//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "arm-isel"
#include "ARM.h"
#include "ARMAddressingModes.h"
#include "ARMTargetMachine.h"
using namespace llvm;
-static cl::opt<bool>
-UseRegSeq("neon-reg-sequence", cl::Hidden,
- cl::desc("Use reg_sequence to model ld / st of multiple neon regs"));
-
//===--------------------------------------------------------------------===//
/// ARMDAGToDAGISel - ARM specific code to select ARM machine
/// instructions for SelectionDAG operations.
char ConstraintCode,
std::vector<SDValue> &OutOps);
- /// PairDRegs - Form a quad register from a pair of D registers.
- ///
+ // Form pairs of consecutive S, D, or Q registers.
+ SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
-
- /// PairDRegs - Form a quad register pair from a pair of Q registers.
- ///
SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
- /// QuadDRegs - Form a quad register pair from a quad of D registers.
- ///
+ // Form sequences of 4 consecutive S, D, or Q registers.
+ SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
+ SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
- /// OctoDRegs - Form 8 consecutive D registers.
- ///
+ // Form sequences of 8 consecutive D registers.
SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
SDValue V4, SDValue V5, SDValue V6, SDValue V7);
};
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)
+ if (!NC || !NC->isNullValue())
return false;
Base = Offset = N;
if (N.getOpcode() == ISD::ADD) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
+ // 8 bits.
if (((RHSC & 0x3) == 0) &&
- ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
+ ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) {
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.
+ // 8 bits.
+ if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) {
Base = N.getOperand(0);
OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
return true;
return NULL;
}
+/// PairSRegs - Form a D register from a pair of S registers.
+///
+SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
+}
+
/// PairDRegs - Form a quad register from a pair of D registers.
///
SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
DebugLoc dl = V0.getNode()->getDebugLoc();
- SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
- SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
- if (llvm::ModelWithRegSequence()) {
- const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
- return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
- }
- SDValue Undef =
- SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
- 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);
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
}
/// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
///
SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
DebugLoc dl = V0.getNode()->getDebugLoc();
- SDValue SubReg0 = CurDAG->getTargetConstant(ARM::QSUBREG_0, MVT::i32);
- SDValue SubReg1 = CurDAG->getTargetConstant(ARM::QSUBREG_1, MVT::i32);
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
}
+/// QuadSRegs - Form 4 consecutive S registers.
+///
+SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
+ SDValue V2, SDValue V3) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
+}
+
/// QuadDRegs - Form 4 consecutive D registers.
///
SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
SDValue V2, SDValue V3) {
DebugLoc dl = V0.getNode()->getDebugLoc();
- SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
- SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
- SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
- SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
+}
+
+/// QuadQRegs - Form 4 consecutive Q registers.
+///
+SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
+ SDValue V2, SDValue V3) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
}
SDValue V4, SDValue V5,
SDValue V6, SDValue V7) {
DebugLoc dl = V0.getNode()->getDebugLoc();
- SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
- SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
- SDValue SubReg2 = CurDAG->getTargetConstant(ARM::DSUBREG_2, MVT::i32);
- SDValue SubReg3 = CurDAG->getTargetConstant(ARM::DSUBREG_3, MVT::i32);
- SDValue SubReg4 = CurDAG->getTargetConstant(ARM::DSUBREG_4, MVT::i32);
- SDValue SubReg5 = CurDAG->getTargetConstant(ARM::DSUBREG_5, MVT::i32);
- SDValue SubReg6 = CurDAG->getTargetConstant(ARM::DSUBREG_6, MVT::i32);
- SDValue SubReg7 = CurDAG->getTargetConstant(ARM::DSUBREG_7, MVT::i32);
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
+ SDValue SubReg4 = CurDAG->getTargetConstant(ARM::dsub_4, MVT::i32);
+ SDValue SubReg5 = CurDAG->getTargetConstant(ARM::dsub_5, MVT::i32);
+ SDValue SubReg6 = CurDAG->getTargetConstant(ARM::dsub_6, MVT::i32);
+ SDValue SubReg7 = CurDAG->getTargetConstant(ARM::dsub_7, MVT::i32);
const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
std::vector<EVT> ResTys(NumVecs, VT);
ResTys.push_back(MVT::Other);
SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
- if (!llvm::ModelWithRegSequence() || NumVecs < 2)
+ if (NumVecs < 2)
return VLd;
SDValue RegSeq;
RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
}
+ assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
- SDValue D = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec,
+ SDValue D = CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec,
dl, VT, RegSeq);
ReplaceUses(SDValue(N, Vec), D);
}
Chain = SDValue(VLd, 2 * NumVecs);
// Combine the even and odd subregs to produce the result.
