llvm-6502/lib/Target/ARM/ARMISelDAGToDAG.cpp
Bob Wilson df9a4f0591 Fix VLDMQ and VSTMQ instructions to use the correct encoding and address modes.
These instructions are only needed for codegen, so I've removed all the
explicit encoding bits for now; they should be set in the same way as the for
VLDMD and VSTMD whenever we add encodings for VFP.  The use of addrmode5
requires that the instructions be custom-selected so that the number of
registers can be set in the AM5Opc value.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@99309 91177308-0d34-0410-b5e6-96231b3b80d8
2010-03-23 18:54:46 +00:00

2014 lines
75 KiB
C++

//===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines an instruction selector for the ARM target.
//
//===----------------------------------------------------------------------===//
#include "ARM.h"
#include "ARMAddressingModes.h"
#include "ARMISelLowering.h"
#include "ARMTargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Intrinsics.h"
#include "llvm/LLVMContext.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
//===--------------------------------------------------------------------===//
/// ARMDAGToDAGISel - ARM specific code to select ARM machine
/// instructions for SelectionDAG operations.
///
namespace {
class ARMDAGToDAGISel : public SelectionDAGISel {
ARMBaseTargetMachine &TM;
/// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
/// make the right decision when generating code for different targets.
const ARMSubtarget *Subtarget;
public:
explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
CodeGenOpt::Level OptLevel)
: SelectionDAGISel(tm, OptLevel), TM(tm),
Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
}
virtual const char *getPassName() const {
return "ARM Instruction Selection";
}
/// getI32Imm - Return a target constant of type i32 with the specified
/// value.
inline SDValue getI32Imm(unsigned Imm) {
return CurDAG->getTargetConstant(Imm, MVT::i32);
}
SDNode *Select(SDNode *N);
bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
SDValue &B, SDValue &C);
bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset, SDValue &Opc);
bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc);
bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset, SDValue &Opc);
bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc);
bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
SDValue &Mode);
bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset);
bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
SDValue &Label);
bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset);
bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
SDValue &Base, SDValue &OffImm,
SDValue &Offset);
bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm, SDValue &Offset);
bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm, SDValue &Offset);
bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm, SDValue &Offset);
bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
SDValue &BaseReg, SDValue &Opc);
bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
SDValue &OffImm);
bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffReg, SDValue &ShImm);
// Include the pieces autogenerated from the target description.
#include "ARMGenDAGISel.inc"
private:
/// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
/// ARM.
SDNode *SelectARMIndexedLoad(SDNode *N);
SDNode *SelectT2IndexedLoad(SDNode *N);
/// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
SDNode *SelectDYN_ALLOC(SDNode *N);
/// SelectVLD - Select NEON load intrinsics. NumVecs should be
/// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
/// loads of D registers and even subregs and odd subregs of Q registers.
/// For NumVecs <= 2, QOpcodes1 is not used.
SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
unsigned *QOpcodes0, unsigned *QOpcodes1);
/// SelectVST - Select NEON store intrinsics. NumVecs should
/// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
/// stores of D registers and even subregs and odd subregs of Q registers.
/// For NumVecs <= 2, QOpcodes1 is not used.
SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
unsigned *QOpcodes0, unsigned *QOpcodes1);
/// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
/// be 2, 3 or 4. The opcode arrays specify the instructions used for
/// load/store of D registers and even subregs and odd subregs of Q registers.
SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
unsigned *DOpcodes, unsigned *QOpcodes0,
unsigned *QOpcodes1);
/// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, unsigned Opc);
/// SelectCMOVOp - Select CMOV instructions for ARM.
SDNode *SelectCMOVOp(SDNode *N);
SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
/// inline asm expressions.
virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
char ConstraintCode,
std::vector<SDValue> &OutOps);
/// PairDRegs - Insert a pair of double registers into an implicit def to
/// form a quad register.
SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
};
}
/// isInt32Immediate - This method tests to see if the node is a 32-bit constant
/// operand. If so Imm will receive the 32-bit value.
static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
Imm = cast<ConstantSDNode>(N)->getZExtValue();
return true;
}
return false;
}
// isInt32Immediate - This method tests to see if a constant operand.
// If so Imm will receive the 32 bit value.
static bool isInt32Immediate(SDValue N, unsigned &Imm) {
return isInt32Immediate(N.getNode(), Imm);
}
// isOpcWithIntImmediate - This method tests to see if the node is a specific
// opcode and that it has a immediate integer right operand.
// If so Imm will receive the 32 bit value.
static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
return N->getOpcode() == Opc &&
isInt32Immediate(N->getOperand(1).getNode(), Imm);
}
bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
SDValue N,
SDValue &BaseReg,
SDValue &ShReg,
SDValue &Opc) {
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
// Don't match base register only case. That is matched to a separate
// lower complexity pattern with explicit register operand.
if (ShOpcVal == ARM_AM::no_shift) return false;
BaseReg = N.getOperand(0);
unsigned ShImmVal = 0;
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
ShReg = CurDAG->getRegister(0, MVT::i32);
ShImmVal = RHS->getZExtValue() & 31;
} else {
ShReg = N.getOperand(1);
}
Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset,
SDValue &Opc) {
if (N.getOpcode() == ISD::MUL) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
// X * [3,5,9] -> X + X * [2,4,8] etc.
