-//===-- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass ----*- C++ -*-=//
+//===-- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass ------------===//
//
// The LLVM Compiler Infrastructure
//
#define DEBUG_TYPE "arm-ldst-opt"
#include "ARM.h"
-#include "ARMAddressingModes.h"
#include "ARMBaseInstrInfo.h"
+#include "ARMBaseRegisterInfo.h"
#include "ARMMachineFunctionInfo.h"
-#include "ARMRegisterInfo.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
+#include "MCTargetDesc/ARMAddressingModes.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/RegisterScavenging.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/CodeGen/SelectionDAGNodes.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Function.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Support/ErrorHandling.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/STLExtras.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/Statistic.h"
using namespace llvm;
STATISTIC(NumLDMGened , "Number of ldm instructions generated");
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
+ const ARMSubtarget *STI;
ARMFunctionInfo *AFI;
RegScavenger *RS;
bool isThumb2;
unsigned Position;
MachineBasicBlock::iterator MBBI;
bool Merged;
- MemOpQueueEntry(int o, unsigned r, bool k, unsigned p,
+ MemOpQueueEntry(int o, unsigned r, bool k, unsigned p,
MachineBasicBlock::iterator i)
: Offset(o), Reg(r), isKill(k), Position(p), MBBI(i), Merged(false) {}
};
bool MergeOps(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
int Offset, unsigned Base, bool BaseKill, int Opcode,
ARMCC::CondCodes Pred, unsigned PredReg, unsigned Scratch,
- DebugLoc dl, SmallVector<std::pair<unsigned, bool>, 8> &Regs);
+ DebugLoc dl,
+ ArrayRef<std::pair<unsigned, bool> > Regs,
+ ArrayRef<unsigned> ImpDefs);
void MergeOpsUpdate(MachineBasicBlock &MBB,
MemOpQueue &MemOps,
unsigned memOpsBegin,
case ARM_AM::db: return ARM::LDMDB;
case ARM_AM::ib: return ARM::LDMIB;
}
- break;
case ARM::STRi12:
++NumSTMGened;
switch (Mode) {
case ARM_AM::db: return ARM::STMDB;
case ARM_AM::ib: return ARM::STMIB;
}
- break;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
++NumLDMGened;
case ARM_AM::ia: return ARM::t2LDMIA;
case ARM_AM::db: return ARM::t2LDMDB;
}
- break;
case ARM::t2STRi8:
case ARM::t2STRi12:
++NumSTMGened;
case ARM_AM::ia: return ARM::t2STMIA;
case ARM_AM::db: return ARM::t2STMDB;
}
- break;
case ARM::VLDRS:
++NumVLDMGened;
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VLDMSIA;
- case ARM_AM::db: return ARM::VLDMSDB;
+ case ARM_AM::db: return 0; // Only VLDMSDB_UPD exists.
}
- break;
case ARM::VSTRS:
++NumVSTMGened;
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VSTMSIA;
- case ARM_AM::db: return ARM::VSTMSDB;
+ case ARM_AM::db: return 0; // Only VSTMSDB_UPD exists.
}
- break;
case ARM::VLDRD:
++NumVLDMGened;
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VLDMDIA;
- case ARM_AM::db: return ARM::VLDMDDB;
+ case ARM_AM::db: return 0; // Only VLDMDDB_UPD exists.
}
- break;
case ARM::VSTRD:
++NumVSTMGened;
switch (Mode) {
default: llvm_unreachable("Unhandled submode!");
case ARM_AM::ia: return ARM::VSTMDIA;
- case ARM_AM::db: return ARM::VSTMDDB;
+ case ARM_AM::db: return 0; // Only VSTMDDB_UPD exists.
