//===-- AArch64ISelDAGToDAG.cpp - A dag to dag inst selector for AArch64 --===//
//
// 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 AArch64 target.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "aarch64-isel"
#include "AArch64.h"
#include "AArch64InstrInfo.h"
#include "AArch64Subtarget.h"
#include "AArch64TargetMachine.h"
#include "Utils/AArch64BaseInfo.h"
#include "llvm/ADT/APSInt.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
//===--------------------------------------------------------------------===//
/// AArch64 specific code to select AArch64 machine instructions for
/// SelectionDAG operations.
///
namespace {
class AArch64DAGToDAGISel : public SelectionDAGISel {
AArch64TargetMachine &TM;
/// Keep a pointer to the AArch64Subtarget around so that we can
/// make the right decision when generating code for different targets.
const AArch64Subtarget *Subtarget;
public:
explicit AArch64DAGToDAGISel(AArch64TargetMachine &tm,
CodeGenOpt::Level OptLevel)
: SelectionDAGISel(tm, OptLevel), TM(tm),
Subtarget(&TM.getSubtarget<AArch64Subtarget>()) {
}
virtual const char *getPassName() const {
return "AArch64 Instruction Selection";
}
// Include the pieces autogenerated from the target description.
#include "AArch64GenDAGISel.inc"
template<unsigned MemSize>
bool SelectOffsetUImm12(SDValue N, SDValue &UImm12) {
const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
if (!CN || CN->getZExtValue() % MemSize != 0
|| CN->getZExtValue() / MemSize > 0xfff)
return false;
UImm12 = CurDAG->getTargetConstant(CN->getZExtValue() / MemSize, MVT::i64);
return true;
}
template<unsigned RegWidth>
bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos) {
return SelectCVTFixedPosOperand(N, FixedPos, RegWidth);
}
/// Used for pre-lowered address-reference nodes, so we already know
/// the fields match. This operand's job is simply to add an
/// appropriate shift operand to the MOVZ/MOVK instruction.
template<unsigned LogShift>
bool SelectMOVWAddressRef(SDValue N, SDValue &Imm, SDValue &Shift) {
Imm = N;
Shift = CurDAG->getTargetConstant(LogShift, MVT::i32);
return true;
}
bool SelectFPZeroOperand(SDValue N, SDValue &Dummy);
bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos,
unsigned RegWidth);
bool SelectInlineAsmMemoryOperand(const SDValue &Op,
char ConstraintCode,
std::vector<SDValue> &OutOps);
bool SelectLogicalImm(SDValue N, SDValue &Imm);
template<unsigned RegWidth>
bool SelectTSTBOperand(SDValue N, SDValue &FixedPos) {
return SelectTSTBOperand(N, FixedPos, RegWidth);
}
bool SelectTSTBOperand(SDValue N, SDValue &FixedPos, unsigned RegWidth);
SDNode *SelectAtomic(SDNode *N, unsigned Op8, unsigned Op16, unsigned Op32,
unsigned Op64);
/// Put the given constant into a pool and return a DAG which will give its
/// address.
SDValue getConstantPoolItemAddress(SDLoc DL, const Constant *CV);
SDNode *TrySelectToMoveImm(SDNode *N);
SDNode *LowerToFPLitPool(SDNode *Node);
SDNode *SelectToLitPool(SDNode *N);
SDNode* Select(SDNode*);
private:
};
}
bool
AArch64DAGToDAGISel::SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos,
unsigned RegWidth) {
const ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N);
if (!CN) return false;
// An FCVT[SU] instruction performs: convertToInt(Val * 2^fbits) where fbits
// is between 1 and 32 for a destination w-register, or 1 and 64 for an
// x-register.
//
// By this stage, we've detected (fp_to_[su]int (fmul Val, THIS_NODE)) so we
// want THIS_NODE to be 2^fbits. This is much easier to deal with using
// integers.
bool IsExact;
// fbits is between 1 and 64 in the worst-case, which means the fmul
// could have 2^64 as an actual operand. Need 65 bits of precision.
