#include "ARM.h"
#include "ARMBaseInstrInfo.h"
#include "ARMCallingConv.h"
-#include "ARMRegisterInfo.h"
#include "ARMTargetMachine.h"
#include "ARMSubtarget.h"
#include "ARMConstantPoolValue.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/DataLayout.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;
-static cl::opt<bool>
-DisableARMFastISel("disable-arm-fast-isel",
- cl::desc("Turn off experimental ARM fast-isel support"),
- cl::init(false), cl::Hidden);
-
extern cl::opt<bool> EnableARMLongCalls;
namespace {
LLVMContext *Context;
public:
- explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
- : FastISel(funcInfo),
+ explicit ARMFastISel(FunctionLoweringInfo &funcInfo,
+ const TargetLibraryInfo *libInfo)
+ : FastISel(funcInfo, libInfo),
TM(funcInfo.MF->getTarget()),
TII(*TM.getInstrInfo()),
TLI(*TM.getTargetLowering()) {
}
// Code from FastISel.cpp.
- virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC);
- virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill);
- virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- unsigned Op1, bool Op1IsKill);
- virtual unsigned FastEmitInst_rrr(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- unsigned Op1, bool Op1IsKill,
- unsigned Op2, bool Op2IsKill);
- virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- uint64_t Imm);
- virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- const ConstantFP *FPImm);
- virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- unsigned Op1, bool Op1IsKill,
- uint64_t Imm);
- virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- uint64_t Imm);
- virtual unsigned FastEmitInst_ii(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- uint64_t Imm1, uint64_t Imm2);
-
- virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
- unsigned Op0, bool Op0IsKill,
- uint32_t Idx);
+ private:
+ unsigned FastEmitInst_(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC);
+ unsigned FastEmitInst_r(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill);
+ unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill);
+ unsigned FastEmitInst_rrr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ unsigned Op2, bool Op2IsKill);
+ unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ uint64_t Imm);
+ unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ const ConstantFP *FPImm);
+ unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ uint64_t Imm);
+ unsigned FastEmitInst_i(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ uint64_t Imm);
+ unsigned FastEmitInst_ii(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ uint64_t Imm1, uint64_t Imm2);
+
+ unsigned FastEmitInst_extractsubreg(MVT RetVT,
+ unsigned Op0, bool Op0IsKill,
+ uint32_t Idx);
// Backend specific FastISel code.
+ private:
virtual bool TargetSelectInstruction(const Instruction *I);
virtual unsigned TargetMaterializeConstant(const Constant *C);
virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
const LoadInst *LI);
-
+ private:
#include "ARMGenFastISel.inc"
// Instruction selection routines.
bool SelectRet(const Instruction *I);
bool SelectTrunc(const Instruction *I);
bool SelectIntExt(const Instruction *I);
+ bool SelectShift(const Instruction *I, ARM_AM::ShiftOpc ShiftTy);
// Utility routines.
private:
bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr,
unsigned Alignment = 0, bool isZExt = true,
bool allocReg = true);
-
bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
unsigned Alignment = 0);
bool ARMComputeAddress(const Value *Obj, Address &Addr);
unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
- unsigned ARMSelectCallOp(const GlobalValue *GV);
+ unsigned ARMSelectCallOp(bool UseReg);
+ unsigned ARMLowerPICELF(const GlobalValue *GV, unsigned Align, EVT VT);
// Call handling routines.
private:
- CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
+ CCAssignFn *CCAssignFnForCall(CallingConv::ID CC,
+ bool Return,
+ bool isVarArg);
bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
SmallVectorImpl<unsigned> &ArgRegs,
SmallVectorImpl<MVT> &ArgVTs,
SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs,
CallingConv::ID CC,
- unsigned &NumBytes);
+ unsigned &NumBytes,
+ bool isVarArg);
+ unsigned getLibcallReg(const Twine &Name);
bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC,
- unsigned &NumBytes);
+ unsigned &NumBytes, bool isVarArg);
bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
// OptionalDef handling routines.
if (VT != MVT::i32) return 0;
Reloc::Model RelocM = TM.getRelocationModel();
-
- // TODO: Need more magic for ARM PIC.
- if (!isThumb2 && (RelocM == Reloc::PIC_)) return 0;
-
- unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
+ bool IsIndirect = Subtarget->GVIsIndirectSymbol(GV, RelocM);
+ const TargetRegisterClass *RC = isThumb2 ?
