#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLocation.h"
+#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
+#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/STLExtras.h"
+#include "llvm/Support/CommandLine.h"
#include <algorithm>
-#include <iostream>
using namespace llvm;
+static cl::opt<bool> ThumbRegScavenging("enable-thumb-reg-scavenging",
+ cl::Hidden,
+ cl::desc("Enable register scavenging on Thumb"));
+
unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum) {
using namespace ARM;
switch (RegEnum) {
case S30: return 30;
case S31: return 31;
default:
- std::cerr << "Unknown ARM register!\n";
+ assert(0 && "Unknown ARM register!");
abort();
}
}
const ARMSubtarget &sti)
: ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
TII(tii), STI(sti),
- FramePtr(STI.useThumbBacktraces() ? ARM::R7 : ARM::R11) {
+ FramePtr((STI.useThumbBacktraces() || STI.isThumb()) ? ARM::R7 : ARM::R11) {
}
bool ARMRegisterInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
return false;
MachineInstrBuilder MIB = BuildMI(MBB, MI, TII.get(ARM::tPUSH));
- for (unsigned i = CSI.size(); i != 0; --i)
- MIB.addReg(CSI[i-1].getReg());
+ for (unsigned i = CSI.size(); i != 0; --i) {
+ unsigned Reg = CSI[i-1].getReg();
+ // Add the callee-saved register as live-in. It's killed at the spill.
+ MBB.addLiveIn(Reg);
+ MIB.addReg(Reg, false/*isDef*/,false/*isImp*/,true/*isKill*/);
+ }
return true;
}
return true;
}
+static inline
+const MachineInstrBuilder &AddDefaultPred(const MachineInstrBuilder &MIB) {
+ return MIB.addImm((int64_t)ARMCC::AL).addReg(0);
+}
+
+static inline
+const MachineInstrBuilder &AddDefaultCC(const MachineInstrBuilder &MIB) {
+ return MIB.addReg(0);
+}
+
+static const MachineInstrBuilder &ARMInstrAddOperand(MachineInstrBuilder &MIB,
+ MachineOperand &MO) {
+ if (MO.isRegister())
+ MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit());
+ else if (MO.isImmediate())
+ MIB = MIB.addImm(MO.getImm());
+ else if (MO.isFrameIndex())
+ MIB = MIB.addFrameIndex(MO.getFrameIndex());
+ else
+ assert(0 && "Unknown operand for ARMInstrAddOperand!");
+
+ return MIB;
+}
+
void ARMRegisterInfo::
storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
- unsigned SrcReg, int FI,
+ unsigned SrcReg, bool isKill, int FI,
const TargetRegisterClass *RC) const {
if (RC == ARM::GPRRegisterClass) {
MachineFunction &MF = *MBB.getParent();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
if (AFI->isThumbFunction())
- BuildMI(MBB, I, TII.get(ARM::tSTRspi)).addReg(SrcReg)
+ BuildMI(MBB, I, TII.get(ARM::tSpill)).addReg(SrcReg, false, false, isKill)
.addFrameIndex(FI).addImm(0);
else
- BuildMI(MBB, I, TII.get(ARM::STR)).addReg(SrcReg)
- .addFrameIndex(FI).addReg(0).addImm(0);
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::STR))
+ .addReg(SrcReg, false, false, isKill)
+ .addFrameIndex(FI).addReg(0).addImm(0));
+ } else if (RC == ARM::DPRRegisterClass) {
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FSTD))
+ .addReg(SrcReg, false, false, isKill)
+ .addFrameIndex(FI).addImm(0));
+ } else {
+ assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FSTS))
+ .addReg(SrcReg, false, false, isKill)
+ .addFrameIndex(FI).addImm(0));
+ }
+}
+
+void ARMRegisterInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
+ bool isKill,
+ SmallVectorImpl<MachineOperand> &Addr,
+ const TargetRegisterClass *RC,
+ SmallVectorImpl<MachineInstr*> &NewMIs) const {
+ unsigned Opc = 0;
+ if (RC == ARM::GPRRegisterClass) {
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ if (AFI->isThumbFunction()) {
+ Opc = Addr[0].isFrameIndex() ? ARM::tSpill : ARM::tSTR;
+ MachineInstrBuilder MIB =
+ BuildMI(TII.get(Opc)).addReg(SrcReg, false, false, isKill);
+ for (unsigned i = 0, e = Addr.size(); i != e; ++i)
+ MIB = ARMInstrAddOperand(MIB, Addr[i]);
+ NewMIs.push_back(MIB);
+ return;
+ }
+ Opc = ARM::STR;
} else if (RC == ARM::DPRRegisterClass) {
- BuildMI(MBB, I, TII.get(ARM::FSTD)).addReg(SrcReg)
- .addFrameIndex(FI).addImm(0);
+ Opc = ARM::FSTD;
} else {
assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
- BuildMI(MBB, I, TII.get(ARM::FSTS)).addReg(SrcReg)
- .addFrameIndex(FI).addImm(0);
+ Opc = ARM::FSTS;
}
+
+ MachineInstrBuilder MIB =
+ BuildMI(TII.get(Opc)).addReg(SrcReg, false, false, isKill);
+ for (unsigned i = 0, e = Addr.size(); i != e; ++i)
+ MIB = ARMInstrAddOperand(MIB, Addr[i]);
+ AddDefaultPred(MIB);
+ NewMIs.push_back(MIB);
+ return;
}
void ARMRegisterInfo::
MachineFunction &MF = *MBB.getParent();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
if (AFI->isThumbFunction())
- BuildMI(MBB, I, TII.get(ARM::tLDRspi), DestReg)
+ BuildMI(MBB, I, TII.get(ARM::tRestore), DestReg)
.addFrameIndex(FI).addImm(0);
else
- BuildMI(MBB, I, TII.get(ARM::LDR), DestReg)
- .addFrameIndex(FI).addReg(0).addImm(0);
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::LDR), DestReg)
+ .addFrameIndex(FI).addReg(0).addImm(0));
} else if (RC == ARM::DPRRegisterClass) {
- BuildMI(MBB, I, TII.get(ARM::FLDD), DestReg)
- .addFrameIndex(FI).addImm(0);
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FLDD), DestReg)
+ .addFrameIndex(FI).addImm(0));
} else {
assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
- BuildMI(MBB, I, TII.get(ARM::FLDS), DestReg)
- .addFrameIndex(FI).addImm(0);
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FLDS), DestReg)
+ .addFrameIndex(FI).addImm(0));
}
}
+void ARMRegisterInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
+ SmallVectorImpl<MachineOperand> &Addr,
+ const TargetRegisterClass *RC,
+ SmallVectorImpl<MachineInstr*> &NewMIs) const {
+ unsigned Opc = 0;
+ if (RC == ARM::GPRRegisterClass) {
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ if (AFI->isThumbFunction()) {
+ Opc = Addr[0].isFrameIndex() ? ARM::tRestore : ARM::tLDR;
+ MachineInstrBuilder MIB = BuildMI(TII.get(Opc), DestReg);
+ for (unsigned i = 0, e = Addr.size(); i != e; ++i)
+ MIB = ARMInstrAddOperand(MIB, Addr[i]);
+ NewMIs.push_back(MIB);
+ return;
+ }
+ Opc = ARM::LDR;
+ } else if (RC == ARM::DPRRegisterClass) {
+ Opc = ARM::FLDD;
+ } else {
+ assert(RC == ARM::SPRRegisterClass && "Unknown regclass!");
+ Opc = ARM::FLDS;
+ }
+
+ MachineInstrBuilder MIB = BuildMI(TII.get(Opc), DestReg);
+ for (unsigned i = 0, e = Addr.size(); i != e; ++i)
+ MIB = ARMInstrAddOperand(MIB, Addr[i]);
+ AddDefaultPred(MIB);
+ NewMIs.push_back(MIB);
+ return;
+}
+
void ARMRegisterInfo::copyRegToReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I,
unsigned DestReg, unsigned SrcReg,
- const TargetRegisterClass *RC) const {
- if (RC == ARM::GPRRegisterClass) {
+ const TargetRegisterClass *DestRC,
+ const TargetRegisterClass *SrcRC) const {
+ if (DestRC != SrcRC) {
+ cerr << "Not yet supported!";
+ abort();
+ }
+
+ if (DestRC == ARM::GPRRegisterClass) {
MachineFunction &MF = *MBB.getParent();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
- BuildMI(MBB, I, TII.get(AFI->isThumbFunction() ? ARM::tMOVrr : ARM::MOVrr),
- DestReg).addReg(SrcReg);
- } else if (RC == ARM::SPRRegisterClass)
- BuildMI(MBB, I, TII.get(ARM::FCPYS), DestReg).addReg(SrcReg);
- else if (RC == ARM::DPRRegisterClass)
- BuildMI(MBB, I, TII.get(ARM::FCPYD), DestReg).addReg(SrcReg);
+ if (AFI->isThumbFunction())
+ BuildMI(MBB, I, TII.get(ARM::tMOVr), DestReg).addReg(SrcReg);
+ else
+ AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::MOVr), DestReg)
+ .addReg(SrcReg)));
+ } else if (DestRC == ARM::SPRRegisterClass)
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FCPYS), DestReg)
+ .addReg(SrcReg));
+ else if (DestRC == ARM::DPRRegisterClass)
+ AddDefaultPred(BuildMI(MBB, I, TII.get(ARM::FCPYD), DestReg)
+ .addReg(SrcReg));
else
abort();
}
+/// emitLoadConstPool - Emits a load from constpool to materialize the
+/// specified immediate.
