb2
- ? ARM::SUBri
- : ((Base == ARM::SP) ? ARM::t2SUBrSPi : ARM::t2SUBri);
+ BaseOpc = !isThumb2 ? ARM::SUBri : ARM::t2SUBri;
Offset = - Offset;
}
int ImmedOffset = isThumb2
BaseKill = true; // New base is always killed right its use.
}
- bool isDPR = Opcode == ARM::VLDRD || Opcode == ARM::VSTRD;
- bool isDef = isi32Load(Opcode) || Opcode == ARM::VLDRS || Opcode == ARM::VLDRD;
- Opcode = getLoadStoreMultipleOpcode(Opcode);
- MachineInstrBuilder MIB = (isAM4)
- ? BuildMI(MBB, MBBI, dl, TII->get(Opcode))
- .addReg(Base, getKillRegState(BaseKill))
- .addImm(ARM_AM::getAM4ModeImm(Mode)).addImm(Pred).addReg(PredReg)
- : BuildMI(MBB, MBBI, dl, TII->get(Opcode))
- .addReg(Base, getKillRegState(BaseKill))
- .addImm(ARM_AM::getAM5Opc(Mode, false, isDPR ? NumRegs<<1 : NumRegs))
- .addImm(Pred).addReg(PredReg);
- MIB.addReg(0); // Add optional writeback (0 for now).
+ bool isDef = (isi32Load(Opcode) || Opcode == ARM::VLDRS ||
+ Opcode == ARM::VLDRD);
+ Opcode = getLoadStoreMultipleOpcode(Opcode, Mode);
+ if (!Opcode) return false;
+ MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII->get(Opcode))
+ .addReg(Base, getKillRegState(BaseKill))
+ .addImm(Pred).addReg(PredReg);
for (unsigned i = 0; i != NumRegs; ++i)
MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
| getKillRegState(Regs[i].second));
return true;
}
+// MergeOpsUpdate - call MergeOps and update MemOps and merges accordingly on
+// success.
+void ARMLoadStoreOpt::MergeOpsUpdate(MachineBasicBlock &MBB,
+ MemOpQueue &memOps,
+ unsigned memOpsBegin, unsigned memOpsEnd,
+ unsigned insertAfter, int Offset,
+ unsigned Base, bool BaseKill,
+ int Opcode,
+ ARMCC::CondCodes Pred, unsigned PredReg,
+ unsigned Scratch,
+ DebugLoc dl,
+ SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
+ // First calculate which of the registers should be killed by the merged
+ // instruction.
+ const unsigned insertPos = memOps[insertAfter].Position;
+ SmallSet<unsigned, 4> KilledRegs;
+ DenseMap<unsigned, unsigned> Killer;
+ for (unsigned i = 0, e = memOps.size(); i != e; ++i) {
+ if (i == memOpsBegin) {
+ i = memOpsEnd;
+ if (i == e)
+ break;
+ }
+ if (memOps[i].Position < insertPos && memOps[i].isKill) {
+ unsigned Reg = memOps[i].Reg;
+ KilledRegs.insert(Reg);
+ Killer[Reg] = i;
+ }
+ }
+
+ SmallVector<std::pair<unsigned, bool>, 8> Regs;
+ for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
+ unsigned Reg = memOps[i].Reg;
+ // If we are inserting the merged operation after an operation that
+ // uses the same register, make sure to transfer any kill flag.
+ bool isKill = memOps[i].isKill || KilledRegs.count(Reg);
+ Regs.push_back(std::make_pair(Reg, isKill));
+ }
+
+ // Try to do the merge.
+ MachineBasicBlock::iterator Loc = memOps[insertAfter].MBBI;
+ ++Loc;
+ if (!MergeOps(MBB, Loc, Offset, Base, BaseKill, Opcode,
+ Pred, PredReg, Scratch, dl, Regs))
+ return;
+
+ // Merge succeeded, update records.
+ Merges.push_back(prior(Loc));
+ for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
+ // Remove kill flags from any memops that come before insertPos.
+ if (Regs[i-memOpsBegin].second) {
+ unsigned Reg = Regs[i-memOpsBegin].first;
+ if (KilledRegs.count(Reg)) {
+ unsigned j = Killer[Reg];
+ int Idx = memOps[j].MBBI->findRegisterUseOperandIdx(Reg, true);
+ assert(Idx >= 0 && "Cannot find killing operand");
+ memOps[j].MBBI->getOperand(Idx).setIsKill(false);
+ memOps[j].isKill = false;
+ }
+ memOps[i].isKill = true;
+ }
+ MBB.erase(memOps[i].MBBI);
+ // Update this memop to refer to the merged instruction.
+ // We may need to move kill flags again.
+ memOps[i].Merged = true;
+ memOps[i].MBBI = Merges.back();
+ memOps[i].Position = insertPos;
+ }
+}
+
/// MergeLDR_STR - Merge a number of load / store instructions into one or more
/// load / store multiple instructions.
void
ARMCC::CondCodes Pred, unsigned PredReg,
unsigned Scratch, MemOpQueue &MemOps,
SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
- bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
+ bool isNotVFP = isi32Load(Opcode) || isi32Store(Opcode);
int Offset = MemOps[SIndex].Offset;
int SOffset = Offset;
- unsigned Pos = MemOps[SIndex].Position;
+ unsigned insertAfter = SIndex;
MachineBasicBlock::iterator Loc = MemOps[SIndex].MBBI;
DebugLoc dl = Loc->getDebugLoc();
- unsigned PReg = Loc->getOperand(0).getReg();
- unsigned PRegNum = ARMRegisterInfo::getRegisterNumbering(PReg);
- bool isKill = Loc->getOperand(0).isKill();
+ const MachineOperand &PMO = Loc->getOperand(0);
+ unsigned PReg = PMO.getReg();
+ unsigned PRegNum = PMO.isUndef() ? UINT_MAX
+ : getARMRegisterNumbering(PReg);
+ unsigned Count = 1;
+ unsigned Limit = ~0U;
+
+ // vldm / vstm limit are 32 for S variants, 16 for D variants.
+
+ switch (Opcode) {
+ default: break;
+ case ARM::VSTRS:
+ Limit = 32;
+ break;
+ case ARM::VSTRD:
+ Limit = 16;
+ break;
+ case ARM::VLDRD:
+ Limit = 16;
+ break;
+ case ARM::VLDRS:
+ Limit = 32;
+ break;
+ }
- SmallVector<std::pair<unsigned,bool>, 8> Regs;
- Regs.push_back(std::make_pair(PReg, isKill));
for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
int NewOffset = MemOps[i].Offset;
- unsigned Reg = MemOps[i].MBBI->getOperand(0).getReg();
- unsigned RegNum = ARMRegisterInfo::getRegisterNumbering(Reg);
- isKill = MemOps[i].MBBI->getOperand(0).isKill();
- // AM4 - register numbers in ascending order.
