//===----------------------------------------------------------------------===//
#include "AArch64ISelLowering.h"
+#include "AArch64CallingConvention.h"
#include "AArch64MachineFunctionInfo.h"
#include "AArch64PerfectShuffle.h"
#include "AArch64Subtarget.h"
cl::desc("Allow AArch64 SLI/SRI formation"),
cl::init(false));
-//===----------------------------------------------------------------------===//
-// AArch64 Lowering public interface.
-//===----------------------------------------------------------------------===//
-static TargetLoweringObjectFile *createTLOF(const Triple &TT) {
- if (TT.isOSBinFormatMachO())
- return new AArch64_MachoTargetObjectFile();
-
- return new AArch64_ELFTargetObjectFile();
-}
-AArch64TargetLowering::AArch64TargetLowering(TargetMachine &TM)
- : TargetLowering(TM, createTLOF(Triple(TM.getTargetTriple()))) {
+AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM)
+ : TargetLowering(TM) {
Subtarget = &TM.getSubtarget<AArch64Subtarget>();
// AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
addDRTypeForNEON(MVT::v2i32);
addDRTypeForNEON(MVT::v1i64);
addDRTypeForNEON(MVT::v1f64);
+ addDRTypeForNEON(MVT::v4f16);
addQRTypeForNEON(MVT::v4f32);
addQRTypeForNEON(MVT::v2f64);
addQRTypeForNEON(MVT::v8i16);
addQRTypeForNEON(MVT::v4i32);
addQRTypeForNEON(MVT::v2i64);
+ addQRTypeForNEON(MVT::v8f16);
}
// Compute derived properties from the register classes
setOperationAction(ISD::FMUL, MVT::f16, Promote);
setOperationAction(ISD::FSUB, MVT::f16, Promote);
+ // v4f16 is also a storage-only type, so promote it to v4f32 when that is
+ // known to be safe.
+ setOperationAction(ISD::FADD, MVT::v4f16, Promote);
+ setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
+ setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
+ setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
+ setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
+ setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
+ AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
+ AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
+ AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
+ AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
+ AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
+ AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
+
+ // Expand all other v4f16 operations.
+ // FIXME: We could generate better code by promoting some operations to
+ // a pair of v4f32s
+ setOperationAction(ISD::FABS, MVT::v4f16, Expand);
+ setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
+ setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
+ setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
+ setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
+ setOperationAction(ISD::FMA, MVT::v4f16, Expand);
+ setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
+ setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
+ setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
+ setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
+ setOperationAction(ISD::FREM, MVT::v4f16, Expand);
+ setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
+ setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
+ setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
+ setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
+ setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
+ setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
+ setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
+ setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
+ setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
+ setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
+ setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
+ setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
+ setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
+ setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
+
+
+ // v8f16 is also a storage-only type, so expand it.
+ setOperationAction(ISD::FABS, MVT::v8f16, Expand);
+ setOperationAction(ISD::FADD, MVT::v8f16, Expand);
+ setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
+ setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
+ setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
+ setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
+ setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
+ setOperationAction(ISD::FMA, MVT::v8f16, Expand);
+ setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
+ setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
+ setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
+ setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
+ setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
+ setOperationAction(ISD::FREM, MVT::v8f16, Expand);
+ setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
+ setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
+ setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
+ setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
+ setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
+ setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
+ setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
+ setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
+ setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
+ setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
+ setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
+ setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
+ setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
+ setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
+ setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
+ setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
+
// AArch64 has implementations of a lot of rounding-like FP operations.
static MVT RoundingTypes[] = { MVT::f32, MVT::f64};
for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
}
+ // Make floating-point constants legal for the large code model, so they don't
+ // become loads from the constant pool.
+ if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
+ setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
+ setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
+ }
+
// AArch64 does not have floating-point extending loads, i1 sign-extending
// load, floating-point truncating stores, or v2i32->v2i16 truncating store.
setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
// AArch64 doesn't have MUL.2d:
setOperationAction(ISD::MUL, MVT::v2i64, Expand);
+ // Custom handling for some quad-vector types to detect MULL.