int RHSC = (int)RHS->getZExtValue();
if (RHSC & 1) {
RHSC = RHSC & ~1;
ARM_AM::AddrOpc AddSub = ARM_AM::add;
if (RHSC < 0) {
AddSub = ARM_AM::sub;
RHSC = - RHSC;
}
if (isPowerOf2_32(RHSC)) {
unsigned ShAmt = Log2_32(RHSC);
Base = Offset = N.getOperand(0);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
ARM_AM::lsl),
MVT::i32);
return true;
}
}
}
}
if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
Base = N;
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
} else if (N.getOpcode() == ARMISD::Wrapper &&
!(Subtarget->useMovt() &&
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
}
Offset = CurDAG->getRegister(0, MVT::i32);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
ARM_AM::no_shift),
MVT::i32);
return true;
}
// Match simple R +/- imm12 operands.
if (N.getOpcode() == ISD::ADD)
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if ((RHSC >= 0 && RHSC < 0x1000) ||
(RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
Offset = CurDAG->getRegister(0, MVT::i32);
ARM_AM::AddrOpc AddSub = ARM_AM::add;
if (RHSC < 0) {
AddSub = ARM_AM::sub;
RHSC = - RHSC;
}
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
ARM_AM::no_shift),
MVT::i32);
return true;
}
}
// Otherwise this is R +/- [possibly shifted] R.
ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
unsigned ShAmt = 0;
Base = N.getOperand(0);
Offset = N.getOperand(1);
if (ShOpcVal != ARM_AM::no_shift) {
// Check to see if the RHS of the shift is a constant, if not, we can't fold
// it.
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(1).getOperand(0);
} else {
ShOpcVal = ARM_AM::no_shift;
}
}
// Try matching (R shl C) + (R).
if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
if (ShOpcVal != ARM_AM::no_shift) {
// Check to see if the RHS of the shift is a constant, if not, we can't
// fold it.
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(0).getOperand(0);
Base = N.getOperand(1);
} else {
ShOpcVal = ARM_AM::no_shift;
}
}
}
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc) {
unsigned Opcode = Op->getOpcode();
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
? cast<LoadSDNode>(Op)->getAddressingMode()
: cast<StoreSDNode>(Op)->getAddressingMode();
ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
? ARM_AM::add : ARM_AM::sub;
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
int Val = (int)C->getZExtValue();
if (Val >= 0 && Val < 0x1000) { // 12 bits.
Offset = CurDAG->getRegister(0, MVT::i32);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
ARM_AM::no_shift),
MVT::i32);
return true;
}
}
Offset = N;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
unsigned ShAmt = 0;
if (ShOpcVal != ARM_AM::no_shift) {
// Check to see if the RHS of the shift is a constant, if not, we can't fold
// it.
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(0);
} else {
ShOpcVal = ARM_AM::no_shift;
}
}
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset,
SDValue &Opc) {
if (N.getOpcode() == ISD::SUB) {
// X - C is canonicalize to X + -C, no need to handle it here.
Base = N.getOperand(0);
Offset = N.getOperand(1);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
return true;
}
if (N.getOpcode() != ISD::ADD) {
Base = N;
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
Offset = CurDAG->getRegister(0, MVT::i32);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
return true;
}
// If the RHS is +/- imm8, fold into addr mode.
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if ((RHSC >= 0 && RHSC < 256) ||
(RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
Offset = CurDAG->getRegister(0, MVT::i32);
ARM_AM::AddrOpc AddSub = ARM_AM::add;
if (RHSC < 0) {
AddSub = ARM_AM::sub;
RHSC = - RHSC;
}
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
return true;
}
}
Base = N.getOperand(0);
Offset = N.getOperand(1);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc) {
unsigned Opcode = Op->getOpcode();
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
? cast<LoadSDNode>(Op)->getAddressingMode()
: cast<StoreSDNode>(Op)->getAddressingMode();
ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
? ARM_AM::add : ARM_AM::sub;
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
int Val = (int)C->getZExtValue();
if (Val >= 0 && Val < 256) {
Offset = CurDAG->getRegister(0, MVT::i32);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
return true;
}
}
Offset = N;
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
SDValue &Addr, SDValue &Mode) {
Addr = N;
Mode = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset) {
if (N.getOpcode() != ISD::ADD) {
Base = N;
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
} else if (N.getOpcode() == ARMISD::Wrapper &&
!(Subtarget->useMovt() &&
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
}
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
MVT::i32);
return true;
}
// If the RHS is +/- imm8, fold into addr mode.
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if ((RHSC & 3) == 0) { // The constant is implicitly multiplied by 4.
RHSC >>= 2;
if ((RHSC >= 0 && RHSC < 256) ||
(RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
ARM_AM::AddrOpc AddSub = ARM_AM::add;
if (RHSC < 0) {
AddSub = ARM_AM::sub;
RHSC = - RHSC;
}
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
MVT::i32);
return true;
}
}
}
Base = N;
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
SDValue &Addr, SDValue &Align) {
Addr = N;
// Default to no alignment.
Align = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Label) {
if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
Offset = N.getOperand(0);
SDValue N1 = N.getOperand(1);
Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
MVT::i32);
return true;
}
return false;
}
bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset){
// FIXME dl should come from the parent load or store, not the address
DebugLoc dl = Op->getDebugLoc();
if (N.getOpcode() != ISD::ADD) {
ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
if (!NC || NC->getZExtValue() != 0)
return false;
Base = Offset = N;
return true;
}
Base = N.getOperand(0);
Offset = N.getOperand(1);
return true;
}
bool
ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
unsigned Scale, SDValue &Base,
SDValue &OffImm, SDValue &Offset) {
if (Scale == 4) {
SDValue TmpBase, TmpOffImm;
if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
return false; // We want to select tLDRspi / tSTRspi instead.
if (N.getOpcode() == ARMISD::Wrapper &&
N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
return false; // We want to select tLDRpci instead.