}
- break;
}
-
- return 0;
}
namespace llvm {
case ARM::t2LDMDB_UPD:
case ARM::t2STMDB:
case ARM::t2STMDB_UPD:
- case ARM::VLDMSDB:
case ARM::VLDMSDB_UPD:
- case ARM::VSTMSDB:
case ARM::VSTMSDB_UPD:
- case ARM::VLDMDDB:
case ARM::VLDMDDB_UPD:
- case ARM::VSTMDDB:
case ARM::VSTMDDB_UPD:
return ARM_AM::db;
case ARM::STMIB_UPD:
return ARM_AM::ib;
}
-
- return ARM_AM::bad_am_submode;
}
} // end namespace ARM_AM
int Offset, unsigned Base, bool BaseKill,
int Opcode, ARMCC::CondCodes Pred,
unsigned PredReg, unsigned Scratch, DebugLoc dl,
- SmallVector<std::pair<unsigned, bool>, 8> &Regs) {
+ ArrayRef<std::pair<unsigned, bool> > Regs,
+ ArrayRef<unsigned> ImpDefs) {
// Only a single register to load / store. Don't bother.
unsigned NumRegs = Regs.size();
if (NumRegs <= 1)
// VLDM/VSTM do not support DB mode without also updating the base reg.
Mode = ARM_AM::db;
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
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);
MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
| getKillRegState(Regs[i].second));
+ // Add implicit defs for super-registers.
+ for (unsigned i = 0, e = ImpDefs.size(); i != e; ++i)
+ MIB.addReg(ImpDefs[i], RegState::ImplicitDefine);
+
return true;
}
// First calculate which of the registers should be killed by the merged
// instruction.
const unsigned insertPos = memOps[insertAfter].Position;
-
- SmallSet<unsigned, 4> UnavailRegs;
SmallSet<unsigned, 4> KilledRegs;
DenseMap<unsigned, unsigned> Killer;
- for (unsigned i = 0; i < memOpsBegin; ++i) {
- if (memOps[i].Position < insertPos && memOps[i].isKill) {
- unsigned Reg = memOps[i].Reg;
- if (memOps[i].Merged)
- UnavailRegs.insert(Reg);
- else {
- KilledRegs.insert(Reg);
- Killer[Reg] = i;
- }
+ for (unsigned i = 0, e = memOps.size(); i != e; ++i) {
+ if (i == memOpsBegin) {
+ i = memOpsEnd;
+ if (i == e)
+ break;
}
- }
- for (unsigned i = memOpsEnd, e = memOps.size(); i != e; ++i) {
if (memOps[i].Position < insertPos && memOps[i].isKill) {
unsigned Reg = memOps[i].Reg;
KilledRegs.insert(Reg);
}
SmallVector<std::pair<unsigned, bool>, 8> Regs;
+ SmallVector<unsigned, 8> ImpDefs;
for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
unsigned Reg = memOps[i].Reg;
- if (UnavailRegs.count(Reg))
- // Register is killed before and it's not easy / possible to update the
- // kill marker on already merged instructions. Abort.
- return;
-
- // If we are inserting the merged operation after an unmerged operation that
+ // 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));
+
+ // Collect any implicit defs of super-registers. They must be preserved.
+ for (MIOperands MO(memOps[i].MBBI); MO.isValid(); ++MO) {
+ if (!MO->isReg() || !MO->isDef() || !MO->isImplicit() || MO->isDead())
+ continue;
+ unsigned DefReg = MO->getReg();
+ if (std::find(ImpDefs.begin(), ImpDefs.end(), DefReg) == ImpDefs.end())
+ ImpDefs.push_back(DefReg);
+ }
}
// 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))
+ Pred, PredReg, Scratch, dl, Regs, ImpDefs))
return;
// Merge succeeded, update records.
Merges.push_back(prior(Loc));
for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
- // Remove kill flags from any unmerged memops that come before insertPos.
+ // 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];
- memOps[j].MBBI->getOperand(0).setIsKill(false);
+ 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;
}
}
DebugLoc dl = Loc->getDebugLoc();
const MachineOperand &PMO = Loc->getOperand(0);
unsigned PReg = PMO.getReg();
- unsigned PRegNum = PMO.isUndef() ? UINT_MAX
- : getARMRegisterNumbering(PReg);
+ unsigned PRegNum = PMO.isUndef() ? UINT_MAX : TRI->getEncodingValue(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;
+ }
for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
int NewOffset = MemOps[i].Offset;
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, the registers
- // must also be consecutive and there is a limit of 16 double-word
- // registers per instruction.