+static void emitLoadConstPool(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator &MBBI,
+ unsigned DestReg, int Val,
+ ARMCC::CondCodes Pred, unsigned PredReg,
+ const TargetInstrInfo &TII, bool isThumb) {
+ MachineFunction &MF = *MBB.getParent();
+ MachineConstantPool *ConstantPool = MF.getConstantPool();
+ Constant *C = ConstantInt::get(Type::Int32Ty, Val);
+ unsigned Idx = ConstantPool->getConstantPoolIndex(C, 2);
+ if (isThumb)
+ BuildMI(MBB, MBBI, TII.get(ARM::tLDRcp), DestReg).addConstantPoolIndex(Idx);
+ else
+ BuildMI(MBB, MBBI, TII.get(ARM::LDRcp), DestReg).addConstantPoolIndex(Idx)
+ .addReg(0).addImm(0).addImm((unsigned)Pred).addReg(PredReg);
+}
+
+void ARMRegisterInfo::reMaterialize(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I,
+ unsigned DestReg,
+ const MachineInstr *Orig) const {
+ if (Orig->getOpcode() == ARM::MOVi2pieces) {
+ emitLoadConstPool(MBB, I, DestReg,
+ Orig->getOperand(1).getImmedValue(),
+ (ARMCC::CondCodes)Orig->getOperand(2).getImmedValue(),
+ Orig->getOperand(3).getReg(),
+ TII, false);
+ return;
+ }
+
+ MachineInstr *MI = Orig->clone();
+ MI->getOperand(0).setReg(DestReg);
+ MBB.insert(I, MI);
+}
+
+/// isLowRegister - Returns true if the register is low register r0-r7.
+///
+static bool isLowRegister(unsigned Reg) {
+ using namespace ARM;
+ switch (Reg) {
+ case R0: case R1: case R2: case R3:
+ case R4: case R5: case R6: case R7:
+ return true;
+ default:
+ return false;
+ }
+}
+
MachineInstr *ARMRegisterInfo::foldMemoryOperand(MachineInstr *MI,
- unsigned OpNum, int FI) const {
+ SmallVectorImpl<unsigned> &Ops,
+ int FI) const {
+ if (Ops.size() != 1) return NULL;
+
+ unsigned OpNum = Ops[0];
unsigned Opc = MI->getOpcode();
MachineInstr *NewMI = NULL;
switch (Opc) {
default: break;
- case ARM::MOVrr: {
+ case ARM::MOVr: {
+ if (MI->getOperand(4).getReg() == ARM::CPSR)
+ // If it is updating CPSR, then it cannot be foled.
+ break;
+ unsigned Pred = MI->getOperand(2).getImmedValue();
+ unsigned PredReg = MI->getOperand(3).getReg();
if (OpNum == 0) { // move -> store
unsigned SrcReg = MI->getOperand(1).getReg();
NewMI = BuildMI(TII.get(ARM::STR)).addReg(SrcReg).addFrameIndex(FI)
- .addReg(0).addImm(0);
+ .addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
} else { // move -> load
unsigned DstReg = MI->getOperand(0).getReg();
NewMI = BuildMI(TII.get(ARM::LDR), DstReg).addFrameIndex(FI).addReg(0)
- .addImm(0);
+ .addImm(0).addImm(Pred).addReg(PredReg);
}
break;
}
- case ARM::tMOVrr: {
+ case ARM::tMOVr: {
if (OpNum == 0) { // move -> store
unsigned SrcReg = MI->getOperand(1).getReg();
- NewMI = BuildMI(TII.get(ARM::tSTRspi)).addReg(SrcReg).addFrameIndex(FI)
+ if (isPhysicalRegister(SrcReg) && !isLowRegister(SrcReg))
+ // tSpill cannot take a high register operand.
+ break;
+ NewMI = BuildMI(TII.get(ARM::tSpill)).addReg(SrcReg).addFrameIndex(FI)
.addImm(0);
} else { // move -> load
unsigned DstReg = MI->getOperand(0).getReg();
- NewMI = BuildMI(TII.get(ARM::tLDRspi), DstReg).addFrameIndex(FI)
+ if (isPhysicalRegister(DstReg) && !isLowRegister(DstReg))
+ // tRestore cannot target a high register operand.