- // AM5 - consecutive register numbers in ascending order.
- if (NewOffset == Offset + (int)Size &&
- ((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
+ const MachineOperand &MO = MemOps[i].MBBI->getOperand(0);
+ unsigned Reg = MO.getReg();
+ unsigned RegNum = MO.isUndef() ? UINT_MAX
+ : getARMRegisterNumbering(Reg);
+ // Register numbers must be in ascending order. For VFP / NEON load and
+ // store multiples, the registers must also be consecutive and within the
+ // limit on the number of registers per instruction.
+ if (Reg != ARM::SP &&
+ NewOffset == Offset + (int)Size &&
+ ((isNotVFP && RegNum > PRegNum) ||
+ ((Count < Limit) && RegNum == PRegNum+1))) {
Offset += Size;
- Regs.push_back(std::make_pair(Reg, isKill));
PRegNum = RegNum;
+ ++Count;
} else {
// Can't merge this in. Try merge the earlier ones first.
- if (MergeOps(MBB, ++Loc, SOffset, Base, false, Opcode, Pred, PredReg,
- Scratch, dl, Regs)) {
- Merges.push_back(prior(Loc));
- for (unsigned j = SIndex; j < i; ++j) {
- MBB.erase(MemOps[j].MBBI);
- MemOps[j].Merged = true;
- }
- }
+ MergeOpsUpdate(MBB, MemOps, SIndex, i, insertAfter, SOffset,
+ Base, false, Opcode, Pred, PredReg, Scratch, dl, Merges);
MergeLDR_STR(MBB, i, Base, Opcode, Size, Pred, PredReg, Scratch,
MemOps, Merges);
return;
}
- if (MemOps[i].Position > Pos) {
- Pos = MemOps[i].Position;
- Loc = MemOps[i].MBBI;
- }
+ if (MemOps[i].Position > MemOps[insertAfter].Position)
+ insertAfter = i;
}
bool BaseKill = Loc->findRegisterUseOperandIdx(Base, true) != -1;
- if (MergeOps(MBB, ++Loc, SOffset, Base, BaseKill, Opcode, Pred, PredReg,
- Scratch, dl, Regs)) {
- Merges.push_back(prior(Loc));
- for (unsigned i = SIndex, e = MemOps.size(); i != e; ++i) {
- MBB.erase(MemOps[i].MBBI);
- MemOps[i].Merged = true;
- }
- }
-
+ MergeOpsUpdate(MBB, MemOps, SIndex, MemOps.size(), insertAfter, SOffset,
+ Base, BaseKill, Opcode, Pred, PredReg, Scratch, dl, Merges);
return;
}
if (!MI)
return false;
if (MI->getOpcode() != ARM::t2SUBri &&
- MI->getOpcode() != ARM::t2SUBrSPi &&
- MI->getOpcode() != ARM::t2SUBrSPi12 &&
MI->getOpcode() != ARM::tSUBspi &&
MI->getOpcode() != ARM::SUBri)
return false;
// Make sure the offset fits in 8 bits.
- if (Bytes <= 0 || (Limit && Bytes >= Limit))
+ if (Bytes == 0 || (Limit && Bytes >= Limit))
return false;
unsigned Scale = (MI->getOpcode() == ARM::tSUBspi) ? 4 : 1; // FIXME
if (!MI)
return false;
if (MI->getOpcode() != ARM::t2ADDri &&
- MI->getOpcode() != ARM::t2ADDrSPi &&
- MI->getOpcode() != ARM::t2ADDrSPi12 &&
MI->getOpcode() != ARM::tADDspi &&
MI->getOpcode() != ARM::ADDri)
return false;
- if (Bytes <= 0 || (Limit && Bytes >= Limit))
+ if (Bytes == 0 || (Limit && Bytes >= Limit))
// Make sure the offset fits in 8 bits.
return false;
static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
switch (MI->getOpcode()) {
default: return 0;
- case ARM::LDR:
- case ARM::STR:
+ case ARM::LDRi12:
+ case ARM::STRi12:
case ARM::t2LDRi8:
case ARM::t2LDRi12:
case ARM::t2STRi8:
case ARM::VLDRD:
case ARM::VSTRD:
return 8;
- case ARM::LDM:
- case ARM::STM:
- case ARM::t2LDM:
- case ARM::t2STM:
- return (MI->getNumOperands() - 5) * 4;
- case ARM::VLDMS:
- case ARM::VSTMS:
- case ARM::VLDMD:
- case ARM::VSTMD:
- return ARM_AM::getAM5Offset(MI->getOperand(1).getImm()) * 4;
+ case ARM::LDMIA:
+ case ARM::LDMDA:
+ case ARM::LDMDB:
+ case ARM::LDMIB:
+ case ARM::STMIA:
+ case ARM::STMDA:
+ case ARM::STMDB:
+ case ARM::STMIB:
+ case ARM::t2LDMIA:
+ case ARM::t2LDMDB:
+ case ARM::t2STMIA:
+ case ARM::t2STMDB:
+ case ARM::VLDMSIA:
+ case ARM::VSTMSIA:
+ return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 4;
+ case ARM::VLDMDIA:
+ case ARM::VSTMDIA:
+ return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 8;
}
}
+static unsigned getUpdatingLSMultipleOpcode(unsigned Opc,
+ ARM_AM::AMSubMode Mode) {
+ switch (Opc) {
+ default: llvm_unreachable("Unhandled opcode!");
+ case ARM::LDMIA:
+ case ARM::LDMDA:
+ case ARM::LDMDB:
+ case ARM::LDMIB:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::LDMIA_UPD;
+ case ARM_AM::ib: return ARM::LDMIB_UPD;
+ case ARM_AM::da: return ARM::LDMDA_UPD;
+ case ARM_AM::db: return ARM::LDMDB_UPD;
+ }
+ break;
+ case ARM::STMIA:
+ case ARM::STMDA:
+ case ARM::STMDB:
+ case ARM::STMIB:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::STMIA_UPD;
+ case ARM_AM::ib: return ARM::STMIB_UPD;
+ case ARM_AM::da: return ARM::STMDA_UPD;
+ case ARM_AM::db: return ARM::STMDB_UPD;
+ }
+ break;
+ case ARM::t2LDMIA:
+ case ARM::t2LDMDB:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::t2LDMIA_UPD;
+ case ARM_AM::db: return ARM::t2LDMDB_UPD;
+ }
+ break;
+ case ARM::t2STMIA:
+ case ARM::t2STMDB:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::t2STMIA_UPD;
+ case ARM_AM::db: return ARM::t2STMDB_UPD;
+ }
+ break;
+ case ARM::VLDMSIA:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VLDMSIA_UPD;
+ case ARM_AM::db: return ARM::VLDMSDB_UPD;
+ }
+ break;
+ case ARM::VLDMDIA:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VLDMDIA_UPD;
+ case ARM_AM::db: return ARM::VLDMDDB_UPD;
+ }
+ break;
+ case ARM::VSTMSIA:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VSTMSIA_UPD;
