SystemZISelLowering.cpp

Go to the documentation of this file.

//===-- SystemZISelLowering.cpp - SystemZ DAG lowering implementation -----===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements the SystemZTargetLowering class.
//
//===----------------------------------------------------------------------===//
 
#include "SystemZISelLowering.h"
#include "SystemZCallingConv.h"
#include "SystemZConstantPoolValue.h"
#include "SystemZMachineFunctionInfo.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
#include "llvm/IR/GlobalAlias.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsS390.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Target/TargetMachine.h"
#include <cctype>
#include <optional>
 
using namespace llvm;
 
#define DEBUG_TYPE "systemz-lower"
 
// Temporarily let this be disabled by default until all known problems
// related to argument extensions are fixed.
static cl::opt<bool> EnableIntArgExtCheck(
    "argext-abi-check", cl::init(false),
    cl::desc("Verify that narrow int args are properly extended per the "
             "SystemZ ABI."));
 
namespace {
// Represents information about a comparison.
struct Comparison {
  Comparison(SDValue Op0In, SDValue Op1In, SDValue ChainIn)
    : Op0(Op0In), Op1(Op1In), Chain(ChainIn),
      Opcode(0), ICmpType(0), CCValid(0), CCMask(0) {}
 
  // The operands to the comparison.
  SDValue Op0, Op1;
 
  // Chain if this is a strict floating-point comparison.
  SDValue Chain;
 
  // The opcode that should be used to compare Op0 and Op1.
  unsigned Opcode;
 
  // A SystemZICMP value.  Only used for integer comparisons.
  unsigned ICmpType;
 
  // The mask of CC values that Opcode can produce.
  unsigned CCValid;
 
  // The mask of CC values for which the original condition is true.
  unsigned CCMask;
};
} // end anonymous namespace
 
// Classify VT as either 32 or 64 bit.
static bool is32Bit(EVT VT) {
  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::i32:
    return true;
  case MVT::i64:
    return false;
  default:
    llvm_unreachable("Unsupported type");
  }
}
 
// Return a version of MachineOperand that can be safely used before the
// final use.
static MachineOperand earlyUseOperand(MachineOperand Op) {
  if (Op.isReg())
    Op.setIsKill(false);
  return Op;
}
 
SystemZTargetLowering::SystemZTargetLowering(const TargetMachine &TM,
                                             const SystemZSubtarget &STI)
    : TargetLowering(TM, STI), Subtarget(STI) {
  MVT PtrVT = MVT::getIntegerVT(TM.getPointerSizeInBits(0));
 
  auto *Regs = STI.getSpecialRegisters();
 
  // Set up the register classes.
  if (Subtarget.hasHighWord())
    addRegisterClass(MVT::i32, &SystemZ::GRX32BitRegClass);
  else
    addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass);
  addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
  if (!useSoftFloat()) {
    if (Subtarget.hasVector()) {
      addRegisterClass(MVT::f16, &SystemZ::VR16BitRegClass);
      addRegisterClass(MVT::f32, &SystemZ::VR32BitRegClass);
      addRegisterClass(MVT::f64, &SystemZ::VR64BitRegClass);
    } else {
      addRegisterClass(MVT::f16, &SystemZ::FP16BitRegClass);
      addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
      addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
    }
    if (Subtarget.hasVectorEnhancements1())
      addRegisterClass(MVT::f128, &SystemZ::VR128BitRegClass);
    else
      addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
 
    if (Subtarget.hasVector()) {
      addRegisterClass(MVT::v16i8, &SystemZ::VR128BitRegClass);
      addRegisterClass(MVT::v8i16, &SystemZ::VR128BitRegClass);
      addRegisterClass(MVT::v4i32, &SystemZ::VR128BitRegClass);
      addRegisterClass(MVT::v2i64, &SystemZ::VR128BitRegClass);
      addRegisterClass(MVT::v8f16, &SystemZ::VR128BitRegClass);
      addRegisterClass(MVT::v4f32, &SystemZ::VR128BitRegClass);
      addRegisterClass(MVT::v2f64, &SystemZ::VR128BitRegClass);
    }
 
    if (Subtarget.hasVector())
      addRegisterClass(MVT::i128, &SystemZ::VR128BitRegClass);
  }
 
