#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/Support/CommandLine.h"
-#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/Statistic.h"
using namespace llvm;
{ ARM::t2ADDri, ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 0,0, 0 },
{ ARM::t2ADDrr, ARM::tADDrr, ARM::tADDhirr, 0, 0, 1, 0, 0,1, 0 },
// Note: immediate scale is 4.
- { ARM::t2ADDrSPi,ARM::tADDrSPi,0, 8, 0, 1, 0, 1,0, 0 },
+ { ARM::t2ADDrSPi,ARM::tADDrSPi,0, 8, 0, 1, 0, 1,0, 1 },
{ ARM::t2ADDSri,ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 2,2, 1 },
{ ARM::t2ADDSrr,ARM::tADDrr, 0, 0, 0, 1, 0, 2,0, 1 },
{ ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 0 },
{ ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 0 },
{ ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 0 },
{ ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 0 },
- { ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0 },
+ //FIXME: Disable CMN, as CCodes are backwards from compare expectations
+ //{ ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0 },
{ ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0 },
- { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0 },
- { ARM::t2CMPzri,ARM::tCMPzi8, 0, 8, 0, 1, 0, 2,0, 0 },
- { ARM::t2CMPzrr,ARM::tCMPzhir,0, 0, 0, 0, 0, 2,0, 0 },
+ { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 1 },
{ ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 0 },
// FIXME: adr.n immediate offset must be multiple of 4.
//{ ARM::t2LEApcrelJT,ARM::tLEApcrelJT, 0, 0, 0, 1, 0, 1,0, 0 },
{ ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 0 },
{ ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 0 },
{ ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 0 },
+ { ARM::t2MOVi16,ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 1 },
// FIXME: Do we need the 16-bit 'S' variant?
{ ARM::t2MOVr,ARM::tMOVgpr2gpr,0, 0, 0, 0, 0, 1,0, 0 },
{ ARM::t2MOVCCr,0, ARM::tMOVCCr, 0, 0, 0, 0, 0,1, 0 },
- { ARM::t2MOVCCi,0, ARM::tMOVCCi, 0, 8, 0, 0, 0,1, 0 },
+ { ARM::t2MOVCCi,0, ARM::tMOVCCi, 0, 8, 0, 1, 0,1, 0 },
{ ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 0 },
{ ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0 },
{ ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 0 },
// FIXME: Clean this up after splitting each Thumb load / store opcode
// into multiple ones.
- { ARM::t2LDRi12,ARM::tLDR, 0, 5, 0, 1, 0, 0,0, 1 },
- { ARM::t2LDRs, ARM::tLDR, 0, 0, 0, 1, 0, 0,0, 1 },
- { ARM::t2LDRBi12,ARM::tLDRB, 0, 5, 0, 1, 0, 0,0, 1 },
- { ARM::t2LDRBs, ARM::tLDRB, 0, 0, 0, 1, 0, 0,0, 1 },