}
if (N.getOpcode() != ISD::ADD) {
if (N.getOpcode() == ARMISD::Wrapper &&
!(Subtarget->useMovt() &&
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
} else
Base = N;
Offset = CurDAG->getRegister(0, MVT::i32);
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
// Thumb does not have [sp, r] address mode.
RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
if ((LHSR && LHSR->getReg() == ARM::SP) ||
(RHSR && RHSR->getReg() == ARM::SP)) {
Base = N;
Offset = CurDAG->getRegister(0, MVT::i32);
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
// If the RHS is + imm5 * scale, fold into addr mode.
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if ((RHSC & (Scale-1)) == 0) { // The constant is implicitly multiplied.
RHSC /= Scale;
if (RHSC >= 0 && RHSC < 32) {
Base = N.getOperand(0);
Offset = CurDAG->getRegister(0, MVT::i32);
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
return true;
}
}
}
Base = N.getOperand(0);
Offset = N.getOperand(1);
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm,
SDValue &Offset) {
return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
}
bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm,
SDValue &Offset) {
return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
}
bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm,
SDValue &Offset) {
return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
}
bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
if (N.getOpcode() != ISD::ADD)
return false;
RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
(LHSR && LHSR->getReg() == ARM::SP)) {
// If the RHS is + imm8 * scale, fold into addr mode.
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if ((RHSC & 3) == 0) { // The constant is implicitly multiplied.
RHSC >>= 2;
if (RHSC >= 0 && RHSC < 256) {
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
return true;
}
}
}
}
return false;
}
bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
SDValue &BaseReg,
SDValue &Opc) {
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
// Don't match base register only case. That is matched to a separate
// lower complexity pattern with explicit register operand.
if (ShOpcVal == ARM_AM::no_shift) return false;
BaseReg = N.getOperand(0);
unsigned ShImmVal = 0;
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
ShImmVal = RHS->getZExtValue() & 31;
Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
return true;
}
return false;
}
bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
// Match simple R + imm12 operands.
// Base only.
if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
if (N.getOpcode() == ISD::FrameIndex) {
// Match frame index...
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
} else if (N.getOpcode() == ARMISD::Wrapper &&
!(Subtarget->useMovt() &&
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::TargetConstantPool)
return false; // We want to select t2LDRpci instead.
} else
Base = N;
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
// Let t2LDRi8 handle (R - imm8).
return false;
int RHSC = (int)RHS->getZExtValue();
if (N.getOpcode() == ISD::SUB)
RHSC = -RHSC;
if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
return true;
}
}
// Base only.
Base = N;
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
// Match simple R - imm8 operands.
if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getSExtValue();
if (N.getOpcode() == ISD::SUB)
RHSC = -RHSC;
if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(Base)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
return true;
}
}
}
return false;
}
bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
SDValue &OffImm){
unsigned Opcode = Op->getOpcode();
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
? cast<LoadSDNode>(Op)->getAddressingMode()
: cast<StoreSDNode>(Op)->getAddressingMode();
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
int RHSC = (int)RHS->getZExtValue();
if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
? CurDAG->getTargetConstant(RHSC, MVT::i32)
: CurDAG->getTargetConstant(-RHSC, MVT::i32);
return true;
}
}
return false;
}
bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
if (N.getOpcode() == ISD::ADD) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if (((RHSC & 0x3) == 0) &&
((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
Base = N.getOperand(0);
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
return true;
}
}
} else if (N.getOpcode() == ISD::SUB) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
Base = N.getOperand(0);
OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
return true;
}
}
}
return false;
}
bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
SDValue &Base,
SDValue &OffReg, SDValue &ShImm) {
// (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
if (N.getOpcode() != ISD::ADD)
return false;
// Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
return false;
else if (RHSC < 0 && RHSC >= -255) // 8 bits
return false;
}
// Look for (R + R) or (R + (R << [1,2,3])).
unsigned ShAmt = 0;
Base = N.getOperand(0);
OffReg = N.getOperand(1);
// Swap if it is ((R << c) + R).
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
if (ShOpcVal != ARM_AM::lsl) {
ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
if (ShOpcVal == ARM_AM::lsl)
std::swap(Base, OffReg);
}
if (ShOpcVal == ARM_AM::lsl) {
// Check to see if the RHS of the shift is a constant, if not, we can't fold
// it.
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
ShAmt = Sh->getZExtValue();
if (ShAmt >= 4) {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
} else
OffReg = OffReg.getOperand(0);
} else {
ShOpcVal = ARM_AM::no_shift;
}
}
ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
return true;
}
//===--------------------------------------------------------------------===//
/// getAL - Returns a ARMCC::AL immediate node.