+ setOperationAction(ISD::MUL, MVT::v8i16, Custom);
+ setOperationAction(ISD::MUL, MVT::v4i32, Custom);
+ setOperationAction(ISD::MUL, MVT::v2i64, Custom);
+
setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
// Likewise, narrowing and extending vector loads/stores aren't handled
}
void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
- if (VT == MVT::v2f32) {
+ if (VT == MVT::v2f32 || VT == MVT::v4f16) {
setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
- } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
+ } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
+ case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
+ case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
+ case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
}
}
static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
+ SDValue Cmp;
+ AArch64CC::CondCode AArch64CC;
if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
EVT VT = RHS.getValueType();
uint64_t C = RHSC->getZExtValue();
break;
case ISD::SETLE:
case ISD::SETGT:
- if ((VT == MVT::i32 && C != 0x7fffffff &&
+ if ((VT == MVT::i32 && C != INT32_MAX &&
isLegalArithImmed((uint32_t)(C + 1))) ||
- (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
+ (VT == MVT::i64 && C != INT64_MAX &&
isLegalArithImmed(C + 1ULL))) {
CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
break;
case ISD::SETULE:
case ISD::SETUGT:
- if ((VT == MVT::i32 && C != 0xffffffff &&
+ if ((VT == MVT::i32 && C != UINT32_MAX &&
isLegalArithImmed((uint32_t)(C + 1))) ||
- (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
+ (VT == MVT::i64 && C != UINT64_MAX &&
isLegalArithImmed(C + 1ULL))) {
CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
}
}
}
-
- SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
- AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
+ // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
+ // For the i8 operand, the largest immediate is 255, so this can be easily
+ // encoded in the compare instruction. For the i16 operand, however, the
+ // largest immediate cannot be encoded in the compare.
+ // Therefore, use a sign extending load and cmn to avoid materializing the -1
+ // constant. For example,
+ // movz w1, #65535
+ // ldrh w0, [x0, #0]
+ // cmp w0, w1
+ // >
+ // ldrsh w0, [x0, #0]
+ // cmn w0, #1
+ // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
+ // if and only if (sext LHS) == (sext RHS). The checks are in place to ensure
+ // both the LHS and RHS are truely zero extended and to make sure the
+ // transformation is profitable.
+ if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
+ if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
+ isa<LoadSDNode>(LHS)) {
+ if (cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
+ cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
+ LHS.getNode()->hasNUsesOfValue(1, 0)) {
+ int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
+ if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
+ SDValue SExt =
+ DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
+ DAG.getValueType(MVT::i16));
+ Cmp = emitComparison(SExt,
+ DAG.getConstant(ValueofRHS, RHS.getValueType()),
+ CC, dl, DAG);
+ AArch64CC = changeIntCCToAArch64CC(CC);
+ AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
+ return Cmp;
+ }
+ }
+ }
+ }
+ Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
+ AArch64CC = changeIntCCToAArch64CC(CC);
AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
return Cmp;
}
if (VT.getSizeInBits() > InVT.getSizeInBits()) {
SDLoc dl(Op);
- SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Op.getOperand(0));
+ MVT ExtVT =
+ MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
+ VT.getVectorNumElements());
+ SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
}
(ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
- StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
+ StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
TargetLowering::CallLoweringInfo CLI(DAG);
CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
.setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
0);
}
+static EVT getExtensionTo64Bits(const EVT &OrigVT) {
+ if (OrigVT.getSizeInBits() >= 64)
+ return OrigVT;
+
+ assert(OrigVT.isSimple() && "Expecting a simple value type");
+
+ MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
+ switch (OrigSimpleTy) {
+ default: llvm_unreachable("Unexpected Vector Type");
+ case MVT::v2i8:
+ case MVT::v2i16:
+ return MVT::v2i32;
+ case MVT::v4i8:
+ return MVT::v4i16;
+ }
+}
+
+static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
+ const EVT &OrigTy,
+ const EVT &ExtTy,
+ unsigned ExtOpcode) {
+ // The vector originally had a size of OrigTy. It was then extended to ExtTy.
+ // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
+ // 64-bits we need to insert a new extension so that it will be 64-bits.
+ assert(ExtTy.is128BitVector() && "Unexpected extension size");
+ if (OrigTy.getSizeInBits() >= 64)
+ return N;
+
+ // Must extend size to at least 64 bits to be used as an operand for VMULL.