  // Compute derived properties from the register classes
  computeRegisterProperties(Subtarget.getRegisterInfo());
 
  // Set up special registers.
  setStackPointerRegisterToSaveRestore(Regs->getStackPointerRegister());
 
  // TODO: It may be better to default to latency-oriented scheduling, however
  // LLVM's current latency-oriented scheduler can't handle physreg definitions
  // such as SystemZ has with CC, so set this to the register-pressure
  // scheduler, because it can.
  setSchedulingPreference(Sched::RegPressure);
 
  setBooleanContents(ZeroOrOneBooleanContent);
  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
 
  setMaxAtomicSizeInBitsSupported(128);
 
  // Instructions are strings of 2-byte aligned 2-byte values.
  setMinFunctionAlignment(Align(2));
  // For performance reasons we prefer 16-byte alignment.
  setPrefFunctionAlignment(Align(16));
 
  // Handle operations that are handled in a similar way for all types.
  for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
       I <= MVT::LAST_FP_VALUETYPE;
       ++I) {
    MVT VT = MVT::SimpleValueType(I);
    if (isTypeLegal(VT)) {
      // Lower SET_CC into an IPM-based sequence.
      setOperationAction(ISD::SETCC, VT, Custom);
      setOperationAction(ISD::STRICT_FSETCC, VT, Custom);
      setOperationAction(ISD::STRICT_FSETCCS, VT, Custom);
 
      // Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
      setOperationAction(ISD::SELECT, VT, Expand);
 
      // Lower SELECT_CC and BR_CC into separate comparisons and branches.
      setOperationAction(ISD::SELECT_CC, VT, Custom);
      setOperationAction(ISD::BR_CC,     VT, Custom);
    }
  }
 
  // Expand jump table branches as address arithmetic followed by an
  // indirect jump.
  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
 
  // Expand BRCOND into a BR_CC (see above).
  setOperationAction(ISD::BRCOND, MVT::Other, Expand);
 
  // Handle integer types except i128.
  for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE;
       I <= MVT::LAST_INTEGER_VALUETYPE;
       ++I) {
    MVT VT = MVT::SimpleValueType(I);
    if (isTypeLegal(VT) && VT != MVT::i128) {
      setOperationAction(ISD::ABS, VT, Legal);
 
      // Expand individual DIV and REMs into DIVREMs.
      setOperationAction(ISD::SDIV, VT, Expand);
      setOperationAction(ISD::UDIV, VT, Expand);
      setOperationAction(ISD::SREM, VT, Expand);
      setOperationAction(ISD::UREM, VT, Expand);
      setOperationAction(ISD::SDIVREM, VT, Custom);
      setOperationAction(ISD::UDIVREM, VT, Custom);
 
      // Support addition/subtraction with overflow.
      setOperationAction(ISD::SADDO, VT, Custom);
      setOperationAction(ISD::SSUBO, VT, Custom);
 
      // Support addition/subtraction with carry.
      setOperationAction(ISD::UADDO, VT, Custom);
      setOperationAction(ISD::USUBO, VT, Custom);
 
      // Support carry in as value rather than glue.
      setOperationAction(ISD::UADDO_CARRY, VT, Custom);
      setOperationAction(ISD::USUBO_CARRY, VT, Custom);
 
      // Lower ATOMIC_LOAD_SUB into ATOMIC_LOAD_ADD if LAA and LAAG are
      // available, or if the operand is constant.
      setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
 
      // Use POPCNT on z196 and above.
      if (Subtarget.hasPopulationCount())
        setOperationAction(ISD::CTPOP, VT, Custom);
      else
        setOperationAction(ISD::CTPOP, VT, Expand);
 
      // No special instructions for these.
      setOperationAction(ISD::CTTZ,            VT, Expand);
      setOperationAction(ISD::ROTR,            VT, Expand);
 
      // Use *MUL_LOHI where possible instead of MULH*.
      setOperationAction(ISD::MULHS, VT, Expand);
      setOperationAction(ISD::MULHU, VT, Expand);
      setOperationAction(ISD::SMUL_LOHI, VT, Custom);
      setOperationAction(ISD::UMUL_LOHI, VT, Custom);
 
