const TargetLowering &TLI;
const ARMFunctionInfo *AFI;
+ // Convenience variable to avoid checking all the time.
+ bool isThumb;
+
public:
explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
: FastISel(funcInfo),
TLI(*TM.getTargetLowering()) {
Subtarget = &TM.getSubtarget<ARMSubtarget>();
AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
+ isThumb = AFI->isThumbFunction();
}
// Code from FastISel.cpp.
// Backend specific FastISel code.
virtual bool TargetSelectInstruction(const Instruction *I);
+ virtual unsigned TargetMaterializeConstant(const Constant *C);
#include "ARMGenFastISel.inc"
// Instruction selection routines.
virtual bool ARMSelectLoad(const Instruction *I);
+ virtual bool ARMSelectStore(const Instruction *I);
+ virtual bool ARMSelectBranch(const Instruction *I);
+ virtual bool ARMSelectCmp(const Instruction *I);
+ virtual bool ARMSelectFPExt(const Instruction *I);
+ virtual bool ARMSelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
// 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 ARMLoadAlloca(const Instruction *I);
+ bool ARMEmitStore(EVT VT, unsigned SrcReg, unsigned Reg, int Offset);
+ bool ARMLoadAlloca(const Instruction *I, EVT VT);
+ bool ARMStoreAlloca(const Instruction *I, unsigned SrcReg, EVT VT);
bool ARMComputeRegOffset(const Value *Obj, unsigned &Reg, int &Offset);
+ unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
+ unsigned ARMMaterializeInt(const Constant *C);
bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
return ResultReg;
}
+// For double width floating point we need to materialize two constants
+// (the high and the low) into integer registers then use a move to get
+// the combined constant into an FP reg.
+unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
+ const APFloat Val = CFP->getValueAPF();
+ bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64;
+
+ // This checks to see if we can use VFP3 instructions to materialize
+ // a constant, otherwise we have to go through the constant pool.
+ if (TLI.isFPImmLegal(Val, VT)) {
+ unsigned Opc = is64bit ? ARM::FCONSTD : ARM::FCONSTS;
+ unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
+ DestReg)
+ .addFPImm(CFP));
+ return DestReg;
+ }
+
+ // No 64-bit at the moment.
+ if (is64bit) return 0;
+
+ // Load this from the constant pool.
+ unsigned DestReg = ARMMaterializeInt(cast<Constant>(CFP));
+
+ // If we have a floating point constant we expect it in a floating point
+ // register.
+ unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::VMOVRS), MoveReg)
+ .addReg(DestReg));
+ return MoveReg;
+}
+
+unsigned ARMFastISel::ARMMaterializeInt(const Constant *C) {
+ // MachineConstantPool wants an explicit alignment.
+ unsigned Align = TD.getPrefTypeAlignment(C->getType());
+ if (Align == 0) {
+ // TODO: Figure out if this is correct.
+ Align = TD.getTypeAllocSize(C->getType());
+ }
+ unsigned Idx = MCP.getConstantPoolIndex(C, Align);
+
+ unsigned DestReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ if (isThumb)
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::t2LDRpci))
+ .addReg(DestReg).addConstantPoolIndex(Idx));
+ else
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::LDRcp))
+ .addReg(DestReg).addConstantPoolIndex(Idx)
+ .addReg(0).addImm(0));
+
+ return DestReg;
+}
+
+unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
+ EVT VT = TLI.getValueType(C->getType(), true);
+
+ // Only handle simple types.
+ if (!VT.isSimple()) return 0;
+
+ if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
+ return ARMMaterializeFP(CFP, VT);
+ return ARMMaterializeInt(C);
+}
+
bool ARMFastISel::isTypeLegal(const Type *Ty, EVT &VT) {
VT = TLI.getValueType(Ty, true);
// 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 (!isThumb)
+ 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::ARMLoadAlloca(const Instruction *I) {
+bool ARMFastISel::ARMLoadAlloca(const Instruction *I, EVT VT) {
Value *Op0 = I->getOperand(0);
// Verify it's an alloca.
FuncInfo.StaticAllocaMap.find(AI);
if (SI != FuncInfo.StaticAllocaMap.end()) {
- TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
+ TargetRegisterClass* RC = TLI.getRegClassFor(VT);
unsigned ResultReg = createResultReg(RC);
TII.loadRegFromStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
ResultReg, SI->second, RC,
unsigned Reg, int Offset) {
assert(VT.isSimple() && "Non-simple types are invalid here!");
-
- bool isThumb = AFI->isThumbFunction();
unsigned Opc;
switch (VT.getSimpleVT().SimpleTy) {
return true;
}
-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))
+bool ARMFastISel::ARMStoreAlloca(const Instruction *I, unsigned SrcReg, EVT VT){
+ 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(VT);
+ 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) {
+ 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;
+ case MVT::f32:
+ if (!Subtarget->hasVFP2()) return false;
+ StrOpc = ARM::VSTRS;
+ break;
+ case MVT::f64:
+ if (!Subtarget->hasVFP2()) return false;
+ StrOpc = ARM::VSTRD;
+ 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;
+
+ // Get the value to be stored into a register.
+ 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, VT))
return true;
+ // 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(1), Reg, Offset))
+ return false;
+
+ if (!ARMEmitStore(VT, SrcReg, Reg, Offset /* 0 */)) return false;
+
+ return false;
+
+}
+
+bool ARMFastISel::ARMSelectLoad(const Instruction *I) {
// Verify we have a legal type before going any further.