APSInt IntVal(65, true);
CN->getValueAPF().convertToInteger(IntVal, APFloat::rmTowardZero, &IsExact);
// N.b. isPowerOf2 also checks for > 0.
if (!IsExact || !IntVal.isPowerOf2()) return false;
unsigned FBits = IntVal.logBase2();
// Checks above should have guaranteed that we haven't lost information in
// finding FBits, but it must still be in range.
if (FBits == 0 || FBits > RegWidth) return false;
FixedPos = CurDAG->getTargetConstant(64 - FBits, MVT::i32);
return true;
}
bool
AArch64DAGToDAGISel::SelectInlineAsmMemoryOperand(const SDValue &Op,
char ConstraintCode,
std::vector<SDValue> &OutOps) {
switch (ConstraintCode) {
default: llvm_unreachable("Unrecognised AArch64 memory constraint");
case 'm':
// FIXME: more freedom is actually permitted for 'm'. We can go
// hunting for a base and an offset if we want. Of course, since
// we don't really know how the operand is going to be used we're
// probably restricted to the load/store pair's simm7 as an offset
// range anyway.
case 'Q':
OutOps.push_back(Op);
}
return false;
}
bool
AArch64DAGToDAGISel::SelectFPZeroOperand(SDValue N, SDValue &Dummy) {
ConstantFPSDNode *Imm = dyn_cast<ConstantFPSDNode>(N);
if (!Imm || !Imm->getValueAPF().isPosZero())
return false;
// Doesn't actually carry any information, but keeps TableGen quiet.
Dummy = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
bool AArch64DAGToDAGISel::SelectLogicalImm(SDValue N, SDValue &Imm) {
uint32_t Bits;
uint32_t RegWidth = N.getValueType().getSizeInBits();
ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
if (!CN) return false;
if (!A64Imms::isLogicalImm(RegWidth, CN->getZExtValue(), Bits))
return false;
Imm = CurDAG->getTargetConstant(Bits, MVT::i32);
return true;
}
SDNode *AArch64DAGToDAGISel::TrySelectToMoveImm(SDNode *Node) {
SDNode *ResNode;
SDLoc dl(Node);
EVT DestType = Node->getValueType(0);
unsigned DestWidth = DestType.getSizeInBits();
unsigned MOVOpcode;
EVT MOVType;
int UImm16, Shift;
uint32_t LogicalBits;
uint64_t BitPat = cast<ConstantSDNode>(Node)->getZExtValue();
if (A64Imms::isMOVZImm(DestWidth, BitPat, UImm16, Shift)) {
MOVType = DestType;
MOVOpcode = DestWidth == 64 ? AArch64::MOVZxii : AArch64::MOVZwii;
} else if (A64Imms::isMOVNImm(DestWidth, BitPat, UImm16, Shift)) {
MOVType = DestType;
MOVOpcode = DestWidth == 64 ? AArch64::MOVNxii : AArch64::MOVNwii;
} else if (DestWidth == 64 && A64Imms::isMOVNImm(32, BitPat, UImm16, Shift)) {
// To get something like 0x0000_0000_ffff_1234 into a 64-bit register we can
// use a 32-bit instruction: "movn w0, 0xedbc".
MOVType = MVT::i32;
MOVOpcode = AArch64::MOVNwii;
} else if (A64Imms::isLogicalImm(DestWidth, BitPat, LogicalBits)) {
MOVOpcode = DestWidth == 64 ? AArch64::ORRxxi : AArch64::ORRwwi;
uint16_t ZR = DestWidth == 64 ? AArch64::XZR : AArch64::WZR;
return CurDAG->getMachineNode(MOVOpcode, dl, DestType,
CurDAG->getRegister(ZR, DestType),
CurDAG->getTargetConstant(LogicalBits, MVT::i32));
} else {
// Can't handle it in one instruction. There's scope for permitting two (or
// more) instructions, but that'll need more thought.