static inline SDValue getAL(SelectionDAG *CurDAG) {
return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
}
SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
LoadSDNode *LD = cast<LoadSDNode>(N);
ISD::MemIndexedMode AM = LD->getAddressingMode();
if (AM == ISD::UNINDEXED)
return NULL;
EVT LoadedVT = LD->getMemoryVT();
SDValue Offset, AMOpc;
bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
unsigned Opcode = 0;
bool Match = false;
if (LoadedVT == MVT::i32 &&
SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
Match = true;
} else if (LoadedVT == MVT::i16 &&
SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
Match = true;
Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
: (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
} else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
if (LD->getExtensionType() == ISD::SEXTLOAD) {
if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
Match = true;
Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
}
} else {
if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
Match = true;
Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
}
}
}
if (Match) {
SDValue Chain = LD->getChain();
SDValue Base = LD->getBasePtr();
SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32), Chain };
return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
MVT::Other, Ops, 6);
}
return NULL;
}
SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
LoadSDNode *LD = cast<LoadSDNode>(N);
ISD::MemIndexedMode AM = LD->getAddressingMode();
if (AM == ISD::UNINDEXED)
return NULL;
EVT LoadedVT = LD->getMemoryVT();
bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
SDValue Offset;
bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
unsigned Opcode = 0;
bool Match = false;
if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
switch (LoadedVT.getSimpleVT().SimpleTy) {
case MVT::i32:
Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
break;
case MVT::i16:
if (isSExtLd)
Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
else
Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
break;
case MVT::i8:
case MVT::i1:
if (isSExtLd)
Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
else
Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
break;
default:
return NULL;
}
Match = true;
}
if (Match) {
SDValue Chain = LD->getChain();
SDValue Base = LD->getBasePtr();
SDValue Ops[]= { Base, Offset, getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32), Chain };
return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
MVT::Other, Ops, 5);
}
return NULL;
}
SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDNode *N) {
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
SDValue Chain = N->getOperand(0);
SDValue Size = N->getOperand(1);
SDValue Align = N->getOperand(2);
SDValue SP = CurDAG->getRegister(ARM::SP, MVT::i32);
int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
if (AlignVal < 0)
// We need to align the stack. Use Thumb1 tAND which is the only thumb
// instruction that can read and write SP. This matches to a pseudo
// instruction that has a chain to ensure the result is written back to
// the stack pointer.
SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);
bool isC = isa<ConstantSDNode>(Size);
uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
// Handle the most common case for both Thumb1 and Thumb2:
// tSUBspi - immediate is between 0 ... 508 inclusive.
if (C <= 508 && ((C & 3) == 0))
// FIXME: tSUBspi encode scale 4 implicitly.
return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
CurDAG->getTargetConstant(C/4, MVT::i32),
Chain);
if (Subtarget->isThumb1Only()) {
// Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
// should have negated the size operand already. FIXME: We can't insert
// new target independent node at this stage so we are forced to negate
// it earlier. Is there a better solution?
return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
Chain);
} else if (Subtarget->isThumb2()) {
if (isC && Predicate_t2_so_imm(Size.getNode())) {
// t2SUBrSPi
SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
} else if (isC && Predicate_imm0_4095(Size.getNode())) {
// t2SUBrSPi12
SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
} else {
// t2SUBrSPs
SDValue Ops[] = { SP, Size,
getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
}
}
// FIXME: Add ADD / SUB sp instructions for ARM.
return 0;
}
/// PairDRegs - Insert a pair of double registers into an implicit def to
/// form a quad register.
SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
DebugLoc dl = V0.getNode()->getDebugLoc();
SDValue Undef =
SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
VT, Undef, V0, SubReg0);
return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
VT, SDValue(Pair, 0), V1, SubReg1);
}
/// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
/// for a 64-bit subregister of the vector.
static EVT GetNEONSubregVT(EVT VT) {
switch (VT.getSimpleVT().SimpleTy) {
default: llvm_unreachable("unhandled NEON type");
case MVT::v16i8: return MVT::v8i8;
case MVT::v8i16: return MVT::v4i16;
case MVT::v4f32: return MVT::v2f32;
case MVT::v4i32: return MVT::v2i32;
case MVT::v2i64: return MVT::v1i64;
}
}
SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
unsigned *DOpcodes, unsigned *QOpcodes0,
unsigned *QOpcodes1) {
assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
return NULL;
SDValue Chain = N->getOperand(0);
EVT VT = N->getValueType(0);
bool is64BitVector = VT.is64BitVector();
unsigned OpcodeIndex;
switch (VT.getSimpleVT().SimpleTy) {
default: llvm_unreachable("unhandled vld type");
// Double-register operations:
case MVT::v8i8: OpcodeIndex = 0; break;
case MVT::v4i16: OpcodeIndex = 1; break;
case MVT::v2f32:
case MVT::v2i32: OpcodeIndex = 2; break;
case MVT::v1i64: OpcodeIndex = 3; break;
// Quad-register operations:
case MVT::v16i8: OpcodeIndex = 0; break;
case MVT::v8i16: OpcodeIndex = 1; break;
case MVT::v4f32:
case MVT::v4i32: OpcodeIndex = 2; break;
case MVT::v2i64: OpcodeIndex = 3;
assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
break;
}
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
if (is64BitVector) {
unsigned Opc = DOpcodes[OpcodeIndex];
const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
std::vector<EVT> ResTys(NumVecs, VT);
ResTys.push_back(MVT::Other);
return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
}
EVT RegVT = GetNEONSubregVT(VT);
if (NumVecs <= 2) {
// Quad registers are directly supported for VLD1 and VLD2,
// loading pairs of D regs.
unsigned Opc = QOpcodes0[OpcodeIndex];
const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
std::vector<EVT> ResTys(2 * NumVecs, RegVT);
ResTys.push_back(MVT::Other);
SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
Chain = SDValue(VLd, 2 * NumVecs);
// Combine the even and odd subregs to produce the result.