+ EVT NewVT = getExtensionTo64Bits(OrigTy);
+
+ return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
+}
+
+static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
+ bool isSigned) {
+ EVT VT = N->getValueType(0);
+
+ if (N->getOpcode() != ISD::BUILD_VECTOR)
+ return false;
+
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ SDNode *Elt = N->getOperand(i).getNode();
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits();
+ unsigned HalfSize = EltSize / 2;
+ if (isSigned) {
+ if (!isIntN(HalfSize, C->getSExtValue()))
+ return false;
+ } else {
+ if (!isUIntN(HalfSize, C->getZExtValue()))
+ return false;
+ }
+ continue;
+ }
+ return false;
+ }
+
+ return true;
+}
+
+static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
+ return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
+ N->getOperand(0)->getValueType(0),
+ N->getValueType(0),
+ N->getOpcode());
+
+ assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
+ EVT VT = N->getValueType(0);
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
+ unsigned NumElts = VT.getVectorNumElements();
+ MVT TruncVT = MVT::getIntegerVT(EltSize);
+ SmallVector<SDValue, 8> Ops;
+ for (unsigned i = 0; i != NumElts; ++i) {
+ ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
+ const APInt &CInt = C->getAPIntValue();
+ // Element types smaller than 32 bits are not legal, so use i32 elements.
+ // The values are implicitly truncated so sext vs. zext doesn't matter.
+ Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
+ }
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
+ MVT::getVectorVT(TruncVT, NumElts), Ops);
+}
+
+static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::SIGN_EXTEND)
+ return true;
+ if (isExtendedBUILD_VECTOR(N, DAG, true))
+ return true;
+ return false;
+}
+
+static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::ZERO_EXTEND)
+ return true;
+ if (isExtendedBUILD_VECTOR(N, DAG, false))
+ return true;
+ return false;
+}
+
+static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
+ unsigned Opcode = N->getOpcode();
+ if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
+ SDNode *N0 = N->getOperand(0).getNode();
+ SDNode *N1 = N->getOperand(1).getNode();
+ return N0->hasOneUse() && N1->hasOneUse() &&
+ isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
+ }
+ return false;
+}
+
+static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
+ unsigned Opcode = N->getOpcode();
+ if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
+ SDNode *N0 = N->getOperand(0).getNode();
+ SDNode *N1 = N->getOperand(1).getNode();
+ return N0->hasOneUse() && N1->hasOneUse() &&
+ isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
+ }
+ return false;
+}
+
+static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
+ // Multiplications are only custom-lowered for 128-bit vectors so that
+ // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
+ EVT VT = Op.getValueType();
+ assert(VT.is128BitVector() && VT.isInteger() &&
+ "unexpected type for custom-lowering ISD::MUL");
+ SDNode *N0 = Op.getOperand(0).getNode();
+ SDNode *N1 = Op.getOperand(1).getNode();
+ unsigned NewOpc = 0;
+ bool isMLA = false;
+ bool isN0SExt = isSignExtended(N0, DAG);
+ bool isN1SExt = isSignExtended(N1, DAG);
+ if (isN0SExt && isN1SExt)
+ NewOpc = AArch64ISD::SMULL;
+ else {
+ bool isN0ZExt = isZeroExtended(N0, DAG);
+ bool isN1ZExt = isZeroExtended(N1, DAG);
+ if (isN0ZExt && isN1ZExt)
+ NewOpc = AArch64ISD::UMULL;
+ else if (isN1SExt || isN1ZExt) {
+ // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
+ // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
+ if (isN1SExt && isAddSubSExt(N0, DAG)) {
+ NewOpc = AArch64ISD::SMULL;
+ isMLA = true;
+ } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
+ NewOpc = AArch64ISD::UMULL;
+ isMLA = true;
+ } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
+ std::swap(N0, N1);
+ NewOpc = AArch64ISD::UMULL;
+ isMLA = true;
+ }
+ }
+
+ if (!NewOpc) {
+ if (VT == MVT::v2i64)
+ // Fall through to expand this. It is not legal.
+ return SDValue();
+ else
+ // Other vector multiplications are legal.
+ return Op;
+ }
+ }
+
+ // Legalize to a S/UMULL instruction
+ SDLoc DL(Op);
+ SDValue Op0;
+ SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
+ if (!isMLA) {
+ Op0 = skipExtensionForVectorMULL(N0, DAG);
+ assert(Op0.getValueType().is64BitVector() &&
+ Op1.getValueType().is64BitVector() &&
+ "unexpected types for extended operands to VMULL");
+ return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
+ }
+ // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
+ // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
+ // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
+ SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
+ SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
+ EVT Op1VT = Op1.getValueType();
+ return DAG.getNode(N0->getOpcode(), DL, VT,
+ DAG.getNode(NewOpc, DL, VT,
+ DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
+ DAG.getNode(NewOpc, DL, VT,
+ DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
+}
SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
SelectionDAG &DAG) const {
return LowerFP_TO_INT(Op, DAG);
case ISD::FSINCOS:
return LowerFSINCOS(Op, DAG);
+ case ISD::MUL:
+ return LowerMUL(Op, DAG);
}
}
#include "AArch64GenCallingConv.inc"
-/// Selects the correct CCAssignFn for a the given CallingConvention
-/// value.