return NULL;
}
ResNode = CurDAG->getMachineNode(MOVOpcode, dl, MOVType,
CurDAG->getTargetConstant(UImm16, MVT::i32),
CurDAG->getTargetConstant(Shift, MVT::i32));
if (MOVType != DestType) {
ResNode = CurDAG->getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
MVT::i64, MVT::i32, MVT::Other,
CurDAG->getTargetConstant(0, MVT::i64),
SDValue(ResNode, 0),
CurDAG->getTargetConstant(AArch64::sub_32, MVT::i32));
}
return ResNode;
}
SDValue
AArch64DAGToDAGISel::getConstantPoolItemAddress(SDLoc DL,
const Constant *CV) {
EVT PtrVT = getTargetLowering()->getPointerTy();
switch (getTargetLowering()->getTargetMachine().getCodeModel()) {
case CodeModel::Small: {
unsigned Alignment =
getTargetLowering()->getDataLayout()->getABITypeAlignment(CV->getType());
return CurDAG->getNode(
AArch64ISD::WrapperSmall, DL, PtrVT,
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_NO_FLAG),
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_LO12),
CurDAG->getConstant(Alignment, MVT::i32));
}
case CodeModel::Large: {
SDNode *LitAddr;
LitAddr = CurDAG->getMachineNode(
AArch64::MOVZxii, DL, PtrVT,
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G3),
CurDAG->getTargetConstant(3, MVT::i32));
LitAddr = CurDAG->getMachineNode(
AArch64::MOVKxii, DL, PtrVT, SDValue(LitAddr, 0),
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G2_NC),
CurDAG->getTargetConstant(2, MVT::i32));
LitAddr = CurDAG->getMachineNode(
AArch64::MOVKxii, DL, PtrVT, SDValue(LitAddr, 0),
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G1_NC),
CurDAG->getTargetConstant(1, MVT::i32));
LitAddr = CurDAG->getMachineNode(
AArch64::MOVKxii, DL, PtrVT, SDValue(LitAddr, 0),
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G0_NC),
CurDAG->getTargetConstant(0, MVT::i32));
return SDValue(LitAddr, 0);
}
default:
llvm_unreachable("Only small and large code models supported now");
}
}
SDNode *AArch64DAGToDAGISel::SelectToLitPool(SDNode *Node) {
SDLoc DL(Node);
uint64_t UnsignedVal = cast<ConstantSDNode>(Node)->getZExtValue();
int64_t SignedVal = cast<ConstantSDNode>(Node)->getSExtValue();
EVT DestType = Node->getValueType(0);
// Since we may end up loading a 64-bit constant from a 32-bit entry the
// constant in the pool may have a different type to the eventual node.
ISD::LoadExtType Extension;
EVT MemType;
assert((DestType == MVT::i64 || DestType == MVT::i32)
&& "Only expect integer constants at the moment");
if (DestType == MVT::i32) {
Extension = ISD::NON_EXTLOAD;
MemType = MVT::i32;
} else if (UnsignedVal <= UINT32_MAX) {
Extension = ISD::ZEXTLOAD;
MemType = MVT::i32;
} else if (SignedVal >= INT32_MIN && SignedVal <= INT32_MAX) {
Extension = ISD::SEXTLOAD;
MemType = MVT::i32;
} else {
Extension = ISD::NON_EXTLOAD;
MemType = MVT::i64;
}
Constant *CV = ConstantInt::get(Type::getIntNTy(*CurDAG->getContext(),
MemType.getSizeInBits()),
UnsignedVal);
SDValue PoolAddr = getConstantPoolItemAddress(DL, CV);
unsigned Alignment =
getTargetLowering()->getDataLayout()->getABITypeAlignment(CV->getType());
return CurDAG->getExtLoad(Extension, DL, DestType, CurDAG->getEntryNode(),
PoolAddr,
MachinePointerInfo::getConstantPool(), MemType,
/* isVolatile = */ false,
/* isNonTemporal = */ false,
Alignment).getNode();
}
SDNode *AArch64DAGToDAGISel::LowerToFPLitPool(SDNode *Node) {
SDLoc DL(Node);
const ConstantFP *FV = cast<ConstantFPSDNode>(Node)->getConstantFPValue();
EVT DestType = Node->getValueType(0);
unsigned Alignment =
getTargetLowering()->getDataLayout()->getABITypeAlignment(FV->getType());
SDValue PoolAddr = getConstantPoolItemAddress(DL, FV);
return CurDAG->getLoad(DestType, DL, CurDAG->getEntryNode(), PoolAddr,
MachinePointerInfo::getConstantPool(),
/* isVolatile = */ false,
/* isNonTemporal = */ false,
/* isInvariant = */ true,
Alignment).getNode();
}
bool
AArch64DAGToDAGISel::SelectTSTBOperand(SDValue N, SDValue &FixedPos,
unsigned RegWidth) {
const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
if (!CN) return false;
uint64_t Val = CN->getZExtValue();
if (!isPowerOf2_64(Val)) return false;
unsigned TestedBit = Log2_64(Val);
// Checks above should have guaranteed that we haven't lost information in
// finding TestedBit, but it must still be in range.