+ unsigned RegNum = MO.isUndef() ? UINT_MAX : TRI->getEncodingValue(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)
- || ((Size < 8 || Count < 16) && RegNum == PRegNum+1))) {
+ ((isNotVFP && RegNum > PRegNum) ||
+ ((Count < Limit) && RegNum == PRegNum+1))) {
Offset += Size;
PRegNum = RegNum;
++Count;
return;
}
-static inline bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
- unsigned Bytes, unsigned Limit,
- ARMCC::CondCodes Pred, unsigned PredReg){
+static bool definesCPSR(MachineInstr *MI) {
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ if (MO.isDef() && MO.getReg() == ARM::CPSR && !MO.isDead())
+ // If the instruction has live CPSR def, then it's not safe to fold it
+ // into load / store.
+ return true;
+ }
+
+ return false;
+}
+
+static bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
+ unsigned Bytes, unsigned Limit,
+ ARMCC::CondCodes Pred, unsigned PredReg) {
unsigned MyPredReg = 0;
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;
+
+ bool CheckCPSRDef = false;
+ switch (MI->getOpcode()) {
+ default: return false;
+ case ARM::t2SUBri:
+ case ARM::SUBri:
+ CheckCPSRDef = true;
+ // fallthrough
+ case ARM::tSUBspi:
+ break;
+ }
// Make sure the offset fits in 8 bits.
if (Bytes == 0 || (Limit && Bytes >= Limit))
return false;
unsigned Scale = (MI->getOpcode() == ARM::tSUBspi) ? 4 : 1; // FIXME
- return (MI->getOperand(0).getReg() == Base &&
- MI->getOperand(1).getReg() == Base &&
- (MI->getOperand(2).getImm()*Scale) == Bytes &&
- llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
- MyPredReg == PredReg);
+ if (!(MI->getOperand(0).getReg() == Base &&
+ MI->getOperand(1).getReg() == Base &&
+ (MI->getOperand(2).getImm()*Scale) == Bytes &&
+ getInstrPredicate(MI, MyPredReg) == Pred &&
+ MyPredReg == PredReg))
+ return false;
+
+ return CheckCPSRDef ? !definesCPSR(MI) : true;
}
-static inline bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
- unsigned Bytes, unsigned Limit,
- ARMCC::CondCodes Pred, unsigned PredReg){
+static bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
+ unsigned Bytes, unsigned Limit,
+ ARMCC::CondCodes Pred, unsigned PredReg) {
unsigned MyPredReg = 0;
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;
+
+ bool CheckCPSRDef = false;
+ switch (MI->getOpcode()) {
+ default: return false;
+ case ARM::t2ADDri:
+ case ARM::ADDri:
+ CheckCPSRDef = true;
+ // fallthrough
+ case ARM::tADDspi:
+ break;
+ }
if (Bytes == 0 || (Limit && Bytes >= Limit))
// Make sure the offset fits in 8 bits.
