#include "ARM.h"
#include "ARMAddressingModes.h"
-#include "ARMISelLowering.h"
#include "ARMTargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetOptions.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.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.
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 2, 3 or 4. The opcode arrays specify the instructions used for
+ /// 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.
+ /// 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 2, 3 or 4. The opcode arrays specify the instructions used for
+ /// 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.
+ /// For NumVecs <= 2, QOpcodes1 is not used.
SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
unsigned *QOpcodes0, unsigned *QOpcodes1);
unsigned *QOpcodes1);
/// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
- SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, unsigned Opc);
+ SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
/// SelectCMOVOp - Select CMOV instructions for ARM.
SDNode *SelectCMOVOp(SDNode *N);
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
+ SDNode *SelectConcatVector(SDNode *N);
+
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
/// inline asm expressions.
virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
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);
+ 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,
SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
unsigned *DOpcodes, unsigned *QOpcodes0,
unsigned *QOpcodes1) {
- assert(NumVecs >=2 && NumVecs <= 4 && "VLD NumVecs out-of-range");
+ assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
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 Pred = getAL(CurDAG);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
if (is64BitVector) {
unsigned Opc = DOpcodes[OpcodeIndex];
}
EVT RegVT = GetNEONSubregVT(VT);
- if (NumVecs == 2) {
- // Quad registers are directly supported for VLD2,
- // loading 2 pairs of D regs.
+ 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(4, VT);
+ std::vector<EVT> ResTys(2 * NumVecs, RegVT);
ResTys.push_back(MVT::Other);
SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
- Chain = SDValue(VLd, 4);
+ Chain = SDValue(VLd, 2 * NumVecs);
// Combine the even and odd subregs to produce the result.
for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
unsigned *DOpcodes, unsigned *QOpcodes0,
unsigned *QOpcodes1) {
- assert(NumVecs >=2 && NumVecs <= 4 && "VST NumVecs out-of-range");
+ assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
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 Pred = getAL(CurDAG);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
SmallVector<SDValue, 10> Ops;
}
EVT RegVT = GetNEONSubregVT(VT);
- if (NumVecs == 2) {
- // Quad registers are directly supported for VST2,
- // storing 2 pairs of D regs.
+ 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,
Ops.push_back(Pred);
Ops.push_back(Reg0); // predicate register
Ops.push_back(Chain);
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 9);
+ return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
+ 5 + 2 * NumVecs);
}
// Otherwise, quad registers are stored with two separate instructions,
case MVT::v4i32: OpcodeIndex = 1; break;
}
- SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
+ SDValue Pred = getAL(CurDAG);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
SmallVector<SDValue, 10> Ops;
}
SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
- unsigned Opc) {
+ bool isSigned) {
if (!Subtarget->hasV6T2Ops())
return NULL;
+ unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
+ : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
+
+
+ // For unsigned extracts, check for a shift right and mask
+ unsigned And_imm = 0;
+ if (N->getOpcode() == ISD::AND) {
+ if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
+
+ // The immediate is a mask of the low bits iff imm & (imm+1) == 0
+ if (And_imm & (And_imm + 1))
+ return NULL;
+
+ unsigned Srl_imm = 0;
+ if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
+ Srl_imm)) {
+ assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
+
+ unsigned Width = CountTrailingOnes_32(And_imm);
+ unsigned LSB = Srl_imm;
+ 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;
+ }
+
+ // Otherwise, we're looking for a shift of a shift
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!");
return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
}
+SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
+ // The only time a CONCAT_VECTORS operation can have legal types is when
+ // two 64-bit vectors are concatenated to a 128-bit vector.
+ EVT VT = N->getValueType(0);
+ if (!VT.is128BitVector() || N->getNumOperands() != 2)
+ llvm_unreachable("unexpected CONCAT_VECTORS");
+ 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);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
+}
+
SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
DebugLoc dl = N->getDebugLoc();
SDNode *ResNode;
if (Subtarget->isThumb1Only()) {
- SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
+ SDValue Pred = getAL(CurDAG);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
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))
+ if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
return I;
break;
case ISD::SRA:
- if (SDNode *I = SelectV6T2BitfieldExtractOp(N,
- Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
+ if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
return I;
break;
case ISD::MUL:
}
break;
case ISD::AND: {
+ // Check for unsigned bitfield extract
+ if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
+ return I;
+
// (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
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 = 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);
+ }
+ // 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 = getAL(CurDAG);
+ 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 MVT::v4f32:
case MVT::v4i32: Opc = ARM::VZIPq32; break;
}
- SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
+ SDValue Pred = getAL(CurDAG);
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 MVT::v4f32:
case MVT::v4i32: Opc = ARM::VUZPq32; break;
}
- SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
+ SDValue Pred = getAL(CurDAG);
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 MVT::v4f32:
case MVT::v4i32: Opc = ARM::VTRNq32; break;
}
- SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
+ SDValue Pred = getAL(CurDAG);
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);
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::VLD2d64 };
+ ARM::VLD2d32, ARM::VLD1q64 };
unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
}
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::VST2d64 };
+ ARM::VST2d32, ARM::VST1q64 };
unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
}
}
}
}
+
+ case ISD::CONCAT_VECTORS: {
+ return SelectConcatVector(N);
+ }
}
return SelectCode(N);
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;
+}