#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Module.h"
+#include "llvm/Operator.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
using namespace llvm;
static cl::opt<bool>
-EnableARMFastISel("arm-fast-isel",
- cl::desc("Turn on experimental ARM fast-isel support"),
- cl::init(false), cl::Hidden);
+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 {
+ // All possible address modes, plus some.
+ typedef struct Address {
+ enum {
+ RegBase,
+ FrameIndexBase
+ } BaseType;
+
+ union {
+ unsigned Reg;
+ int FI;
+ } Base;
+
+ int Offset;
+ unsigned Scale;
+ unsigned PlusReg;
+
+ // Innocuous defaults for our address.
+ Address()
+ : BaseType(RegBase), Offset(0), Scale(0), PlusReg(0) {
+ Base.Reg = 0;
+ }
+ } Address;
+
class ARMFastISel : public FastISel {
/// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
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,
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill,
const ConstantFP *FPImm);
- virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- uint64_t Imm);
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_extractsubreg(MVT RetVT,
unsigned Op0, bool Op0IsKill,
uint32_t Idx);
// Instruction selection routines.
private:
- virtual bool SelectLoad(const Instruction *I);
- virtual bool SelectStore(const Instruction *I);
- virtual bool SelectBranch(const Instruction *I);
- virtual bool SelectCmp(const Instruction *I);
- virtual bool SelectFPExt(const Instruction *I);
- virtual bool SelectFPTrunc(const Instruction *I);
- virtual bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
- virtual bool SelectSIToFP(const Instruction *I);
- virtual bool SelectFPToSI(const Instruction *I);
- virtual bool SelectSDiv(const Instruction *I);
- virtual bool SelectCall(const Instruction *I);
+ bool SelectLoad(const Instruction *I);
+ bool SelectStore(const Instruction *I);
+ bool SelectBranch(const Instruction *I);
+ bool SelectCmp(const Instruction *I);
+ bool SelectFPExt(const Instruction *I);
+ bool SelectFPTrunc(const Instruction *I);
+ bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
+ bool SelectSIToFP(const Instruction *I);
+ bool SelectFPToSI(const Instruction *I);
+ bool SelectSDiv(const Instruction *I);
+ bool SelectSRem(const Instruction *I);
+ bool SelectCall(const Instruction *I);
+ bool SelectSelect(const Instruction *I);
+ bool SelectRet(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, EVT VT);
- bool ARMStoreAlloca(const Instruction *I, unsigned SrcReg, EVT VT);
- bool ARMComputeRegOffset(const Value *Obj, unsigned &Reg, int &Offset);
+ bool isTypeLegal(const Type *Ty, MVT &VT);
+ bool isLoadTypeLegal(const Type *Ty, MVT &VT);
+ bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr);
+ bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
+ bool ARMComputeAddress(const Value *Obj, Address &Addr);
+ void ARMSimplifyAddress(Address &Addr, EVT VT);
unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
unsigned ARMMaterializeInt(const Constant *C, EVT VT);
unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
+ unsigned ARMSelectCallOp(const GlobalValue *GV);
// Call handling routines.
private:
+ bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
+ unsigned &ResultReg);
CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
- bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
+ bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
SmallVectorImpl<unsigned> &ArgRegs,
- SmallVectorImpl<EVT> &ArgVTs,
+ SmallVectorImpl<MVT> &ArgVTs,
SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs,
CallingConv::ID CC,
unsigned &NumBytes);
- bool FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
+ bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC,
unsigned &NumBytes);
bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
// OptionalDef handling routines.
private:
+ bool isARMNEONPred(const MachineInstr *MI);
bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
+ void AddLoadStoreOperands(EVT VT, Address &Addr,
+ const MachineInstrBuilder &MIB);
};
} // end anonymous namespace
return true;
}
+bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
+ const TargetInstrDesc &TID = MI->getDesc();
+
+ // If we're a thumb2 or not NEON function we were handled via isPredicable.
+ if ((TID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
+ AFI->isThumb2Function())
+ return false;
+
+ for (unsigned i = 0, e = TID.getNumOperands(); i != e; ++i)
+ if (TID.OpInfo[i].isPredicate())
+ return true;
+
+ return false;
+}
+
// If the machine is predicable go ahead and add the predicate operands, if
// it needs default CC operands add those.
+// TODO: If we want to support thumb1 then we'll need to deal with optional
+// CPSR defs that need to be added before the remaining operands. See s_cc_out
+// for descriptions why.
const MachineInstrBuilder &
ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
MachineInstr *MI = &*MIB;
- // Do we use a predicate?
- if (TII.isPredicable(MI))
+ // Do we use a predicate? or...
+ // Are we NEON in ARM mode and have a predicate operand? If so, I know
+ // we're not predicable but add it anyways.
+ if (TII.isPredicable(MI) || isARMNEONPred(MI))
AddDefaultPred(MIB);
-
+
// Do we optionally set a predicate? Preds is size > 0 iff the predicate
// defines CPSR. All other OptionalDefines in ARM are the CCR register.
bool CPSR = false;
return ResultReg;
}
+unsigned ARMFastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ unsigned Op2, bool Op2IsKill) {
+ unsigned ResultReg = createResultReg(RC);
+ const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+
+ if (II.getNumDefs() >= 1)
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ .addReg(Op0, Op0IsKill * RegState::Kill)
+ .addReg(Op1, Op1IsKill * RegState::Kill)
+ .addReg(Op2, Op2IsKill * RegState::Kill));
+ else {
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ .addReg(Op0, Op0IsKill * RegState::Kill)
+ .addReg(Op1, Op1IsKill * RegState::Kill)
+ .addReg(Op2, Op2IsKill * RegState::Kill));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(TargetOpcode::COPY), ResultReg)
+ .addReg(II.ImplicitDefs[0]));
+ }
+ return ResultReg;
+}
+
unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill,
// TODO: Don't worry about 64-bit now, but when this is fixed remove the
// checks from the various callers.
unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
- if (VT.getSimpleVT().SimpleTy == MVT::f64) return 0;
-
+ if (VT == MVT::f64) return 0;
+
unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VMOVRS), MoveReg)
}
unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
- if (VT.getSimpleVT().SimpleTy == MVT::i64) return 0;
-
+ if (VT == MVT::i64) return 0;
+
unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VMOVSR), MoveReg)
// 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;
+ bool is64bit = VT == 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.