- if (llvm::ModelWithRegSequence()) {
- if (NumVecs == 1) {
- SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
- ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
- } else {
- SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
- SDValue(VLd, 0), SDValue(VLd, 1),
- SDValue(VLd, 2), SDValue(VLd, 3)), 0);
- SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0, dl, VT, QQ);
- SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_1, dl, VT, QQ);
- ReplaceUses(SDValue(N, 0), Q0);
- ReplaceUses(SDValue(N, 1), Q1);
- }
+ if (NumVecs == 1) {
+ SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
+ ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
} else {
- 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));
- }
+ SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
+ SDValue(VLd, 0), SDValue(VLd, 1),
+ SDValue(VLd, 2), SDValue(VLd, 3)), 0);
+ SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::qsub_0, dl, VT, QQ);
+ SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::qsub_1, dl, VT, QQ);
+ ReplaceUses(SDValue(N, 0), Q0);
+ ReplaceUses(SDValue(N, 1), Q1);
}
} else {
// Otherwise, quad registers are loaded with two separate instructions,
SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
Chain = SDValue(VLdB, NumVecs+1);
- if (llvm::ModelWithRegSequence()) {
- SDValue V0 = SDValue(VLdA, 0);
- SDValue V1 = SDValue(VLdB, 0);
- SDValue V2 = SDValue(VLdA, 1);
- SDValue V3 = SDValue(VLdB, 1);
- SDValue V4 = SDValue(VLdA, 2);
- SDValue V5 = SDValue(VLdB, 2);
- SDValue V6 = (NumVecs == 3)
- ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
- 0)
- : SDValue(VLdA, 3);
- SDValue V7 = (NumVecs == 3)
- ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT),
- 0)
- : SDValue(VLdB, 3);
- SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
- V4, V5, V6, V7), 0);
-
- // Extract out the 3 / 4 Q registers.
- for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
- SDValue Q = CurDAG->getTargetExtractSubreg(ARM::QSUBREG_0+Vec,
- dl, VT, RegSeq);
- ReplaceUses(SDValue(N, Vec), Q);
- }
- } else {
- // 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));
- }
+ SDValue V0 = SDValue(VLdA, 0);
+ SDValue V1 = SDValue(VLdB, 0);
+ SDValue V2 = SDValue(VLdA, 1);
+ SDValue V3 = SDValue(VLdB, 1);
+ SDValue V4 = SDValue(VLdA, 2);
+ SDValue V5 = SDValue(VLdB, 2);
+ SDValue V6 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
+ : SDValue(VLdA, 3);
+ SDValue V7 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
+ : SDValue(VLdB, 3);
+ SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
+ V4, V5, V6, V7), 0);
+
+ // Extract out the 3 / 4 Q registers.
+ assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
+ SDValue Q = CurDAG->getTargetExtractSubreg(ARM::qsub_0+Vec,
+ dl, VT, RegSeq);
+ ReplaceUses(SDValue(N, Vec), Q);
}
}
ReplaceUses(SDValue(N, NumVecs), Chain);
Ops.push_back(Align);
if (is64BitVector) {
- if (llvm::ModelWithRegSequence() && NumVecs >= 2) {
+ if (NumVecs >= 2) {
SDValue RegSeq;
SDValue V0 = N->getOperand(0+3);
SDValue V1 = N->getOperand(1+3);
}
// Now extract the D registers back out.