+ case ARM_AM::db: return ARM::VSTMSDB_UPD;
+ }
+ break;
+ case ARM::VSTMDIA:
+ switch (Mode) {
+ default: llvm_unreachable("Unhandled submode!");
+ case ARM_AM::ia: return ARM::VSTMDIA_UPD;
+ case ARM_AM::db: return ARM::VSTMDDB_UPD;
+ }
+ break;
+ }
+
+ return 0;
+}
+
/// MergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
/// register into the LDM/STM/VLDM{D|S}/VSTM{D|S} op when possible:
///
MachineBasicBlock::iterator &I) {
MachineInstr *MI = MBBI;
unsigned Base = MI->getOperand(0).getReg();
+ bool BaseKill = MI->getOperand(0).isKill();
unsigned Bytes = getLSMultipleTransferSize(MI);
unsigned PredReg = 0;
ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
int Opcode = MI->getOpcode();
- bool isAM4 = Opcode == ARM::LDM || Opcode == ARM::t2LDM ||
- Opcode == ARM::STM || Opcode == ARM::t2STM;
+ DebugLoc dl = MI->getDebugLoc();
- if (isAM4) {
- if (ARM_AM::getAM4WBFlag(MI->getOperand(1).getImm()))
+ // Can't use an updating ld/st if the base register is also a dest
+ // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
+ for (unsigned i = 2, e = MI->getNumOperands(); i != e; ++i)
+ if (MI->getOperand(i).getReg() == Base)
return false;
- // Can't use the updating AM4 sub-mode if the base register is also a dest
- // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
- for (unsigned i = 3, e = MI->getNumOperands(); i != e; ++i) {
- if (MI->getOperand(i).getReg() == Base)
- return false;
- }
-
- ARM_AM::AMSubMode Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
- if (MBBI != MBB.begin()) {
- MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
- if (Mode == ARM_AM::ia &&
- isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::db, true));
- MI->getOperand(4).setReg(Base);
- MI->getOperand(4).setIsDef();
- MBB.erase(PrevMBBI);
- return true;
- } else if (Mode == ARM_AM::ib &&
- isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::da, true));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- MBB.erase(PrevMBBI);
- return true;
- }
- }
+ bool DoMerge = false;
+ ARM_AM::AMSubMode Mode = ARM_AM::getLoadStoreMultipleSubMode(Opcode);
- if (MBBI != MBB.end()) {
- MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
- if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
- isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- if (NextMBBI == I) {
- Advance = true;
- ++I;
- }
- MBB.erase(NextMBBI);
- return true;
- } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
- isMatchingDecrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- if (NextMBBI == I) {
- Advance = true;
- ++I;
- }
- MBB.erase(NextMBBI);
- return true;
- }
+ // Try merging with the previous instruction.
+ MachineBasicBlock::iterator BeginMBBI = MBB.begin();
+ if (MBBI != BeginMBBI) {
+ MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
+ while (PrevMBBI != BeginMBBI && PrevMBBI->isDebugValue())
+ --PrevMBBI;
+ if (Mode == ARM_AM::ia &&
+ isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ Mode = ARM_AM::db;
+ DoMerge = true;
+ } else if (Mode == ARM_AM::ib &&
+ isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ Mode = ARM_AM::da;
+ DoMerge = true;
}
- } else {
- // VLDM{D|S}, VSTM{D|S} addressing mode 5 ops.
- if (ARM_AM::getAM5WBFlag(MI->getOperand(1).getImm()))
- return false;
+ if (DoMerge)
+ MBB.erase(PrevMBBI);
+ }
- ARM_AM::AMSubMode Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
- unsigned Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
- if (MBBI != MBB.begin()) {
- MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
- if (Mode == ARM_AM::ia &&
- isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::db, true, Offset));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- MBB.erase(PrevMBBI);
- return true;
- }
+ // Try merging with the next instruction.
+ MachineBasicBlock::iterator EndMBBI = MBB.end();
+ if (!DoMerge && MBBI != EndMBBI) {
+ MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
+ while (NextMBBI != EndMBBI && NextMBBI->isDebugValue())
+ ++NextMBBI;
+ if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
+ isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ DoMerge = true;
+ } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
+ isMatchingDecrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ DoMerge = true;
}
-
- if (MBBI != MBB.end()) {
- MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
- if (Mode == ARM_AM::ia &&
- isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::ia, true, Offset));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- if (NextMBBI == I) {
- Advance = true;
- ++I;
- }
- MBB.erase(NextMBBI);
+ if (DoMerge) {
+ if (NextMBBI == I) {
+ Advance = true;
+ ++I;
}
- return true;
+ MBB.erase(NextMBBI);
}
}
- return false;
+ if (!DoMerge)
+ return false;
+
+ unsigned NewOpc = getUpdatingLSMultipleOpcode(Opcode, Mode);
+ MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
+ .addReg(Base, getDefRegState(true)) // WB base register
+ .addReg(Base, getKillRegState(BaseKill))
+ .addImm(Pred).addReg(PredReg);
+
+ // Transfer the rest of operands.
+ for (unsigned OpNum = 3, e = MI->getNumOperands(); OpNum != e; ++OpNum)
+ MIB.addOperand(MI->getOperand(OpNum));
+
+ // Transfer memoperands.
+ MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+
+ MBB.erase(MBBI);
+ return true;
}
-static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
+static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc,
+ ARM_AM::AddrOpc Mode) {
switch (Opc) {
- case ARM::LDR: return ARM::LDR_PRE;
- case ARM::STR: return ARM::STR_PRE;
- case ARM::VLDRS: return ARM::VLDMS;
- case ARM::VLDRD: return ARM::VLDMD;
- case ARM::VSTRS: return ARM::VSTMS;
- case ARM::VSTRD: return ARM::VSTMD;
+ case ARM::LDRi12:
+ return ARM::LDR_PRE_IMM;
+ case ARM::STRi12:
+ return ARM::STR_PRE_IMM;
+ case ARM::VLDRS:
+ return Mode == ARM_AM::add ? ARM::VLDMSIA_UPD : ARM::VLDMSDB_UPD;
+ case ARM::VLDRD:
+ return Mode == ARM_AM::add ? ARM::VLDMDIA_UPD : ARM::VLDMDDB_UPD;
+ case ARM::VSTRS:
+ return Mode == ARM_AM::add ? ARM::VSTMSIA_UPD : ARM::VSTMSDB_UPD;
+ case ARM::VSTRD:
+ return Mode == ARM_AM::add ? ARM::VSTMDIA_UPD : ARM::VSTMDDB_UPD;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
return ARM::t2LDR_PRE;
return 0;
}
-static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc) {
+static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc,
+ ARM_AM::AddrOpc Mode) {
switch (Opc) {
- case ARM::LDR: return ARM::LDR_POST;
- case ARM::STR: return ARM::STR_POST;
- case ARM::VLDRS: return ARM::VLDMS;
- case ARM::VLDRD: return ARM::VLDMD;
- case ARM::VSTRS: return ARM::VSTMS;
- case ARM::VSTRD: return ARM::VSTMD;
+ case ARM::LDRi12:
+ return ARM::LDR_POST_IMM;
+ case ARM::STRi12:
+ return ARM::STR_POST_IMM;
+ case ARM::VLDRS:
+ return Mode == ARM_AM::add ? ARM::VLDMSIA_UPD : ARM::VLDMSDB_UPD;
+ case ARM::VLDRD:
+ return Mode == ARM_AM::add ? ARM::VLDMDIA_UPD : ARM::VLDMDDB_UPD;
+ case ARM::VSTRS:
+ return Mode == ARM_AM::add ? ARM::VSTMSIA_UPD : ARM::VSTMSDB_UPD;
+ case ARM::VSTRD:
+ return Mode == ARM_AM::add ? ARM::VSTMDIA_UPD : ARM::VSTMDDB_UPD;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
return ARM::t2LDR_POST;
unsigned Bytes = getLSMultipleTransferSize(MI);
int Opcode = MI->getOpcode();
DebugLoc dl = MI->getDebugLoc();
- bool isAM5 = Opcode == ARM::VLDRD || Opcode == ARM::VLDRS ||
- Opcode == ARM::VSTRD || Opcode == ARM::VSTRS;
- bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
- if (isAM2 && ARM_AM::getAM2Offset(MI->getOperand(3).getImm()) != 0)
- return false;
- else if (isAM5 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0)
- return false;
- else if (isT2i32Load(Opcode) || isT2i32Store(Opcode))
+ bool isAM5 = (Opcode == ARM::VLDRD || Opcode == ARM::VLDRS ||
+ Opcode == ARM::VSTRD || Opcode == ARM::VSTRS);
+ bool isAM2 = (Opcode == ARM::LDRi12 || Opcode == ARM::STRi12);
+ if (isi32Load(Opcode) || isi32Store(Opcode))
if (MI->getOperand(2).getImm() != 0)
return false;
+ if (isAM5 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0)
+ return false;
bool isLd = isi32Load(Opcode) || Opcode == ARM::VLDRS || Opcode == ARM::VLDRD;
// Can't do the merge if the destination register is the same as the would-be
unsigned NewOpc = 0;
// AM2 - 12 bits, thumb2 - 8 bits.
unsigned Limit = isAM5 ? 0 : (isAM2 ? 0x1000 : 0x100);
- if (MBBI != MBB.begin()) {
+
+ // Try merging with the previous instruction.
+ MachineBasicBlock::iterator BeginMBBI = MBB.begin();
+ if (MBBI != BeginMBBI) {
MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
+ while (PrevMBBI != BeginMBBI && PrevMBBI->isDebugValue())
+ --PrevMBBI;
if (isMatchingDecrement(PrevMBBI, Base, Bytes, Limit, Pred, PredReg)) {
DoMerge = true;
AddSub = ARM_AM::sub;
- NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
} else if (!isAM5 &&
isMatchingIncrement(PrevMBBI, Base, Bytes, Limit,Pred,PredReg)) {
DoMerge = true;
- NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
}
- if (DoMerge)
+ if (DoMerge) {
+ NewOpc = getPreIndexedLoadStoreOpcode(Opcode, AddSub);
MBB.erase(PrevMBBI);
+ }
}
- if (!DoMerge && MBBI != MBB.end()) {
+ // Try merging with the next instruction.
+ MachineBasicBlock::iterator EndMBBI = MBB.end();
+ if (!DoMerge && MBBI != EndMBBI) {
MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
+ while (NextMBBI != EndMBBI && NextMBBI->isDebugValue())
+ ++NextMBBI;
if (!isAM5 &&
isMatchingDecrement(NextMBBI, Base, Bytes, Limit, Pred, PredReg)) {
DoMerge = true;
AddSub = ARM_AM::sub;
- NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
} else if (isMatchingIncrement(NextMBBI, Base, Bytes, Limit,Pred,PredReg)) {
DoMerge = true;
- NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
}
if (DoMerge) {
+ NewOpc = getPostIndexedLoadStoreOpcode(Opcode, AddSub);
if (NextMBBI == I) {
Advance = true;
++I;
if (!DoMerge)
return false;
- bool isDPR = NewOpc == ARM::VLDMD || NewOpc == ARM::VSTMD;
- unsigned Offset = 0;
- if (isAM5)
- Offset = ARM_AM::getAM5Opc((AddSub == ARM_AM::sub)
- ? ARM_AM::db
- : ARM_AM::ia, true, (isDPR ? 2 : 1));
- else if (isAM2)
- Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
- else
- Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
- if (isLd) {
- if (isAM5)
- // VLDMS, VLDMD
- BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
- .addReg(Base, getKillRegState(BaseKill))
- .addImm(Offset).addImm(Pred).addReg(PredReg)
- .addReg(Base, getDefRegState(true)) // WB base register
- .addReg(MI->getOperand(0).getReg(), RegState::Define);
- else if (isAM2)
- // LDR_PRE, LDR_POST,
- BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
- .addReg(Base, RegState::Define)
- .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
- else
+ if (isAM5) {
+ // VLDM[SD}_UPD, VSTM[SD]_UPD
+ // (There are no base-updating versions of VLDR/VSTR instructions, but the
+ // updating load/store-multiple instructions can be used with only one
+ // register.)