.addFPImm(CFP));
return DestReg;
}
-
+
// Require VFP2 for loading fp constants.
if (!Subtarget->hasVFP2()) return false;
-
+
// MachineConstantPool wants an explicit alignment.
unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
if (Align == 0) {
unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
-
+
// The extra reg is for addrmode5.
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
DestReg)
}
unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
-
+
// For now 32-bit only.
- if (VT.getSimpleVT().SimpleTy != MVT::i32) return false;
-
+ if (VT != MVT::i32) return false;
+
+ unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
+
+ // If we can do this in a single instruction without a constant pool entry
+ // do so now.
+ const ConstantInt *CI = cast<ConstantInt>(C);
+ if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getSExtValue())) {
+ unsigned Opc = isThumb ? ARM::t2MOVi16 : ARM::MOVi16;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), DestReg)
+ .addImm(CI->getSExtValue()));
+ return DestReg;
+ }
+
// MachineConstantPool wants an explicit alignment.
unsigned Align = TD.getPrefTypeAlignment(C->getType());
if (Align == 0) {
Align = TD.getTypeAllocSize(C->getType());
}
unsigned Idx = MCP.getConstantPoolIndex(C, Align);
- unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
-
+
if (isThumb)
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::t2LDRpci), DestReg)
.addConstantPoolIndex(Idx));
else
- // The extra reg and immediate are for addrmode2.
+ // The extra immediate is for addrmode2.
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::LDRcp), DestReg)
.addConstantPoolIndex(Idx)
- .addReg(0).addImm(0));
+ .addImm(0));
return DestReg;
}
unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
// For now 32-bit only.
- if (VT.getSimpleVT().SimpleTy != MVT::i32) return 0;
-
+ if (VT != MVT::i32) return 0;
+
Reloc::Model RelocM = TM.getRelocationModel();
-
+
// TODO: No external globals for now.
if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) return 0;
-
+
// TODO: Need more magic for ARM PIC.
if (!isThumb && (RelocM == Reloc::PIC_)) return 0;
-
+
// MachineConstantPool wants an explicit alignment.
unsigned Align = TD.getPrefTypeAlignment(GV->getType());
if (Align == 0) {
// TODO: Figure out if this is correct.
Align = TD.getTypeAllocSize(GV->getType());
}
-
+
// Grab index.
unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
- unsigned Id = AFI->createConstPoolEntryUId();
+ unsigned Id = AFI->createPICLabelUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, Id,
ARMCP::CPValue, PCAdj);
unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
-
+
// Load value.
MachineInstrBuilder MIB;
unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
if (RelocM == Reloc::PIC_)
MIB.addImm(Id);
} else {
- // The extra reg and immediate are for addrmode2.
+ // The extra immediate is for addrmode2.
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
DestReg)
.addConstantPoolIndex(Idx)
- .addReg(0).addImm(0);
+ .addImm(0);
}
AddOptionalDefs(MIB);
return DestReg;
return ARMMaterializeGV(GV, VT);
else if (isa<ConstantInt>(C))
return ARMMaterializeInt(C, VT);
-
+
return 0;
}
unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
// Don't handle dynamic allocas.
if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
-
- EVT VT;
- if (!isTypeLegal(AI->getType(), VT)) return false;
-
+
+ MVT VT;
+ if (!isLoadTypeLegal(AI->getType(), VT)) return false;
+
DenseMap<const AllocaInst*, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
.addImm(0));
return ResultReg;
}
-
+
return 0;
}
-bool ARMFastISel::isTypeLegal(const Type *Ty, EVT &VT) {
- VT = TLI.getValueType(Ty, true);
+bool ARMFastISel::isTypeLegal(const Type *Ty, MVT &VT) {
+ EVT evt = TLI.getValueType(Ty, true);
// Only handle simple types.
- if (VT == MVT::Other || !VT.isSimple()) return false;
+ if (evt == MVT::Other || !evt.isSimple()) return false;
+ VT = evt.getSimpleVT();
// 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) {
+bool ARMFastISel::isLoadTypeLegal(const Type *Ty, MVT &VT) {
if (isTypeLegal(Ty, VT)) return true;
// If this is a type than can be sign or zero-extended to a basic operation
return false;
}
-// Computes the Reg+Offset to get to an object.
-bool ARMFastISel::ARMComputeRegOffset(const Value *Obj, unsigned &Reg,
- int &Offset) {
+// Computes the address to get to an object.
+bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
// Some boilerplate from the X86 FastISel.
const User *U = NULL;
unsigned Opcode = Instruction::UserOp1;
if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
- // Don't walk into other basic blocks; it's possible we haven't
- // visited them yet, so the instructions may not yet be assigned
- // virtual registers.
- if (FuncInfo.MBBMap[I->getParent()] != FuncInfo.MBB)
- return false;
- Opcode = I->getOpcode();
- U = I;
+ // Don't walk into other basic blocks unless the object is an alloca from
+ // another block, otherwise it may not have a virtual register assigned.
+ if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
+ FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
+ Opcode = I->getOpcode();
+ U = I;
+ }
} else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
Opcode = C->getOpcode();
U = C;
switch (Opcode) {
default:
break;
+ case Instruction::BitCast: {
+ // Look through bitcasts.
+ return ARMComputeAddress(U->getOperand(0), Addr);
+ }
+ case Instruction::IntToPtr: {
+ // Look past no-op inttoptrs.