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
SDNode *Q = PairDRegs(VT, SDValue(VLd, 2*Vec), SDValue(VLd, 2*Vec+1));
ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
}
} else {
// Otherwise, quad registers are loaded with two separate instructions,
// where one loads the even registers and the other loads the odd registers.
std::vector<EVT> ResTys(NumVecs, RegVT);
ResTys.push_back(MemAddr.getValueType());
ResTys.push_back(MVT::Other);
// Load the even subregs.
unsigned Opc = QOpcodes0[OpcodeIndex];
const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
Chain = SDValue(VLdA, NumVecs+1);
// Load the odd subregs.
Opc = QOpcodes1[OpcodeIndex];
const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
Align, Reg0, Pred, Reg0, Chain };
SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
Chain = SDValue(VLdB, NumVecs+1);
// Combine the even and odd subregs to produce the result.
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
SDNode *Q = PairDRegs(VT, SDValue(VLdA, Vec), SDValue(VLdB, Vec));
ReplaceUses(SDValue(N, Vec), SDValue(Q, 0));
}
}
ReplaceUses(SDValue(N, NumVecs), Chain);
return NULL;
}
SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
unsigned *DOpcodes, unsigned *QOpcodes0,
unsigned *QOpcodes1) {
assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
return NULL;
SDValue Chain = N->getOperand(0);
EVT VT = N->getOperand(3).getValueType();
bool is64BitVector = VT.is64BitVector();
unsigned OpcodeIndex;
switch (VT.getSimpleVT().SimpleTy) {
default: llvm_unreachable("unhandled vst type");
// Double-register operations:
case MVT::v8i8: OpcodeIndex = 0; break;
case MVT::v4i16: OpcodeIndex = 1; break;
case MVT::v2f32:
case MVT::v2i32: OpcodeIndex = 2; break;
case MVT::v1i64: OpcodeIndex = 3; break;
// Quad-register operations:
case MVT::v16i8: OpcodeIndex = 0; break;
case MVT::v8i16: OpcodeIndex = 1; break;
case MVT::v4f32:
case MVT::v4i32: OpcodeIndex = 2; break;
case MVT::v2i64: OpcodeIndex = 3;
assert(NumVecs == 1 && "v2i64 type only supported for VST1");
break;
}
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
SmallVector<SDValue, 10> Ops;
Ops.push_back(MemAddr);
Ops.push_back(Align);
if (is64BitVector) {
unsigned Opc = DOpcodes[OpcodeIndex];
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
Ops.push_back(N->getOperand(Vec+3));
Ops.push_back(Pred);
Ops.push_back(Reg0); // predicate register
Ops.push_back(Chain);
return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
}
EVT RegVT = GetNEONSubregVT(VT);
if (NumVecs <= 2) {
// Quad registers are directly supported for VST1 and VST2,
// storing pairs of D regs.
unsigned Opc = QOpcodes0[OpcodeIndex];
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
N->getOperand(Vec+3)));
Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
N->getOperand(Vec+3)));
}
Ops.push_back(Pred);
Ops.push_back(Reg0); // predicate register
Ops.push_back(Chain);
return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
5 + 2 * NumVecs);
}
// Otherwise, quad registers are stored with two separate instructions,
// where one stores the even registers and the other stores the odd registers.
Ops.push_back(Reg0); // post-access address offset
// Store the even subregs.
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
N->getOperand(Vec+3)));
Ops.push_back(Pred);
Ops.push_back(Reg0); // predicate register
Ops.push_back(Chain);
unsigned Opc = QOpcodes0[OpcodeIndex];
SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
MVT::Other, Ops.data(), NumVecs+6);
Chain = SDValue(VStA, 1);
// Store the odd subregs.
Ops[0] = SDValue(VStA, 0); // MemAddr
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
N->getOperand(Vec+3));
Ops[NumVecs+5] = Chain;
Opc = QOpcodes1[OpcodeIndex];
SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
MVT::Other, Ops.data(), NumVecs+6);
Chain = SDValue(VStB, 1);
ReplaceUses(SDValue(N, 0), Chain);
return NULL;
}
SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
unsigned NumVecs, unsigned *DOpcodes,
unsigned *QOpcodes0,
unsigned *QOpcodes1) {
assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
return NULL;
SDValue Chain = N->getOperand(0);
unsigned Lane =
cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
bool is64BitVector = VT.is64BitVector();
// Quad registers are handled by load/store of subregs. Find the subreg info.
unsigned NumElts = 0;
int SubregIdx = 0;
EVT RegVT = VT;
if (!is64BitVector) {
RegVT = GetNEONSubregVT(VT);
NumElts = RegVT.getVectorNumElements();
SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
}
unsigned OpcodeIndex;
switch (VT.getSimpleVT().SimpleTy) {
default: llvm_unreachable("unhandled vld/vst lane type");
// Double-register operations:
case MVT::v8i8: OpcodeIndex = 0; break;
case MVT::v4i16: OpcodeIndex = 1; break;
case MVT::v2f32:
case MVT::v2i32: OpcodeIndex = 2; break;
// Quad-register operations:
case MVT::v8i16: OpcodeIndex = 0; break;
case MVT::v4f32:
case MVT::v4i32: OpcodeIndex = 1; break;
}
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
SmallVector<SDValue, 10> Ops;
Ops.push_back(MemAddr);
Ops.push_back(Align);
unsigned Opc = 0;
if (is64BitVector) {
Opc = DOpcodes[OpcodeIndex];
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
Ops.push_back(N->getOperand(Vec+3));
} else {
// Check if this is loading the even or odd subreg of a Q register.
if (Lane < NumElts) {
Opc = QOpcodes0[OpcodeIndex];
} else {
Lane -= NumElts;
Opc = QOpcodes1[OpcodeIndex];
}
// Extract the subregs of the input vector.