+/// Selects the correct CCAssignFn for a given CallingConvention value.
CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
bool IsVarArg) const {
switch (CC) {
unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
uint32_t BEAlign = 0;
- if (ArgSize < 8 && !Subtarget->isLittleEndian())
+ if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
+ !Ins[i].Flags.isInConsecutiveRegs())
BEAlign = 8 - ArgSize;
int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
MachinePointerInfo::getFixedStack(FI),
- MemVT, false, false, false, 0, nullptr);
+ MemVT, false, false, false, 0);
InVals.push_back(ArgValue);
}
// cannot rely on the linker replacing the tail call with a return.
if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
const GlobalValue *GV = G->getGlobal();
- if (GV->hasExternalWeakLinkage())
+ const Triple TT(getTargetMachine().getTargetTriple());
+ if (GV->hasExternalWeakLinkage() &&
+ (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
return false;
}
unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
: VA.getValVT().getSizeInBits();
OpSize = (OpSize + 7) / 8;
- if (!Subtarget->isLittleEndian() && !Flags.isByVal()) {
+ if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
+ !Flags.isInConsecutiveRegs()) {
if (OpSize < 8)
BEAlign = 8 - OpSize;
}
SelectionDAG &DAG) const {
EVT PtrVT = getPointerTy();
SDLoc DL(Op);
- const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
+ const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
+ const GlobalValue *GV = GN->getGlobal();
unsigned char OpFlags =
Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
}
+ if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
+ assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
+ "use of MO_CONSTPOOL only supported on small model");
+ SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
+ SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
+ unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
+ SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
+ SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
+ SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr,
+ MachinePointerInfo::getConstantPool(),
+ /*isVolatile=*/ false,
+ /*isNonTemporal=*/ true,
+ /*isInvariant=*/ true, 8);
+ if (GN->getOffset() != 0)
+ return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
+ DAG.getConstant(GN->getOffset(), PtrVT));
+ return GlobalAddr;
+ }
+
if (getTargetMachine().getCodeModel() == CodeModel::Large) {
const unsigned char MO_NC = AArch64II::MO_NC;
return DAG.getNode(
AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
return SDValue();
+ if (!Subtarget->hasNEON())
+ return SDValue();
+
// While there is no integer popcount instruction, it can
// be more efficiently lowered to the following sequence that uses
// AdvSIMD registers/instructions as long as the copies to/from
SDValue SourceVec = V.getOperand(0);
auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
if (Source == Sources.end())
- Sources.push_back(ShuffleSourceInfo(SourceVec));
+ Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
// Update the minimum and maximum lane number seen.
unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
// This stage of the search produces a source with the same element type as
// the original, but with a total width matching the BUILD_VECTOR output.
EVT EltVT = SrcVT.getVectorElementType();
- EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT,
- VT.getSizeInBits() / EltVT.getSizeInBits());
+ unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
+ EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
- if (Src.MaxElt - Src.MinElt >= NumElts) {
+ if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
// Span too large for a VEXT to cope
return SDValue();
}
- if (Src.MinElt >= NumElts) {
+ if (Src.MinElt >= NumSrcElts) {
// The extraction can just take the second half
Src.ShuffleVec =
DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
- DAG.getIntPtrConstant(NumElts));
- Src.WindowBase = -NumElts;
- } else if (Src.MaxElt < NumElts) {
+ DAG.getConstant(NumSrcElts, MVT::i64));
+ Src.WindowBase = -NumSrcElts;
+ } else if (Src.MaxElt < NumSrcElts) {
// The extraction can just take the first half
- Src.ShuffleVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT,
- Src.ShuffleVec, DAG.getIntPtrConstant(0));
+ Src.ShuffleVec =
+ DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
+ DAG.getConstant(0, MVT::i64));
} else {
// An actual VEXT is needed
- SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT,
- Src.ShuffleVec, DAG.getIntPtrConstant(0));
+ SDValue VEXTSrc1 =
+ DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
+ DAG.getConstant(0, MVT::i64));
SDValue VEXTSrc2 =
DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
- DAG.getIntPtrConstant(NumElts));
+ DAG.getConstant(NumSrcElts, MVT::i64));
unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
VT.getVectorElementType() == MVT::f32)
return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
// vrev <4 x i16> -> REV32
- if (VT.getVectorElementType() == MVT::i16)
+ if (VT.getVectorElementType() == MVT::i16 ||
+ VT.getVectorElementType() == MVT::f16)
return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
// vrev <4 x i8> -> REV16
assert(VT.getVectorElementType() == MVT::i8);
static unsigned getDUPLANEOp(EVT EltType) {
if (EltType == MVT::i8)
return AArch64ISD::DUPLANE8;
- if (EltType == MVT::i16)
+ if (EltType == MVT::i16 || EltType == MVT::f16)
return AArch64ISD::DUPLANE16;
if (EltType == MVT::i32 || EltType == MVT::f32)
return AArch64ISD::DUPLANE32;
SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
EVT ScalarVT = VT.getVectorElementType();
- if (ScalarVT.getSizeInBits() < 32)
+
+ if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
ScalarVT = MVT::i32;
return DAG.getNode(
SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(16, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(24, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
}
SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(16, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(24, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
}
if (VT.getSizeInBits() == 128) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
DAG.getConstant(CnstVal, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
// Support the V64 version via subregister insertion.
SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
DAG.getConstant(CnstVal, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(16, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(24, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(264, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(272, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
// The few faces of FMOV...
MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
DAG.getConstant(CnstVal, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
// The many faces of MVNI...
SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(16, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(24, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(0, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(8, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(264, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
DAG.getConstant(CnstVal, MVT::i32),
DAG.getConstant(272, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
+ return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
}
}
// Insertion/extraction are legal for V128 types.
if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
- VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
+ VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
+ VT == MVT::v8f16)
return Op;
if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
- VT != MVT::v1i64 && VT != MVT::v2f32)
+ VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
return SDValue();
// For V64 types, we perform insertion by expanding the value
// Insertion/extraction are legal for V128 types.
if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
- VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
+ VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
+ VT == MVT::v8f16)
return Op;
if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
- VT != MVT::v1i64 && VT != MVT::v2f32)
+ VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
return SDValue();
// For V64 types, we perform extraction by expanding the value
SDValue N0 = N->getOperand(0);
unsigned Lg2 = Divisor.countTrailingZeros();
SDValue Zero = DAG.getConstant(0, VT);
- SDValue Pow2MinusOne = DAG.getConstant((1 << Lg2) - 1, VT);
+ SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, VT);
// Add (N0 < 0) ? Pow2 - 1 : 0;
SDValue CCVal;
return SDValue();
}
-static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
+ const AArch64Subtarget *Subtarget) {
// First try to optimize away the conversion when it's conditionally from
// a constant. Vectors only.
SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
// conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
// This eliminates an "integer-to-vector-move UOP and improve throughput.
SDValue N0 = N->getOperand(0);
- if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
+ if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
// Do not change the width of a volatile load.
!cast<LoadSDNode>(N0)->isVolatile()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
// If the vector type isn't a simple VT, it's beyond the scope of what
// we're worried about here. Let legalization do its thing and hope for
// the best.
- if (!ResVT.isSimple())
+ SDValue Src = N->getOperand(0);
+ EVT SrcVT = Src->getValueType(0);
+ if (!ResVT.isSimple() || !SrcVT.isSimple())
return SDValue();
- SDValue Src = N->getOperand(0);
- MVT SrcVT = Src->getValueType(0).getSimpleVT();
// If the source VT is a 64-bit vector, we can play games and get the
// better results we want.
if (SrcVT.getSizeInBits() != 64)
EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
LoVT.getVectorNumElements());
Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
- DAG.getIntPtrConstant(0));
+ DAG.getConstant(0, MVT::i64));
Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
- DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
+ DAG.getConstant(InNVT.getVectorNumElements(), MVT::i64));
Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
EVT HalfVT =
EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
- DAG.getIntPtrConstant(0));
+ DAG.getConstant(0, MVT::i64));
SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
- DAG.getIntPtrConstant(NumElts));
+ DAG.getConstant(NumElts, MVT::i64));
SDValue BasePtr = S->getBasePtr();
SDValue NewST1 =
DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
Ops.push_back(Inc);
EVT Tys[3] = { VT, MVT::i64, MVT::Other };
- SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, 3));
+ SDVTList SDTys = DAG.getVTList(Tys);
unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
MemVT,
Tys[n] = VecTy;
Tys[n++] = MVT::i64; // Type of write back register
Tys[n] = MVT::Other; // Type of the chain
- SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs + 2));
+ SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
return SDValue();
}
+// Checks to see if the value is the prescribed width and returns information
+// about its extension mode.