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, VT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT,
RegSeq));
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, VT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT,
RegSeq));
if (NumVecs > 2)
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, VT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,
RegSeq));
if (NumVecs > 3)
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, VT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,
RegSeq));
} else {
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
// Quad registers are directly supported for VST1 and VST2,
// storing pairs of D regs.
unsigned Opc = QOpcodes0[OpcodeIndex];
- if (llvm::ModelWithRegSequence() && NumVecs == 2) {
+ if (NumVecs == 2) {
// First extract the pair of Q registers.
SDValue Q0 = N->getOperand(3);
SDValue Q1 = N->getOperand(4);
SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
// Now extract the D registers back out.
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
QQ));
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
QQ));
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_2, dl, RegVT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, RegVT,
QQ));
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_3, dl, RegVT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, RegVT,
QQ));
Ops.push_back(Pred);
Ops.push_back(Reg0); // predicate register
return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
} else {
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
N->getOperand(Vec+3)));
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
N->getOperand(Vec+3)));
}
Ops.push_back(Pred);
// Otherwise, quad registers are stored with two separate instructions,
// where one stores the even registers and the other stores the odd registers.
- if (llvm::ModelWithRegSequence()) {
- // Form the QQQQ REG_SEQUENCE.
- SDValue V[8];
- for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
- V[i] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(Vec+3));
- V[i+1] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(Vec+3));
- }
- if (NumVecs == 3)
- V[6] = V[7] =
- SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0);
- SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
- V[4], V[5], V[6], V[7]), 0);
-
- // Store the even D registers.
- Ops.push_back(Reg0); // post-access address offset
- for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
- Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec*2, dl,
- RegVT, RegSeq));
- 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 D registers.
- Ops[0] = SDValue(VStA, 0); // MemAddr
- for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
- Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1+Vec*2, dl,
- RegVT, RegSeq);
- 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;
- } else {
- 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;
- }
+ // Form the QQQQ REG_SEQUENCE.
+ SDValue V[8];
+ for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
+ V[i] = CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
+ N->getOperand(Vec+3));
+ V[i+1] = CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
+ N->getOperand(Vec+3));
+ }
+ if (NumVecs == 3)
+ V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
+
+ SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
+ V[4], V[5], V[6], V[7]), 0);
+
+ // Store the even D registers.
+ assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
+ Ops.push_back(Reg0); // post-access address offset
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec*2, dl,
+ RegVT, RegSeq));
+ 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 D registers.
+ Ops[0] = SDValue(VStA, 0); // MemAddr
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::dsub_1+Vec*2, dl,
+ RegVT, RegSeq);
+ 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,
// Quad registers are handled by load/store of subregs. Find the subreg info.
unsigned NumElts = 0;
- int SubregIdx = 0;
+ bool Even = false;
EVT RegVT = VT;
if (!is64BitVector) {
RegVT = GetNEONSubregVT(VT);
NumElts = RegVT.getVectorNumElements();
- SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
+ Even = Lane < NumElts;
}
unsigned OpcodeIndex;
unsigned Opc = 0;
if (is64BitVector) {
Opc = DOpcodes[OpcodeIndex];
- for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
- Ops.push_back(N->getOperand(Vec+3));
+ SDValue RegSeq;
+ SDValue V0 = N->getOperand(0+3);
+ SDValue V1 = N->getOperand(1+3);
+ if (NumVecs == 2) {
+ RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
+ } else {
+ SDValue V2 = N->getOperand(2+3);
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : N->getOperand(3+3);
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
+ }
+
+ // Now extract the D registers back out.
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
+ if (NumVecs > 2)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,RegSeq));
+ if (NumVecs > 3)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,RegSeq));
} else {
// Check if this is loading the even or odd subreg of a Q register.
if (Lane < NumElts) {
Lane -= NumElts;
Opc = QOpcodes1[OpcodeIndex];
}
+
+ SDValue RegSeq;
+ SDValue V0 = N->getOperand(0+3);
+ SDValue V1 = N->getOperand(1+3);
+ if (NumVecs == 2) {
+ RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
+ } else {
+ SDValue V2 = N->getOperand(2+3);
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : N->getOperand(3+3);
+ RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
+ }
+
// Extract the subregs of the input vector.