      // The fp<=>i32/i64 conversions are all Legal except for f16 and for
      // unsigned on z10 (only z196 and above have native support for
      // unsigned conversions).
      for (auto Op : {ISD::FP_TO_SINT, ISD::STRICT_FP_TO_SINT,
                      ISD::SINT_TO_FP, ISD::STRICT_SINT_TO_FP})
        setOperationAction(Op, VT, Custom);
      for (auto Op : {ISD::FP_TO_UINT, ISD::STRICT_FP_TO_UINT})
        setOperationAction(Op, VT, Custom);
      for (auto Op : {ISD::UINT_TO_FP, ISD::STRICT_UINT_TO_FP}) {
        // Handle unsigned 32-bit input types as signed 64-bit types on z10.
        auto OpAction =
            (!Subtarget.hasFPExtension() && VT == MVT::i32) ? Promote : Custom;
        setOperationAction(Op, VT, OpAction);
      }
    }
  }
 
  // Handle i128 if legal.
  if (isTypeLegal(MVT::i128)) {
    // No special instructions for these.
    setOperationAction(ISD::SDIVREM,   MVT::i128, Expand);
    setOperationAction(ISD::UDIVREM,   MVT::i128, Expand);
    setOperationAction(ISD::SMUL_LOHI, MVT::i128, Expand);
    setOperationAction(ISD::UMUL_LOHI, MVT::i128, Expand);
    setOperationAction(ISD::ROTR,      MVT::i128, Expand);
    setOperationAction(ISD::ROTL,      MVT::i128, Expand);
 
    // We may be able to use VSLDB/VSLD/VSRD for these.
    setOperationAction(ISD::FSHL,      MVT::i128, Custom);
    setOperationAction(ISD::FSHR,      MVT::i128, Custom);
 
    // No special instructions for these before z17.
    if (!Subtarget.hasVectorEnhancements3()) {
      setOperationAction(ISD::MUL,   MVT::i128, Expand);
      setOperationAction(ISD::MULHS, MVT::i128, Expand);
      setOperationAction(ISD::MULHU, MVT::i128, Expand);
      setOperationAction(ISD::SDIV,  MVT::i128, Expand);
      setOperationAction(ISD::UDIV,  MVT::i128, Expand);
      setOperationAction(ISD::SREM,  MVT::i128, Expand);
      setOperationAction(ISD::UREM,  MVT::i128, Expand);
      setOperationAction(ISD::CTLZ,  MVT::i128, Expand);
      setOperationAction(ISD::CTTZ,  MVT::i128, Expand);
    } else {
      // Even if we do have a legal 128-bit multiply, we do not
      // want 64-bit multiply-high operations to use it.
      setOperationAction(ISD::MULHS, MVT::i64, Custom);
      setOperationAction(ISD::MULHU, MVT::i64, Custom);
    }
 
    // Support addition/subtraction with carry.
    setOperationAction(ISD::UADDO, MVT::i128, Custom);
    setOperationAction(ISD::USUBO, MVT::i128, Custom);
    setOperationAction(ISD::UADDO_CARRY, MVT::i128, Custom);
    setOperationAction(ISD::USUBO_CARRY, MVT::i128, Custom);
 
    // Use VPOPCT and add up partial results.
    setOperationAction(ISD::CTPOP, MVT::i128, Custom);
 
    // Additional instructions available with z17.
    if (Subtarget.hasVectorEnhancements3()) {
      setOperationAction(ISD::ABS, MVT::i128, Legal);
 
      setOperationAction({ISD::SMIN, ISD::UMIN, ISD::SMAX, ISD::UMAX},
                         MVT::i128, Legal);
    }
  }
 
  // These need custom handling in order to handle the f16 conversions.
  setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
  setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
  setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i128, Custom);
 
  // Type legalization will convert 8- and 16-bit atomic operations into
  // forms that operate on i32s (but still keeping the original memory VT).
  // Lower them into full i32 operations.
  setOperationAction(ISD::ATOMIC_SWAP,      MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_ADD,  MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_SUB,  MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_AND,  MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_OR,   MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_XOR,  MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_MIN,  MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_MAX,  MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
 