return false;
unsigned Scale = (MI->getOpcode() == ARM::tADDspi) ? 4 : 1; // FIXME
- return (MI->getOperand(0).getReg() == Base &&
- MI->getOperand(1).getReg() == Base &&
- (MI->getOperand(2).getImm()*Scale) == Bytes &&
- llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
- MyPredReg == PredReg);
+ if (!(MI->getOperand(0).getReg() == Base &&
+ MI->getOperand(1).getReg() == Base &&
+ (MI->getOperand(2).getImm()*Scale) == Bytes &&
+ getInstrPredicate(MI, MyPredReg) == Pred &&
+ MyPredReg == PredReg))
+ return false;
+
+ return CheckCPSRDef ? !definesCPSR(MI) : true;
}
static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
case ARM::t2STMIA:
case ARM::t2STMDB:
case ARM::VLDMSIA:
- case ARM::VLDMSDB:
case ARM::VSTMSIA:
- case ARM::VSTMSDB:
return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 4;
case ARM::VLDMDIA:
- case ARM::VLDMDDB:
case ARM::VSTMDIA:
- case ARM::VSTMDDB:
return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 8;
}
}
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_AM::da: return ARM::STMDA_UPD;
case ARM_AM::db: return ARM::STMDB_UPD;
}
- break;
case ARM::t2LDMIA:
case ARM::t2LDMDB:
switch (Mode) {
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) {
case ARM_AM::ia: return ARM::t2STMIA_UPD;
case ARM_AM::db: return ARM::t2STMDB_UPD;
}
- break;
case ARM::VLDMSIA:
- case ARM::VLDMSDB:
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:
- case ARM::VLDMDDB:
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:
- case ARM::VSTMSDB:
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:
- case ARM::VSTMDDB:
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
bool BaseKill = MI->getOperand(0).isKill();
unsigned Bytes = getLSMultipleTransferSize(MI);
unsigned PredReg = 0;
- ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
+ ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
int Opcode = MI->getOpcode();
DebugLoc dl = MI->getDebugLoc();
MIB.addOperand(MI->getOperand(OpNum));
// Transfer memoperands.
- (*MIB).setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+ MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
MBB.erase(MBBI);
return true;
ARM_AM::AddrOpc Mode) {
switch (Opc) {
case ARM::LDRi12:
- return ARM::LDR_PRE;
+ return ARM::LDR_PRE_IMM;
case ARM::STRi12:
- return ARM::STR_PRE;
+ return ARM::STR_PRE_IMM;
case ARM::VLDRS:
return Mode == ARM_AM::add ? ARM::VLDMSIA_UPD : ARM::VLDMSDB_UPD;
case ARM::VLDRD:
return ARM::t2STR_PRE;
default: llvm_unreachable("Unhandled opcode!");
}
- return 0;
}
static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc,
ARM_AM::AddrOpc Mode) {
switch (Opc) {
case ARM::LDRi12:
- return ARM::LDR_POST;
+ return ARM::LDR_POST_IMM;
case ARM::STRi12:
- return ARM::STR_POST;
+ return ARM::STR_POST_IMM;
case ARM::VLDRS:
return Mode == ARM_AM::add ? ARM::VLDMSIA_UPD : ARM::VLDMSDB_UPD;
case ARM::VLDRD:
return ARM::t2STR_POST;
default: llvm_unreachable("Unhandled opcode!");
}
- return 0;
}
/// MergeBaseUpdateLoadStore - Fold proceeding/trailing inc/dec of base
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
// writeback register.
- if (isLd && MI->getOperand(0).getReg() == Base)
+ if (MI->getOperand(0).getReg() == Base)
return false;
unsigned PredReg = 0;
- ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
+ ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
bool DoMerge = false;
ARM_AM::AddrOpc AddSub = ARM_AM::add;
unsigned NewOpc = 0;
if (!DoMerge)
return false;
- unsigned Offset = 0;
- if (isAM2)
- Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
- else if (!isAM5)
- Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
-
if (isAM5) {
// VLDM[SD}_UPD, VSTM[SD]_UPD
// (There are no base-updating versions of VLDR/VSTR instructions, but the
.addReg(MO.getReg(), (isLd ? getDefRegState(true) :
getKillRegState(MO.isKill())));
} else if (isLd) {
- 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 (isAM2) {
+ // LDR_PRE, LDR_POST
+ if (NewOpc == ARM::LDR_PRE_IMM || NewOpc == ARM::LDRB_PRE_IMM) {
+ int Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
+ BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
+ .addReg(Base, RegState::Define)
+ .addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
+ } else {
+ int Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
+ 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 (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.
unsigned Opcode = MI->getOpcode();
if (Opcode == ARM::LDRD || Opcode == ARM::STRD ||
Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8) {
+ const MachineOperand &BaseOp = MI->getOperand(2);
+ unsigned BaseReg = BaseOp.getReg();
unsigned EvenReg = MI->getOperand(0).getReg();
unsigned OddReg = MI->getOperand(1).getReg();
unsigned EvenRegNum = TRI->getDwarfRegNum(EvenReg, false);
unsigned OddRegNum = TRI->getDwarfRegNum(OddReg, false);
- if ((EvenRegNum & 1) == 0 && (EvenRegNum + 1) == OddRegNum)
+ // ARM errata 602117: LDRD with base in list may result in incorrect base
+ // register when interrupted or faulted.