+ (const TargetRegisterClass*)&ARM::rGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass;
+ unsigned DestReg = createResultReg(RC);
// Use movw+movt when possible, it avoids constant pool entries.
// Darwin targets don't support movt with Reloc::Static, see
Align = TD.getTypeAllocSize(GV->getType());
}
+ if (Subtarget->isTargetELF() && RelocM == Reloc::PIC_)
+ return ARMLowerPICELF(GV, Align, VT);
+
// Grab index.
unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 :
(Subtarget->isThumb() ? 4 : 8);
.addConstantPoolIndex(Idx);
if (RelocM == Reloc::PIC_)
MIB.addImm(Id);
+ AddOptionalDefs(MIB);
} else {
// The extra immediate is for addrmode2.
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
DestReg)
.addConstantPoolIndex(Idx)
.addImm(0);
+ AddOptionalDefs(MIB);
+
+ if (RelocM == Reloc::PIC_) {
+ unsigned Opc = IsIndirect ? ARM::PICLDR : ARM::PICADD;
+ unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
+
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(Opc), NewDestReg)
+ .addReg(DestReg)
+ .addImm(Id);
+ AddOptionalDefs(MIB);
+ return NewDestReg;
+ }
}
- AddOptionalDefs(MIB);
}
- if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
+ if (IsIndirect) {
MachineInstrBuilder MIB;
unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
if (isThumb2)
if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
MVT VT;
- if (!isLoadTypeLegal(AI->getType(), VT)) return false;
+ if (!isLoadTypeLegal(AI->getType(), VT)) return 0;
DenseMap<const AllocaInst*, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
// put the alloca address into a register, set the base type back to
// register and continue. This should almost never happen.
if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
- const TargetRegisterClass *RC = isThumb2 ? ARM::tGPRRegisterClass
- : ARM::GPRRegisterClass;
+ const TargetRegisterClass *RC = isThumb2 ?
+ (const TargetRegisterClass*)&ARM::tGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass;
unsigned ResultReg = createResultReg(RC);
unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
useAM3 = true;
}
}
- RC = ARM::GPRRegisterClass;
+ RC = &ARM::GPRRegClass;
break;
case MVT::i16:
+ if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
Opc = isZExt ? ARM::t2LDRHi8 : ARM::t2LDRSHi8;
Opc = isZExt ? ARM::LDRH : ARM::LDRSH;
useAM3 = true;
}
- RC = ARM::GPRRegisterClass;
+ RC = &ARM::GPRRegClass;
break;
case MVT::i32:
+ if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
Opc = ARM::t2LDRi8;
} else {
Opc = ARM::LDRi12;
}
- RC = ARM::GPRRegisterClass;
+ RC = &ARM::GPRRegClass;
break;
case MVT::f32:
if (!Subtarget->hasVFP2()) return false;
needVMOV = true;
VT = MVT::i32;
Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
- RC = ARM::GPRRegisterClass;
+ RC = &ARM::GPRRegClass;
} else {
Opc = ARM::VLDRS;
RC = TLI.getRegClassFor(VT);
// This is mostly going to be Neon/vector support.
default: return false;
case MVT::i1: {
- unsigned Res = createResultReg(isThumb2 ? ARM::tGPRRegisterClass :
- ARM::GPRRegisterClass);
+ unsigned Res = createResultReg(isThumb2 ?
+ (const TargetRegisterClass*)&ARM::tGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass);
unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), Res)
}
break;
case MVT::i16:
+ if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
StrOpc = ARM::t2STRHi8;
}
break;
case MVT::i32:
+ if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
StrOpc = ARM::t2STRi8;
unsigned Opc = isThumb2 ? ARM::tBRIND : ARM::BX;
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc))
.addReg(AddrReg));
- return true;
+
+ const IndirectBrInst *IB = cast<IndirectBrInst>(I);
+ for (unsigned i = 0, e = IB->getNumSuccessors(); i != e; ++i)
+ FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[IB->getSuccessor(i)]);
+
+ return true;
}
bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
SrcVT == MVT::i1) {
const APInt &CIVal = ConstInt->getValue();
Imm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
- if (Imm < 0) {
+ // For INT_MIN/LONG_MIN (i.e., 0x80000000) we need to use a cmp, rather
+ // then a cmn, because there is no way to represent 2147483648 as a
+ // signed 32-bit int.