  // Whether or not i128 is not a legal type, we need to custom lower
  // the atomic operations in order to exploit SystemZ instructions.
  setOperationAction(ISD::ATOMIC_LOAD,     MVT::i128, Custom);
  setOperationAction(ISD::ATOMIC_STORE,    MVT::i128, Custom);
  setOperationAction(ISD::ATOMIC_LOAD, MVT::f128, Custom);
  setOperationAction(ISD::ATOMIC_STORE, MVT::f128, Custom);
 
  // Mark sign/zero extending atomic loads as legal, which will make
  // DAGCombiner fold extensions into atomic loads if possible.
  setAtomicLoadExtAction({ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i64,
                         {MVT::i8, MVT::i16, MVT::i32}, Legal);
  setAtomicLoadExtAction({ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i32,
                         {MVT::i8, MVT::i16}, Legal);
  setAtomicLoadExtAction({ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i16,
                         MVT::i8, Legal);
 
  // We can use the CC result of compare-and-swap to implement
  // the "success" result of ATOMIC_CMP_SWAP_WITH_SUCCESS.
  setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i64, Custom);
  setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
 
  setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
 
  // Traps are legal, as we will convert them to "j .+2".
  setOperationAction(ISD::TRAP, MVT::Other, Legal);
 
  // We have native support for a 64-bit CTLZ, via FLOGR.
  setOperationAction(ISD::CTLZ, MVT::i32, Promote);
  setOperationAction(ISD::CTLZ_ZERO_POISON, MVT::i32, Promote);
  setOperationAction(ISD::CTLZ, MVT::i64, Legal);
 
  // On z17 we have native support for a 64-bit CTTZ.
  if (Subtarget.hasMiscellaneousExtensions4()) {
    setOperationAction(ISD::CTTZ, MVT::i32, Promote);
    setOperationAction(ISD::CTTZ_ZERO_POISON, MVT::i32, Promote);
    setOperationAction(ISD::CTTZ, MVT::i64, Legal);
  }
 
  // On z15 we have native support for a 64-bit CTPOP.
  if (Subtarget.hasMiscellaneousExtensions3()) {
    setOperationAction(ISD::CTPOP, MVT::i32, Promote);
    setOperationAction(ISD::CTPOP, MVT::i64, Legal);
  }
 
  // Give LowerOperation the chance to replace 64-bit ORs with subregs.
  setOperationAction(ISD::OR, MVT::i64, Custom);
 
  // Expand 128 bit shifts without using a libcall.
  setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
  setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
  setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
 
  // Also expand 256 bit shifts if i128 is a legal type.
  if (isTypeLegal(MVT::i128)) {
    setOperationAction(ISD::SRL_PARTS, MVT::i128, Expand);
    setOperationAction(ISD::SHL_PARTS, MVT::i128, Expand);
    setOperationAction(ISD::SRA_PARTS, MVT::i128, Expand);
  }
 
  // Handle bitcast from fp128 to i128.
  if (!isTypeLegal(MVT::i128))
    setOperationAction(ISD::BITCAST, MVT::i128, Custom);
 
  // We have native instructions for i8, i16 and i32 extensions, but not i1.
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
  for (MVT VT : MVT::integer_valuetypes()) {
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
    setLoadExtAction(ISD::EXTLOAD,  VT, MVT::i1, Promote);
  }
 
  // Handle the various types of symbolic address.
  setOperationAction(ISD::ConstantPool,     PtrVT, Custom);
  setOperationAction(ISD::GlobalAddress,    PtrVT, Custom);
  setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom);
  setOperationAction(ISD::BlockAddress,     PtrVT, Custom);
  setOperationAction(ISD::JumpTable,        PtrVT, Custom);
 
  // We need to handle dynamic allocations specially because of the
  // 160-byte area at the bottom of the stack.
  setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
  setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, PtrVT, Custom);
 
  setOperationAction(ISD::STACKSAVE,    MVT::Other, Custom);
  setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom);
 
  // Handle prefetches with PFD or PFDRL.
  setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
 
  // Handle readcyclecounter with STCKF.
  setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
 
  for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
    // Assume by default that all vector operations need to be expanded.
    for (unsigned Opcode = 0; Opcode < ISD::BUILTIN_OP_END; ++Opcode)
      if (getOperationAction(Opcode, VT) == Legal)
        setOperationAction(Opcode, VT, Expand);
 