- { ARM::t2LDRHi12,ARM::tLDRH, 0, 5, 0, 1, 0, 0,0, 1 },
- { ARM::t2LDRHs, ARM::tLDRH, 0, 0, 0, 1, 0, 0,0, 1 },
+ { ARM::t2LDRi12,ARM::tLDRi, ARM::tLDRspi, 5, 8, 1, 0, 0,0, 1 },
+ { ARM::t2LDRs, ARM::tLDRr, 0, 0, 0, 1, 0, 0,0, 1 },
+ { ARM::t2LDRBi12,ARM::tLDRBi, 0, 5, 0, 1, 0, 0,0, 1 },
+ { ARM::t2LDRBs, ARM::tLDRBr, 0, 0, 0, 1, 0, 0,0, 1 },
+ { ARM::t2LDRHi12,ARM::tLDRHi, 0, 5, 0, 1, 0, 0,0, 1 },
+ { ARM::t2LDRHs, ARM::tLDRHr, 0, 0, 0, 1, 0, 0,0, 1 },
{ ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 1 },
{ ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 1 },
- { ARM::t2STRi12,ARM::tSTR, 0, 5, 0, 1, 0, 0,0, 1 },
- { ARM::t2STRs, ARM::tSTR, 0, 0, 0, 1, 0, 0,0, 1 },
- { ARM::t2STRBi12,ARM::tSTRB, 0, 5, 0, 1, 0, 0,0, 1 },
- { ARM::t2STRBs, ARM::tSTRB, 0, 0, 0, 1, 0, 0,0, 1 },
- { ARM::t2STRHi12,ARM::tSTRH, 0, 5, 0, 1, 0, 0,0, 1 },
- { ARM::t2STRHs, ARM::tSTRH, 0, 0, 0, 1, 0, 0,0, 1 },
-
- { ARM::t2LDM_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 1 },
- { ARM::t2LDM, ARM::tLDM, ARM::tPOP, 0, 0, 1, 1, 1,1, 1 },
- { ARM::t2STM, ARM::tSTM, ARM::tPUSH, 0, 0, 1, 1, 1,1, 1 },
+ { ARM::t2STRi12,ARM::tSTRi, ARM::tSTRspi, 5, 8, 1, 0, 0,0, 1 },
+ { ARM::t2STRs, ARM::tSTRr, 0, 0, 0, 1, 0, 0,0, 1 },
+ { ARM::t2STRBi12,ARM::tSTRBi, 0, 5, 0, 1, 0, 0,0, 1 },
+ { ARM::t2STRBs, ARM::tSTRBr, 0, 0, 0, 1, 0, 0,0, 1 },
+ { ARM::t2STRHi12,ARM::tSTRHi, 0, 5, 0, 1, 0, 0,0, 1 },
+ { ARM::t2STRHs, ARM::tSTRHr, 0, 0, 0, 1, 0, 0,0, 1 },
+
+ { ARM::t2LDMIA, ARM::tLDMIA, 0, 0, 0, 1, 1, 1,1, 1 },
+ { ARM::t2LDMIA_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 1 },
+ { ARM::t2LDMIA_UPD,ARM::tLDMIA_UPD,ARM::tPOP,0, 0, 1, 1, 1,1, 1 },
+ // ARM::t2STM (with no basereg writeback) has no Thumb1 equivalent
+ { ARM::t2STMIA_UPD,ARM::tSTMIA_UPD, 0, 0, 0, 1, 1, 1,1, 1 },
+ { ARM::t2STMDB_UPD, 0, ARM::tPUSH, 0, 0, 1, 1, 1,1, 1 },
};
- class VISIBILITY_HIDDEN Thumb2SizeReduce : public MachineFunctionPass {
+ class Thumb2SizeReduce : public MachineFunctionPass {
public:
static char ID;
Thumb2SizeReduce();
char Thumb2SizeReduce::ID = 0;
}
-Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(&ID) {
+Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(ID) {
for (unsigned i = 0, e = array_lengthof(ReduceTable); i != e; ++i) {
unsigned FromOpc = ReduceTable[i].WideOpc;
if (!ReduceOpcodeMap.insert(std::make_pair(FromOpc, i)).second)
/// Old opcode has an optional def of CPSR.
if (HasCC)
return true;
- // If both old opcode does not implicit CPSR def, then it's not ok since
- // these new opcodes CPSR def is not meant to be thrown away. e.g. CMP.