+ unsigned SubIdx = Even ? ARM::dsub_0 : ARM::dsub_1;
for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
- Ops.push_back(CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
- N->getOperand(Vec+3)));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
+ RegSeq));
}
Ops.push_back(getI32Imm(Lane));
Ops.push_back(Pred);
ResTys.push_back(MVT::Other);
SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
- if (llvm::ModelWithRegSequence() && is64BitVector) {
- SDValue RegSeq;
+ // Form a REG_SEQUENCE to force register allocation.
+ SDValue RegSeq;
+ if (is64BitVector) {
SDValue V0 = SDValue(VLdLn, 0);
SDValue V1 = SDValue(VLdLn, 1);
-
- // Form a REG_SEQUENCE to force register allocation.
if (NumVecs == 2) {
RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
} else {
SDValue V2 = SDValue(VLdLn, 2);
// If it's a vld3, form a quad D-register but discard the last part.
SDValue V3 = (NumVecs == 3)
- ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
- : SDValue(VLdLn, 3);
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : SDValue(VLdLn, 3);
RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
}
-
- for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
- SDValue D = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0+Vec, dl, VT,
- RegSeq);
- ReplaceUses(SDValue(N, Vec), D);
+ } else {
+ // For 128-bit vectors, take the 64-bit results of the load and insert
+ // them as subregs into the result.
+ SDValue V[8];
+ for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
+ if (Even) {
+ V[i] = SDValue(VLdLn, Vec);
+ V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
+ } else {
+ V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
+ V[i+1] = SDValue(VLdLn, Vec);
+ }
}
- ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
- return NULL;
- }
-
- // For a 64-bit vector load to D registers, nothing more needs to be done.
- if (is64BitVector)
- return VLdLn;
+ if (NumVecs == 3)
+ V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
- // 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);
+ if (NumVecs == 2)
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
+ else
+ RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
+ V[4], V[5], V[6], V[7]), 0);
}
- Chain = SDValue(VLdLn, NumVecs);
- ReplaceUses(SDValue(N, NumVecs), Chain);
+ assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
+ assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
+ unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ ReplaceUses(SDValue(N, Vec),
+ CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
+ ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
return NULL;
}
DebugLoc dl = N->getDebugLoc();
SDValue V0 = N->getOperand(0);
SDValue V1 = N->getOperand(1);
- SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
- SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
}
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);
+ SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
+ return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
} else {
SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
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);
+ 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),
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);
+ 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),
SDValue Pred = getAL(CurDAG);
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);
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
+ SDNode *Ret = CurDAG->getMachineNode(ARM::VLDMQ, dl,
+ MVT::v2f64, MVT::Other, Ops, 5);
+ cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
+ return Ret;
}
// Other cases are autogenerated.
break;
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);
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
+ SDNode *Ret = CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
+ cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
+ return Ret;
}
// Other cases are autogenerated.
break;
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
}
+ case ARMISD::BUILD_VECTOR: {
+ EVT VecVT = N->getValueType(0);
+ EVT EltVT = VecVT.getVectorElementType();
+ unsigned NumElts = VecVT.getVectorNumElements();
+ if (EltVT.getSimpleVT() == MVT::f64) {
+ assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
+ return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
+ }
+ assert(EltVT.getSimpleVT() == MVT::f32 &&
+ "unexpected type for BUILD_VECTOR");
+ if (NumElts == 2)
+ return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
+ assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
+ return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
+ N->getOperand(2), N->getOperand(3));
+ }
case ISD::INTRINSIC_VOID:
case ISD::INTRINSIC_W_CHAIN: {
CodeGenOpt::Level OptLevel) {
return new ARMDAGToDAGISel(TM, OptLevel);
}
-
-/// ModelWithRegSequence - Return true if isel should use REG_SEQUENCE to model
-/// operations involving sub-registers.
-bool llvm::ModelWithRegSequence() {
- return UseRegSeq;
-}
projects / oota-llvm.git / blobdiff