+static
+bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
+ ExtType = ISD::NON_EXTLOAD;
+ switch(V.getNode()->getOpcode()) {
+ default:
+ return false;
+ case ISD::LOAD: {
+ LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
+ if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
+ || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
+ ExtType = LoadNode->getExtensionType();
+ return true;
+ }
+ return false;
+ }
+ case ISD::AssertSext: {
+ VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
+ if ((TypeNode->getVT() == MVT::i8 && width == 8)
+ || (TypeNode->getVT() == MVT::i16 && width == 16)) {
+ ExtType = ISD::SEXTLOAD;
+ return true;
+ }
+ return false;
+ }
+ case ISD::AssertZext: {
+ VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
+ if ((TypeNode->getVT() == MVT::i8 && width == 8)
+ || (TypeNode->getVT() == MVT::i16 && width == 16)) {
+ ExtType = ISD::ZEXTLOAD;
+ return true;
+ }
+ return false;
+ }
+ case ISD::Constant:
+ case ISD::TargetConstant: {
+ if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
+ 1LL << (width - 1))
+ return true;
+ return false;
+ }
+ }
+
+ return true;
+}
+
+// This function does a whole lot of voodoo to determine if the tests are
+// equivalent without and with a mask. Essentially what happens is that given a
+// DAG resembling:
+//
+// +-------------+ +-------------+ +-------------+ +-------------+
+// | Input | | AddConstant | | CompConstant| | CC |
+// +-------------+ +-------------+ +-------------+ +-------------+
+// | | | |
+// V V | +----------+
+// +-------------+ +----+ | |
+// | ADD | |0xff| | |
+// +-------------+ +----+ | |
+// | | | |
+// V V | |
+// +-------------+ | |
+// | AND | | |
+// +-------------+ | |
+// | | |
+// +-----+ | |
+// | | |
+// V V V
+// +-------------+
+// | CMP |
+// +-------------+
+//
+// The AND node may be safely removed for some combinations of inputs. In
+// particular we need to take into account the extension type of the Input,
+// the exact values of AddConstant, CompConstant, and CC, along with the nominal
+// width of the input (this can work for any width inputs, the above graph is
+// specific to 8 bits.
+//
+// The specific equations were worked out by generating output tables for each
+// AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
+// problem was simplified by working with 4 bit inputs, which means we only
+// needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
+// extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
+// patterns present in both extensions (0,7). For every distinct set of
+// AddConstant and CompConstants bit patterns we can consider the masked and
+// unmasked versions to be equivalent if the result of this function is true for
+// all 16 distinct bit patterns of for the current extension type of Input (w0).
+//
+// sub w8, w0, w1
+// and w10, w8, #0x0f
+// cmp w8, w2
+// cset w9, AArch64CC
+// cmp w10, w2
+// cset w11, AArch64CC
+// cmp w9, w11
+// cset w0, eq
+// ret
+//
+// Since the above function shows when the outputs are equivalent it defines
+// when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
+// would be expensive to run during compiles. The equations below were written
+// in a test harness that confirmed they gave equivalent outputs to the above
+// for all inputs function, so they can be used determine if the removal is
+// legal instead.
+//
+// isEquivalentMaskless() is the code for testing if the AND can be removed
+// factored out of the DAG recognition as the DAG can take several forms.
+
+static
+bool isEquivalentMaskless(unsigned CC, unsigned width,
+ ISD::LoadExtType ExtType, signed AddConstant,
+ signed CompConstant) {
+ // By being careful about our equations and only writing the in term
+ // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
+ // make them generally applicable to all bit widths.
+ signed MaxUInt = (1 << width);
+
+ // For the purposes of these comparisons sign extending the type is
+ // equivalent to zero extending the add and displacing it by half the integer
+ // width. Provided we are careful and make sure our equations are valid over
+ // the whole range we can just adjust the input and avoid writing equations
+ // for sign extended inputs.