+ if (Imm < 0 && Imm != (int)0x80000000) {
isNegativeImm = true;
Imm = -Imm;
}
if (!UseImm)
CmpOpc = ARM::t2CMPrr;
else
- CmpOpc = isNegativeImm ? ARM::t2CMNzri : ARM::t2CMPri;
+ CmpOpc = isNegativeImm ? ARM::t2CMNri : ARM::t2CMPri;
} else {
if (!UseImm)
CmpOpc = ARM::CMPrr;
else
- CmpOpc = isNegativeImm ? ARM::CMNzri : ARM::CMPri;
+ CmpOpc = isNegativeImm ? ARM::CMNri : ARM::CMPri;
}
break;
}
// Now set a register based on the comparison. Explicitly set the predicates
// here.
unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
- const TargetRegisterClass *RC = isThumb2 ? ARM::rGPRRegisterClass
- : ARM::GPRRegisterClass;
+ const TargetRegisterClass *RC = isThumb2 ?
+ (const TargetRegisterClass*)&ARM::rGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass;
unsigned DestReg = createResultReg(RC);
Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
unsigned ZeroReg = TargetMaterializeConstant(Zero);
unsigned Op = getRegForValue(V);
if (Op == 0) return false;
- unsigned Result = createResultReg(ARM::DPRRegisterClass);
+ unsigned Result = createResultReg(&ARM::DPRRegClass);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VCVTDS), Result)
.addReg(Op));
unsigned Op = getRegForValue(V);
if (Op == 0) return false;
- unsigned Result = createResultReg(ARM::SPRRegisterClass);
+ unsigned Result = createResultReg(&ARM::SPRRegClass);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VCVTSD), Result)
.addReg(Op));
// type and the target independent selector doesn't know how to handle it.
if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
return false;
-
+
unsigned Opc;
switch (ISDOpcode) {
default: return false;
// Call Handling Code
-// This is largely taken directly from CCAssignFnForNode - we don't support
-// varargs in FastISel so that part has been removed.
+// This is largely taken directly from CCAssignFnForNode
// TODO: We may not support all of this.
-CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
+CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC,
+ bool Return,
+ bool isVarArg) {
switch (CC) {
default:
llvm_unreachable("Unsupported calling convention");
case CallingConv::Fast:
- // Ignore fastcc. Silence compiler warnings.
- (void)RetFastCC_ARM_APCS;
- (void)FastCC_ARM_APCS;
+ if (Subtarget->hasVFP2() && !isVarArg) {
+ if (!Subtarget->isAAPCS_ABI())
+ return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
+ // For AAPCS ABI targets, just use VFP variant of the calling convention.
+ return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
+ }
// Fallthrough
case CallingConv::C:
// Use target triple & subtarget features to do actual dispatch.
if (Subtarget->isAAPCS_ABI()) {
if (Subtarget->hasVFP2() &&
- TM.Options.FloatABIType == FloatABI::Hard)
+ TM.Options.FloatABIType == FloatABI::Hard && !isVarArg)
return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
else
return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
} else
return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
case CallingConv::ARM_AAPCS_VFP:
- return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
+ if (!isVarArg)
+ return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
+ // Fall through to soft float variant, variadic functions don't
+ // use hard floating point ABI.
case CallingConv::ARM_AAPCS:
return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
case CallingConv::ARM_APCS:
return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
+ case CallingConv::GHC:
+ if (Return)
+ llvm_unreachable("Can't return in GHC call convention");
+ else
+ return CC_ARM_APCS_GHC;
}
}
SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs,
CallingConv::ID CC,
- unsigned &NumBytes) {
+ unsigned &NumBytes,
+ bool isVarArg) {
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
- CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
+ CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs, *Context);
+ CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags,
+ CCAssignFnForCall(CC, false, isVarArg));
+
+ // Check that we can handle all of the arguments. If we can't, then bail out
+ // now before we add code to the MBB.
+ for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ CCValAssign &VA = ArgLocs[i];
+ MVT ArgVT = ArgVTs[VA.getValNo()];
+
+ // We don't handle NEON/vector parameters yet.
+ if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
+ return false;
+
+ // Now copy/store arg to correct locations.
+ if (VA.isRegLoc() && !VA.needsCustom()) {
+ continue;
+ } else if (VA.needsCustom()) {
+ // TODO: We need custom lowering for vector (v2f64) args.
+ if (VA.getLocVT() != MVT::f64 ||
+ // TODO: Only handle register args for now.