+ bool Errata602117 = EvenReg == BaseReg && STI->isCortexM3();
+ if (!Errata602117 &&
+ ((EvenRegNum & 1) == 0 && (EvenRegNum + 1) == OddRegNum))
return false;
MachineBasicBlock::iterator NewBBI = MBBI;
bool OddDeadKill = isLd ?
MI->getOperand(1).isDead() : MI->getOperand(1).isKill();
bool OddUndef = MI->getOperand(1).isUndef();
- const MachineOperand &BaseOp = MI->getOperand(2);
- unsigned BaseReg = BaseOp.getReg();
bool BaseKill = BaseOp.isKill();
bool BaseUndef = BaseOp.isUndef();
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);
+ ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
if (OddRegNum > EvenRegNum && OffImm == 0) {
// Ascending register numbers and no offset. It's safe to change it to a
unsigned NewOpc = (isLd)
? (isT2 ? (OffImm < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
: (isT2 ? (OffImm < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STRi12);
+ // Be extra careful for thumb2. t2LDRi8 can't reference a zero offset,
+ // so adjust and use t2LDRi12 here for that.
+ unsigned NewOpc2 = (isLd)
+ ? (isT2 ? (OffImm+4 < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
+ : (isT2 ? (OffImm+4 < 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.
(BaseKill || OffKill) &&
(TRI->regsOverlap(EvenReg, BaseReg))) {
assert(!TRI->regsOverlap(OddReg, BaseReg));
- InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
+ InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc2,
OddReg, OddDeadKill, false,
BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
EvenDeadKill = false;
OddDeadKill = true;
}
+ // Never kill the base register in the first instruction.
+ if (EvenReg == BaseReg)
+ EvenDeadKill = false;
InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
EvenReg, EvenDeadKill, EvenUndef,
BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
NewBBI = llvm::prior(MBBI);
- InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
+ InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc2,
OddReg, OddDeadKill, OddUndef,
BaseReg, BaseKill, BaseUndef, OffKill, OffUndef,
Pred, PredReg, TII, isT2);
bool isKill = MO.isDef() ? false : MO.isKill();
unsigned Base = MBBI->getOperand(1).getReg();
unsigned PredReg = 0;
- ARMCC::CondCodes Pred = llvm::getInstrPredicate(MBBI, PredReg);
+ ARMCC::CondCodes Pred = getInstrPredicate(MBBI, PredReg);
int Offset = getMemoryOpOffset(MBBI);
// Watch out for:
// r4 := ldr [r5]
// merge the ldr's so far, including this one. But don't try to
// combine the following ldr(s).
Clobber = (isi32Load(Opcode) && Base == MBBI->getOperand(0).getReg());
+
+ // Watch out for:
+ // r4 := ldr [r0, #8]
+ // r4 := ldr [r0, #4]
+ //
+ // The optimization may reorder the second ldr in front of the first
+ // ldr, which violates write after write(WAW) dependence. The same as
+ // str. Try to merge inst(s) already in MemOps.
+ bool Overlap = false;
+ for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end(); I != E; ++I) {
+ if (TRI->regsOverlap(Reg, I->MBBI->getOperand(0).getReg())) {
+ Overlap = true;
+ break;
+ }
+ }
+
if (CurrBase == 0 && !Clobber) {
// Start of a new chain.
CurrBase = Base;
MemOps.push_back(MemOpQueueEntry(Offset, Reg, isKill, Position, MBBI));
++NumMemOps;
Advance = true;
- } else {
+ } else if (!Overlap) {
if (Clobber) {
TryMerge = true;
Advance = true;
// First advance to the instruction just before the start of the chain.
AdvanceRS(MBB, MemOps);
// Find a scratch register.
- unsigned Scratch = RS->FindUnusedReg(ARM::GPRRegisterClass);
+ unsigned Scratch = RS->FindUnusedReg(&ARM::GPRRegClass);
// Process the load / store instructions.