+ MachineOperand &MO = MI->getOperand(0);
+ BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
+ .addReg(Base, getDefRegState(true)) // WB base register
+ .addReg(Base, getKillRegState(isLd ? BaseKill : false))
+ .addImm(Pred).addReg(PredReg)
+ .addReg(MO.getReg(), (isLd ? getDefRegState(true) :
+ getKillRegState(MO.isKill())));
+ } else if (isLd) {
+ if (isAM2) {
+ int Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
+ // LDR_PRE, LDR_POST
+ if (NewOpc == ARM::LDR_PRE_IMM || NewOpc == ARM::LDRB_PRE_IMM) {
+ BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
+ .addReg(Base, RegState::Define)
+ .addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
+ } else {
+ BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
+ .addReg(Base, RegState::Define)
+ .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
+ }
+ } else {
+ int Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
// t2LDR_PRE, t2LDR_POST
BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
.addReg(Base, RegState::Define)
.addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
+ }
} else {
MachineOperand &MO = MI->getOperand(0);
- if (isAM5)
- // VSTMS, VSTMD
- BuildMI(MBB, MBBI, dl, TII->get(NewOpc)).addReg(Base).addImm(Offset)
- .addImm(Pred).addReg(PredReg)
- .addReg(Base, getDefRegState(true)) // WB base register
- .addReg(MO.getReg(), getKillRegState(MO.isKill()));
- else if (isAM2)
+ // FIXME: post-indexed stores use am2offset_imm, which still encodes
+ // the vestigal zero-reg offset register. When that's fixed, this clause
+ // can be removed entirely.
+ if (isAM2 && NewOpc == ARM::STR_POST_IMM) {
+ int Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
// STR_PRE, STR_POST
BuildMI(MBB, MBBI, dl, TII->get(NewOpc), Base)
.addReg(MO.getReg(), getKillRegState(MO.isKill()))
.addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
- else
+ } else {
+ int Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
// t2STR_PRE, t2STR_POST
BuildMI(MBB, MBBI, dl, TII->get(NewOpc), Base)
.addReg(MO.getReg(), getKillRegState(MO.isKill()))
.addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
+ }
}
MBB.erase(MBBI);
return true;
}
-/// isMemoryOp - Returns true if instruction is a memory operations (that this
-/// pass is capable of operating on).
+/// isMemoryOp - Returns true if instruction is a memory operation that this
+/// pass is capable of operating on.
static bool isMemoryOp(const MachineInstr *MI) {
+ // When no memory operands are present, conservatively assume unaligned,
+ // volatile, unfoldable.
+ if (!MI->hasOneMemOperand())
+ return false;
+
+ const MachineMemOperand *MMO = *MI->memoperands_begin();
+
+ // Don't touch volatile memory accesses - we may be changing their order.
+ if (MMO->isVolatile())
+ return false;
+
+ // Unaligned ldr/str is emulated by some kernels, but unaligned ldm/stm is
+ // not.
+ if (MMO->getAlignment() < 4)
+ return false;
+
+ // str <undef> could probably be eliminated entirely, but for now we just want
+ // to avoid making a mess of it.
+ // FIXME: Use str <undef> as a wildcard to enable better stm folding.
+ if (MI->getNumOperands() > 0 && MI->getOperand(0).isReg() &&
+ MI->getOperand(0).isUndef())
+ return false;
+
+ // Likewise don't mess with references to undefined addresses.
+ if (MI->getNumOperands() > 1 && MI->getOperand(1).isReg() &&
+ MI->getOperand(1).isUndef())
+ return false;
+
int Opcode = MI->getOpcode();
switch (Opcode) {
default: break;
- case ARM::LDR:
- case ARM::STR:
- return MI->getOperand(1).isReg() && MI->getOperand(2).getReg() == 0;
case ARM::VLDRS:
case ARM::VSTRS:
return MI->getOperand(1).isReg();
case ARM::VLDRD:
case ARM::VSTRD:
return MI->getOperand(1).isReg();
+ case ARM::LDRi12:
+ case ARM::STRi12:
case ARM::t2LDRi8:
case ARM::t2LDRi12:
case ARM::t2STRi8:
static int getMemoryOpOffset(const MachineInstr *MI) {
int Opcode = MI->getOpcode();
- bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
bool isAM3 = Opcode == ARM::LDRD || Opcode == ARM::STRD;
unsigned NumOperands = MI->getDesc().getNumOperands();
unsigned OffField = MI->getOperand(NumOperands-3).getImm();
if (Opcode == ARM::t2LDRi12 || Opcode == ARM::t2LDRi8 ||
Opcode == ARM::t2STRi12 || Opcode == ARM::t2STRi8 ||
- Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8)
+ Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8 ||
+ Opcode == ARM::LDRi12 || Opcode == ARM::STRi12)
return OffField;
- int Offset = isAM2
- ? ARM_AM::getAM2Offset(OffField)
- : (isAM3 ? ARM_AM::getAM3Offset(OffField)
- : ARM_AM::getAM5Offset(OffField) * 4);
- if (isAM2) {
- if (ARM_AM::getAM2Op(OffField) == ARM_AM::sub)
- Offset = -Offset;
- } else if (isAM3) {
+ int Offset = isAM3 ? ARM_AM::getAM3Offset(OffField)
+ : ARM_AM::getAM5Offset(OffField) * 4;
+ if (isAM3) {
if (ARM_AM::getAM3Op(OffField) == ARM_AM::sub)
Offset = -Offset;
} else {
static void InsertLDR_STR(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI,
- int OffImm, bool isDef,
+ int Offset, bool isDef,
DebugLoc dl, unsigned NewOpc,
unsigned Reg, bool RegDeadKill, bool RegUndef,
unsigned BaseReg, bool BaseKill, bool BaseUndef,
- unsigned OffReg, bool OffKill, bool OffUndef,
+ bool OffKill, bool OffUndef,
ARMCC::CondCodes Pred, unsigned PredReg,
const TargetInstrInfo *TII, bool isT2) {
- int Offset = OffImm;
- if (!isT2) {
- if (OffImm < 0)
- Offset = ARM_AM::getAM2Opc(ARM_AM::sub, -OffImm, ARM_AM::no_shift);
- else
- Offset = ARM_AM::getAM2Opc(ARM_AM::add, OffImm, ARM_AM::no_shift);
- }
if (isDef) {
MachineInstrBuilder MIB = BuildMI(MBB, MBBI, MBBI->getDebugLoc(),
TII->get(NewOpc))
.addReg(Reg, getDefRegState(true) | getDeadRegState(RegDeadKill))
.addReg(BaseReg, getKillRegState(BaseKill)|getUndefRegState(BaseUndef));
- if (!isT2)
- MIB.addReg(OffReg, getKillRegState(OffKill)|getUndefRegState(OffUndef));
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
} else {
MachineInstrBuilder MIB = BuildMI(MBB, MBBI, MBBI->getDebugLoc(),
TII->get(NewOpc))
.addReg(Reg, getKillRegState(RegDeadKill) | getUndefRegState(RegUndef))
.addReg(BaseReg, getKillRegState(BaseKill)|getUndefRegState(BaseUndef));
- if (!isT2)
- MIB.addReg(OffReg, getKillRegState(OffKill)|getUndefRegState(OffUndef));
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
}
}
if ((EvenRegNum & 1) == 0 && (EvenRegNum + 1) == OddRegNum)
return false;
+ MachineBasicBlock::iterator NewBBI = MBBI;
bool isT2 = Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8;
bool isLd = Opcode == ARM::LDRD || Opcode == ARM::t2LDRDi8;
bool EvenDeadKill = isLd ?