+ !VA.isRegLoc() || !ArgLocs[++i].isRegLoc())
+ return false;
+ } else {
+ switch (static_cast<EVT>(ArgVT).getSimpleVT().SimpleTy) {
+ default:
+ return false;
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ break;
+ case MVT::f32:
+ if (!Subtarget->hasVFP2())
+ return false;
+ break;
+ case MVT::f64:
+ if (!Subtarget->hasVFP2())
+ return false;
+ break;
+ }
+ }
+ }
+
+ // At the point, we are able to handle the call's arguments in fast isel.
// Get a count of how many bytes are to be pushed on the stack.
NumBytes = CCInfo.getNextStackOffset();
unsigned Arg = ArgRegs[VA.getValNo()];
MVT ArgVT = ArgVTs[VA.getValNo()];
- // We don't handle NEON/vector parameters yet.
- if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
- return false;
+ assert((!ArgVT.isVector() && ArgVT.getSizeInBits() <= 64) &&
+ "We don't handle NEON/vector parameters yet.");
// Handle arg promotion, etc.
switch (VA.getLocInfo()) {
RegArgs.push_back(VA.getLocReg());
} else if (VA.needsCustom()) {
// TODO: We need custom lowering for vector (v2f64) args.
- if (VA.getLocVT() != MVT::f64) return false;
+ assert(VA.getLocVT() == MVT::f64 &&
+ "Custom lowering for v2f64 args not available");
CCValAssign &NextVA = ArgLocs[++i];
- // TODO: Only handle register args for now.
- if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
+ assert(VA.isRegLoc() && NextVA.isRegLoc() &&
+ "We only handle register args!");
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VMOVRRD), VA.getLocReg())
Addr.Base.Reg = ARM::SP;
Addr.Offset = VA.getLocMemOffset();
- if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
+ bool EmitRet = ARMEmitStore(ArgVT, Arg, Addr); (void)EmitRet;
+ assert(EmitRet && "Could not emit a store for argument!");
}
}
+
return true;
}
bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC,
- unsigned &NumBytes) {
+ unsigned &NumBytes, bool isVarArg) {
// Issue CALLSEQ_END
unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
// Now the return value.
if (RetVT != MVT::isVoid) {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
- CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
+ CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
// Copy all of the result registers out of their specified physreg.
if (RVLocs.size() == 2 && RetVT == MVT::f64) {
if (!FuncInfo.CanLowerReturn)
return false;
- if (F.isVarArg())
- return false;
-
CallingConv::ID CC = F.getCallingConv();
if (Ret->getNumOperands() > 0) {
SmallVector<ISD::OutputArg, 4> Outs;
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ValLocs;
CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
- CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
+ CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */,
+ F.isVarArg()));
const Value *RV = Ret->getOperand(0);
unsigned Reg = getRegForValue(RV);
return true;
}
-unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
+unsigned ARMFastISel::ARMSelectCallOp(bool UseReg) {
+ if (UseReg)
+ return isThumb2 ? ARM::tBLXr : ARM::BLX;
+ else
+ return isThumb2 ? ARM::tBL : ARM::BL;
+}
- // iOS needs the r9 versions of the opcodes.
- bool isiOS = Subtarget->isTargetIOS();
- if (isThumb2) {
- return isiOS ? ARM::tBLr9 : ARM::tBL;
- } else {
- return isiOS ? ARM::BLr9 : ARM::BL;
- }
+unsigned ARMFastISel::getLibcallReg(const Twine &Name) {
+ GlobalValue *GV = new GlobalVariable(Type::getInt32Ty(*Context), false,
+ GlobalValue::ExternalLinkage, 0, Name);
+ return ARMMaterializeGV(GV, TLI.getValueType(GV->getType()));
}
// A quick function that will emit a call for a named libcall in F with the
else if (!isTypeLegal(RetTy, RetVT))
return false;
- // TODO: For now if we have long calls specified we don't handle the call.
- if (EnableARMLongCalls) return false;
+ // Can't handle non-double multi-reg retvals.
+ if (RetVT != MVT::isVoid && RetVT != MVT::i32) {
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, false));
+ if (RVLocs.size() >= 2 && RetVT != MVT::f64)
+ return false;
+ }
// Set up the argument vectors.
SmallVector<Value*, 8> Args;
// Handle the arguments now that we've gotten them.
