// Instruction selection routines.
virtual bool ARMSelectLoad(const Instruction *I);
+ virtual bool ARMSelectStore(const Instruction *I);
// Utility routines.
private:
+ bool isTypeLegal(const Type *Ty, EVT &VT);
+ bool isLoadTypeLegal(const Type *Ty, EVT &VT);
+ bool ARMEmitLoad(EVT VT, unsigned &ResultReg, unsigned Reg, int Offset);
+ bool ARMEmitStore(EVT VT, unsigned SrcReg, unsigned Reg, int Offset);
+ bool ARMLoadAlloca(const Instruction *I);
+ bool ARMStoreAlloca(const Instruction *I, unsigned SrcReg);
bool ARMComputeRegOffset(const Value *Obj, unsigned &Reg, int &Offset);
+ bool ARMMaterializeConstant(const ConstantInt *Val, unsigned &Reg);
bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
return ResultReg;
}
+bool ARMFastISel::isTypeLegal(const Type *Ty, EVT &VT) {
+ VT = TLI.getValueType(Ty, true);
+
+ // Only handle simple types.
+ if (VT == MVT::Other || !VT.isSimple()) return false;
+
+ // Handle all legal types, i.e. a register that will directly hold this
+ // value.
+ return TLI.isTypeLegal(VT);
+}
+
+bool ARMFastISel::isLoadTypeLegal(const Type *Ty, EVT &VT) {
+ if (isTypeLegal(Ty, VT)) return true;
+
+ // If this is a type than can be sign or zero-extended to a basic operation
+ // go ahead and accept it now.
+ if (VT == MVT::i8 || VT == MVT::i16)
+ return true;
+
+ return false;
+}
+
// Computes the Reg+Offset to get to an object.
bool ARMFastISel::ARMComputeRegOffset(const Value *Obj, unsigned &Reg,
int &Offset) {
//errs() << "Failing Opcode is: " << *Op1 << "\n";
break;
case Instruction::Alloca: {
- // Do static allocas.
- const AllocaInst *A = cast<AllocaInst>(Obj);
- DenseMap<const AllocaInst*, int>::iterator SI =
- FuncInfo.StaticAllocaMap.find(A);
- if (SI != FuncInfo.StaticAllocaMap.end())
- Offset =
- TM.getRegisterInfo()->getFrameIndexReference(*FuncInfo.MF,
- SI->second, Reg);
- else
- return false;
- return true;
+ assert(false && "Alloca should have been handled earlier!");
+ return false;
}
}
if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
//errs() << "Failing GV is: " << GV << "\n";
+ (void)GV;
return false;
}
// Try to get this in a register if nothing else has worked.
Reg = getRegForValue(Obj);
- return Reg != 0;
+ if (Reg == 0) return false;
+
+ // Since the offset may be too large for the load instruction
+ // get the reg+offset into a register.
+ // TODO: Verify the additions work, otherwise we'll need to add the
+ // offset instead of 0 to the instructions and do all sorts of operand
+ // munging.
+ // TODO: Optimize this somewhat.
+ if (Offset != 0) {
+ ARMCC::CondCodes Pred = ARMCC::AL;
+ unsigned PredReg = 0;
+
+ if (!AFI->isThumbFunction())
+ emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ Reg, Reg, Offset, Pred, PredReg,
+ static_cast<const ARMBaseInstrInfo&>(TII));
+ else {
+ assert(AFI->isThumb2Function());
+ emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ Reg, Reg, Offset, Pred, PredReg,
+ static_cast<const ARMBaseInstrInfo&>(TII));
+ }
+ }
+
+ return true;
}
-bool ARMFastISel::ARMSelectLoad(const Instruction *I) {
+bool ARMFastISel::ARMLoadAlloca(const Instruction *I) {
+ Value *Op0 = I->getOperand(0);
+
+ // Verify it's an alloca.
+ if (const AllocaInst *AI = dyn_cast<AllocaInst>(Op0)) {
+ DenseMap<const AllocaInst*, int>::iterator SI =
+ FuncInfo.StaticAllocaMap.find(AI);
+
+ if (SI != FuncInfo.StaticAllocaMap.end()) {
+ TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
+ unsigned ResultReg = createResultReg(RC);
+ TII.loadRegFromStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
+ ResultReg, SI->second, RC,
+ TM.getRegisterInfo());
+ UpdateValueMap(I, ResultReg);
+ return true;
+ }
+ }
+ return false;
+}
+
+bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg,
+ unsigned Reg, int Offset) {
+
+ assert(VT.isSimple() && "Non-simple types are invalid here!");
+
+ bool isThumb = AFI->isThumbFunction();
+ unsigned Opc;
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ default:
+ assert(false && "Trying to emit for an unhandled type!");
+ return false;
+ case MVT::i16:
+ Opc = isThumb ? ARM::tLDRH : ARM::LDRH;
+ VT = MVT::i32;
+ break;
+ case MVT::i8:
+ Opc = isThumb ? ARM::tLDRB : ARM::LDRB;
+ VT = MVT::i32;
+ break;
+ case MVT::i32:
+ Opc = isThumb ? ARM::tLDR : ARM::LDR;
+ break;
+ }
+
+ ResultReg = createResultReg(TLI.getRegClassFor(VT));
+
+ // TODO: Fix the Addressing modes so that these can share some code.
+ // Since this is a Thumb1 load this will work in Thumb1 or 2 mode.