unsigned BaseReg = BaseOp.getReg();
bool BaseKill = BaseOp.isKill();
bool BaseUndef = BaseOp.isUndef();
- unsigned OffReg = isT2 ? 0 : MI->getOperand(3).getReg();
bool OffKill = isT2 ? false : MI->getOperand(3).isKill();
bool OffUndef = isT2 ? false : MI->getOperand(3).isUndef();
int OffImm = getMemoryOpOffset(MI);
unsigned PredReg = 0;
ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
- if (OddRegNum > EvenRegNum && OffReg == 0 && OffImm == 0) {
+ if (OddRegNum > EvenRegNum && OffImm == 0) {
// Ascending register numbers and no offset. It's safe to change it to a
// ldm or stm.
unsigned NewOpc = (isLd)
- ? (isT2 ? ARM::t2LDM : ARM::LDM)
- : (isT2 ? ARM::t2STM : ARM::STM);
+ ? (isT2 ? ARM::t2LDMIA : ARM::LDMIA)
+ : (isT2 ? ARM::t2STMIA : ARM::STMIA);
if (isLd) {
BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
.addReg(BaseReg, getKillRegState(BaseKill))
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
.addImm(Pred).addReg(PredReg)
- .addReg(0)
.addReg(EvenReg, getDefRegState(isLd) | getDeadRegState(EvenDeadKill))
.addReg(OddReg, getDefRegState(isLd) | getDeadRegState(OddDeadKill));
++NumLDRD2LDM;
} else {
BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
.addReg(BaseReg, getKillRegState(BaseKill))
- .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
.addImm(Pred).addReg(PredReg)
- .addReg(0)
.addReg(EvenReg,
getKillRegState(EvenDeadKill) | getUndefRegState(EvenUndef))
.addReg(OddReg,
getKillRegState(OddDeadKill) | getUndefRegState(OddUndef));
++NumSTRD2STM;
}
+ NewBBI = llvm::prior(MBBI);
} else {
// Split into two instructions.
- assert((!isT2 || !OffReg) &&
- "Thumb2 ldrd / strd does not encode offset register!");
unsigned NewOpc = (isLd)
- ? (isT2 ? (OffImm < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDR)
- : (isT2 ? (OffImm < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STR);
+ ? (isT2 ? (OffImm < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
+ : (isT2 ? (OffImm < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STRi12);
DebugLoc dl = MBBI->getDebugLoc();
// If this is a load and base register is killed, it may have been
// re-defed by the load, make sure the first load does not clobber it.
if (isLd &&
(BaseKill || OffKill) &&
- (TRI->regsOverlap(EvenReg, BaseReg) ||
- (OffReg && TRI->regsOverlap(EvenReg, OffReg)))) {
- assert(!TRI->regsOverlap(OddReg, BaseReg) &&
- (!OffReg || !TRI->regsOverlap(OddReg, OffReg)));
+ (TRI->regsOverlap(EvenReg, BaseReg))) {
+ assert(!TRI->regsOverlap(OddReg, BaseReg));
InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
OddReg, OddDeadKill, false,
- BaseReg, false, BaseUndef, OffReg, false, OffUndef,
+ BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
+ NewBBI = llvm::prior(MBBI);
InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
EvenReg, EvenDeadKill, false,
- BaseReg, BaseKill, BaseUndef, OffReg, OffKill, OffUndef,
+ BaseReg, BaseKill, BaseUndef, OffKill, OffUndef,
Pred, PredReg, TII, isT2);
} else {
if (OddReg == EvenReg && EvenDeadKill) {
- // If the two source operands are the same, the kill marker is probably
- // on the first one. e.g.
+ // If the two source operands are the same, the kill marker is
+ // probably on the first one. e.g.
// t2STRDi8 %R5<kill>, %R5, %R9<kill>, 0, 14, %reg0
EvenDeadKill = false;
OddDeadKill = true;
}
InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
EvenReg, EvenDeadKill, EvenUndef,
- BaseReg, false, BaseUndef, OffReg, false, OffUndef,
+ BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
+ NewBBI = llvm::prior(MBBI);
InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
OddReg, OddDeadKill, OddUndef,
- BaseReg, BaseKill, BaseUndef, OffReg, OffKill, OffUndef,
+ BaseReg, BaseKill, BaseUndef, OffKill, OffUndef,
Pred, PredReg, TII, isT2);
}
if (isLd)
++NumSTRD2STR;
}
- MBBI = prior(MBBI);
MBB.erase(MI);
+ MBBI = NewBBI;
+ return true;
}
return false;
}
if (isMemOp) {
int Opcode = MBBI->getOpcode();
unsigned Size = getLSMultipleTransferSize(MBBI);
+ const MachineOperand &MO = MBBI->getOperand(0);
+ unsigned Reg = MO.getReg();
+ bool isKill = MO.isDef() ? false : MO.isKill();
unsigned Base = MBBI->getOperand(1).getReg();
unsigned PredReg = 0;
ARMCC::CondCodes Pred = llvm::getInstrPredicate(MBBI, PredReg);
CurrSize = Size;
CurrPred = Pred;
CurrPredReg = PredReg;
- MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
- NumMemOps++;
+ MemOps.push_back(MemOpQueueEntry(Offset, Reg, isKill, Position, MBBI));
+ ++NumMemOps;
Advance = true;
} else {
if (Clobber) {
// No need to match PredReg.
// Continue adding to the queue.
if (Offset > MemOps.back().Offset) {
- MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
- NumMemOps++;
+ MemOps.push_back(MemOpQueueEntry(Offset, Reg, isKill,
+ Position, MBBI));
+ ++NumMemOps;
Advance = true;
} else {
for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
I != E; ++I) {
if (Offset < I->Offset) {
- MemOps.insert(I, MemOpQueueEntry(Offset, Position, MBBI));
- NumMemOps++;
+ MemOps.insert(I, MemOpQueueEntry(Offset, Reg, isKill,
+ Position, MBBI));
+ ++NumMemOps;
Advance = true;
break;
} else if (Offset == I->Offset) {
}
}
- if (Advance) {
+ if (MBBI->isDebugValue()) {
+ ++MBBI;
+ if (MBBI == E)
+ // Reach the end of the block, try merging the memory instructions.