+ break;
+ NewMI = BuildMI(TII.get(ARM::tRestore), DstReg).addFrameIndex(FI)
.addImm(0);
}
break;
}
case ARM::FCPYS: {
+ unsigned Pred = MI->getOperand(2).getImmedValue();
+ unsigned PredReg = MI->getOperand(3).getReg();
if (OpNum == 0) { // move -> store
unsigned SrcReg = MI->getOperand(1).getReg();
NewMI = BuildMI(TII.get(ARM::FSTS)).addReg(SrcReg).addFrameIndex(FI)
- .addImm(0);
+ .addImm(0).addImm(Pred).addReg(PredReg);
} else { // move -> load
unsigned DstReg = MI->getOperand(0).getReg();
- NewMI = BuildMI(TII.get(ARM::FLDS), DstReg).addFrameIndex(FI).addImm(0);
+ NewMI = BuildMI(TII.get(ARM::FLDS), DstReg).addFrameIndex(FI)
+ .addImm(0).addImm(Pred).addReg(PredReg);
}
break;
}
case ARM::FCPYD: {
+ unsigned Pred = MI->getOperand(2).getImmedValue();
+ unsigned PredReg = MI->getOperand(3).getReg();
if (OpNum == 0) { // move -> store
unsigned SrcReg = MI->getOperand(1).getReg();
NewMI = BuildMI(TII.get(ARM::FSTD)).addReg(SrcReg).addFrameIndex(FI)
- .addImm(0);
+ .addImm(0).addImm(Pred).addReg(PredReg);
} else { // move -> load
unsigned DstReg = MI->getOperand(0).getReg();
- NewMI = BuildMI(TII.get(ARM::FLDD), DstReg).addFrameIndex(FI).addImm(0);
+ NewMI = BuildMI(TII.get(ARM::FLDD), DstReg).addFrameIndex(FI)
+ .addImm(0).addImm(Pred).addReg(PredReg);
}
break;
}
return NewMI;
}
-const unsigned* ARMRegisterInfo::getCalleeSavedRegs() const {
+bool ARMRegisterInfo::canFoldMemoryOperand(MachineInstr *MI,
+ SmallVectorImpl<unsigned> &Ops) const {
+ if (Ops.size() != 1) return 0;
+
+ unsigned OpNum = Ops[0];
+ unsigned Opc = MI->getOpcode();
+ switch (Opc) {
+ default: break;
+ case ARM::MOVr:
+ // If it is updating CPSR, then it cannot be foled.
+ return MI->getOperand(4).getReg() != ARM::CPSR;
+ case ARM::tMOVr: {
+ if (OpNum == 0) { // move -> store
+ unsigned SrcReg = MI->getOperand(1).getReg();
+ if (isPhysicalRegister(SrcReg) && !isLowRegister(SrcReg))
+ // tSpill cannot take a high register operand.
+ return false;
+ } else { // move -> load
+ unsigned DstReg = MI->getOperand(0).getReg();
+ if (isPhysicalRegister(DstReg) && !isLowRegister(DstReg))
+ // tRestore cannot target a high register operand.
+ return false;
+ }
+ return true;
+ }
+ case ARM::FCPYS:
+ case ARM::FCPYD:
+ return true;
+ }
+
+ return false;
+}
+
+const unsigned*
+ARMRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
static const unsigned CalleeSavedRegs[] = {
ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8,
ARM::R7, ARM::R6, ARM::R5, ARM::R4,
}
const TargetRegisterClass* const *
-ARMRegisterInfo::getCalleeSavedRegClasses() const {
+ARMRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const {
static const TargetRegisterClass * const CalleeSavedRegClasses[] = {
&ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
&ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass,
return CalleeSavedRegClasses;
}
+BitVector ARMRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
+ // FIXME: avoid re-calculating this everytime.
+ BitVector Reserved(getNumRegs());
+ Reserved.set(ARM::SP);
+ Reserved.set(ARM::PC);
+ if (STI.isTargetDarwin() || hasFP(MF))
+ Reserved.set(FramePtr);
+ // Some targets reserve R9.
+ if (STI.isR9Reserved())
+ Reserved.set(ARM::R9);
+ return Reserved;
+}
+
+bool
+ARMRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const {
+ switch (Reg) {
+ default: break;
+ case ARM::SP:
+ case ARM::PC:
+ return true;
+ case ARM::R7:
+ case ARM::R11:
+ if (FramePtr == Reg && (STI.isTargetDarwin() || hasFP(MF)))
+ return true;
+ break;
+ case ARM::R9:
+ return STI.isR9Reserved();
+ }
+
+ return false;
+}
+
+bool
+ARMRegisterInfo::requiresRegisterScavenging(const MachineFunction &MF) const {
+ const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ return ThumbRegScavenging || !AFI->isThumbFunction();
+}
+
/// hasFP - Return true if the specified function should have a dedicated frame
/// pointer register. This is true if the function has variable sized allocas
/// or if frame pointer elimination is disabled.
return NoFramePointerElim || MF.getFrameInfo()->hasVarSizedObjects();
}
+// hasReservedCallFrame - Under normal circumstances, when a frame pointer is
+// not required, we reserve argument space for call sites in the function
+// immediately on entry to the current function. This eliminates the need for
+// add/sub sp brackets around call sites. Returns true if the call frame is
+// included as part of the stack frame.
+bool ARMRegisterInfo::hasReservedCallFrame(MachineFunction &MF) const {
+ const MachineFrameInfo *FFI = MF.getFrameInfo();
+ unsigned CFSize = FFI->getMaxCallFrameSize();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ // It's not always a good idea to include the call frame as part of the
+ // stack frame. ARM (especially Thumb) has small immediate offset to
+ // address the stack frame. So a large call frame can cause poor codegen
+ // and may even makes it impossible to scavenge a register.
+ if (AFI->isThumbFunction()) {
+ if (CFSize >= ((1 << 8) - 1) * 4 / 2) // Half of imm8 * 4
+ return false;
+ } else {
+ if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12
+ return false;
+ }
+ return !MF.getFrameInfo()->hasVarSizedObjects();
+}
+
/// emitARMRegPlusImmediate - Emits a series of instructions to materialize
/// a destreg = basereg + immediate in ARM code.
static
void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI,
- unsigned DestReg, unsigned BaseReg,
- int NumBytes, const TargetInstrInfo &TII) {
+ unsigned DestReg, unsigned BaseReg, int NumBytes,
+ ARMCC::CondCodes Pred, unsigned PredReg,
+ const TargetInstrInfo &TII) {
bool isSub = NumBytes < 0;
if (isSub) NumBytes = -NumBytes;
// Build the new ADD / SUB.
BuildMI(MBB, MBBI, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg)
- .addReg(BaseReg).addImm(SOImmVal);
+ .addReg(BaseReg, false, false, true).addImm(SOImmVal)
+ .addImm((unsigned)Pred).addReg(PredReg).addReg(0);
BaseReg = DestReg;
}
}
-/// isLowRegister - Returns true if the register is low register r0-r7.
-///
-static bool isLowRegister(unsigned Reg) {
- using namespace ARM;
- switch (Reg) {
- case R0: case R1: case R2: case R3:
- case R4: case R5: case R6: case R7:
- return true;
- default:
- return false;
- }
-}
-
/// calcNumMI - Returns the number of instructions required to materialize
/// the specific add / sub r, c instruction.
static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes,
Bytes -= ThisVal;
NumMIs++;
NumBits = 8;
- Scale = 1;
+ Scale = 1; // Followed by a number of tADDi8.
Chunk = ((1 << NumBits) - 1) * Scale;
}
return NumMIs;
}
-/// emitThumbRegPlusConstPool - Emits a series of instructions to materialize
-/// a destreg = basereg + immediate in Thumb code. Load the immediate from a
+/// emitThumbRegPlusImmInReg - Emits a series of instructions to materialize
+/// a destreg = basereg + immediate in Thumb code. Materialize the immediate
+/// in a register using mov / mvn sequences or load the immediate from a
/// constpool entry.
static
-void emitThumbRegPlusConstPool(MachineBasicBlock &MBB,
+void emitThumbRegPlusImmInReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI,
unsigned DestReg, unsigned BaseReg,
- int NumBytes, const TargetInstrInfo &TII) {
- MachineFunction &MF = *MBB.getParent();
- MachineConstantPool *ConstantPool = MF.getConstantPool();
- bool isHigh = !isLowRegister(DestReg) || !isLowRegister(BaseReg);
+ int NumBytes, bool CanChangeCC,
+ const TargetInstrInfo &TII) {
+ bool isHigh = !isLowRegister(DestReg) ||
+ (BaseReg != 0 && !isLowRegister(BaseReg));
bool isSub = false;
// Subtract doesn't have high register version. Load the negative value
- // if either base or dest register is a high register.