    // Likewise all truncating stores and extending loads.
    for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
      setTruncStoreAction(VT, InnerVT, Expand);
      setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
      setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
      setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
    }
 
    if (isTypeLegal(VT)) {
      // These operations are legal for anything that can be stored in a
      // vector register, even if there is no native support for the format
      // as such.  In particular, we can do these for v4f32 even though there
      // are no specific instructions for that format.
      setOperationAction(ISD::LOAD, VT, Legal);
      setOperationAction(ISD::STORE, VT, Legal);
      setOperationAction(ISD::VSELECT, VT, Legal);
      setOperationAction(ISD::BITCAST, VT, Legal);
      setOperationAction({ISD::UNDEF, ISD::POISON}, VT, Legal);
 
      // Likewise, except that we need to replace the nodes with something
      // more specific.
      setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
    }
  }
 
  // Handle integer vector types.
  for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
    if (isTypeLegal(VT)) {
      // These operations have direct equivalents.
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Legal);
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Legal);
      setOperationAction(ISD::ADD, VT, Legal);
      setOperationAction(ISD::SUB, VT, Legal);
      if (VT != MVT::v2i64 || Subtarget.hasVectorEnhancements3()) {
        setOperationAction(ISD::MUL, VT, Legal);
        setOperationAction(ISD::MULHS, VT, Legal);
        setOperationAction(ISD::MULHU, VT, Legal);
      }
      if (Subtarget.hasVectorEnhancements3() &&
          VT != MVT::v16i8 && VT != MVT::v8i16) {
        setOperationAction(ISD::SDIV, VT, Legal);
        setOperationAction(ISD::UDIV, VT, Legal);
        setOperationAction(ISD::SREM, VT, Legal);
        setOperationAction(ISD::UREM, VT, Legal);
      }
      setOperationAction(ISD::ABS, VT, Legal);
      setOperationAction(ISD::AND, VT, Legal);
      setOperationAction(ISD::OR, VT, Legal);
      setOperationAction(ISD::XOR, VT, Legal);
      if (Subtarget.hasVectorEnhancements1())
        setOperationAction(ISD::CTPOP, VT, Legal);
      else
        setOperationAction(ISD::CTPOP, VT, Custom);
      setOperationAction(ISD::CTTZ, VT, Legal);
      setOperationAction(ISD::CTLZ, VT, Legal);
 
      // Convert a GPR scalar to a vector by inserting it into element 0.
      setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
 
      // Use a series of unpacks for extensions.
      setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom);
      setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom);
 
      // Detect shifts/rotates by a scalar amount and convert them into
      // V*_BY_SCALAR.
      setOperationAction(ISD::SHL, VT, Custom);
      setOperationAction(ISD::SRA, VT, Custom);
      setOperationAction(ISD::SRL, VT, Custom);
      setOperationAction(ISD::ROTL, VT, Custom);
 
      // Add ISD::VECREDUCE_ADD as custom in order to implement
      // it with VZERO+VSUM
      setOperationAction(ISD::VECREDUCE_ADD, VT, Custom);
 
      // Map SETCCs onto one of VCE, VCH or VCHL, swapping the operands
      // and inverting the result as necessary.
      setOperationAction(ISD::SETCC, VT, Custom);
 
      setOperationAction({ISD::SMIN, ISD::UMIN, ISD::SMAX, ISD::UMAX}, VT,
                         Legal);
    }
  }
 
  if (Subtarget.hasVector()) {
    // There should be no need to check for float types other than v2f64
    // since <2 x f32> isn't a legal type.
    setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
    setOperationAction(ISD::FP_TO_SINT, MVT::v2f64, Legal);
    setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
    setOperationAction(ISD::FP_TO_UINT, MVT::v2f64, Legal);
    setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
    setOperationAction(ISD::SINT_TO_FP, MVT::v2f64, Legal);
    setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
    setOperationAction(ISD::UINT_TO_FP, MVT::v2f64, Legal);
 
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2i64, Legal);
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2i64, Legal);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i64, Legal);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i64, Legal);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2f64, Legal);
  }
 
  if (Subtarget.hasVectorEnhancements2()) {
    setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
    setOperationAction(ISD::FP_TO_SINT, MVT::v4f32, Legal);
    setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
    setOperationAction(ISD::FP_TO_UINT, MVT::v4f32, Legal);
    setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
    setOperationAction(ISD::SINT_TO_FP, MVT::v4f32, Legal);
    setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
    setOperationAction(ISD::UINT_TO_FP, MVT::v4f32, Legal);
 