+ if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
+ return ARMComputeAddress(U->getOperand(0), Addr);
+ break;
+ }
+ case Instruction::PtrToInt: {
+ // Look past no-op ptrtoints.
+ if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
+ return ARMComputeAddress(U->getOperand(0), Addr);
+ break;
+ }
+ case Instruction::GetElementPtr: {
+ Address SavedAddr = Addr;
+ int TmpOffset = Addr.Offset;
+
+ // Iterate through the GEP folding the constants into offsets where
+ // we can.
+ gep_type_iterator GTI = gep_type_begin(U);
+ for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
+ i != e; ++i, ++GTI) {
+ const Value *Op = *i;
+ if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
+ const StructLayout *SL = TD.getStructLayout(STy);
+ unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
+ TmpOffset += SL->getElementOffset(Idx);
+ } else {
+ uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
+ for (;;) {
+ if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
+ // Constant-offset addressing.
+ TmpOffset += CI->getSExtValue() * S;
+ break;
+ }
+ if (isa<AddOperator>(Op) &&
+ (!isa<Instruction>(Op) ||
+ FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
+ == FuncInfo.MBB) &&
+ isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
+ // An add (in the same block) with a constant operand. Fold the
+ // constant.
+ ConstantInt *CI =
+ cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
+ TmpOffset += CI->getSExtValue() * S;
+ // Iterate on the other operand.
+ Op = cast<AddOperator>(Op)->getOperand(0);
+ continue;
+ }
+ // Unsupported
+ goto unsupported_gep;
+ }
+ }
+ }
+
+ // Try to grab the base operand now.
+ Addr.Offset = TmpOffset;
+ if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
+
+ // We failed, restore everything and try the other options.
+ Addr = SavedAddr;
+
+ unsupported_gep:
+ break;
+ }
case Instruction::Alloca: {
- assert(false && "Alloca should have been handled earlier!");
- return false;
+ const AllocaInst *AI = cast<AllocaInst>(Obj);
+ DenseMap<const AllocaInst*, int>::iterator SI =
+ FuncInfo.StaticAllocaMap.find(AI);
+ if (SI != FuncInfo.StaticAllocaMap.end()) {
+ Addr.BaseType = Address::FrameIndexBase;
+ Addr.Base.FI = SI->second;
+ return true;
+ }
+ break;
}
}
- // FIXME: Handle global variables.
+ // Materialize the global variable's address into a reg which can
+ // then be used later to load the variable.
if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
- (void)GV;
- return false;
+ unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
+ if (Tmp == 0) return false;
+
+ Addr.Base.Reg = Tmp;
+ return true;
}
// Try to get this in a register if nothing else has worked.
- Reg = getRegForValue(Obj);
- if (Reg == 0) return false;
+ if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
+ return Addr.Base.Reg != 0;
+}
- // Since the offset may be too large for the load instruction
+void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
+
+ assert(VT.isSimple() && "Non-simple types are invalid here!");
+
+ bool needsLowering = false;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default:
+ assert(false && "Unhandled load/store type!");
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ // Integer loads/stores handle 12-bit offsets.
+ needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
+ break;
+ case MVT::f32:
+ case MVT::f64:
+ // Floating point operands handle 8-bit offsets.
+ needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
+ break;
+ }
+
+ // If this is a stack pointer and the offset needs to be simplified then
+ // 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) {
+ TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
+ ARM::GPRRegisterClass;
+ unsigned ResultReg = createResultReg(RC);
+ unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg)
+ .addFrameIndex(Addr.Base.FI)
+ .addImm(0));
+ Addr.Base.Reg = ResultReg;
+ Addr.BaseType = Address::RegBase;
+ }
+
+ // Since the offset is too large for the load/store 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) {
+ if (needsLowering) {
ARMCC::CondCodes Pred = ARMCC::AL;
unsigned PredReg = 0;
+ TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
+ ARM::GPRRegisterClass;
+ unsigned BaseReg = createResultReg(RC);
+
if (!isThumb)
emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- Reg, Reg, Offset, Pred, PredReg,
+ BaseReg, Addr.Base.Reg, Addr.Offset,
+ Pred, PredReg,
static_cast<const ARMBaseInstrInfo&>(TII));
else {
assert(AFI->isThumb2Function());
emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- Reg, Reg, Offset, Pred, PredReg,
+ BaseReg, Addr.Base.Reg, Addr.Offset, Pred, PredReg,
static_cast<const ARMBaseInstrInfo&>(TII));
}
+ Addr.Offset = 0;
+ Addr.Base.Reg = BaseReg;
}
- return true;
}
-bool ARMFastISel::ARMLoadAlloca(const Instruction *I, EVT VT) {
- Value *Op0 = I->getOperand(0);
+void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
+ const MachineInstrBuilder &MIB) {
+ // addrmode5 output depends on the selection dag addressing dividing the
+ // offset by 4 that it then later multiplies. Do this here as well.
+ if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
+ VT.getSimpleVT().SimpleTy == MVT::f64)
+ Addr.Offset /= 4;
+
+ // Frame base works a bit differently. Handle it separately.
+ if (Addr.BaseType == Address::FrameIndexBase) {
+ int FI = Addr.Base.FI;
+ int Offset = Addr.Offset;
+ MachineMemOperand *MMO =
+ FuncInfo.MF->getMachineMemOperand(
+ MachinePointerInfo::getFixedStack(FI, Offset),
+ MachineMemOperand::MOLoad,
+ MFI.getObjectSize(FI),
+ MFI.getObjectAlignment(FI));
+ // Now add the rest of the operands.
+ MIB.addFrameIndex(FI);
+
+ // ARM halfword load/stores need an additional operand.
+ if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
+
+ MIB.addImm(Addr.Offset);
+ MIB.addMemOperand(MMO);
+ } else {
+ // Now add the rest of the operands.