- if (NumBytes < 0 && !isHigh) {
+ // if either base or dest register is a high register. Also, if do not
+ // issue sub as part of the sequence if condition register is to be
+ // preserved.
+ if (NumBytes < 0 && !isHigh && CanChangeCC) {
isSub = true;
NumBytes = -NumBytes;
}
- Constant *C = ConstantInt::get(Type::Int32Ty, NumBytes);
- unsigned Idx = ConstantPool->getConstantPoolIndex(C, 2);
unsigned LdReg = DestReg;
if (DestReg == ARM::SP) {
assert(BaseReg == ARM::SP && "Unexpected!");
LdReg = ARM::R3;
- BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R12).addReg(ARM::R3);
+ BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::R12)
+ .addReg(ARM::R3, false, false, true);
}
- // Load the constant.
- BuildMI(MBB, MBBI, TII.get(ARM::tLDRpci), LdReg).addConstantPoolIndex(Idx);
+
+ if (NumBytes <= 255 && NumBytes >= 0)
+ BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
+ else if (NumBytes < 0 && NumBytes >= -255) {
+ BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes);
+ BuildMI(MBB, MBBI, TII.get(ARM::tNEG), LdReg)
+ .addReg(LdReg, false, false, true);
+ } else
+ emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, ARMCC::AL, 0, TII, true);
+
// Emit add / sub.
int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr);
const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, TII.get(Opc), DestReg);
- if (DestReg == ARM::SP)
- MIB.addReg(BaseReg).addReg(LdReg);
+ if (DestReg == ARM::SP || isSub)
+ MIB.addReg(BaseReg).addReg(LdReg, false, false, true);
else
- MIB.addReg(LdReg).addReg(BaseReg);
+ MIB.addReg(LdReg).addReg(BaseReg, false, false, true);
if (DestReg == ARM::SP)
- BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), ARM::R3).addReg(ARM::R12);
+ BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::R3)
+ .addReg(ARM::R12, false, false, true);
}
/// emitThumbRegPlusImmediate - Emits a series of instructions to materialize
if (isSub) Bytes = -NumBytes;
bool isMul4 = (Bytes & 3) == 0;
bool isTwoAddr = false;
- bool DstNeBase = false;
+ bool DstNotEqBase = false;
unsigned NumBits = 1;
unsigned Scale = 1;
int Opc = 0;
// r1 = sub sp, c
// r8 = sub sp, c
if (DestReg != BaseReg)
- DstNeBase = true;
+ DstNotEqBase = true;
NumBits = 8;
Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8;
isTwoAddr = true;
}
unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale);
- unsigned Threshold = (DestReg == ARM::SP) ? 4 : 3;
+ unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2;
if (NumMIs > Threshold) {
// This will expand into too many instructions. Load the immediate from a
// constpool entry.
- emitThumbRegPlusConstPool(MBB, MBBI, DestReg, BaseReg, NumBytes, TII);
+ emitThumbRegPlusImmInReg(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII);
return;
}
- if (DstNeBase) {
+ if (DstNotEqBase) {
if (isLowRegister(DestReg) && isLowRegister(BaseReg)) {
// If both are low registers, emit DestReg = add BaseReg, max(Imm, 7)
unsigned Chunk = (1 << 3) - 1;
unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes;
Bytes -= ThisVal;
BuildMI(MBB, MBBI, TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg)
- .addReg(BaseReg).addImm(ThisVal);
+ .addReg(BaseReg, false, false, true).addImm(ThisVal);
} else {
- BuildMI(MBB, MBBI, TII.get(ARM::tMOVrr), DestReg).addReg(BaseReg);
+ BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), DestReg)
+ .addReg(BaseReg, false, false, true);
}
BaseReg = DestReg;
}
if (isTwoAddr)
BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(DestReg).addImm(ThisVal);
else {
- BuildMI(MBB, MBBI, TII.get(Opc), DestReg).addReg(BaseReg).addImm(ThisVal);
+ bool isKill = BaseReg != ARM::SP;
+ BuildMI(MBB, MBBI, TII.get(Opc), DestReg)
+ .addReg(BaseReg, false, false, isKill).addImm(ThisVal);
BaseReg = DestReg;
if (Opc == ARM::tADDrSPi) {
}
if (ExtraOpc)
- BuildMI(MBB, MBBI, TII.get(ExtraOpc), DestReg).addReg(DestReg)
+ BuildMI(MBB, MBBI, TII.get(ExtraOpc), DestReg)
+ .addReg(DestReg, false, false, true)
.addImm(((unsigned)NumBytes) & 3);
}
static
void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI,
- int NumBytes, bool isThumb, const TargetInstrInfo &TII) {
+ int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg,
+ bool isThumb, const TargetInstrInfo &TII) {
if (isThumb)
emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
else
- emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII);
+ emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes,
+ Pred, PredReg, TII);
}
void ARMRegisterInfo::
eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
MachineBasicBlock::iterator I) const {
- if (hasFP(MF)) {
+ if (!hasReservedCallFrame(MF)) {
// If we have alloca, convert as follows:
// ADJCALLSTACKDOWN -> sub, sp, sp, amount
// ADJCALLSTACKUP -> add, sp, sp, amount
Amount = (Amount+Align-1)/Align*Align;
// Replace the pseudo instruction with a new instruction...
- if (Old->getOpcode() == ARM::ADJCALLSTACKDOWN) {
- emitSPUpdate(MBB, I, -Amount, AFI->isThumbFunction(), TII);
+ unsigned Opc = Old->getOpcode();
+ bool isThumb = AFI->isThumbFunction();
+ ARMCC::CondCodes Pred = isThumb
+ ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(1).getImmedValue();
+ if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) {
+ // Note: PredReg is operand 2 for ADJCALLSTACKDOWN.
+ unsigned PredReg = isThumb ? 0 : Old->getOperand(2).getReg();
+ emitSPUpdate(MBB, I, -Amount, Pred, PredReg, isThumb, TII);
} else {
- assert(Old->getOpcode() == ARM::ADJCALLSTACKUP);
- emitSPUpdate(MBB, I, Amount, AFI->isThumbFunction(), TII);
+ // Note: PredReg is operand 3 for ADJCALLSTACKUP.
+ unsigned PredReg = isThumb ? 0 : Old->getOperand(3).getReg();
+ assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP);
+ emitSPUpdate(MBB, I, Amount, Pred, PredReg, isThumb, TII);
}
}
}
int Chunk = (1 << 8) - 1;
int ThisVal = (Imm > Chunk) ? Chunk : Imm;
Imm -= ThisVal;
- BuildMI(MBB, MBBI, TII.get(ARM::tMOVri8), DestReg).addImm(ThisVal);
+ BuildMI(MBB, MBBI, TII.get(ARM::tMOVi8), DestReg).addImm(ThisVal);
if (Imm > 0)
emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII);
if (isSub)
- BuildMI(MBB, MBBI, TII.get(ARM::tNEG), DestReg).addReg(DestReg);
+ BuildMI(MBB, MBBI, TII.get(ARM::tNEG), DestReg)
+ .addReg(DestReg, false, false, true);
}
-void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II) const{
+/// findScratchRegister - Find a 'free' ARM register. If register scavenger
+/// is not being used, R12 is available. Otherwise, try for a call-clobbered
+/// register first and then a spilled callee-saved register if that fails.