+ // If old opcode does not implicitly define CPSR, then it's not ok since
+ // these new opcodes' CPSR def is not meant to be thrown away. e.g. CMP.
if (!HasImplicitCPSRDef(MI->getDesc()))
return false;
HasCC = true;
static bool VerifyLowRegs(MachineInstr *MI) {
unsigned Opc = MI->getOpcode();
- bool isPCOk = (Opc == ARM::t2LDM_RET) || (Opc == ARM::t2LDM);
- bool isLROk = (Opc == ARM::t2STM);
+ bool isPCOk = (Opc == ARM::t2LDMIA_RET || Opc == ARM::t2LDMIA ||
+ Opc == ARM::t2LDMDB || Opc == ARM::t2LDMIA_UPD ||
+ Opc == ARM::t2LDMDB_UPD);
+ bool isLROk = (Opc == ARM::t2STMIA_UPD || Opc == ARM::t2STMDB_UPD);
bool isSPOk = isPCOk || isLROk || (Opc == ARM::t2ADDrSPi);
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
continue;
if (isLROk && Reg == ARM::LR)
continue;
- if (isSPOk && Reg == ARM::SP)
- continue;
+ if (Reg == ARM::SP) {
+ if (isSPOk)
+ continue;
+ if (i == 1 && (Opc == ARM::t2LDRi12 || Opc == ARM::t2STRi12))
+ // Special case for these ldr / str with sp as base register.
+ continue;
+ }
if (!isARMLowRegister(Reg))
return false;
}
unsigned Scale = 1;
bool HasImmOffset = false;
bool HasShift = false;
+ bool HasOffReg = true;
bool isLdStMul = false;
- bool isPopPush = false;
unsigned Opc = Entry.NarrowOpc1;
unsigned OpNum = 3; // First 'rest' of operands.
+ uint8_t ImmLimit = Entry.Imm1Limit;
+
switch (Entry.WideOpc) {
default:
llvm_unreachable("Unexpected Thumb2 load / store opcode!");
case ARM::t2LDRi12:
case ARM::t2STRi12:
+ if (MI->getOperand(1).getReg() == ARM::SP) {
+ Opc = Entry.NarrowOpc2;
+ ImmLimit = Entry.Imm2Limit;
+ HasOffReg = false;
+ }
+
Scale = 4;
HasImmOffset = true;
+ HasOffReg = false;
break;
case ARM::t2LDRBi12:
case ARM::t2STRBi12:
HasImmOffset = true;
+ HasOffReg = false;
break;
case ARM::t2LDRHi12:
case ARM::t2STRHi12:
Scale = 2;
HasImmOffset = true;
+ HasOffReg = false;
break;
case ARM::t2LDRs:
case ARM::t2LDRBs:
HasShift = true;
OpNum = 4;
break;
- case ARM::t2LDM_RET:
- case ARM::t2LDM:
- case ARM::t2STM: {
- OpNum = 0;
+ case ARM::t2LDMIA:
+ case ARM::t2LDMDB: {
unsigned BaseReg = MI->getOperand(0).getReg();
- unsigned Mode = MI->getOperand(1).getImm();
- if (BaseReg == ARM::SP && ARM_AM::getAM4WBFlag(Mode)) {
- Opc = Entry.NarrowOpc2;
- isPopPush = true;
+ if (!isARMLowRegister(BaseReg) || Entry.WideOpc != ARM::t2LDMIA)
+ return false;
+
+ // For the non-writeback version (this one), the base register must be
+ // one of the registers being loaded.