+ TryMerge = true;
+ } else if (Advance) {
++Position;
++MBBI;
if (MBBI == E)
MergeLDR_STR(MBB, 0, CurrBase, CurrOpc, CurrSize,
CurrPred, CurrPredReg, Scratch, MemOps, Merges);
- // Try folding preceeding/trailing base inc/dec into the generated
+ // Try folding preceding/trailing base inc/dec into the generated
// LDM/STM ops.
for (unsigned i = 0, e = Merges.size(); i < e; ++i)
if (MergeBaseUpdateLSMultiple(MBB, Merges[i], Advance, MBBI))
++NumMerges;
NumMerges += Merges.size();
- // Try folding preceeding/trailing base inc/dec into those load/store
+ // Try folding preceding/trailing base inc/dec into those load/store
// that were not merged to form LDM/STM ops.
for (unsigned i = 0; i != NumMemOps; ++i)
if (!MemOps[i].Merged)
// RS may be pointing to an instruction that's deleted.
RS->skipTo(prior(MBBI));
} else if (NumMemOps == 1) {
- // Try folding preceeding/trailing base inc/dec into the single
+ // Try folding preceding/trailing base inc/dec into the single
// load/store.
if (MergeBaseUpdateLoadStore(MBB, MemOps[0].MBBI, TII, Advance, MBBI)) {
++NumMerges;
return NumMerges > 0;
}
-namespace {
- struct OffsetCompare {
- bool operator()(const MachineInstr *LHS, const MachineInstr *RHS) const {
- int LOffset = getMemoryOpOffset(LHS);
- int ROffset = getMemoryOpOffset(RHS);
- assert(LHS == RHS || LOffset != ROffset);
- return LOffset > ROffset;
- }
- };
-}
-
-/// MergeReturnIntoLDM - If this is a exit BB, try merging the return op
-/// (bx lr) into the preceeding stack restore so it directly restore the value
-/// of LR into pc.
-/// ldmfd sp!, {r7, lr}
+/// MergeReturnIntoLDM - If this is a exit BB, try merging the return ops
+/// ("bx lr" and "mov pc, lr") into the preceding stack restore so it
+/// directly restore the value of LR into pc.
+/// ldmfd sp!, {..., lr}
/// bx lr
+/// or
+/// ldmfd sp!, {..., lr}
+/// mov pc, lr
/// =>
-/// ldmfd sp!, {r7, pc}
+/// ldmfd sp!, {..., pc}
bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
if (MBB.empty()) return false;
- MachineBasicBlock::iterator MBBI = prior(MBB.end());
+ MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
if (MBBI != MBB.begin() &&
- (MBBI->getOpcode() == ARM::BX_RET || MBBI->getOpcode() == ARM::tBX_RET)) {
+ (MBBI->getOpcode() == ARM::BX_RET ||
+ MBBI->getOpcode() == ARM::tBX_RET ||
+ MBBI->getOpcode() == ARM::MOVPCLR)) {
MachineInstr *PrevMI = prior(MBBI);
- if (PrevMI->getOpcode() == ARM::LDM || PrevMI->getOpcode() == ARM::t2LDM) {
+ unsigned Opcode = PrevMI->getOpcode();
+ if (Opcode == ARM::LDMIA_UPD || Opcode == ARM::LDMDA_UPD ||
+ Opcode == ARM::LDMDB_UPD || Opcode == ARM::LDMIB_UPD ||
+ Opcode == ARM::t2LDMIA_UPD || Opcode == ARM::t2LDMDB_UPD) {
MachineOperand &MO = PrevMI->getOperand(PrevMI->getNumOperands()-1);
if (MO.getReg() != ARM::LR)
return false;
- unsigned NewOpc = isThumb2 ? ARM::t2LDM_RET : ARM::LDM_RET;
+ unsigned NewOpc = (isThumb2 ? ARM::t2LDMIA_RET : ARM::LDMIA_RET);
+ assert(((isThumb2 && Opcode == ARM::t2LDMIA_UPD) ||
+ Opcode == ARM::LDMIA_UPD) && "Unsupported multiple load-return!");
PrevMI->setDesc(TII->get(NewOpc));
MO.setReg(ARM::PC);
+ PrevMI->copyImplicitOps(&*MBBI);
MBB.erase(MBBI);
return true;
}
++MFI) {
MachineBasicBlock &MBB = *MFI;
Modified |= LoadStoreMultipleOpti(MBB);
- Modified |= MergeReturnIntoLDM(MBB);
+ if (TM.getSubtarget<ARMSubtarget>().hasV5TOps())
+ Modified |= MergeReturnIntoLDM(MBB);
}
delete RS;
namespace {
struct ARMPreAllocLoadStoreOpt : public MachineFunctionPass{
static char ID;
- ARMPreAllocLoadStoreOpt() : MachineFunctionPass(&ID) {}
+ ARMPreAllocLoadStoreOpt() : MachineFunctionPass(ID) {}
const TargetData *TD;
const TargetInstrInfo *TII;
bool CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1, DebugLoc &dl,
unsigned &NewOpc, unsigned &EvenReg,
unsigned &OddReg, unsigned &BaseReg,
- unsigned &OffReg, int &Offset,
+ int &Offset,
unsigned &PredReg, ARMCC::CondCodes &Pred,
bool &isT2);
bool RescheduleOps(MachineBasicBlock *MBB,
// some day.
SmallSet<unsigned, 4> AddedRegPressure;
while (++I != E) {
- if (MemOps.count(&*I))
+ if (I->isDebugValue() || MemOps.count(&*I))
continue;
- const TargetInstrDesc &TID = I->getDesc();
- if (TID.isCall() || TID.isTerminator() || TID.hasUnmodeledSideEffects())
+ const MCInstrDesc &MCID = I->getDesc();
+ if (MCID.isCall() || MCID.isTerminator() || I->hasUnmodeledSideEffects())
return false;
- if (isLd && TID.mayStore())
+ if (isLd && MCID.mayStore())
return false;
if (!isLd) {
- if (TID.mayLoad())
+ if (MCID.mayLoad())
return false;
// It's not safe to move the first 'str' down.
// str r1, [r0]
// strh r5, [r0]
// str r4, [r0, #+4]
- if (TID.mayStore())
+ if (MCID.mayStore())
return false;
}
for (unsigned j = 0, NumOps = I->getNumOperands(); j != NumOps; ++j) {
DebugLoc &dl,
unsigned &NewOpc, unsigned &EvenReg,
unsigned &OddReg, unsigned &BaseReg,
- unsigned &OffReg, int &Offset,
- unsigned &PredReg,
+ int &Offset, unsigned &PredReg,
ARMCC::CondCodes &Pred,
bool &isT2) {
// Make sure we're allowed to generate LDRD/STRD.