+static
+unsigned findScratchRegister(RegScavenger *RS, const TargetRegisterClass *RC,
+ ARMFunctionInfo *AFI) {
+ unsigned Reg = RS ? RS->FindUnusedReg(RC, true) : (unsigned) ARM::R12;
+ if (Reg == 0)
+ // Try a already spilled CS register.
+ Reg = RS->FindUnusedReg(RC, AFI->getSpilledCSRegisters());
+
+ return Reg;
+}
+
+void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
+ int SPAdj, RegScavenger *RS) const{
unsigned i = 0;
MachineInstr &MI = *II;
MachineBasicBlock &MBB = *MI.getParent();
unsigned FrameReg = ARM::SP;
int FrameIndex = MI.getOperand(i).getFrameIndex();
int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
- MF.getFrameInfo()->getStackSize();
+ MF.getFrameInfo()->getStackSize() + SPAdj;
if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
Offset -= AFI->getGPRCalleeSavedArea1Offset();
else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
Offset -= AFI->getDPRCalleeSavedAreaOffset();
else if (hasFP(MF)) {
+ assert(SPAdj == 0 && "Unexpected");
// There is alloca()'s in this function, must reference off the frame
// pointer instead.
FrameReg = getFrameRegister(MF);
const TargetInstrDescriptor &Desc = TII.get(Opcode);
unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
bool isSub = false;
-
+
if (Opcode == ARM::ADDri) {
Offset += MI.getOperand(i+1).getImm();
if (Offset == 0) {
// Turn it into a move.
- MI.setInstrDescriptor(TII.get(ARM::MOVrr));
+ MI.setInstrDescriptor(TII.get(ARM::MOVr));
MI.getOperand(i).ChangeToRegister(FrameReg, false);
MI.RemoveOperand(i+1);
return;
// a sequence of ADDri instructions. First though, pull as much of the imm
// into this ADDri as possible.
unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
- unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, (32-RotAmt) & 31);
+ unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
// We will handle these bits from offset, clear them.
Offset &= ~ThisImmVal;
MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal);
} else if (Opcode == ARM::tADDrSPi) {
Offset += MI.getOperand(i+1).getImm();
- assert((Offset & 3) == 0 &&
- "Thumb add/sub sp, #imm immediate must be multiple of 4!");
- Offset >>= 2;
+
+ // Can't use tADDrSPi if it's based off the frame pointer.
+ unsigned NumBits = 0;
+ unsigned Scale = 1;
+ if (FrameReg != ARM::SP) {
+ Opcode = ARM::tADDi3;
+ MI.setInstrDescriptor(TII.get(ARM::tADDi3));
+ NumBits = 3;
+ } else {
+ NumBits = 8;
+ Scale = 4;
+ assert((Offset & 3) == 0 &&
+ "Thumb add/sub sp, #imm immediate must be multiple of 4!");
+ }
+
if (Offset == 0) {
// Turn it into a move.
- MI.setInstrDescriptor(TII.get(ARM::tMOVrr));
+ MI.setInstrDescriptor(TII.get(ARM::tMOVr));
MI.getOperand(i).ChangeToRegister(FrameReg, false);
MI.RemoveOperand(i+1);
return;
}
// Common case: small offset, fits into instruction.
- if ((Offset & ~255U) == 0) {
+ unsigned Mask = (1 << NumBits) - 1;
+ if (((Offset / Scale) & ~Mask) == 0) {
// Replace the FrameIndex with sp / fp
MI.getOperand(i).ChangeToRegister(FrameReg, false);
- MI.getOperand(i+1).ChangeToImmediate(Offset);
+ MI.getOperand(i+1).ChangeToImmediate(Offset / Scale);
return;
}
unsigned DestReg = MI.getOperand(0).getReg();
+ unsigned Bytes = (Offset > 0) ? Offset : -Offset;
+ unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, NumBits, Scale);
+ // MI would expand into a large number of instructions. Don't try to
+ // simplify the immediate.
+ if (NumMIs > 2) {
+ emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII);
+ MBB.erase(II);
+ return;
+ }
+
if (Offset > 0) {
// Translate r0 = add sp, imm to
// r0 = add sp, 255*4
// r0 = add r0, (imm - 255*4)
MI.getOperand(i).ChangeToRegister(FrameReg, false);
- MI.getOperand(i+1).ChangeToImmediate(255);
- Offset = (Offset - 255) << 2;
+ MI.getOperand(i+1).ChangeToImmediate(Mask);
+ Offset = (Offset - Mask * Scale);
MachineBasicBlock::iterator NII = next(II);
emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII);
} else {
// Translate r0 = add sp, -imm to
// r0 = -imm (this is then translated into a series of instructons)
// r0 = add r0, sp
- Offset <<= 2;
emitThumbConstant(MBB, II, DestReg, Offset, TII);
MI.setInstrDescriptor(TII.get(ARM::tADDhirr));
- MI.getOperand(i).ChangeToRegister(DestReg, false);
+ MI.getOperand(i).ChangeToRegister(DestReg, false, false, true);
MI.getOperand(i+1).ChangeToRegister(FrameReg, false);
}
return;
case ARMII::AddrModeTs: {
ImmIdx = i+1;
InstrOffs = MI.getOperand(ImmIdx).getImm();
- NumBits = 8;
+ NumBits = (FrameReg == ARM::SP) ? 8 : 5;
Scale = 4;
break;
}
default:
- std::cerr << "Unsupported addressing mode!\n";
+ assert(0 && "Unsupported addressing mode!");
abort();
break;
}
Offset += InstrOffs * Scale;
- assert((Scale == 1 || (Offset & (Scale-1)) == 0) &&
- "Can't encode this offset!");
- if (Offset < 0) {
+ assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
+ if (Offset < 0 && !isThumb) {
Offset = -Offset;
isSub = true;
}
+ // Common case: small offset, fits into instruction.
MachineOperand &ImmOp = MI.getOperand(ImmIdx);
int ImmedOffset = Offset / Scale;
unsigned Mask = (1 << NumBits) - 1;
return;
}
- // Otherwise, it didn't fit. Pull in what we can to simplify the immediate.
+ bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill;
if (AddrMode == ARMII::AddrModeTs) {
- // Thumb tLDRspi, tSTRspi. These will change to instructions that use a
- // different base register.
+ // Thumb tLDRspi, tSTRspi. These will change to instructions that use
+ // a different base register.
NumBits = 5;
Mask = (1 << NumBits) - 1;
}
-
- ImmedOffset = ImmedOffset & Mask;
- if (isSub)
- ImmedOffset |= 1 << NumBits;
- ImmOp.ChangeToImmediate(ImmedOffset);
- Offset &= ~(Mask*Scale);
+ // If this is a thumb spill / restore, we will be using a constpool load to
+ // materialize the offset.