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
N->getOperand(Vec+3)));
}
Ops.push_back(getI32Imm(Lane));
Ops.push_back(Pred);
Ops.push_back(Reg0);
Ops.push_back(Chain);
if (!IsLoad)
return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
std::vector<EVT> ResTys(NumVecs, RegVT);
ResTys.push_back(MVT::Other);
SDNode *VLdLn =
CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+6);
// For a 64-bit vector load to D registers, nothing more needs to be done.
if (is64BitVector)
return VLdLn;
// For 128-bit vectors, take the 64-bit results of the load and insert them
// as subregs into the result.
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
SDValue QuadVec = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
N->getOperand(Vec+3),
SDValue(VLdLn, Vec));
ReplaceUses(SDValue(N, Vec), QuadVec);
}
Chain = SDValue(VLdLn, NumVecs);
ReplaceUses(SDValue(N, NumVecs), Chain);
return NULL;
}
SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
unsigned Opc) {
if (!Subtarget->hasV6T2Ops())
return NULL;
unsigned Shl_imm = 0;
if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
unsigned Srl_imm = 0;
if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
unsigned Width = 32 - Srl_imm;
int LSB = Srl_imm - Shl_imm;
if (LSB < 0)
return NULL;
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { N->getOperand(0).getOperand(0),
CurDAG->getTargetConstant(LSB, MVT::i32),
CurDAG->getTargetConstant(Width, MVT::i32),
getAL(CurDAG), Reg0 };
return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
}
}
return NULL;
}
SDNode *ARMDAGToDAGISel::
SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
SDValue CPTmp0;
SDValue CPTmp1;
if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
unsigned Opc = 0;
switch (SOShOp) {
case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
default:
llvm_unreachable("Unknown so_reg opcode!");
break;
}
SDValue SOShImm =
CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
}
return 0;
}
SDNode *ARMDAGToDAGISel::
SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
SDValue CPTmp0;
SDValue CPTmp1;
SDValue CPTmp2;
if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
}
return 0;
}
SDNode *ARMDAGToDAGISel::
SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
if (!T)
return 0;
if (Predicate_t2_so_imm(TrueVal.getNode())) {
SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
return CurDAG->SelectNodeTo(N,
ARM::t2MOVCCi, MVT::i32, Ops, 5);
}
return 0;
}
SDNode *ARMDAGToDAGISel::
SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
if (!T)
return 0;
if (Predicate_so_imm(TrueVal.getNode())) {
SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
return CurDAG->SelectNodeTo(N,
ARM::MOVCCi, MVT::i32, Ops, 5);
}
return 0;
}
SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
EVT VT = N->getValueType(0);
SDValue FalseVal = N->getOperand(0);
SDValue TrueVal = N->getOperand(1);
SDValue CC = N->getOperand(2);
SDValue CCR = N->getOperand(3);
SDValue InFlag = N->getOperand(4);
assert(CC.getOpcode() == ISD::Constant);
assert(CCR.getOpcode() == ISD::Register);
ARMCC::CondCodes CCVal =
(ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
// Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
// Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
// Pattern complexity = 18 cost = 1 size = 0
SDValue CPTmp0;
SDValue CPTmp1;
SDValue CPTmp2;
if (Subtarget->isThumb()) {
SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
CCVal, CCR, InFlag);
if (!Res)
Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
if (Res)
return Res;
} else {
SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
CCVal, CCR, InFlag);
if (!Res)
Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
if (Res)
return Res;
}
// Pattern: (ARMcmov:i32 GPR:i32:$false,
// (imm:i32)<<P:Predicate_so_imm>>:$true,
// (imm:i32):$cc)
// Emits: (MOVCCi:i32 GPR:i32:$false,
// (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
// Pattern complexity = 10 cost = 1 size = 0
if (Subtarget->isThumb()) {
SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
CCVal, CCR, InFlag);
if (!Res)
Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
if (Res)
return Res;
} else {
SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
CCVal, CCR, InFlag);
if (!Res)
Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
if (Res)
return Res;
}
}
// Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
// Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
// Pattern complexity = 6 cost = 1 size = 0
//
// Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
// Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
// Pattern complexity = 6 cost = 11 size = 0
//
// Also FCPYScc and FCPYDcc.
SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
unsigned Opc = 0;
switch (VT.getSimpleVT().SimpleTy) {
default: assert(false && "Illegal conditional move type!");
break;
case MVT::i32:
Opc = Subtarget->isThumb()
? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
: ARM::MOVCCr;
break;
case MVT::f32:
Opc = ARM::VMOVScc;
break;
case MVT::f64:
Opc = ARM::VMOVDcc;
break;
}
return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
}
SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
DebugLoc dl = N->getDebugLoc();
if (N->isMachineOpcode())
return NULL; // Already selected.
switch (N->getOpcode()) {
default: break;
case ISD::Constant: {
unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
bool UseCP = true;
if (Subtarget->hasThumb2())
// Thumb2-aware targets have the MOVT instruction, so all immediates can
// be done with MOV + MOVT, at worst.
UseCP = 0;
else {
if (Subtarget->isThumb()) {
UseCP = (Val > 255 && // MOV
~Val > 255 && // MOV + MVN
!ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL
} else
UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
ARM_AM::getSOImmVal(~Val) == -1 && // MVN
!ARM_AM::isSOImmTwoPartVal(Val)); // two instrs.