// FIXME: VLDRS / VSTRS -> VLDRD / VSTRD
unsigned Scale = 1;
unsigned Opcode = Op0->getOpcode();
- if (Opcode == ARM::LDR)
+ if (Opcode == ARM::LDRi12)
NewOpc = ARM::LDRD;
- else if (Opcode == ARM::STR)
+ else if (Opcode == ARM::STRi12)
NewOpc = ARM::STRD;
else if (Opcode == ARM::t2LDRi8 || Opcode == ARM::t2LDRi12) {
NewOpc = ARM::t2LDRDi8;
} else
return false;
- // Make sure the offset registers match.
- if (!isT2 &&
- (Op0->getOperand(2).getReg() != Op1->getOperand(2).getReg()))
- return false;
-
- // Must sure the base address satisfies i64 ld / st alignment requirement.
+ // Make sure the base address satisfies i64 ld / st alignment requirement.
if (!Op0->hasOneMemOperand() ||
!(*Op0->memoperands_begin())->getValue() ||
(*Op0->memoperands_begin())->isVolatile())
return false;
unsigned Align = (*Op0->memoperands_begin())->getAlignment();
- Function *Func = MF->getFunction();
+ const Function *Func = MF->getFunction();
unsigned ReqAlign = STI->hasV6Ops()
- ? TD->getPrefTypeAlignment(Type::getInt64Ty(Func->getContext()))
+ ? TD->getABITypeAlignment(Type::getInt64Ty(Func->getContext()))
: 8; // Pre-v6 need 8-byte align
if (Align < ReqAlign)
return false;
// Then make sure the immediate offset fits.
int OffImm = getMemoryOpOffset(Op0);
if (isT2) {
- if (OffImm < 0) {
- if (OffImm < -255)
- // Can't fall back to t2LDRi8 / t2STRi8.
- return false;
- } else {
- int Limit = (1 << 8) * Scale;
- if (OffImm >= Limit || (OffImm & (Scale-1)))
- return false;
- }
+ int Limit = (1 << 8) * Scale;
+ if (OffImm >= Limit || (OffImm <= -Limit) || (OffImm & (Scale-1)))
+ return false;
Offset = OffImm;
} else {
ARM_AM::AddrOpc AddSub = ARM_AM::add;
if (EvenReg == OddReg)
return false;
BaseReg = Op0->getOperand(1).getReg();
- if (!isT2)
- OffReg = Op0->getOperand(2).getReg();
Pred = llvm::getInstrPredicate(Op0, PredReg);
dl = Op0->getDebugLoc();
return true;
}
+namespace {
+ struct OffsetCompare {
+ bool operator()(const MachineInstr *LHS, const MachineInstr *RHS) const {
+ int LOffset = getMemoryOpOffset(LHS);
+ int ROffset = getMemoryOpOffset(RHS);
+ assert(LHS == RHS || LOffset != ROffset);
+ return LOffset > ROffset;
+ }
+ };
+}
+
bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
SmallVector<MachineInstr*, 4> &Ops,
unsigned Base, bool isLd,
std::sort(Ops.begin(), Ops.end(), OffsetCompare());
// The loads / stores of the same base are in order. Scan them from first to
- // last and check for the followins:
+ // last and check for the following:
// 1. Any def of base.
// 2. Any gaps.
while (Ops.size() > 1) {
} else {
// This is the new location for the loads / stores.
MachineBasicBlock::iterator InsertPos = isLd ? FirstOp : LastOp;
- while (InsertPos != MBB->end() && MemOps.count(InsertPos))
+ while (InsertPos != MBB->end()
+ && (MemOps.count(InsertPos) || InsertPos->isDebugValue()))
++InsertPos;
// If we are moving a pair of loads / stores, see if it makes sense
MachineInstr *Op0 = Ops.back();
MachineInstr *Op1 = Ops[Ops.size()-2];
unsigned EvenReg = 0, OddReg = 0;
- unsigned BaseReg = 0, OffReg = 0, PredReg = 0;
+ unsigned BaseReg = 0, PredReg = 0;
ARMCC::CondCodes Pred = ARMCC::AL;
bool isT2 = false;
unsigned NewOpc = 0;
int Offset = 0;
DebugLoc dl;
if (NumMove == 2 && CanFormLdStDWord(Op0, Op1, dl, NewOpc,
- EvenReg, OddReg, BaseReg, OffReg,
+ EvenReg, OddReg, BaseReg,
Offset, PredReg, Pred, isT2)) {
Ops.pop_back();
Ops.pop_back();
+ const MCInstrDesc &MCID = TII->get(NewOpc);
+ const TargetRegisterClass *TRC = TII->getRegClass(MCID, 0, TRI);
+ MRI->constrainRegClass(EvenReg, TRC);
+ MRI->constrainRegClass(OddReg, TRC);
+
// Form the pair instruction.
if (isLd) {
- MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos,
- dl, TII->get(NewOpc))
+ MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos, dl, MCID)
.addReg(EvenReg, RegState::Define)
.addReg(OddReg, RegState::Define)
.addReg(BaseReg);
+ // FIXME: We're converting from LDRi12 to an insn that still
+ // uses addrmode2, so we need an explicit offset reg. It should
+ // always by reg0 since we're transforming LDRi12s.
if (!isT2)
- MIB.addReg(OffReg);
+ MIB.addReg(0);
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
++NumLDRDFormed;
} else {
- MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos,
- dl, TII->get(NewOpc))
+ MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos, dl, MCID)
.addReg(EvenReg)
.addReg(OddReg)
.addReg(BaseReg);
+ // FIXME: We're converting from LDRi12 to an insn that still
+ // uses addrmode2, so we need an explicit offset reg. It should
+ // always by reg0 since we're transforming STRi12s.
if (!isT2)
- MIB.addReg(OffReg);
+ MIB.addReg(0);
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
++NumSTRDFormed;
}
while (MBBI != E) {
for (; MBBI != E; ++MBBI) {
MachineInstr *MI = MBBI;
- const TargetInstrDesc &TID = MI->getDesc();
- if (TID.isCall() || TID.isTerminator()) {
+ const MCInstrDesc &MCID = MI->getDesc();
+ if (MCID.isCall() || MCID.isTerminator()) {
// Stop at barriers.
++MBBI;
break;
}
- MI2LocMap[MI] = Loc++;
+ if (!MI->isDebugValue())
+ MI2LocMap[MI] = ++Loc;
+
if (!isMemoryOp(MI))
continue;
unsigned PredReg = 0;
projects / oota-llvm.git / blobdiff