SmallVector<unsigned, 4> RegArgs;
unsigned NumBytes;
- if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
+ if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
+ RegArgs, CC, NumBytes, false))
return false;
- // Issue the call, BLr9 for iOS, BL otherwise.
- // TODO: Turn this into the table of arm call ops.
- MachineInstrBuilder MIB;
- unsigned CallOpc = ARMSelectCallOp(NULL);
- if(isThumb2)
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc)))
- .addExternalSymbol(TLI.getLibcallName(Call));
+ unsigned CalleeReg = 0;
+ if (EnableARMLongCalls) {
+ CalleeReg = getLibcallReg(TLI.getLibcallName(Call));
+ if (CalleeReg == 0) return false;
+ }
+
+ // Issue the call.
+ unsigned CallOpc = ARMSelectCallOp(EnableARMLongCalls);
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(CallOpc));
+ // BL / BLX don't take a predicate, but tBL / tBLX do.
+ if (isThumb2)
+ AddDefaultPred(MIB);
+ if (EnableARMLongCalls)
+ MIB.addReg(CalleeReg);
else
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc))
- .addExternalSymbol(TLI.getLibcallName(Call)));
+ MIB.addExternalSymbol(TLI.getLibcallName(Call));
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
- MIB.addReg(RegArgs[i]);
+ MIB.addReg(RegArgs[i], RegState::Implicit);
// Add a register mask with the call-preserved registers.
// Proper defs for return values will be added by setPhysRegsDeadExcept().
// Finish off the call including any return values.
SmallVector<unsigned, 4> UsedRegs;
- if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
+ if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, false)) return false;
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
// Can't handle inline asm.
if (isa<InlineAsm>(Callee)) return false;
- // Only handle global variable Callees.
- const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
- if (!GV)
- return false;
-
// Check the calling convention.
ImmutableCallSite CS(CI);
CallingConv::ID CC = CS.getCallingConv();
// TODO: Avoid some calling conventions?
- // Let SDISel handle vararg functions.
PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
FunctionType *FTy = cast<FunctionType>(PT->getElementType());
- if (FTy->isVarArg())
- return false;
+ bool isVarArg = FTy->isVarArg();
// Handle *simple* calls for now.
Type *RetTy = I->getType();
RetVT != MVT::i8 && RetVT != MVT::i1)
return false;
- // TODO: For now if we have long calls specified we don't handle the call.
- if (EnableARMLongCalls) return false;
+ // Can't handle non-double multi-reg retvals.
+ if (RetVT != MVT::isVoid && RetVT != MVT::i1 && RetVT != MVT::i8 &&
+ RetVT != MVT::i16 && RetVT != MVT::i32) {
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
+ if (RVLocs.size() >= 2 && RetVT != MVT::f64)
+ return false;
+ }
// Set up the argument vectors.
SmallVector<Value*, 8> Args;
ISD::ArgFlagsTy Flags;
unsigned AttrInd = i - CS.arg_begin() + 1;
- if (CS.paramHasAttr(AttrInd, Attribute::SExt))
+ if (CS.paramHasAttr(AttrInd, Attributes::SExt))
Flags.setSExt();
- if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
+ if (CS.paramHasAttr(AttrInd, Attributes::ZExt))
Flags.setZExt();
// FIXME: Only handle *easy* calls for now.
- if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
- CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
- CS.paramHasAttr(AttrInd, Attribute::Nest) ||
- CS.paramHasAttr(AttrInd, Attribute::ByVal))
+ if (CS.paramHasAttr(AttrInd, Attributes::InReg) ||
+ CS.paramHasAttr(AttrInd, Attributes::StructRet) ||
+ CS.paramHasAttr(AttrInd, Attributes::Nest) ||
+ CS.paramHasAttr(AttrInd, Attributes::ByVal))
return false;
Type *ArgTy = (*i)->getType();
// Handle the arguments now that we've gotten them.
SmallVector<unsigned, 4> RegArgs;
unsigned NumBytes;
- if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
+ if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
+ RegArgs, CC, NumBytes, isVarArg))
return false;
- // Issue the call, BLr9 for iOS, BL otherwise.
- // TODO: Turn this into the table of arm call ops.
- MachineInstrBuilder MIB;
- unsigned CallOpc = ARMSelectCallOp(GV);
- // Explicitly adding the predicate here.