+ if (isThumb)
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg)
+ .addReg(Reg).addImm(Offset).addReg(0));
+ else
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg)
+ .addReg(Reg).addReg(0).addImm(Offset));
+
+ return true;
+}
+
+bool ARMFastISel::ARMMaterializeConstant(const ConstantInt *CI, unsigned &Reg) {
+ unsigned Opc;
+ bool Signed = true;
+ bool isThumb = AFI->isThumbFunction();
+ EVT VT = TLI.getValueType(CI->getType(), true);
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return false;
+ case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ Opc = isThumb ? ARM::t2MOVi32imm : ARM::MOVi32imm; break;
+ }
+
+ Reg = createResultReg(TLI.getRegClassFor(VT));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
+ Reg)
+ .addImm(Signed ? (uint64_t) CI->getSExtValue() :
+ CI->getZExtValue()));
+
+ return true;
+}
+
+bool ARMFastISel::ARMStoreAlloca(const Instruction *I, unsigned SrcReg) {
+ Value *Op1 = I->getOperand(1);
+
+ // Verify it's an alloca.
+ if (const AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
+ DenseMap<const AllocaInst*, int>::iterator SI =
+ FuncInfo.StaticAllocaMap.find(AI);
+
+ if (SI != FuncInfo.StaticAllocaMap.end()) {
+ TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
+ assert(SrcReg != 0 && "Nothing to store!");
+ TII.storeRegToStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
+ SrcReg, true /*isKill*/, SI->second, RC,
+ TM.getRegisterInfo());
+ return true;
+ }
+ }
+ return false;
+}
+
+bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg,
+ unsigned DstReg, int Offset) {
+ bool isThumb = AFI->isThumbFunction();
+
+ unsigned StrOpc;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return false;
+ case MVT::i1:
+ case MVT::i8: StrOpc = isThumb ? ARM::tSTRB : ARM::STRB; break;
+ case MVT::i16: StrOpc = isThumb ? ARM::tSTRH : ARM::STRH; break;
+ case MVT::i32: StrOpc = isThumb ? ARM::tSTR : ARM::STR; break;
+ }
+
+ if (isThumb)
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(StrOpc), SrcReg)
+ .addReg(DstReg).addImm(Offset).addReg(0));
+ else
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(StrOpc), SrcReg)
+ .addReg(DstReg).addReg(0).addImm(Offset));
+
+ return true;
+}
+
+bool ARMFastISel::ARMSelectStore(const Instruction *I) {
+ Value *Op0 = I->getOperand(0);
+ unsigned SrcReg = 0;
+
+ // Yay type legalization
+ EVT VT;
+ if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
+ return false;
+
+ // First see if we're a constant that we want to store, we'll need to
+ // materialize that into a register.
+ // Handle 'null' like i32/i64 0.
+ if (isa<ConstantPointerNull>(Op0))
+ Op0 = Constant::getNullValue(TD.getIntPtrType(Op0->getContext()));
+
+ // If this is a store of a simple constant, materialize the constant into
+ // a register then emit the store into the location.
+ if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op0))
+ if (!ARMMaterializeConstant(CI, SrcReg))
+ return false;
+
+ // If Reg is still 0, try to get the value into a register.
+ if (SrcReg == 0)
+ SrcReg = getRegForValue(Op0);
+ if (SrcReg == 0)
+ return false;
+
+ // If we're an alloca we know we have a frame index and can emit the store
+ // quickly.
+ if (ARMStoreAlloca(I, SrcReg))
+ return true;
+
// Our register and offset with innocuous defaults.
unsigned Reg = 0;
int Offset = 0;
- // TODO: Think about using loadRegFromStackSlot() here when we can.
+ // See if we can handle this as Reg + Offset
+ if (!ARMComputeRegOffset(I->getOperand(1), Reg, Offset))
+ return false;
+
+ if (!ARMEmitStore(VT, SrcReg, Reg, Offset /* 0 */)) return false;
+
+ return false;
+
+}
+
+bool ARMFastISel::ARMSelectLoad(const Instruction *I) {
+ // If we're an alloca we know we have a frame index and can emit the load
+ // directly in short order.
+ if (ARMLoadAlloca(I))
+ return true;
+
+ // Verify we have a legal type before going any further.
+ EVT VT;
+ if (!isLoadTypeLegal(I->getType(), VT))
+ return false;
+
+ // Our register and offset with innocuous defaults.
+ unsigned Reg = 0;
+ int Offset = 0;
// See if we can handle this as Reg + Offset
if (!ARMComputeRegOffset(I->getOperand(0), Reg, Offset))
return false;
-
- // Since the offset may be too large for the load instruction
- // get the reg+offset into a register.
- // TODO: Optimize this somewhat.
- ARMCC::CondCodes Pred = ARMCC::AL;
- unsigned PredReg = 0;
- if (!AFI->isThumbFunction())
- emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- Reg, Reg, Offset, Pred, PredReg,
- static_cast<const ARMBaseInstrInfo&>(TII));
- else {
- assert(AFI->isThumb2Function());
- emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- Reg, Reg, Offset, Pred, PredReg,
- static_cast<const ARMBaseInstrInfo&>(TII));
- }
+ unsigned ResultReg;
+ if (!ARMEmitLoad(VT, ResultReg, Reg, Offset /* 0 */)) return false;
- // FIXME: There is more than one register class in the world...
- unsigned ResultReg = createResultReg(ARM::GPRRegisterClass);
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(ARM::LDR), ResultReg)
- .addImm(0).addReg(Reg).addImm(Offset));
-
+ UpdateValueMap(I, ResultReg);
return true;
}
switch (I->getOpcode()) {
case Instruction::Load:
return ARMSelectLoad(I);
+ case Instruction::Store:
+ return ARMSelectStore(I);
default: break;
}
return false;