+ if (ExtType == ISD::SEXTLOAD)
+ AddConstant -= (1 << (width-1));
+
+ switch(CC) {
+ case AArch64CC::LE:
+ case AArch64CC::GT: {
+ if ((AddConstant == 0) ||
+ (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
+ (AddConstant >= 0 && CompConstant < 0) ||
+ (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
+ return true;
+ } break;
+ case AArch64CC::LT:
+ case AArch64CC::GE: {
+ if ((AddConstant == 0) ||
+ (AddConstant >= 0 && CompConstant <= 0) ||
+ (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
+ return true;
+ } break;
+ case AArch64CC::HI:
+ case AArch64CC::LS: {
+ if ((AddConstant >= 0 && CompConstant < 0) ||
+ (AddConstant <= 0 && CompConstant >= -1 &&
+ CompConstant < AddConstant + MaxUInt))
+ return true;
+ } break;
+ case AArch64CC::PL:
+ case AArch64CC::MI: {
+ if ((AddConstant == 0) ||
+ (AddConstant > 0 && CompConstant <= 0) ||
+ (AddConstant < 0 && CompConstant <= AddConstant))
+ return true;
+ } break;
+ case AArch64CC::LO:
+ case AArch64CC::HS: {
+ if ((AddConstant >= 0 && CompConstant <= 0) ||
+ (AddConstant <= 0 && CompConstant >= 0 &&
+ CompConstant <= AddConstant + MaxUInt))
+ return true;
+ } break;
+ case AArch64CC::EQ:
+ case AArch64CC::NE: {
+ if ((AddConstant > 0 && CompConstant < 0) ||
+ (AddConstant < 0 && CompConstant >= 0 &&
+ CompConstant < AddConstant + MaxUInt) ||
+ (AddConstant >= 0 && CompConstant >= 0 &&
+ CompConstant >= AddConstant) ||
+ (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
+
+ return true;
+ } break;
+ case AArch64CC::VS:
+ case AArch64CC::VC:
+ case AArch64CC::AL:
+ case AArch64CC::NV:
+ return true;
+ case AArch64CC::Invalid:
+ break;
+ }
+
+ return false;
+}
+
+static
+SDValue performCONDCombine(SDNode *N,
+ TargetLowering::DAGCombinerInfo &DCI,
+ SelectionDAG &DAG, unsigned CCIndex,
+ unsigned CmpIndex) {
+ unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
+ SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
+ unsigned CondOpcode = SubsNode->getOpcode();
+
+ if (CondOpcode != AArch64ISD::SUBS)
+ return SDValue();
+
+ // There is a SUBS feeding this condition. Is it fed by a mask we can
+ // use?
+
+ SDNode *AndNode = SubsNode->getOperand(0).getNode();
+ unsigned MaskBits = 0;
+
+ if (AndNode->getOpcode() != ISD::AND)
+ return SDValue();
+
+ if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
+ uint32_t CNV = CN->getZExtValue();
+ if (CNV == 255)
+ MaskBits = 8;
+ else if (CNV == 65535)
+ MaskBits = 16;
+ }
+
+ if (!MaskBits)
+ return SDValue();
+
+ SDValue AddValue = AndNode->getOperand(0);
+
+ if (AddValue.getOpcode() != ISD::ADD)
+ return SDValue();
+
+ // The basic dag structure is correct, grab the inputs and validate them.
+
+ SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
+ SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
+ SDValue SubsInputValue = SubsNode->getOperand(1);
+
+ // The mask is present and the provenance of all the values is a smaller type,
+ // lets see if the mask is superfluous.
+
+ if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
+ !isa<ConstantSDNode>(SubsInputValue.getNode()))
+ return SDValue();
+
+ ISD::LoadExtType ExtType;
+
+ if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
+ !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
+ !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
+ return SDValue();
+
+ if(!isEquivalentMaskless(CC, MaskBits, ExtType,
+ cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
+ cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
+ return SDValue();
+
+ // The AND is not necessary, remove it.
+
+ SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
+ SubsNode->getValueType(1));
+ SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
+
+ SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
+ DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
+
+ return SDValue(N, 0);
+}
+
// Optimize compare with zero and branch.
static SDValue performBRCONDCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
SelectionDAG &DAG) {
+ SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
+ if (NV.getNode())
+ N = NV.getNode();
SDValue Chain = N->getOperand(0);
SDValue Dest = N->getOperand(1);
SDValue CCVal = N->getOperand(2);
SDValue N0 = N->getOperand(0);
EVT ResVT = N->getValueType(0);
- if (!N->getOperand(1).getValueType().isVector())
+ if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
return SDValue();
- if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
+ // If NumMaskElts == 0, the comparison is larger than select result. The
+ // largest real NEON comparison is 64-bits per lane, which means the result is
+ // at most 32-bits and an illegal vector. Just bail out for now.