+ bool isOK = false;
+ for (unsigned i = 4; i < MI->getNumOperands(); ++i) {
+ if (MI->getOperand(i).getReg() == BaseReg) {
+ isOK = true;
+ break;
+ }
+ }
+
+ if (!isOK)
+ return false;
+
+ OpNum = 0;
+ isLdStMul = true;
+ break;
+ }
+ case ARM::t2LDMIA_RET: {
+ unsigned BaseReg = MI->getOperand(1).getReg();
+ if (BaseReg != ARM::SP)
+ return false;
+ Opc = Entry.NarrowOpc2; // tPOP_RET
+ OpNum = 2;
+ isLdStMul = true;
+ break;
+ }
+ case ARM::t2LDMIA_UPD:
+ case ARM::t2LDMDB_UPD:
+ case ARM::t2STMIA_UPD:
+ case ARM::t2STMDB_UPD: {
+ OpNum = 0;
+
+ unsigned BaseReg = MI->getOperand(1).getReg();
+ if (BaseReg == ARM::SP &&
+ (Entry.WideOpc == ARM::t2LDMIA_UPD ||
+ Entry.WideOpc == ARM::t2STMDB_UPD)) {
+ Opc = Entry.NarrowOpc2; // tPOP or tPUSH
OpNum = 2;
- } else if (Entry.WideOpc == ARM::t2LDM_RET ||
- !isARMLowRegister(BaseReg) ||
- !ARM_AM::getAM4WBFlag(Mode) ||
- ARM_AM::getAM4SubMode(Mode) != ARM_AM::ia) {
+ } else if (!isARMLowRegister(BaseReg) ||
+ (Entry.WideOpc != ARM::t2LDMIA_UPD &&
+ Entry.WideOpc != ARM::t2STMIA_UPD)) {
return false;
}
+
isLdStMul = true;
break;
}
if (HasShift) {
OffsetReg = MI->getOperand(2).getReg();
OffsetKill = MI->getOperand(2).isKill();
+
if (MI->getOperand(3).getImm())
// Thumb1 addressing mode doesn't support shift.
return false;
unsigned OffsetImm = 0;
if (HasImmOffset) {
OffsetImm = MI->getOperand(2).getImm();
- unsigned MaxOffset = ((1 << Entry.Imm1Limit) - 1) * Scale;
- if ((OffsetImm & (Scale-1)) || OffsetImm > MaxOffset)
+ unsigned MaxOffset = ((1 << ImmLimit) - 1) * Scale;
+
+ if ((OffsetImm & (Scale - 1)) || OffsetImm > MaxOffset)
// Make sure the immediate field fits.
return false;
}
// Add the 16-bit load / store instruction.
- // FIXME: Thumb1 addressing mode encode both immediate and register offset.
DebugLoc dl = MI->getDebugLoc();
MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Opc));
if (!isLdStMul) {
- MIB.addOperand(MI->getOperand(0)).addOperand(MI->getOperand(1));
- if (Entry.NarrowOpc1 != ARM::tLDRSB && Entry.NarrowOpc1 != ARM::tLDRSH) {
- // tLDRSB and tLDRSH do not have an immediate offset field. On the other
- // hand, it must have an offset register.
- // FIXME: Remove this special case.
- MIB.addImm(OffsetImm/Scale);
- }
+ MIB.addOperand(MI->getOperand(0));
+ MIB.addOperand(MI->getOperand(1));
+
+ if (HasImmOffset)
+ MIB.addImm(OffsetImm / Scale);
+
assert((!HasShift || OffsetReg) && "Invalid so_reg load / store address!");
- MIB.addReg(OffsetReg, getKillRegState(OffsetKill));
+ if (HasOffReg)
+ MIB.addReg(OffsetReg, getKillRegState(OffsetKill));
}
// Transfer the rest of operands.
for (unsigned e = MI->getNumOperands(); OpNum != e; ++OpNum)
MIB.addOperand(MI->getOperand(OpNum));
- DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
+ // Transfer memoperands.
+ (*MIB).setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+
+ DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
MBB.erase(MI);
++NumLdSts;
unsigned Opc = MI->getOpcode();
switch (Opc) {
default: break;
- case ARM::t2ADDSri:
+ case ARM::t2ADDSri:
case ARM::t2ADDSrr: {
unsigned PredReg = 0;
if (getInstrPredicate(MI, PredReg) == ARMCC::AL) {
if (MI->getOperand(2).getImm() == 0)
return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
break;
+ case ARM::t2MOVi16:
+ // Can convert only 'pure' immediate operands, not immediates obtained as
+ // globals' addresses.