RS->forward(prior(MBBI));
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 ops
-/// ("bx lr" and "mov pc, lr") into the preceeding stack restore so it
+/// ("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
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 ||
Opcode == ARM::LDMIA_UPD) && "Unsupported multiple load-return!");
PrevMI->setDesc(TII->get(NewOpc));
MO.setReg(ARM::PC);
- PrevMI->copyImplicitOps(&*MBBI);
+ PrevMI->copyImplicitOps(*MBB.getParent(), &*MBBI);
MBB.erase(MBBI);
return true;
}
AFI = Fn.getInfo<ARMFunctionInfo>();
TII = TM.getInstrInfo();
TRI = TM.getRegisterInfo();
+ STI = &TM.getSubtarget<ARMSubtarget>();
RS = new RegScavenger();
isThumb2 = AFI->isThumb2Function();
static char ID;
ARMPreAllocLoadStoreOpt() : MachineFunctionPass(ID) {}
- const TargetData *TD;
+ const DataLayout *TD;
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
const ARMSubtarget *STI;
unsigned &PredReg, ARMCC::CondCodes &Pred,
bool &isT2);
bool RescheduleOps(MachineBasicBlock *MBB,
- SmallVector<MachineInstr*, 4> &Ops,
+ SmallVectorImpl<MachineInstr *> &Ops,
unsigned Base, bool isLd,
DenseMap<MachineInstr*, unsigned> &MI2LocMap);
bool RescheduleLoadStoreInstrs(MachineBasicBlock *MBB);
}
bool ARMPreAllocLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
- TD = Fn.getTarget().getTargetData();
+ TD = Fn.getTarget().getDataLayout();
TII = Fn.getTarget().getInstrInfo();
TRI = Fn.getTarget().getRegisterInfo();
STI = &Fn.getTarget().getSubtarget<ARMSubtarget>();
while (++I != E) {
if (I->isDebugValue() || MemOps.count(&*I))
continue;
- const TargetInstrDesc &TID = I->getDesc();
- if (TID.isCall() || TID.isTerminator() || TID.hasUnmodeledSideEffects())
+ if (I->isCall() || I->isTerminator() || I->hasUnmodeledSideEffects())
return false;
- if (isLd && TID.mayStore())
+ if (isLd && I->mayStore())
return false;
if (!isLd) {
- if (TID.mayLoad())
+ if (I->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 (I->mayStore())
return false;
}
for (unsigned j = 0, NumOps = I->getNumOperands(); j != NumOps; ++j) {
return AddedRegPressure.size() <= MemRegs.size() * 2;
}
+
+/// Copy Op0 and Op1 operands into a new array assigned to MI.
+static void concatenateMemOperands(MachineInstr *MI, MachineInstr *Op0,
+ MachineInstr *Op1) {
+ assert(MI->memoperands_empty() && "expected a new machineinstr");
+ size_t numMemRefs = (Op0->memoperands_end() - Op0->memoperands_begin())
+ + (Op1->memoperands_end() - Op1->memoperands_begin());
+
+ MachineFunction *MF = MI->getParent()->getParent();
+ MachineSDNode::mmo_iterator MemBegin = MF->allocateMemRefsArray(numMemRefs);
+ MachineSDNode::mmo_iterator MemEnd =
+ std::copy(Op0->memoperands_begin(), Op0->memoperands_end(), MemBegin);
+ MemEnd =
+ std::copy(Op1->memoperands_begin(), Op1->memoperands_end(), MemEnd);
+ MI->setMemRefs(MemBegin, MemEnd);
+}
+
bool
ARMPreAllocLoadStoreOpt::CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1,
DebugLoc &dl,
return false;
// Make sure the base address satisfies i64 ld / st alignment requirement.
+ // At the moment, we ignore the memoryoperand's value.