+ MIB.addReg(Addr.Base.Reg);
+
+ // ARM halfword load/stores need an additional operand.
+ if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(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(VT);
- unsigned ResultReg = createResultReg(RC);
- TII.loadRegFromStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
- ResultReg, SI->second, RC,
- TM.getRegisterInfo());
- UpdateValueMap(I, ResultReg);
- return true;
- }
+ MIB.addImm(Addr.Offset);
}
- return false;
+ AddOptionalDefs(MIB);
}
-bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg,
- unsigned Reg, int Offset) {
+bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
assert(VT.isSimple() && "Non-simple types are invalid here!");
unsigned Opc;
- bool isFloat = false;
+ TargetRegisterClass *RC;
switch (VT.getSimpleVT().SimpleTy) {
- default:
- // This is mostly going to be Neon/vector support.
- return false;
+ // This is mostly going to be Neon/vector support.
+ default: return false;
case MVT::i16:
- Opc = isThumb ? ARM::tLDRH : ARM::LDRH;
- VT = MVT::i32;
+ Opc = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
+ RC = ARM::GPRRegisterClass;
break;
case MVT::i8:
- Opc = isThumb ? ARM::tLDRB : ARM::LDRB;
- VT = MVT::i32;
+ Opc = isThumb ? ARM::t2LDRBi12 : ARM::LDRBi12;
+ RC = ARM::GPRRegisterClass;
break;
case MVT::i32:
- Opc = isThumb ? ARM::tLDR : ARM::LDR;
+ Opc = isThumb ? ARM::t2LDRi12 : ARM::LDRi12;
+ RC = ARM::GPRRegisterClass;
break;
case MVT::f32:
Opc = ARM::VLDRS;
- isFloat = true;
+ RC = TLI.getRegClassFor(VT);
break;
case MVT::f64:
Opc = ARM::VLDRD;
- isFloat = true;
+ RC = TLI.getRegClassFor(VT);
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.
- // The thumb addressing mode has operands swapped from the arm addressing
- // mode, the floating point one only has two operands.
- if (isFloat)
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(Opc), ResultReg)
- .addReg(Reg).addImm(Offset));
- else 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));
+ // Simplify this down to something we can handle.
+ ARMSimplifyAddress(Addr, VT);
+
+ // Create the base instruction, then add the operands.
+ ResultReg = createResultReg(RC);
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg);
+ AddLoadStoreOperands(VT, Addr, MIB);
return true;
}
bool ARMFastISel::SelectLoad(const Instruction *I) {
// Verify we have a legal type before going any further.
- EVT VT;
+ MVT 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;
+ // See if we can handle this address.
+ Address Addr;
+ if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
unsigned ResultReg;
- if (!ARMEmitLoad(VT, ResultReg, Reg, Offset /* 0 */)) return false;
-
+ if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
UpdateValueMap(I, ResultReg);
return true;
}
-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) {
+bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
unsigned StrOpc;
- bool isFloat = false;
switch (VT.getSimpleVT().SimpleTy) {
+ // This is mostly going to be Neon/vector support.
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::i1: {
+ unsigned Res = createResultReg(isThumb ? ARM::tGPRRegisterClass :
+ ARM::GPRRegisterClass);
+ unsigned Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), Res)
+ .addReg(SrcReg).addImm(1));
+ SrcReg = Res;
+ } // Fallthrough here.
+ case MVT::i8:
+ StrOpc = isThumb ? ARM::t2STRBi12 : ARM::STRBi12;
+ break;
+ case MVT::i16:
+ StrOpc = isThumb ? ARM::t2STRHi12 : ARM::STRH;
+ break;
+ case MVT::i32:
+ StrOpc = isThumb ? ARM::t2STRi12 : ARM::STRi12;
+ break;
case MVT::f32:
if (!Subtarget->hasVFP2()) return false;
StrOpc = ARM::VSTRS;
- isFloat = true;
break;
case MVT::f64:
if (!Subtarget->hasVFP2()) return false;
StrOpc = ARM::VSTRD;
- isFloat = true;
break;
}
-
- // The thumb addressing mode has operands swapped from the arm addressing
- // mode, the floating point one only has two operands.
- if (isFloat)
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(StrOpc))
- .addReg(SrcReg).addReg(DstReg).addImm(Offset));
- else if (isThumb)
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(StrOpc))
- .addReg(SrcReg).addReg(DstReg).addImm(Offset).addReg(0));
-
- else
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(StrOpc))
- .addReg(SrcReg).addReg(DstReg).addReg(0).addImm(Offset));
-
+ // Simplify this down to something we can handle.
+ ARMSimplifyAddress(Addr, VT);
+
+ // Create the base instruction, then add the operands.
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(StrOpc))
+ .addReg(SrcReg, getKillRegState(true));
+ AddLoadStoreOperands(VT, Addr, MIB);
return true;
}
Value *Op0 = I->getOperand(0);
unsigned SrcReg = 0;
- // Yay type legalization
- EVT VT;
+ // Verify we have a legal type before going any further.
+ MVT 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;
+ if (SrcReg == 0) return false;
- // See if we can handle this as Reg + Offset
- if (!ARMComputeRegOffset(I->getOperand(1), Reg, Offset))
+ // See if we can handle this address.
+ Address Addr;
+ if (!ARMComputeAddress(I->getOperand(1), Addr))
return false;
- if (!ARMEmitStore(VT, SrcReg, Reg, Offset /* 0 */)) return false;
-
+ if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
return true;
}
switch (Pred) {
// Needs two compares...
case CmpInst::FCMP_ONE:
- case CmpInst::FCMP_UEQ:
+ case CmpInst::FCMP_UEQ:
default:
- assert(false && "Unhandled CmpInst::Predicate!");
+ // AL is our "false" for now. The other two need more compares.
return ARMCC::AL;
case CmpInst::ICMP_EQ:
case CmpInst::FCMP_OEQ:
return ARMCC::PL;
case CmpInst::ICMP_SLT:
case CmpInst::FCMP_ULT:
- return ARMCC::LT;
+ return ARMCC::LT;
case CmpInst::ICMP_SLE:
case CmpInst::FCMP_ULE:
return ARMCC::LE;
MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
// Simple branch support.