+ if (MI->getOperand(1).isImm())
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
+ break;
+ case ARM::t2CMPrr: {
+ // Try to reduce to the lo-reg only version first. Why there are two
+ // versions of the instruction is a mystery.
+ // It would be nice to just have two entries in the master table that
+ // are prioritized, but the table assumes a unique entry for each
+ // source insn opcode. So for now, we hack a local entry record to use.
+ static const ReduceEntry NarrowEntry =
+ { ARM::t2CMPrr,ARM::tCMPr, 0, 0, 0, 1, 1,2, 0, 1 };
+ if (ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR))
+ return true;
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
+ }
+ case ARM::t2ADDrSPi: {
+ static const ReduceEntry NarrowEntry =
+ { ARM::t2ADDrSPi,ARM::tADDspi, 0, 7, 0, 1, 0, 1, 0, 1 };
+ if (MI->getOperand(0).getReg() == ARM::SP)
+ return ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
+ }
}
return false;
}
if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr))
return false;
- const TargetInstrDesc &TID = MI->getDesc();
unsigned Reg0 = MI->getOperand(0).getReg();
unsigned Reg1 = MI->getOperand(1).getReg();
- if (Reg0 != Reg1)
- return false;
+ if (Reg0 != Reg1) {
+ // Try to commute the operands to make it a 2-address instruction.
+ unsigned CommOpIdx1, CommOpIdx2;
+ if (!TII->findCommutedOpIndices(MI, CommOpIdx1, CommOpIdx2) ||
+ CommOpIdx1 != 1 || MI->getOperand(CommOpIdx2).getReg() != Reg0)
+ return false;
+ MachineInstr *CommutedMI = TII->commuteInstruction(MI);
+ if (!CommutedMI)
+ return false;
+ }
if (Entry.LowRegs2 && !isARMLowRegister(Reg0))
return false;
if (Entry.Imm2Limit) {
bool HasCC = false;
bool CCDead = false;
+ const TargetInstrDesc &TID = MI->getDesc();
if (TID.hasOptionalDef()) {
unsigned NumOps = TID.getNumOperands();
HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
MIB.addOperand(MI->getOperand(i));
}
- DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
+ DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
MBB.erase(MI);
++Num2Addrs;
return false;
} else if (MO.isImm() &&
!TID.OpInfo[i].isPredicate()) {
- if (MO.getImm() > Limit || (MO.getImm() & (Scale-1)) != 0)
+ if (((unsigned)MO.getImm()) > Limit || (MO.getImm() & (Scale-1)) != 0)
return false;
}
}
if (!TID.isPredicable() && NewTID.isPredicable())
AddDefaultPred(MIB);
- DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
+ DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
MBB.erase(MI);
++NumNarrows;
bool Thumb2SizeReduce::ReduceMBB(MachineBasicBlock &MBB) {
bool Modified = false;
- bool LiveCPSR = false;
// Yes, CPSR could be livein.
- for (MachineBasicBlock::const_livein_iterator I = MBB.livein_begin(),
- E = MBB.livein_end(); I != E; ++I) {
- if (*I == ARM::CPSR) {
- LiveCPSR = true;
- break;
- }
- }
+ bool LiveCPSR = MBB.isLiveIn(ARM::CPSR);
MachineBasicBlock::iterator MII = MBB.begin(), E = MBB.end();
MachineBasicBlock::iterator NextMII;
for (; MII != E; MII = NextMII) {
- NextMII = next(MII);
+ NextMII = llvm::next(MII);
MachineInstr *MI = &*MII;
LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR);
goto ProcessNext;
}
- // Try to transform ro a 16-bit non-two-address instruction.
+ // Try to transform to a 16-bit non-two-address instruction.
if (Entry.NarrowOpc1 && ReduceToNarrow(MBB, MI, Entry, LiveCPSR)) {
Modified = true;
MachineBasicBlock::iterator I = prior(NextMII);
projects / oota-llvm.git / blobdiff