}
if (UseCP) {
SDValue CPIdx =
CurDAG->getTargetConstantPool(ConstantInt::get(
Type::getInt32Ty(*CurDAG->getContext()), Val),
TLI.getPointerTy());
SDNode *ResNode;
if (Subtarget->isThumb1Only()) {
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
Ops, 4);
} else {
SDValue Ops[] = {
CPIdx,
CurDAG->getRegister(0, MVT::i32),
CurDAG->getTargetConstant(0, MVT::i32),
getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32),
CurDAG->getEntryNode()
};
ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
Ops, 6);
}
ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
return NULL;
}
// Other cases are autogenerated.
break;
}
case ISD::FrameIndex: {
// Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
int FI = cast<FrameIndexSDNode>(N)->getIndex();
SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
if (Subtarget->isThumb1Only()) {
return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
CurDAG->getTargetConstant(0, MVT::i32));
} else {
unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
ARM::t2ADDri : ARM::ADDri);
SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
CurDAG->getRegister(0, MVT::i32) };
return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
}
}
case ARMISD::DYN_ALLOC:
return SelectDYN_ALLOC(N);
case ISD::SRL:
if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))
return I;
break;
case ISD::SRA:
if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
return I;
break;
case ISD::MUL:
if (Subtarget->isThumb1Only())
break;
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
unsigned RHSV = C->getZExtValue();
if (!RHSV) break;
if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
unsigned ShImm = Log2_32(RHSV-1);
if (ShImm >= 32)
break;
SDValue V = N->getOperand(0);
ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
if (Subtarget->isThumb()) {
SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
} else {
SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
}
}
if (isPowerOf2_32(RHSV+1)) { // 2^n-1?
unsigned ShImm = Log2_32(RHSV+1);
if (ShImm >= 32)
break;
SDValue V = N->getOperand(0);
ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
if (Subtarget->isThumb()) {
SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
} else {
SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
}
}
}
break;
case ISD::AND: {
// (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
// of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
// are entirely contributed by c2 and lower 16-bits are entirely contributed
// by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
// Select it to: "movt x, ((c1 & 0xffff) >> 16)
EVT VT = N->getValueType(0);
if (VT != MVT::i32)
break;
unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
? ARM::t2MOVTi16
: (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
if (!Opc)
break;
SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
if (!N1C)
break;
if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
SDValue N2 = N0.getOperand(1);
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
if (!N2C)
break;
unsigned N1CVal = N1C->getZExtValue();
unsigned N2CVal = N2C->getZExtValue();
if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
(N1CVal & 0xffffU) == 0xffffU &&
(N2CVal & 0xffffU) == 0x0U) {
SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
MVT::i32);
SDValue Ops[] = { N0.getOperand(0), Imm16,
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
}
}
break;
}
case ARMISD::VMOVRRD:
return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
N->getOperand(0), getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32));
case ISD::UMUL_LOHI: {
if (Subtarget->isThumb1Only())
break;
if (Subtarget->isThumb()) {
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
CurDAG->getRegister(0, MVT::i32) };
return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
} else {
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
CurDAG->getRegister(0, MVT::i32) };
return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
}
}
case ISD::SMUL_LOHI: {
if (Subtarget->isThumb1Only())
break;
if (Subtarget->isThumb()) {
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
} else {
SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
CurDAG->getRegister(0, MVT::i32) };
return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
}
}
case ISD::LOAD: {
SDNode *ResNode = 0;
if (Subtarget->isThumb() && Subtarget->hasThumb2())
ResNode = SelectT2IndexedLoad(N);
else
ResNode = SelectARMIndexedLoad(N);
if (ResNode)
return ResNode;
// VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
if (Subtarget->hasVFP2() &&
N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
SDValue Chain = N->getOperand(0);
SDValue AM5Opc =
CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
return CurDAG->getMachineNode(ARM::VLDMQ, dl, MVT::v2f64, MVT::Other,
Ops, 5);
}
// Other cases are autogenerated.
break;
}
case ISD::STORE: {
// VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
if (Subtarget->hasVFP2() &&
N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
SDValue Chain = N->getOperand(0);
SDValue AM5Opc =
CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
AM5Opc, Pred, PredReg, Chain };
return CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
}
// Other cases are autogenerated.
break;
}
case ARMISD::BRCOND: {
// Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
// Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
// Pattern complexity = 6 cost = 1 size = 0
// Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
// Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
// Pattern complexity = 6 cost = 1 size = 0
// Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
// Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
// Pattern complexity = 6 cost = 1 size = 0
unsigned Opc = Subtarget->isThumb() ?