- if(isThumb2) {
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc)));
- if (!IntrMemName)
- MIB.addGlobalAddress(GV, 0, 0);
- else
- MIB.addExternalSymbol(IntrMemName, 0);
- } else {
- if (!IntrMemName)
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc))
- .addGlobalAddress(GV, 0, 0));
+ bool UseReg = false;
+ const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
+ if (!GV || EnableARMLongCalls) UseReg = true;
+
+ unsigned CalleeReg = 0;
+ if (UseReg) {
+ if (IntrMemName)
+ CalleeReg = getLibcallReg(IntrMemName);
else
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc))
- .addExternalSymbol(IntrMemName, 0));
+ CalleeReg = getRegForValue(Callee);
+
+ if (CalleeReg == 0) return false;
}
-
+
+ // Issue the call.
+ unsigned CallOpc = ARMSelectCallOp(UseReg);
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(CallOpc));
+
+ // ARM calls don't take a predicate, but tBL / tBLX do.
+ if(isThumb2)
+ AddDefaultPred(MIB);
+ if (UseReg)
+ MIB.addReg(CalleeReg);
+ else if (!IntrMemName)
+ MIB.addGlobalAddress(GV, 0, 0);
+ else
+ MIB.addExternalSymbol(IntrMemName, 0);
+
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
- MIB.addReg(RegArgs[i]);
+ MIB.addReg(RegArgs[i], RegState::Implicit);
// Add a register mask with the call-preserved registers.
// Proper defs for return values will be added by setPhysRegsDeadExcept().
// Finish off the call including any return values.
SmallVector<unsigned, 4> UsedRegs;
- if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
+ if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, isVarArg))
+ return false;
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
return Len <= 16;
}
-bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len) {
+bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src,
+ uint64_t Len) {
// Make sure we don't bloat code by inlining very large memcpy's.
if (!ARMIsMemCpySmall(Len))
return false;
// FIXME: Handle more intrinsics.
switch (I.getIntrinsicID()) {
default: return false;
+ case Intrinsic::frameaddress: {
+ MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
+ MFI->setFrameAddressIsTaken(true);
+
+ unsigned LdrOpc;
+ const TargetRegisterClass *RC;
+ if (isThumb2) {
+ LdrOpc = ARM::t2LDRi12;
+ RC = (const TargetRegisterClass*)&ARM::tGPRRegClass;
+ } else {
+ LdrOpc = ARM::LDRi12;
+ RC = (const TargetRegisterClass*)&ARM::GPRRegClass;
+ }
+
+ const ARMBaseRegisterInfo *RegInfo =
+ static_cast<const ARMBaseRegisterInfo*>(TM.getRegisterInfo());
+ unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
+ unsigned SrcReg = FramePtr;
+
+ // Recursively load frame address
+ // ldr r0 [fp]
+ // ldr r0 [r0]
+ // ldr r0 [r0]
+ // ...
+ unsigned DestReg;
+ unsigned Depth = cast<ConstantInt>(I.getOperand(0))->getZExtValue();
+ while (Depth--) {
+ DestReg = createResultReg(RC);
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(LdrOpc), DestReg)
+ .addReg(SrcReg).addImm(0));
+ SrcReg = DestReg;
+ }
+ UpdateValueMap(&I, SrcReg);
+ return true;
+ }
case Intrinsic::memcpy:
case Intrinsic::memmove: {
const MemTransferInst &MTI = cast<MemTransferInst>(I);
return true;
}
}
-
+
if (!MTI.getLength()->getType()->isIntegerTy(32))
return false;
-
+
if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
return false;
// Don't handle volatile.
if (MSI.isVolatile())
return false;
-
+
if (!MSI.getLength()->getType()->isIntegerTy(32))
return false;
-
+
if (MSI.getDestAddressSpace() > 255)
return false;
-
+
return SelectCall(&I, "memset");
}
+ case Intrinsic::trap: {
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::TRAP));
+ return true;
+ }
}
}
bool ARMFastISel::SelectTrunc(const Instruction *I) {
- // The high bits for a type smaller than the register size are assumed to be
+ // The high bits for a type smaller than the register size are assumed to be
// undefined.
Value *Op = I->getOperand(0);
return true;
}
+bool ARMFastISel::SelectShift(const Instruction *I,
+ ARM_AM::ShiftOpc ShiftTy) {
+ // We handle thumb2 mode by target independent selector
+ // or SelectionDAG ISel.
+ if (isThumb2)
+ return false;
+
+ // Only handle i32 now.