+ EVT SrcVT = N0.getOperand(0).getValueType();
+
+ // Don't try to do this optimization when the setcc itself has i1 operands.
+ // There are no legal vectors of i1, so this would be pointless.
+ if (SrcVT == MVT::i1)
return SDValue();
- SDLoc DL(N0);
+ int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
+ if (!ResVT.isVector() || NumMaskElts == 0)
+ return SDValue();
- EVT SrcVT = N0.getOperand(0).getValueType();
- SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT,
- ResVT.getSizeInBits() / SrcVT.getSizeInBits());
+ SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
// First perform a vector comparison, where lane 0 is the one we're interested
// in.
+ SDLoc DL(N0);
SDValue LHS =
DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
SDValue RHS =
// Now duplicate the comparison mask we want across all other lanes.
SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
- Mask = DAG.getNode(ISD::BITCAST, DL, ResVT.changeVectorElementTypeToInteger(),
- Mask);
+ Mask = DAG.getNode(ISD::BITCAST, DL,
+ ResVT.changeVectorElementTypeToInteger(), Mask);
return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
}
return performMulCombine(N, DAG, DCI, Subtarget);
case ISD::SINT_TO_FP:
case ISD::UINT_TO_FP:
- return performIntToFpCombine(N, DAG);
+ return performIntToFpCombine(N, DAG, Subtarget);
case ISD::OR:
return performORCombine(N, DCI, Subtarget);
case ISD::INTRINSIC_WO_CHAIN:
return performSTORECombine(N, DCI, DAG, Subtarget);
case AArch64ISD::BRCOND:
return performBRCONDCombine(N, DCI, DAG);
+ case AArch64ISD::CSEL:
+ return performCONDCombine(N, DCI, DAG, 2, 3);
case AArch64ISD::DUP:
return performPostLD1Combine(N, DCI, false);
case ISD::INSERT_VECTOR_ELT:
static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
SelectionDAG &DAG) {
- if (N->getValueType(0) != MVT::i16)
- return;
-
SDLoc DL(N);
SDValue Op = N->getOperand(0);
- assert(Op.getValueType() == MVT::f16 &&
- "Inconsistent bitcast? Only 16-bit types should be i16 or f16");
+
+ if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
+ return;
+
Op = SDValue(
DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
DAG.getUNDEF(MVT::i32), Op,
}
}
-bool AArch64TargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
- // Loads and stores less than 128-bits are already atomic; ones above that
- // are doomed anyway, so defer to the default libcall and blame the OS when
- // things go wrong:
- if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
- return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128;
- else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
- return LI->getType()->getPrimitiveSizeInBits() == 128;
-
- // For the real atomic operations, we have ldxr/stxr up to 128 bits.
- return Inst->getType()->getPrimitiveSizeInBits() <= 128;
-}
-
bool AArch64TargetLowering::useLoadStackGuardNode() const {
return true;
}
+bool AArch64TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
+ // Combine multiple FDIVs with the same divisor into multiple FMULs by the
+ // reciprocal if there are three or more FDIVs.
+ return NumUsers > 2;
+}
+
TargetLoweringBase::LegalizeTypeAction
AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
MVT SVT = VT.getSimpleVT();
return TargetLoweringBase::getPreferredVectorAction(VT);
}
+// Loads and stores less than 128-bits are already atomic; ones above that
+// are doomed anyway, so defer to the default libcall and blame the OS when
+// things go wrong.
+bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
+ unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
+ return Size == 128;
+}
+
+// Loads and stores less than 128-bits are already atomic; ones above that
+// are doomed anyway, so defer to the default libcall and blame the OS when
+// things go wrong.
+bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
+ unsigned Size = LI->getType()->getPrimitiveSizeInBits();
+ return Size == 128;
+}
+
+// For the real atomic operations, we have ldxr/stxr up to 128 bits,
+bool AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
+ unsigned Size = AI->getType()->getPrimitiveSizeInBits();
+ return Size <= 128;
+}
+
+bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
+ return true;
+}
+
Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
AtomicOrdering Ord) const {
Module *M = Builder.GetInsertBlock()->getParent()->getParent();
Val, Stxr->getFunctionType()->getParamType(0)),
Addr);
}
+
+bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
+ Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
+ return Ty->isArrayTy();
+}