EVT VT;
if (!isLoadTypeLegal(I->getType(), VT))
return false;
+ // If we're an alloca we know we have a frame index and can emit the load
+ // directly in short order.
+ if (ARMLoadAlloca(I, VT))
+ return true;
+
// 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;
- // TODO: Verify the additions above work, otherwise we'll need to add the
- // offset instead of 0 and do all sorts of operand munging.
- if (!ARMEmitLoad(VT, ResultReg, Reg, 0)) return false;
+ if (!ARMEmitLoad(VT, ResultReg, Reg, Offset /* 0 */)) return false;
UpdateValueMap(I, ResultReg);
return true;
}
+bool ARMFastISel::ARMSelectBranch(const Instruction *I) {
+ const BranchInst *BI = cast<BranchInst>(I);
+ MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
+ MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
+
+ // Simple branch support.
+ unsigned CondReg = getRegForValue(BI->getCondition());
+ if (CondReg == 0) return false;
+
+ unsigned CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
+ unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
+ .addReg(CondReg).addReg(CondReg));
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
+ .addMBB(TBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
+ FastEmitBranch(FBB, DL);
+ FuncInfo.MBB->addSuccessor(TBB);
+ return true;
+}
+
+bool ARMFastISel::ARMSelectCmp(const Instruction *I) {
+ const CmpInst *CI = cast<CmpInst>(I);
+
+ EVT VT;
+ const Type *Ty = CI->getOperand(0)->getType();
+ if (!isTypeLegal(Ty, VT))
+ return false;
+
+ bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
+ if (isFloat && !Subtarget->hasVFP2())
+ return false;
+
+ unsigned CmpOpc;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return false;
+ // TODO: Verify compares.
+ case MVT::f32:
+ CmpOpc = ARM::VCMPES;
+ break;
+ case MVT::f64:
+ CmpOpc = ARM::VCMPED;
+ break;
+ case MVT::i32:
+ CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
+ break;
+ }
+
+ unsigned Arg1 = getRegForValue(CI->getOperand(0));
+ if (Arg1 == 0) return false;
+
+ unsigned Arg2 = getRegForValue(CI->getOperand(1));
+ if (Arg2 == 0) return false;
+
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
+ .addReg(Arg1).addReg(Arg2));
+
+ // For floating point we need to move the result to a register we can
+ // actually do something with.
+ if (isFloat)
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::FMSTAT)));
+ return true;
+}
+
+bool ARMFastISel::ARMSelectFPExt(const Instruction *I) {
+ // Make sure we have VFP and that we're extending float to double.
+ if (!Subtarget->hasVFP2()) return false;
+
+ Value *V = I->getOperand(0);
+ if (!I->getType()->isDoubleTy() ||
+ !V->getType()->isFloatTy()) return false;
+
+ unsigned Op = getRegForValue(V);
+ if (Op == 0) return false;
+
+ unsigned Result = createResultReg(ARM::DPRRegisterClass);
+
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::VCVTDS), Result)
+ .addReg(Op));
+ UpdateValueMap(I, Result);
+ return true;
+}
+
+bool ARMFastISel::ARMSelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
+ EVT VT = TLI.getValueType(I->getType(), true);
+
+ // We can get here in the case when we want to use NEON for our fp
+ // operations, but can't figure out how to. Just use the vfp instructions
+ // if we have them.
+ // FIXME: It'd be nice to use NEON instructions.
+ const Type *Ty = I->getType();
+ bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
+ if (isFloat && !Subtarget->hasVFP2())
+ return false;
+
+ unsigned Op1 = getRegForValue(I->getOperand(0));
+ if (Op1 == 0) return false;
+
+ unsigned Op2 = getRegForValue(I->getOperand(1));
+ if (Op2 == 0) return false;
+
+ unsigned Opc;
+ bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64 ||
+ VT.getSimpleVT().SimpleTy == MVT::i64;
+ switch (ISDOpcode) {
+ default: return false;
+ case ISD::FADD:
+ Opc = is64bit ? ARM::VADDD : ARM::VADDS;
+ break;
+ case ISD::FSUB:
+ Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
+ break;
+ case ISD::FMUL:
+ Opc = is64bit ? ARM::VMULD : ARM::VMULS;
+ break;
+ }
+ unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg)
+ .addReg(Op1).addReg(Op2));
+ return true;
+}
+
+// TODO: SoftFP support.
bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
// No Thumb-1 for now.
- if (AFI->isThumbFunction() && !AFI->isThumb2Function()) return false;
+ if (isThumb && !AFI->isThumb2Function()) return false;
switch (I->getOpcode()) {
case Instruction::Load:
return ARMSelectLoad(I);
+ case Instruction::Store:
+ return ARMSelectStore(I);
+ case Instruction::Br:
+ return ARMSelectBranch(I);
+ case Instruction::ICmp:
+ case Instruction::FCmp:
+ return ARMSelectCmp(I);
+ case Instruction::FPExt:
+ return ARMSelectFPExt(I);
+ case Instruction::FAdd:
+ return ARMSelectBinaryOp(I, ISD::FADD);
+ case Instruction::FSub:
+ return ARMSelectBinaryOp(I, ISD::FSUB);
+ case Instruction::FMul:
+ return ARMSelectBinaryOp(I, ISD::FMUL);
default: break;
}
return false;
projects / oota-llvm.git / blobdiff