+ EVT DestVT = TLI.getValueType(I->getType(), true);
+ if (DestVT != MVT::i32)
+ return false;
+
+ unsigned Opc = ARM::MOVsr;
+ unsigned ShiftImm;
+ Value *Src2Value = I->getOperand(1);
+ if (const ConstantInt *CI = dyn_cast<ConstantInt>(Src2Value)) {
+ ShiftImm = CI->getZExtValue();
+
+ // Fall back to selection DAG isel if the shift amount
+ // is zero or greater than the width of the value type.
+ if (ShiftImm == 0 || ShiftImm >=32)
+ return false;
+
+ Opc = ARM::MOVsi;
+ }
+
+ Value *Src1Value = I->getOperand(0);
+ unsigned Reg1 = getRegForValue(Src1Value);
+ if (Reg1 == 0) return false;
+
+ unsigned Reg2 = 0;
+ if (Opc == ARM::MOVsr) {
+ Reg2 = getRegForValue(Src2Value);
+ if (Reg2 == 0) return false;
+ }
+
+ unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ if(ResultReg == 0) return false;
+
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg)
+ .addReg(Reg1);
+
+ if (Opc == ARM::MOVsi)
+ MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, ShiftImm));
+ else if (Opc == ARM::MOVsr) {
+ MIB.addReg(Reg2);
+ MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, 0));
+ }
+
+ AddOptionalDefs(MIB);
+ UpdateValueMap(I, ResultReg);
+ return true;
+}
+
// TODO: SoftFP support.
bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
case Instruction::ZExt:
case Instruction::SExt:
return SelectIntExt(I);
+ case Instruction::Shl:
+ return SelectShift(I, ARM_AM::lsl);
+ case Instruction::LShr:
+ return SelectShift(I, ARM_AM::lsr);
+ case Instruction::AShr:
+ return SelectShift(I, ARM_AM::asr);
default: break;
}
return false;
// See if we can handle this address.
Address Addr;
if (!ARMComputeAddress(LI->getOperand(0), Addr)) return false;
-
+
unsigned ResultReg = MI->getOperand(0).getReg();
if (!ARMEmitLoad(VT, ResultReg, Addr, LI->getAlignment(), isZExt, false))
return false;
return true;
}
+unsigned ARMFastISel::ARMLowerPICELF(const GlobalValue *GV,
+ unsigned Align, EVT VT) {
+ bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
+ ARMConstantPoolConstant *CPV =
+ ARMConstantPoolConstant::Create(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
+ unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
+
+ unsigned Opc;
+ unsigned DestReg1 = createResultReg(TLI.getRegClassFor(VT));
+ // Load value.
+ if (isThumb2) {
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::t2LDRpci), DestReg1)
+ .addConstantPoolIndex(Idx));
+ Opc = UseGOTOFF ? ARM::t2ADDrr : ARM::t2LDRs;
+ } else {
+ // The extra immediate is for addrmode2.
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(ARM::LDRcp), DestReg1)
+ .addConstantPoolIndex(Idx).addImm(0));
+ Opc = UseGOTOFF ? ARM::ADDrr : ARM::LDRrs;
+ }
+
+ unsigned GlobalBaseReg = AFI->getGlobalBaseReg();
+ if (GlobalBaseReg == 0) {
+ GlobalBaseReg = MRI.createVirtualRegister(TLI.getRegClassFor(VT));
+ AFI->setGlobalBaseReg(GlobalBaseReg);
+ }
+
+ unsigned DestReg2 = createResultReg(TLI.getRegClassFor(VT));
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(Opc), DestReg2)
+ .addReg(DestReg1)
+ .addReg(GlobalBaseReg);
+ if (!UseGOTOFF)
+ MIB.addImm(0);
+ AddOptionalDefs(MIB);
+
+ return DestReg2;
+}
+
namespace llvm {
- llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
+ FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo,
+ const TargetLibraryInfo *libInfo) {
// Completely untested on non-iOS.
const TargetMachine &TM = funcInfo.MF->getTarget();
// Darwin and thumb1 only for now.
const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
- if (Subtarget->isTargetIOS() && !Subtarget->isThumb1Only() &&
- !DisableARMFastISel)
- return new ARMFastISel(funcInfo);
+ if (Subtarget->isTargetIOS() && !Subtarget->isThumb1Only())
+ return new ARMFastISel(funcInfo, libInfo);
return 0;
}
}