- // TODO: Try to avoid the re-computation in some places.
- unsigned CondReg = getRegForValue(BI->getCondition());
- if (CondReg == 0) return false;
- // Re-set the flags just in case.
- unsigned CmpOpc = isThumb ? ARM::t2CMPri : ARM::CMPri;
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
- .addReg(CondReg).addImm(1));
-
+ // If we can, avoid recomputing the compare - redoing it could lead to wonky
+ // behavior.
+ // TODO: Factor this out.
+ if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
+ if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
+ MVT 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.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;
+ }
+
+ // Get the compare predicate.
+ ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
+
+ // We may not handle every CC for now.
+ if (ARMPred == ARMCC::AL) return false;
+
+ 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 comparison register
+ // that we can then use for branches.
+ if (isFloat)
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::FMSTAT)));
+
+ unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
+ .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
+ FastEmitBranch(FBB, DL);
+ FuncInfo.MBB->addSuccessor(TBB);
+ return true;
+ }
+ }
+
+ unsigned CmpReg = getRegForValue(BI->getCondition());
+ if (CmpReg == 0) return false;
+
+ // We've been divorced from our compare! Our block was split, and
+ // now our compare lives in a predecessor block. We musn't
+ // re-compare here, as the children of the compare aren't guaranteed
+ // live across the block boundary (we *could* check for this).
+ // Regardless, the compare has been done in the predecessor block,
+ // and it left a value for us in a virtual register. Ergo, we test
+ // the one-bit value left in the virtual register.
+ unsigned TstOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
+ .addReg(CmpReg).addImm(1));
+
unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
- .addMBB(TBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
+ .addMBB(TBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
FastEmitBranch(FBB, DL);
FuncInfo.MBB->addSuccessor(TBB);
- return true;
+ return true;
}
bool ARMFastISel::SelectCmp(const Instruction *I) {
const CmpInst *CI = cast<CmpInst>(I);
- EVT VT;
+ MVT VT;
const Type *Ty = CI->getOperand(0)->getType();
if (!isTypeLegal(Ty, VT))
return false;
unsigned CmpOpc;
unsigned CondReg;
- switch (VT.getSimpleVT().SimpleTy) {
+ switch (VT.SimpleTy) {
default: return false;
// TODO: Verify compares.
case MVT::f32:
// Get the compare predicate.
ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
-
+
// We may not handle every CC for now.
if (ARMPred == ARMCC::AL) return false;
// Now set a register based on the comparison. Explicitly set the predicates
// here.
- unsigned MovCCOpc = isThumb ? ARM::tMOVCCi : ARM::MOVCCi;
- unsigned DestReg = createResultReg(ARM::GPRRegisterClass);
- Constant *Zero
+ unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
+ TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
+ : ARM::GPRRegisterClass;
+ unsigned DestReg = createResultReg(RC);
+ Constant *Zero
= ConstantInt::get(Type::getInt32Ty(*Context), 0);
unsigned ZeroReg = TargetMaterializeConstant(Zero);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
bool ARMFastISel::SelectSIToFP(const Instruction *I) {
// Make sure we have VFP.
if (!Subtarget->hasVFP2()) return false;
-
- EVT DstVT;
+
+ MVT DstVT;
const Type *Ty = I->getType();
if (!isTypeLegal(Ty, DstVT))
return false;
-
+
unsigned Op = getRegForValue(I->getOperand(0));
if (Op == 0) return false;
-
+
// The conversion routine works on fp-reg to fp-reg and the operand above
// was an integer, move it to the fp registers if possible.
unsigned FP = ARMMoveToFPReg(MVT::f32, Op);
if (FP == 0) return false;
-
+
unsigned Opc;
if (Ty->isFloatTy()) Opc = ARM::VSITOS;
else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
else return 0;
-
+
unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
ResultReg)
bool ARMFastISel::SelectFPToSI(const Instruction *I) {
// Make sure we have VFP.
if (!Subtarget->hasVFP2()) return false;
-
- EVT DstVT;
+
+ MVT DstVT;
const Type *RetTy = I->getType();
if (!isTypeLegal(RetTy, DstVT))
return false;
-
+
unsigned Op = getRegForValue(I->getOperand(0));
if (Op == 0) return false;
-
+
unsigned Opc;
const Type *OpTy = I->getOperand(0)->getType();
if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
else return 0;
-
+
// f64->s32 or f32->s32 both need an intermediate f32 reg.
unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
ResultReg)
.addReg(Op));
-
+
// This result needs to be in an integer register, but the conversion only
// takes place in fp-regs.
unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
if (IntReg == 0) return false;
-
+
UpdateValueMap(I, IntReg);
return true;
}
+bool ARMFastISel::SelectSelect(const Instruction *I) {
+ MVT VT;
+ if (!isTypeLegal(I->getType(), VT))
+ return false;
+
+ // Things need to be register sized for register moves.
+ if (VT != MVT::i32) return false;
+ const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
+
+ unsigned CondReg = getRegForValue(I->getOperand(0));
+ if (CondReg == 0) return false;
+ unsigned Op1Reg = getRegForValue(I->getOperand(1));
+ if (Op1Reg == 0) return false;
+ unsigned Op2Reg = getRegForValue(I->getOperand(2));
+ if (Op2Reg == 0) return false;
+
+ unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
+ .addReg(CondReg).addImm(1));
+ unsigned ResultReg = createResultReg(RC);
+ unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
+ .addReg(Op1Reg).addReg(Op2Reg)
+ .addImm(ARMCC::EQ).addReg(ARM::CPSR);
+ UpdateValueMap(I, ResultReg);
+ return true;
+}
+
bool ARMFastISel::SelectSDiv(const Instruction *I) {
- EVT VT;
+ MVT VT;
const Type *Ty = I->getType();
if (!isTypeLegal(Ty, VT))
return false;
// If we have integer div support we should have selected this automagically.