+ if (AddrMode == ARMII::AddrModeTs && isThumSpillRestore)
+ ImmOp.ChangeToImmediate(0);
+ else {
+ // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
+ ImmedOffset = ImmedOffset & Mask;
+ if (isSub)
+ ImmedOffset |= 1 << NumBits;
+ ImmOp.ChangeToImmediate(ImmedOffset);
+ Offset &= ~(Mask*Scale);
+ }
}
// If we get here, the immediate doesn't fit into the instruction. We folded
if (TII.isLoad(Opcode)) {
// Use the destination register to materialize sp + offset.
unsigned TmpReg = MI.getOperand(0).getReg();
- emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg,
- isSub ? -Offset : Offset, TII);
+ bool UseRR = false;
+ if (Opcode == ARM::tRestore) {
+ if (FrameReg == ARM::SP)
+ emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
+ else {
+ emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, TII, true);
+ UseRR = true;
+ }
+ } else
+ emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
MI.setInstrDescriptor(TII.get(ARM::tLDR));
- MI.getOperand(i).ChangeToRegister(TmpReg, false);
- MI.addRegOperand(0, false); // tLDR has an extra register operand.
+ MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
+ if (UseRR)
+ MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
+ else
+ MI.addRegOperand(0, false); // tLDR has an extra register operand.
} else if (TII.isStore(Opcode)) {
// FIXME! This is horrific!!! We need register scavenging.
// Our temporary workaround has marked r3 unavailable. Of course, r3 is
// being storing here. If that's the case, we do the following:
// r12 = r2
// Use r2 to materialize sp + offset
- // str r12, r2
+ // str r3, r2
// r2 = r12
unsigned ValReg = MI.getOperand(0).getReg();
unsigned TmpReg = ARM::R3;
+ bool UseRR = false;
if (ValReg == ARM::R3) {
- BuildMI(MBB, II, TII.get(ARM::tMOVrr), ARM::R12).addReg(ARM::R2);
+ BuildMI(MBB, II, TII.get(ARM::tMOVr), ARM::R12)
+ .addReg(ARM::R2, false, false, true);
TmpReg = ARM::R2;
}
- emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg,
- isSub ? -Offset : Offset, TII);
+ if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
+ BuildMI(MBB, II, TII.get(ARM::tMOVr), ARM::R12)
+ .addReg(ARM::R3, false, false, true);
+ if (Opcode == ARM::tSpill) {
+ if (FrameReg == ARM::SP)
+ emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg,Offset,false,TII);
+ else {
+ emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, TII, true);
+ UseRR = true;
+ }
+ } else
+ emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII);
MI.setInstrDescriptor(TII.get(ARM::tSTR));
- MI.getOperand(i).ChangeToRegister(TmpReg, false);
- MI.addRegOperand(0, false); // tSTR has an extra register operand.
+ MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true);
+ if (UseRR)
+ MI.addRegOperand(FrameReg, false); // Use [reg, reg] addrmode.
+ else
+ MI.addRegOperand(0, false); // tSTR has an extra register operand.
+
+ MachineBasicBlock::iterator NII = next(II);
if (ValReg == ARM::R3)
- BuildMI(MBB, II, TII.get(ARM::tMOVrr), ARM::R2).addReg(ARM::R12);
+ BuildMI(MBB, NII, TII.get(ARM::tMOVr), ARM::R2)
+ .addReg(ARM::R12, false, false, true);
+ if (TmpReg == ARM::R3 && AFI->isR3LiveIn())
+ BuildMI(MBB, NII, TII.get(ARM::tMOVr), ARM::R3)
+ .addReg(ARM::R12, false, false, true);
} else
assert(false && "Unexpected opcode!");
} else {
// If the offset we have is too large to fit into the instruction, we need
// to form it with a series of ADDri's. Do this by taking 8-bit chunks
// out of 'Offset'.
- emitARMRegPlusImmediate(MBB, II, ARM::R12, FrameReg,
- isSub ? -Offset : Offset, TII);
- MI.getOperand(i).ChangeToRegister(ARM::R12, false);
+ unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI);
+ if (ScratchReg == 0)
+ // No register is "free". Scavenge a register.
+ ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II, SPAdj);
+ int PIdx = MI.findFirstPredOperandIdx();
+ ARMCC::CondCodes Pred = (PIdx == -1)
+ ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImmedValue();
+ unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
+ emitARMRegPlusImmediate(MBB, II, ScratchReg, FrameReg,
+ isSub ? -Offset : Offset, Pred, PredReg, TII);
+ MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
}
}
-void ARMRegisterInfo::
-processFunctionBeforeCalleeSavedScan(MachineFunction &MF) const {
+static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) {
+ const MachineFrameInfo *FFI = MF.getFrameInfo();
+ int Offset = 0;
+ for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
+ int FixedOff = -FFI->getObjectOffset(i);
+ if (FixedOff > Offset) Offset = FixedOff;
+ }
+ for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
+ Offset += FFI->getObjectSize(i);
+ unsigned Align = FFI->getObjectAlignment(i);
+ // Adjust to alignment boundary
+ Offset = (Offset+Align-1)/Align*Align;
+ }
+ return (unsigned)Offset;
+}
+
+void
+ARMRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
+ RegScavenger *RS) const {
// This tells PEI to spill the FP as if it is any other callee-save register
// to take advantage the eliminateFrameIndex machinery. This also ensures it
// is spilled in the order specified by getCalleeSavedRegs() to make it easier
unsigned NumGPRSpills = 0;
SmallVector<unsigned, 4> UnspilledCS1GPRs;
SmallVector<unsigned, 4> UnspilledCS2GPRs;
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
// Don't spill FP if the frame can be eliminated. This is determined
// by scanning the callee-save registers to see if any is used.
unsigned Reg = CSRegs[i];
bool Spilled = false;
if (MF.isPhysRegUsed(Reg)) {
+ AFI->setCSRegisterIsSpilled(Reg);
Spilled = true;
CanEliminateFrame = false;
} else {
if (!STI.isTargetDarwin()) {
if (Reg == ARM::LR)
LRSpilled = true;
- else
- CS1Spilled = true;
+ CS1Spilled = true;
continue;
}
}
}
- ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
bool ForceLRSpill = false;
if (!LRSpilled && AFI->isThumbFunction()) {
unsigned FnSize = ARM::GetFunctionSize(MF);
- // Force LR spill if the Thumb function size is > 2048. This enables the
+ // Force LR to be spilled if the Thumb function size is > 2048. This enables
// use of BL to implement far jump. If it turns out that it's not needed
- // the branch fix up path will undo it.
+ // then the branch fix up path will undo it.
if (FnSize >= (1 << 11)) {
CanEliminateFrame = false;
ForceLRSpill = true;
}
}
+ bool ExtraCSSpill = false;
if (!CanEliminateFrame || hasFP(MF)) {
AFI->setHasStackFrame(true);
// If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled.