+ // If we want to use AliasAnalysis, we should check it accordingly.
if (!Op0->hasOneMemOperand() ||
- !(*Op0->memoperands_begin())->getValue() ||
(*Op0->memoperands_begin())->isVolatile())
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();
- Pred = llvm::getInstrPredicate(Op0, PredReg);
+ Pred = 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,
+ SmallVectorImpl<MachineInstr *> &Ops,
unsigned Base, bool isLd,
DenseMap<MachineInstr*, unsigned> &MI2LocMap) {
bool RetVal = false;
LastOp = Op;
}
- unsigned Opcode = Op->getOpcode();
- if (LastOpcode && Opcode != LastOpcode)
+ unsigned LSMOpcode
+ = getLoadStoreMultipleOpcode(Op->getOpcode(), ARM_AM::ia);
+ if (LastOpcode && LSMOpcode != LastOpcode)
break;
int Offset = getMemoryOpOffset(Op);
}
LastOffset = Offset;
LastBytes = Bytes;
- LastOpcode = Opcode;
+ LastOpcode = LSMOpcode;
if (++NumMove == 8) // FIXME: Tune this limit.
break;
}
Ops.pop_back();
Ops.pop_back();
+ const MCInstrDesc &MCID = TII->get(NewOpc);
+ const TargetRegisterClass *TRC = TII->getRegClass(MCID, 0, TRI, *MF);
+ 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);
if (!isT2)
MIB.addReg(0);
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
+ concatenateMemOperands(MIB, Op0, Op1);
+ DEBUG(dbgs() << "Formed " << *MIB << "\n");
++NumLDRDFormed;
} else {
- MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos,
- dl, TII->get(NewOpc))
+ MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos, dl, MCID)
.addReg(EvenReg)
.addReg(OddReg)
.addReg(BaseReg);
if (!isT2)
MIB.addReg(0);
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
+ concatenateMemOperands(MIB, Op0, Op1);
+ DEBUG(dbgs() << "Formed " << *MIB << "\n");
++NumSTRDFormed;
}
MBB->erase(Op0);
while (MBBI != E) {
for (; MBBI != E; ++MBBI) {
MachineInstr *MI = MBBI;
- const TargetInstrDesc &TID = MI->getDesc();
- if (TID.isCall() || TID.isTerminator()) {
+ if (MI->isCall() || MI->isTerminator()) {
// Stop at barriers.
++MBBI;
break;
if (!isMemoryOp(MI))
continue;
unsigned PredReg = 0;
- if (llvm::getInstrPredicate(MI, PredReg) != ARMCC::AL)
+ if (getInstrPredicate(MI, PredReg) != ARMCC::AL)
continue;
int Opc = MI->getOpcode();
if (!StopHere)
BI->second.push_back(MI);
} else {
- SmallVector<MachineInstr*, 4> MIs;
- MIs.push_back(MI);
- Base2LdsMap[Base] = MIs;
+ Base2LdsMap[Base].push_back(MI);
LdBases.push_back(Base);
}
} else {
if (!StopHere)
BI->second.push_back(MI);
} else {
- SmallVector<MachineInstr*, 4> MIs;
- MIs.push_back(MI);
- Base2StsMap[Base] = MIs;
+ Base2StsMap[Base].push_back(MI);
StBases.push_back(Base);
}
}
// Re-schedule loads.
for (unsigned i = 0, e = LdBases.size(); i != e; ++i) {
unsigned Base = LdBases[i];
- SmallVector<MachineInstr*, 4> &Lds = Base2LdsMap[Base];
+ SmallVectorImpl<MachineInstr *> &Lds = Base2LdsMap[Base];
if (Lds.size() > 1)
RetVal |= RescheduleOps(MBB, Lds, Base, true, MI2LocMap);
}
// Re-schedule stores.
for (unsigned i = 0, e = StBases.size(); i != e; ++i) {
unsigned Base = StBases[i];
- SmallVector<MachineInstr*, 4> &Sts = Base2StsMap[Base];
+ SmallVectorImpl<MachineInstr *> &Sts = Base2StsMap[Base];
if (Sts.size() > 1)
RetVal |= RescheduleOps(MBB, Sts, Base, false, MI2LocMap);
}
projects / oota-llvm.git / blobdiff