// In case we have a real miss go ahead and return false and we'll pick
// it up later.
- if (Subtarget->hasDivide()) return false;
-
+ if (Subtarget->hasDivide()) return false;
+
// Otherwise emit a libcall.
RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
- if (VT == MVT::i16)
+ if (VT == MVT::i8)
+ LC = RTLIB::SDIV_I8;
+ else if (VT == MVT::i16)
LC = RTLIB::SDIV_I16;
else if (VT == MVT::i32)
LC = RTLIB::SDIV_I32;
else if (VT == MVT::i128)
LC = RTLIB::SDIV_I128;
assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
-
+
+ return ARMEmitLibcall(I, LC);
+}
+
+bool ARMFastISel::SelectSRem(const Instruction *I) {
+ MVT VT;
+ const Type *Ty = I->getType();
+ if (!isTypeLegal(Ty, VT))
+ return false;
+
+ RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
+ if (VT == MVT::i8)
+ LC = RTLIB::SREM_I8;
+ else if (VT == MVT::i16)
+ LC = RTLIB::SREM_I16;
+ else if (VT == MVT::i32)
+ LC = RTLIB::SREM_I32;
+ else if (VT == MVT::i64)
+ LC = RTLIB::SREM_I64;
+ else if (VT == MVT::i128)
+ LC = RTLIB::SREM_I128;
+ assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
+
return ARMEmitLibcall(I, LC);
}
if (Op2 == 0) return false;
unsigned Opc;
- bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64 ||
- VT.getSimpleVT().SimpleTy == MVT::i64;
+ bool is64bit = VT == MVT::f64 || VT == MVT::i64;
switch (ISDOpcode) {
default: return false;
case ISD::FADD:
// Call Handling Code
+bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
+ EVT SrcVT, unsigned &ResultReg) {
+ unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
+ Src, /*TODO: Kill=*/false);
+
+ if (RR != 0) {
+ ResultReg = RR;
+ return true;
+ } else
+ return false;
+}
+
// This is largely taken directly from CCAssignFnForNode - we don't support
// varargs in FastISel so that part has been removed.
// TODO: We may not support all of this.
switch (CC) {
default:
llvm_unreachable("Unsupported calling convention");
- case CallingConv::C:
case CallingConv::Fast:
+ // Ignore fastcc. Silence compiler warnings.
+ (void)RetFastCC_ARM_APCS;
+ (void)FastCC_ARM_APCS;
+ // Fallthrough
+ case CallingConv::C:
// Use target triple & subtarget features to do actual dispatch.
if (Subtarget->isAAPCS_ABI()) {
if (Subtarget->hasVFP2() &&
bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
SmallVectorImpl<unsigned> &ArgRegs,
- SmallVectorImpl<EVT> &ArgVTs,
+ SmallVectorImpl<MVT> &ArgVTs,
SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs,
CallingConv::ID CC,
// Issue CALLSEQ_START
unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
- .addImm(NumBytes);
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(AdjStackDown))
+ .addImm(NumBytes));
// Process the args.
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
unsigned Arg = ArgRegs[VA.getValNo()];
- EVT ArgVT = ArgVTs[VA.getValNo()];
+ MVT ArgVT = ArgVTs[VA.getValNo()];
+
+ // We don't handle NEON/vector parameters yet.
+ if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
+ return false;
// Handle arg promotion, etc.
switch (VA.getLocInfo()) {
case CCValAssign::Full: break;
- default:
- assert(false && "Handle arg promotion.");
- return false;
+ case CCValAssign::SExt: {
+ bool Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
+ Arg, ArgVT, Arg);
+ assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
+ Emitted = true;
+ ArgVT = VA.getLocVT();
+ break;
+ }
+ case CCValAssign::ZExt: {
+ bool Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
+ Arg, ArgVT, Arg);
+ assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
+ Emitted = true;
+ ArgVT = VA.getLocVT();
+ break;
+ }
+ case CCValAssign::AExt: {
+ bool Emitted = FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
+ Arg, ArgVT, Arg);
+ if (!Emitted)
+ Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
+ Arg, ArgVT, Arg);
+ if (!Emitted)
+ Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
+ Arg, ArgVT, Arg);
+
+ assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
+ ArgVT = VA.getLocVT();
+ break;
+ }
+ case CCValAssign::BCvt: {
+ unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
+ /*TODO: Kill=*/false);
+ assert(BC != 0 && "Failed to emit a bitcast!");
+ Arg = BC;
+ ArgVT = VA.getLocVT();
+ break;
+ }
+ default: llvm_unreachable("Unknown arg promotion!");
}
// Now copy/store arg to correct locations.
- if (VA.isRegLoc()) {
+ if (VA.isRegLoc() && !VA.needsCustom()) {
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
VA.getLocReg())
.addReg(Arg);
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;
+
+ CCValAssign &NextVA = ArgLocs[++i];
+
+ // TODO: Only handle register args for now.
+ if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
+
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::VMOVRRD), VA.getLocReg())
+ .addReg(NextVA.getLocReg(), RegState::Define)
+ .addReg(Arg));
+ RegArgs.push_back(VA.getLocReg());
+ RegArgs.push_back(NextVA.getLocReg());
} else {
- // Need to store
- return false;
+ assert(VA.isMemLoc());
+ // Need to store on the stack.