((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
SDValue Chain = N->getOperand(0);
SDValue N1 = N->getOperand(1);
SDValue N2 = N->getOperand(2);
SDValue N3 = N->getOperand(3);
SDValue InFlag = N->getOperand(4);
assert(N1.getOpcode() == ISD::BasicBlock);
assert(N2.getOpcode() == ISD::Constant);
assert(N3.getOpcode() == ISD::Register);
SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
cast<ConstantSDNode>(N2)->getZExtValue()),
MVT::i32);
SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
MVT::Flag, Ops, 5);
Chain = SDValue(ResNode, 0);
if (N->getNumValues() == 2) {
InFlag = SDValue(ResNode, 1);
ReplaceUses(SDValue(N, 1), InFlag);
}
ReplaceUses(SDValue(N, 0),
SDValue(Chain.getNode(), Chain.getResNo()));
return NULL;
}
case ARMISD::CMOV:
return SelectCMOVOp(N);
case ARMISD::CNEG: {
EVT VT = N->getValueType(0);
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
SDValue N2 = N->getOperand(2);
SDValue N3 = N->getOperand(3);
SDValue InFlag = N->getOperand(4);
assert(N2.getOpcode() == ISD::Constant);
assert(N3.getOpcode() == ISD::Register);
SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
cast<ConstantSDNode>(N2)->getZExtValue()),
MVT::i32);
SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
unsigned Opc = 0;
switch (VT.getSimpleVT().SimpleTy) {
default: assert(false && "Illegal conditional move type!");
break;
case MVT::f32:
Opc = ARM::VNEGScc;
break;
case MVT::f64:
Opc = ARM::VNEGDcc;
break;
}
return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
}
case ARMISD::VZIP: {
unsigned Opc = 0;
EVT VT = N->getValueType(0);
switch (VT.getSimpleVT().SimpleTy) {
default: return NULL;
case MVT::v8i8: Opc = ARM::VZIPd8; break;
case MVT::v4i16: Opc = ARM::VZIPd16; break;
case MVT::v2f32:
case MVT::v2i32: Opc = ARM::VZIPd32; break;
case MVT::v16i8: Opc = ARM::VZIPq8; break;
case MVT::v8i16: Opc = ARM::VZIPq16; break;
case MVT::v4f32:
case MVT::v4i32: Opc = ARM::VZIPq32; break;
}
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
}
case ARMISD::VUZP: {
unsigned Opc = 0;
EVT VT = N->getValueType(0);
switch (VT.getSimpleVT().SimpleTy) {
default: return NULL;
case MVT::v8i8: Opc = ARM::VUZPd8; break;
case MVT::v4i16: Opc = ARM::VUZPd16; break;
case MVT::v2f32:
case MVT::v2i32: Opc = ARM::VUZPd32; break;
case MVT::v16i8: Opc = ARM::VUZPq8; break;
case MVT::v8i16: Opc = ARM::VUZPq16; break;
case MVT::v4f32:
case MVT::v4i32: Opc = ARM::VUZPq32; break;
}
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
}
case ARMISD::VTRN: {
unsigned Opc = 0;
EVT VT = N->getValueType(0);
switch (VT.getSimpleVT().SimpleTy) {
default: return NULL;
case MVT::v8i8: Opc = ARM::VTRNd8; break;
case MVT::v4i16: Opc = ARM::VTRNd16; break;
case MVT::v2f32:
case MVT::v2i32: Opc = ARM::VTRNd32; break;
case MVT::v16i8: Opc = ARM::VTRNq8; break;
case MVT::v8i16: Opc = ARM::VTRNq16; break;
case MVT::v4f32:
case MVT::v4i32: Opc = ARM::VTRNq32; break;
}
SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
}
case ISD::INTRINSIC_VOID:
case ISD::INTRINSIC_W_CHAIN: {
unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
switch (IntNo) {
default:
break;
case Intrinsic::arm_neon_vld1: {
unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
ARM::VLD1d32, ARM::VLD1d64 };
unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
ARM::VLD1q32, ARM::VLD1q64 };
return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
}
case Intrinsic::arm_neon_vld2: {
unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
ARM::VLD2d32, ARM::VLD1q64 };
unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
}
case Intrinsic::arm_neon_vld3: {
unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
ARM::VLD3d32, ARM::VLD1d64T };
unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
ARM::VLD3q16_UPD,
ARM::VLD3q32_UPD };
unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
ARM::VLD3q16odd_UPD,
ARM::VLD3q32odd_UPD };
return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld4: {
unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
ARM::VLD4d32, ARM::VLD1d64Q };
unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
ARM::VLD4q16_UPD,
ARM::VLD4q32_UPD };
unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
ARM::VLD4q16odd_UPD,
ARM::VLD4q32odd_UPD };
return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld2lane: {
unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld3lane: {
unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld4lane: {
unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst1: {
unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
ARM::VST1d32, ARM::VST1d64 };
unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
ARM::VST1q32, ARM::VST1q64 };
return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
}
case Intrinsic::arm_neon_vst2: {
unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
ARM::VST2d32, ARM::VST1q64 };
unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
}
case Intrinsic::arm_neon_vst3: {
unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
ARM::VST3d32, ARM::VST1d64T };
unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
ARM::VST3q16_UPD,
ARM::VST3q32_UPD };
unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
ARM::VST3q16odd_UPD,
ARM::VST3q32odd_UPD };
return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst4: {
unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
ARM::VST4d32, ARM::VST1d64Q };
unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
ARM::VST4q16_UPD,
ARM::VST4q32_UPD };
unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
ARM::VST4q16odd_UPD,
ARM::VST4q32odd_UPD };
return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst2lane: {
unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst3lane: {
unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst4lane: {
unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
}
}
}
}
return SelectCode(N);
}
bool ARMDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
std::vector<SDValue> &OutOps) {
assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
// Require the address to be in a register. That is safe for all ARM
// variants and it is hard to do anything much smarter without knowing
// how the operand is used.
OutOps.push_back(Op);
return false;
}
/// createARMISelDag - This pass converts a legalized DAG into a
/// ARM-specific DAG, ready for instruction scheduling.
///
FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
CodeGenOpt::Level OptLevel) {
return new ARMDAGToDAGISel(TM, OptLevel);
}