// Spill LR as well so we can fold BX_RET to the registers restore (LDM).
if (!LRSpilled && CS1Spilled) {
- MF.changePhyRegUsed(ARM::LR, true);
+ MF.setPhysRegUsed(ARM::LR);
+ AFI->setCSRegisterIsSpilled(ARM::LR);
NumGPRSpills++;
UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(),
UnspilledCS1GPRs.end(), (unsigned)ARM::LR));
ForceLRSpill = false;
+ ExtraCSSpill = true;
}
// Darwin ABI requires FP to point to the stack slot that contains the
// previous FP.
if (STI.isTargetDarwin() || hasFP(MF)) {
- MF.changePhyRegUsed(FramePtr, true);
+ MF.setPhysRegUsed(FramePtr);
NumGPRSpills++;
}
// the integer and double callee save areas.
unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
if (TargetAlign == 8 && (NumGPRSpills & 1)) {
- if (CS1Spilled && !UnspilledCS1GPRs.empty())
- MF.changePhyRegUsed(UnspilledCS1GPRs.front(), true);
- else if (!UnspilledCS2GPRs.empty())
- MF.changePhyRegUsed(UnspilledCS2GPRs.front(), true);
+ if (CS1Spilled && !UnspilledCS1GPRs.empty()) {
+ for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) {
+ unsigned Reg = UnspilledCS1GPRs[i];
+ // Don't spiil high register if the function is thumb
+ if (!AFI->isThumbFunction() || isLowRegister(Reg) || Reg == ARM::LR) {
+ MF.setPhysRegUsed(Reg);
+ AFI->setCSRegisterIsSpilled(Reg);
+ if (!isReservedReg(MF, Reg))
+ ExtraCSSpill = true;
+ break;
+ }
+ }
+ } else if (!UnspilledCS2GPRs.empty() &&
+ !AFI->isThumbFunction()) {
+ unsigned Reg = UnspilledCS2GPRs.front();
+ MF.setPhysRegUsed(Reg);
+ AFI->setCSRegisterIsSpilled(Reg);
+ if (!isReservedReg(MF, Reg))
+ ExtraCSSpill = true;
+ }
+ }
+
+ // Estimate if we might need to scavenge a register at some point in order
+ // to materialize a stack offset. If so, either spill one additiona
+ // callee-saved register or reserve a special spill slot to facilitate
+ // register scavenging.
+ if (RS && !ExtraCSSpill && !AFI->isThumbFunction()) {
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+ unsigned Size = estimateStackSize(MF, MFI);
+ unsigned Limit = (1 << 12) - 1;
+ for (MachineFunction::iterator BB = MF.begin(),E = MF.end();BB != E; ++BB)
+ for (MachineBasicBlock::iterator I= BB->begin(); I != BB->end(); ++I) {
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
+ if (I->getOperand(i).isFrameIndex()) {
+ unsigned Opcode = I->getOpcode();
+ const TargetInstrDescriptor &Desc = TII.get(Opcode);
+ unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
+ if (AddrMode == ARMII::AddrMode3) {
+ Limit = (1 << 8) - 1;
+ goto DoneEstimating;
+ } else if (AddrMode == ARMII::AddrMode5) {
+ unsigned ThisLimit = ((1 << 8) - 1) * 4;
+ if (ThisLimit < Limit)
+ Limit = ThisLimit;
+ }
+ }
+ }
+ DoneEstimating:
+ if (Size >= Limit) {
+ // If any non-reserved CS register isn't spilled, just spill one or two
+ // extra. That should take care of it!
+ unsigned NumExtras = TargetAlign / 4;
+ SmallVector<unsigned, 2> Extras;
+ while (NumExtras && !UnspilledCS1GPRs.empty()) {
+ unsigned Reg = UnspilledCS1GPRs.back();
+ UnspilledCS1GPRs.pop_back();
+ if (!isReservedReg(MF, Reg)) {
+ Extras.push_back(Reg);
+ NumExtras--;
+ }
+ }
+ while (NumExtras && !UnspilledCS2GPRs.empty()) {
+ unsigned Reg = UnspilledCS2GPRs.back();
+ UnspilledCS2GPRs.pop_back();
+ if (!isReservedReg(MF, Reg)) {
+ Extras.push_back(Reg);
+ NumExtras--;
+ }
+ }
+ if (Extras.size() && NumExtras == 0) {
+ for (unsigned i = 0, e = Extras.size(); i != e; ++i) {
+ MF.setPhysRegUsed(Extras[i]);
+ AFI->setCSRegisterIsSpilled(Extras[i]);
+ }
+ } else {
+ // Reserve a slot closest to SP or frame pointer.
+ const TargetRegisterClass *RC = &ARM::GPRRegClass;
+ RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
+ RC->getAlignment()));
+ }
+ }
}
}
if (ForceLRSpill) {
- MF.changePhyRegUsed(ARM::LR, true);
- AFI->setLRIsForceSpilled(true);
+ MF.setPhysRegUsed(ARM::LR);
+ AFI->setCSRegisterIsSpilled(ARM::LR);
+ AFI->setLRIsSpilledForFarJump(true);
}
}
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
bool isThumb = AFI->isThumbFunction();
unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize();
- unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
unsigned NumBytes = MFI->getStackSize();
const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo();
if (isThumb) {
+ // Check if R3 is live in. It might have to be used as a scratch register.
+ for (MachineFunction::livein_iterator I=MF.livein_begin(),E=MF.livein_end();
+ I != E; ++I) {
+ if ((*I).first == ARM::R3) {
+ AFI->setR3IsLiveIn(true);
+ break;
+ }
+ }
+
// Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4.
NumBytes = (NumBytes + 3) & ~3;
MFI->setStackSize(NumBytes);
// belongs to which callee-save spill areas.
unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
int FramePtrSpillFI = 0;
+
+ if (VARegSaveSize)
+ emitSPUpdate(MBB, MBBI, -VARegSaveSize, ARMCC::AL, 0, isThumb, TII);
+
if (!AFI->hasStackFrame()) {
if (NumBytes != 0)
- emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
+ emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII);
return;
}
- if (VARegSaveSize)
- emitSPUpdate(MBB, MBBI, -VARegSaveSize, isThumb, TII);
-
for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
unsigned Reg = CSI[i].getReg();
int FI = CSI[i].getFrameIdx();
}
}
- if (Align == 8 && (GPRCS1Size & 7) != 0)
- // Pad CS1 to ensure proper alignment.
- GPRCS1Size += 4;
-
if (!isThumb) {
// Build the new SUBri to adjust SP for integer callee-save spill area 1.
- emitSPUpdate(MBB, MBBI, -GPRCS1Size, isThumb, TII);
+ emitSPUpdate(MBB, MBBI, -GPRCS1Size, ARMCC::AL, 0, isThumb, TII);
movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI);
} else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH)
++MBBI;
// Darwin ABI requires FP to point to the stack slot that contains the
// previous FP.
- if (STI.isTargetDarwin() || hasFP(MF))
- BuildMI(MBB, MBBI, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri), FramePtr)
+ if (STI.isTargetDarwin() || hasFP(MF)) {
+ MachineInstrBuilder MIB =
+ BuildMI(MBB, MBBI, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri),FramePtr)
.addFrameIndex(FramePtrSpillFI).addImm(0);
+ if (!isThumb) AddDefaultCC(AddDefaultPred(MIB));
+ }
if (!isThumb) {
// Build the new SUBri to adjust SP for integer callee-save spill area 2.
- emitSPUpdate(MBB, MBBI, -GPRCS2Size, false, TII);
+ emitSPUpdate(MBB, MBBI, -GPRCS2Size, ARMCC::AL, 0, false, TII);
// Build the new SUBri to adjust SP for FP callee-save spill area.
movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI);
- emitSPUpdate(MBB, MBBI, -DPRCSSize, false, TII);
+ emitSPUpdate(MBB, MBBI, -DPRCSSize, ARMCC::AL, 0, false, TII);
}
// Determine starting offsets of spill areas.