if (TestedBit >= RegWidth) return false;
FixedPos = CurDAG->getTargetConstant(TestedBit, MVT::i64);
return true;
}
SDNode *AArch64DAGToDAGISel::SelectAtomic(SDNode *Node, unsigned Op8,
unsigned Op16,unsigned Op32,
unsigned Op64) {
// Mostly direct translation to the given operations, except that we preserve
// the AtomicOrdering for use later on.
AtomicSDNode *AN = cast<AtomicSDNode>(Node);
EVT VT = AN->getMemoryVT();
unsigned Op;
if (VT == MVT::i8)
Op = Op8;
else if (VT == MVT::i16)
Op = Op16;
else if (VT == MVT::i32)
Op = Op32;
else if (VT == MVT::i64)
Op = Op64;
else
llvm_unreachable("Unexpected atomic operation");
SmallVector<SDValue, 4> Ops;
for (unsigned i = 1; i < AN->getNumOperands(); ++i)
Ops.push_back(AN->getOperand(i));
Ops.push_back(CurDAG->getTargetConstant(AN->getOrdering(), MVT::i32));
Ops.push_back(AN->getOperand(0)); // Chain moves to the end
return CurDAG->SelectNodeTo(Node, Op,
AN->getValueType(0), MVT::Other,
&Ops[0], Ops.size());
}
SDNode *AArch64DAGToDAGISel::Select(SDNode *Node) {
// Dump information about the Node being selected
DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << "\n");
if (Node->isMachineOpcode()) {
DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << "\n");
return NULL;
}
switch (Node->getOpcode()) {
case ISD::ATOMIC_LOAD_ADD:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_ADD_I8,
AArch64::ATOMIC_LOAD_ADD_I16,
AArch64::ATOMIC_LOAD_ADD_I32,
AArch64::ATOMIC_LOAD_ADD_I64);
case ISD::ATOMIC_LOAD_SUB:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_SUB_I8,
AArch64::ATOMIC_LOAD_SUB_I16,
AArch64::ATOMIC_LOAD_SUB_I32,
AArch64::ATOMIC_LOAD_SUB_I64);
case ISD::ATOMIC_LOAD_AND:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_AND_I8,
AArch64::ATOMIC_LOAD_AND_I16,
AArch64::ATOMIC_LOAD_AND_I32,
AArch64::ATOMIC_LOAD_AND_I64);
case ISD::ATOMIC_LOAD_OR:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_OR_I8,
AArch64::ATOMIC_LOAD_OR_I16,
AArch64::ATOMIC_LOAD_OR_I32,
AArch64::ATOMIC_LOAD_OR_I64);
case ISD::ATOMIC_LOAD_XOR:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_XOR_I8,
AArch64::ATOMIC_LOAD_XOR_I16,
AArch64::ATOMIC_LOAD_XOR_I32,
AArch64::ATOMIC_LOAD_XOR_I64);
case ISD::ATOMIC_LOAD_NAND:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_NAND_I8,
AArch64::ATOMIC_LOAD_NAND_I16,
AArch64::ATOMIC_LOAD_NAND_I32,
AArch64::ATOMIC_LOAD_NAND_I64);
case ISD::ATOMIC_LOAD_MIN:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_MIN_I8,
AArch64::ATOMIC_LOAD_MIN_I16,
AArch64::ATOMIC_LOAD_MIN_I32,
AArch64::ATOMIC_LOAD_MIN_I64);
case ISD::ATOMIC_LOAD_MAX:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_MAX_I8,
AArch64::ATOMIC_LOAD_MAX_I16,
AArch64::ATOMIC_LOAD_MAX_I32,
AArch64::ATOMIC_LOAD_MAX_I64);
case ISD::ATOMIC_LOAD_UMIN:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_UMIN_I8,
AArch64::ATOMIC_LOAD_UMIN_I16,
AArch64::ATOMIC_LOAD_UMIN_I32,
AArch64::ATOMIC_LOAD_UMIN_I64);
case ISD::ATOMIC_LOAD_UMAX:
return SelectAtomic(Node,
AArch64::ATOMIC_LOAD_UMAX_I8,