+ Address Addr;
+ Addr.BaseType = Address::RegBase;
+ Addr.Base.Reg = ARM::SP;
+ Addr.Offset = VA.getLocMemOffset();
+
+ if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
}
}
-
return true;
}
-bool ARMFastISel::FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
+bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC,
unsigned &NumBytes) {
// Issue CALLSEQ_END
unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
- .addImm(NumBytes).addImm(0);
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(AdjStackUp))
+ .addImm(NumBytes).addImm(0));
// Now the return value.
- if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) {
+ if (RetVT != MVT::isVoid) {
SmallVector<CCValAssign, 16> RVLocs;
CCState CCInfo(CC, false, TM, RVLocs, *Context);
CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
// Copy all of the result registers out of their specified physreg.
- if (RVLocs.size() == 2 && RetVT.getSimpleVT().SimpleTy == MVT::f64) {
+ if (RVLocs.size() == 2 && RetVT == MVT::f64) {
// For this move we copy into two registers and then move into the
// double fp reg we want.
- // TODO: Are the copies necessary?
- TargetRegisterClass *CopyRC = TLI.getRegClassFor(MVT::i32);
- unsigned Copy1 = createResultReg(CopyRC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- Copy1).addReg(RVLocs[0].getLocReg());
- UsedRegs.push_back(RVLocs[0].getLocReg());
-
- unsigned Copy2 = createResultReg(CopyRC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- Copy2).addReg(RVLocs[1].getLocReg());
- UsedRegs.push_back(RVLocs[1].getLocReg());
-
EVT DestVT = RVLocs[0].getValVT();
TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
unsigned ResultReg = createResultReg(DstRC);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VMOVDRR), ResultReg)
- .addReg(Copy1).addReg(Copy2));
-
- // Finally update the result.
+ .addReg(RVLocs[0].getLocReg())
+ .addReg(RVLocs[1].getLocReg()));
+
+ UsedRegs.push_back(RVLocs[0].getLocReg());
+ UsedRegs.push_back(RVLocs[1].getLocReg());
+
+ // Finally update the result.
UpdateValueMap(I, ResultReg);
} else {
- assert(RVLocs.size() == 1 && "Can't handle non-double multi-reg retvals!");
+ assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
EVT CopyVT = RVLocs[0].getValVT();
TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
ResultReg).addReg(RVLocs[0].getLocReg());
UsedRegs.push_back(RVLocs[0].getLocReg());
- // Finally update the result.
+ // Finally update the result.
UpdateValueMap(I, ResultReg);
}
}
- return true;
+ return true;
+}
+
+bool ARMFastISel::SelectRet(const Instruction *I) {
+ const ReturnInst *Ret = cast<ReturnInst>(I);
+ const Function &F = *I->getParent()->getParent();
+
+ if (!FuncInfo.CanLowerReturn)
+ return false;
+
+ if (F.isVarArg())
+ return false;
+
+ CallingConv::ID CC = F.getCallingConv();
+ if (Ret->getNumOperands() > 0) {
+ SmallVector<ISD::OutputArg, 4> Outs;
+ GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
+ Outs, TLI);
+
+ // Analyze operands of the call, assigning locations to each operand.
+ SmallVector<CCValAssign, 16> ValLocs;
+ CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
+ CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
+
+ const Value *RV = Ret->getOperand(0);
+ unsigned Reg = getRegForValue(RV);
+ if (Reg == 0)
+ return false;
+
+ // Only handle a single return value for now.
+ if (ValLocs.size() != 1)
+ return false;
+
+ CCValAssign &VA = ValLocs[0];
+
+ // Don't bother handling odd stuff for now.
+ if (VA.getLocInfo() != CCValAssign::Full)
+ return false;
+ // Only handle register returns for now.
+ if (!VA.isRegLoc())
+ return false;
+ // TODO: For now, don't try to handle cases where getLocInfo()
+ // says Full but the types don't match.
+ if (TLI.getValueType(RV->getType()) != VA.getValVT())
+ return false;
+
+ // Make the copy.
+ unsigned SrcReg = Reg + VA.getValNo();
+ unsigned DstReg = VA.getLocReg();
+ const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
+ // Avoid a cross-class copy. This is very unlikely.
+ if (!SrcRC->contains(DstReg))
+ return false;
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+ DstReg).addReg(SrcReg);
+
+ // Mark the register as live out of the function.
+ MRI.addLiveOut(VA.getLocReg());
+ }
+
+ unsigned RetOpc = isThumb ? ARM::tBX_RET : ARM::BX_RET;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(RetOpc)));
+ return true;
+}
+
+unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
+
+ // Darwin needs the r9 versions of the opcodes.
+ bool isDarwin = Subtarget->isTargetDarwin();
+ if (isThumb) {
+ return isDarwin ? ARM::tBLr9 : ARM::tBL;
+ } else {
+ return isDarwin ? ARM::BLr9 : ARM::BL;
+ }
}
// A quick function that will emit a call for a named libcall in F with the
// vector of passed arguments for the Instruction in I. We can assume that we
-// can emit a call for any libcall we can produce. This is an abridged version
-// of the full call infrastructure since we won't need to worry about things
+// can emit a call for any libcall we can produce. This is an abridged version
+// of the full call infrastructure since we won't need to worry about things
// like computed function pointers or strange arguments at call sites.
// TODO: Try to unify this and the normal call bits for ARM, then try to unify
// with X86.
bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
-
+
// Handle *simple* calls for now.
const Type *RetTy = I->getType();
- EVT RetVT;
+ MVT RetVT;
if (RetTy->isVoidTy())
RetVT = MVT::isVoid;
else if (!isTypeLegal(RetTy, RetVT))
return false;
-
+
// For now we're using BLX etc on the assumption that we have v5t ops.
if (!Subtarget->hasV5TOps()) return false;
-
+
+ // TODO: For now if we have long calls specified we don't handle the call.
+ if (EnableARMLongCalls) return false;
+
// Set up the argument vectors.