// Insert it after all the callee-save spills.
if (!isThumb)
movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI);
- emitSPUpdate(MBB, MBBI, -NumBytes, isThumb, TII);
+ emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII);
+ }
+
+ if(STI.isTargetELF() && hasFP(MF)) {
+ MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() -
+ AFI->getFramePtrSpillOffset());
}
AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) {
return ((MI->getOpcode() == ARM::FLDD ||
MI->getOpcode() == ARM::LDR ||
- MI->getOpcode() == ARM::tLDRspi) &&
+ MI->getOpcode() == ARM::tRestore) &&
MI->getOperand(1).isFrameIndex() &&
isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs));
}
void ARMRegisterInfo::emitEpilogue(MachineFunction &MF,
- MachineBasicBlock &MBB) const {
+ MachineBasicBlock &MBB) const {
MachineBasicBlock::iterator MBBI = prior(MBB.end());
assert((MBBI->getOpcode() == ARM::BX_RET ||
MBBI->getOpcode() == ARM::tBX_RET ||
int NumBytes = (int)MFI->getStackSize();
if (!AFI->hasStackFrame()) {
if (NumBytes != 0)
- emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
- return;
- }
-
- // Unwind MBBI to point to first LDR / FLDD.
- const unsigned *CSRegs = getCalleeSavedRegs();
- if (MBBI != MBB.begin()) {
- do
- --MBBI;
- while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
- if (!isCSRestore(MBBI, CSRegs))
- ++MBBI;
- }
-
- // Move SP to start of FP callee save spill area.
- NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
- AFI->getGPRCalleeSavedArea2Size() +
- AFI->getDPRCalleeSavedAreaSize());
- if (isThumb) {
- if (MBBI->getOpcode() == ARM::tBX_RET &&
- &MBB.front() != MBBI &&
- prior(MBBI)->getOpcode() == ARM::tPOP) {
- MachineBasicBlock::iterator PMBBI = prior(MBBI);
- emitSPUpdate(MBB, PMBBI, NumBytes, isThumb, TII);
- } else
- emitSPUpdate(MBB, MBBI, NumBytes, isThumb, TII);
+ emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
} else {
- // Darwin ABI requires FP to point to the stack slot that contains the
- // previous FP.
- if (STI.isTargetDarwin() || hasFP(MF)) {
- NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
- // Reset SP based on frame pointer only if the stack frame extends beyond
- // frame pointer stack slot.
- if (AFI->getGPRCalleeSavedArea2Size() ||
- AFI->getDPRCalleeSavedAreaSize() ||
- AFI->getDPRCalleeSavedAreaOffset())
+ // Unwind MBBI to point to first LDR / FLDD.
+ const unsigned *CSRegs = getCalleeSavedRegs();
+ if (MBBI != MBB.begin()) {
+ do
+ --MBBI;
+ while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs));
+ if (!isCSRestore(MBBI, CSRegs))
+ ++MBBI;
+ }
+
+ // Move SP to start of FP callee save spill area.
+ NumBytes -= (AFI->getGPRCalleeSavedArea1Size() +
+ AFI->getGPRCalleeSavedArea2Size() +
+ AFI->getDPRCalleeSavedAreaSize());
+ if (isThumb) {
+ if (hasFP(MF)) {
+ NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
+ // Reset SP based on frame pointer only if the stack frame extends beyond
+ // frame pointer stack slot or target is ELF and the function has FP.
if (NumBytes)
- BuildMI(MBB, MBBI, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
- .addImm(NumBytes);
+ emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes, TII);
else
- BuildMI(MBB, MBBI, TII.get(ARM::MOVrr), ARM::SP).addReg(FramePtr);
- } else if (NumBytes) {
- emitSPUpdate(MBB, MBBI, NumBytes, false, TII);
- }
+ BuildMI(MBB, MBBI, TII.get(ARM::tMOVr), ARM::SP).addReg(FramePtr);
+ } else {
+ if (MBBI->getOpcode() == ARM::tBX_RET &&
+ &MBB.front() != MBBI &&
+ prior(MBBI)->getOpcode() == ARM::tPOP) {
+ MachineBasicBlock::iterator PMBBI = prior(MBBI);
+ emitSPUpdate(MBB, PMBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
+ } else
+ emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII);
+ }
+ } else {
+ // Darwin ABI requires FP to point to the stack slot that contains the
+ // previous FP.
+ if ((STI.isTargetDarwin() && NumBytes) || hasFP(MF)) {
+ NumBytes = AFI->getFramePtrSpillOffset() - NumBytes;
+ // Reset SP based on frame pointer only if the stack frame extends beyond
+ // frame pointer stack slot or target is ELF and the function has FP.
+ if (AFI->getGPRCalleeSavedArea2Size() ||
+ AFI->getDPRCalleeSavedAreaSize() ||
+ AFI->getDPRCalleeSavedAreaOffset()||
+ hasFP(MF))
+ if (NumBytes)
+ BuildMI(MBB, MBBI, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr)
+ .addImm(NumBytes)
+ .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
+ else
+ BuildMI(MBB, MBBI, TII.get(ARM::MOVr), ARM::SP).addReg(FramePtr)
+ .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
+ } else if (NumBytes) {
+ emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, false, TII);
+ }
- // Move SP to start of integer callee save spill area 2.
- movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
- emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), false, TII);
+ // Move SP to start of integer callee save spill area 2.
+ movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI);
+ emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), ARMCC::AL, 0,
+ false, TII);
- // Move SP to start of integer callee save spill area 1.
- movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
- emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), false, TII);
+ // Move SP to start of integer callee save spill area 1.
+ movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI);
+ emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), ARMCC::AL, 0,
+ false, TII);
- // Move SP to SP upon entry to the function.
- movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
- emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), false, TII);
+ // Move SP to SP upon entry to the function.
+ movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI);
+ emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), ARMCC::AL, 0,
+ false, TII);
+ }
}
if (VARegSaveSize) {
- // Epilogue for vararg functions: pop LR to R3 and branch off it.
- // FIXME: Verify this is still ok when R3 is no longer being reserved.
- BuildMI(MBB, MBBI, TII.get(ARM::tPOP)).addReg(ARM::R3);
- emitSPUpdate(MBB, MBBI, VARegSaveSize, isThumb, TII);
- BuildMI(MBB, MBBI, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3);
- MBB.erase(MBBI);
+ if (isThumb)
+ // Epilogue for vararg functions: pop LR to R3 and branch off it.
+ // FIXME: Verify this is still ok when R3 is no longer being reserved.
+ BuildMI(MBB, MBBI, TII.get(ARM::tPOP)).addReg(ARM::R3);
+
+ emitSPUpdate(MBB, MBBI, VARegSaveSize, ARMCC::AL, 0, isThumb, TII);
+
+ if (isThumb) {
+ BuildMI(MBB, MBBI, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3);
+ MBB.erase(MBBI);
+ }
}
}
}
unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const {
- return STI.useThumbBacktraces() ? ARM::R7 : ARM::R11;
+ if (STI.isTargetDarwin() || hasFP(MF))
+ return (STI.useThumbBacktraces() || STI.isThumb()) ? ARM::R7 : ARM::R11;
+ else
+ return ARM::SP;
+}
+
+unsigned ARMRegisterInfo::getEHExceptionRegister() const {
+ assert(0 && "What is the exception register");
+ return 0;
+}
+
+unsigned ARMRegisterInfo::getEHHandlerRegister() const {
+ assert(0 && "What is the exception handler register");
+ return 0;
+}
+
+int ARMRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
+ assert(0 && "What is the dwarf register number");
+ return -1;
}
#include "ARMGenRegisterInfo.inc"
projects / oota-llvm.git / blobdiff