AArch64::ATOMIC_LOAD_UMAX_I16,
AArch64::ATOMIC_LOAD_UMAX_I32,
AArch64::ATOMIC_LOAD_UMAX_I64);
case ISD::ATOMIC_SWAP:
return SelectAtomic(Node,
AArch64::ATOMIC_SWAP_I8,
AArch64::ATOMIC_SWAP_I16,
AArch64::ATOMIC_SWAP_I32,
AArch64::ATOMIC_SWAP_I64);
case ISD::ATOMIC_CMP_SWAP:
return SelectAtomic(Node,
AArch64::ATOMIC_CMP_SWAP_I8,
AArch64::ATOMIC_CMP_SWAP_I16,
AArch64::ATOMIC_CMP_SWAP_I32,
AArch64::ATOMIC_CMP_SWAP_I64);
case ISD::FrameIndex: {
int FI = cast<FrameIndexSDNode>(Node)->getIndex();
EVT PtrTy = getTargetLowering()->getPointerTy();
SDValue TFI = CurDAG->getTargetFrameIndex(FI, PtrTy);
return CurDAG->SelectNodeTo(Node, AArch64::ADDxxi_lsl0_s, PtrTy,
TFI, CurDAG->getTargetConstant(0, PtrTy));
}
case ISD::ConstantPool: {
// Constant pools are fine, just create a Target entry.
ConstantPoolSDNode *CN = cast<ConstantPoolSDNode>(Node);
const Constant *C = CN->getConstVal();
SDValue CP = CurDAG->getTargetConstantPool(C, CN->getValueType(0));
ReplaceUses(SDValue(Node, 0), CP);
return NULL;
}
case ISD::Constant: {
SDNode *ResNode = 0;
if (cast<ConstantSDNode>(Node)->getZExtValue() == 0) {
// XZR and WZR are probably even better than an actual move: most of the
// time they can be folded into another instruction with *no* cost.
EVT Ty = Node->getValueType(0);
assert((Ty == MVT::i32 || Ty == MVT::i64) && "unexpected type");
uint16_t Register = Ty == MVT::i32 ? AArch64::WZR : AArch64::XZR;
ResNode = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
SDLoc(Node),
Register, Ty).getNode();
}
// Next best option is a move-immediate, see if we can do that.
if (!ResNode) {
ResNode = TrySelectToMoveImm(Node);
}
if (ResNode)
return ResNode;
// If even that fails we fall back to a lit-pool entry at the moment. Future
// tuning may change this to a sequence of MOVZ/MOVN/MOVK instructions.
ResNode = SelectToLitPool(Node);
assert(ResNode && "We need *some* way to materialise a constant");
// We want to continue selection at this point since the litpool access
// generated used generic nodes for simplicity.
ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
Node = ResNode;
break;
}
case ISD::ConstantFP: {
if (A64Imms::isFPImm(cast<ConstantFPSDNode>(Node)->getValueAPF())) {
// FMOV will take care of it from TableGen
break;
}
SDNode *ResNode = LowerToFPLitPool(Node);
ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
// We want to continue selection at this point since the litpool access
// generated used generic nodes for simplicity.
Node = ResNode;
break;
}
default:
break; // Let generic code handle it
}
SDNode *ResNode = SelectCode(Node);
DEBUG(dbgs() << "=> ";
if (ResNode == NULL || ResNode == Node)
Node->dump(CurDAG);
else
ResNode->dump(CurDAG);
dbgs() << "\n");
return ResNode;
}
/// This pass converts a legalized DAG into a AArch64-specific DAG, ready for
/// instruction scheduling.
FunctionPass *llvm::createAArch64ISelDAG(AArch64TargetMachine &TM,
CodeGenOpt::Level OptLevel) {
return new AArch64DAGToDAGISel(TM, OptLevel);
}