SmallVector<Value*, 8> Args;
SmallVector<unsigned, 8> ArgRegs;
- SmallVector<EVT, 8> ArgVTs;
+ SmallVector<MVT, 8> ArgVTs;
SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
Args.reserve(I->getNumOperands());
ArgRegs.reserve(I->getNumOperands());
Value *Op = I->getOperand(i);
unsigned Arg = getRegForValue(Op);
if (Arg == 0) return false;
-
+
const Type *ArgTy = Op->getType();
- EVT ArgVT;
+ MVT ArgVT;
if (!isTypeLegal(ArgTy, ArgVT)) return false;
-
+
ISD::ArgFlagsTy Flags;
unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
Flags.setOrigAlign(OriginalAlignment);
-
+
Args.push_back(Op);
ArgRegs.push_back(Arg);
ArgVTs.push_back(ArgVT);
ArgFlags.push_back(Flags);
}
-
+
// Handle the arguments now that we've gotten them.
SmallVector<unsigned, 4> RegArgs;
unsigned NumBytes;
if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
return false;
-
+
// Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
- // TODO: Turn this into the table of arm call ops.
+ // TODO: Turn this into the table of arm call ops.
MachineInstrBuilder MIB;
- unsigned CallOpc;
+ unsigned CallOpc = ARMSelectCallOp(NULL);
if(isThumb)
- CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
+ // Explicitly adding the predicate here.
+ MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(CallOpc)))
+ .addExternalSymbol(TLI.getLibcallName(Call));
else
- CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
- .addExternalSymbol(TLI.getLibcallName(Call));
-
+ // Explicitly adding the predicate here.
+ MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(CallOpc))
+ .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]);
-
+
// Finish off the call including any return values.
- SmallVector<unsigned, 4> UsedRegs;
+ SmallVector<unsigned, 4> UsedRegs;
if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
-
+
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
-
+
return true;
}
const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel()))
return false;
-
+
// Check the calling convention.
ImmutableCallSite CS(CI);
CallingConv::ID CC = CS.getCallingConv();
+
// TODO: Avoid some calling conventions?
- if (CC != CallingConv::C) {
- errs() << "Can't handle calling convention: " << CC << "\n";
- return false;
- }
-
+
// Let SDISel handle vararg functions.
const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
if (FTy->isVarArg())
return false;
-
+
// Handle *simple* calls for now.
const Type *RetTy = I->getType();
- EVT RetVT;
+ MVT RetVT;
if (RetTy->isVoidTy())
RetVT = MVT::isVoid;
else if (!isTypeLegal(RetTy, RetVT))
return false;
-
+
// For now we're using BLX etc on the assumption that we have v5t ops.
// TODO: Maybe?
if (!Subtarget->hasV5TOps()) return false;
+
+ // TODO: For now if we have long calls specified we don't handle the call.
+ if (EnableARMLongCalls) return false;
// Set up the argument vectors.
SmallVector<Value*, 8> Args;
SmallVector<unsigned, 8> ArgRegs;
- SmallVector<EVT, 8> ArgVTs;
+ SmallVector<MVT, 8> ArgVTs;
SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
Args.reserve(CS.arg_size());
ArgRegs.reserve(CS.arg_size());
for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
i != e; ++i) {
unsigned Arg = getRegForValue(*i);
-
+
if (Arg == 0)
return false;
ISD::ArgFlagsTy Flags;
return false;
const Type *ArgTy = (*i)->getType();
- EVT ArgVT;
+ MVT ArgVT;
if (!isTypeLegal(ArgTy, ArgVT))
return false;
unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
Flags.setOrigAlign(OriginalAlignment);
-
+
Args.push_back(*i);
ArgRegs.push_back(Arg);
ArgVTs.push_back(ArgVT);
ArgFlags.push_back(Flags);
}
-
+
// Handle the arguments now that we've gotten them.
SmallVector<unsigned, 4> RegArgs;
unsigned NumBytes;
if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
return false;
-
+
// Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
- // TODO: Turn this into the table of arm call ops.
+ // TODO: Turn this into the table of arm call ops.
MachineInstrBuilder MIB;
- unsigned CallOpc;
+ unsigned CallOpc = ARMSelectCallOp(GV);
+ // Explicitly adding the predicate here.
if(isThumb)
- CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
+ // Explicitly adding the predicate here.
+ MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(CallOpc)))
+ .addGlobalAddress(GV, 0, 0);
else
- CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
- .addGlobalAddress(GV, 0, 0);
-
+ // Explicitly adding the predicate here.
+ MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(CallOpc))
+ .addGlobalAddress(GV, 0, 0));
+
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
MIB.addReg(RegArgs[i]);
-
+
// Finish off the call including any return values.
- SmallVector<unsigned, 4> UsedRegs;
+ SmallVector<unsigned, 4> UsedRegs;
if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
-
+
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
-
+
return true;
-
+
}
// TODO: SoftFP support.
bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
- // No Thumb-1 for now.
- if (isThumb && !AFI->isThumb2Function()) return false;
switch (I->getOpcode()) {
case Instruction::Load:
return SelectBinaryOp(I, ISD::FMUL);
case Instruction::SDiv:
return SelectSDiv(I);
+ case Instruction::SRem:
+ return SelectSRem(I);
case Instruction::Call:
return SelectCall(I);
+ case Instruction::Select:
+ return SelectSelect(I);
+ case Instruction::Ret:
+ return SelectRet(I);
default: break;
}
return false;
namespace llvm {
llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
- if (EnableARMFastISel) return new ARMFastISel(funcInfo);
+ // Completely untested on non-darwin.
+ const TargetMachine &TM = funcInfo.MF->getTarget();
+
+ // Darwin and thumb1 only for now.
+ const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
+ if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
+ !DisableARMFastISel)
+ return new ARMFastISel(funcInfo);
return 0;
}
}