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v4i32, Legal);
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v4i32, Legal);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i32, Legal);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i32, Legal);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4f32, Legal);
  }
 
  // Handle floating-point types.
  if (!useSoftFloat()) {
    // Promote all f16 operations to float, with some exceptions below.
    for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
      setOperationAction(Opc, MVT::f16, Promote);
    setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
    for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
      setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
      setTruncStoreAction(VT, MVT::f16, Expand);
    }
    for (auto Op : {ISD::LOAD, ISD::ATOMIC_LOAD, ISD::STORE, ISD::ATOMIC_STORE})
      setOperationAction(Op, MVT::f16, Subtarget.hasVector() ? Legal : Custom);
    setOperationAction(ISD::FP_ROUND, MVT::f16, LibCall);
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, LibCall);
    setOperationAction(ISD::BITCAST, MVT::i16, Custom);
 
    for (auto Op : {ISD::FNEG, ISD::FABS, ISD::FCOPYSIGN})
      setOperationAction(Op, MVT::f16, Legal);
  }
 
  for (unsigned I = MVT::FIRST_FP_VALUETYPE;
       I <= MVT::LAST_FP_VALUETYPE;
       ++I) {
    MVT VT = MVT::SimpleValueType(I);
    if (isTypeLegal(VT) && VT != MVT::f16) {
      // We can use FI for FRINT.
      setOperationAction(ISD::FRINT, VT, Legal);
 
      // We can use the extended form of FI for other rounding operations.
      if (Subtarget.hasFPExtension()) {
        setOperationAction(ISD::FNEARBYINT, VT, Legal);
        setOperationAction(ISD::FFLOOR, VT, Legal);
        setOperationAction(ISD::FCEIL, VT, Legal);
        setOperationAction(ISD::FTRUNC, VT, Legal);
        setOperationAction(ISD::FROUND, VT, Legal);
        setOperationAction(ISD::FROUNDEVEN, VT, Legal);
      }
 
      // No special instructions for these.
      setOperationAction(ISD::FSIN, VT, Expand);
      setOperationAction(ISD::FCOS, VT, Expand);
      setOperationAction(ISD::FSINCOS, VT, Expand);
      setOperationAction(ISD::FREM, VT, LibCall);
      setOperationAction(ISD::FPOW, VT, Expand);
 
      // Special treatment.
      setOperationAction(ISD::IS_FPCLASS, VT, Custom);
 
      // Handle constrained floating-point operations.
      setOperationAction(ISD::STRICT_FADD, VT, Legal);
      setOperationAction(ISD::STRICT_FSUB, VT, Legal);
      setOperationAction(ISD::STRICT_FMUL, VT, Legal);
      setOperationAction(ISD::STRICT_FDIV, VT, Legal);
      setOperationAction(ISD::STRICT_FMA, VT, Legal);
      setOperationAction(ISD::STRICT_FSQRT, VT, Legal);
      setOperationAction(ISD::STRICT_FRINT, VT, Legal);
      setOperationAction(ISD::STRICT_FP_ROUND, VT, Legal);
      if (Subtarget.hasFPExtension()) {
        setOperationAction(ISD::STRICT_FNEARBYINT, VT, Legal);
        setOperationAction(ISD::STRICT_FFLOOR, VT, Legal);
        setOperationAction(ISD::STRICT_FCEIL, VT, Legal);
        setOperationAction(ISD::STRICT_FTRUNC, VT, Legal);
        setOperationAction(ISD::STRICT_FROUND, VT, Legal);
        setOperationAction(ISD::STRICT_FROUNDEVEN, VT, Legal);
      }
 
      // Extension from f16 needs libcall.
      setOperationAction(ISD::FP_EXTEND, VT, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND, VT, Custom);
    }
  }
 
  // Handle floating-point vector types.
  if (Subtarget.hasVector()) {
    // Scalar-to-vector conversion is just a subreg.
    setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8f16, Legal);
    setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal);
    setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal);
 
    // Some insertions and extractions can be done directly but others
    // need to go via integers.
    setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f16, Custom);
    setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
    setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f16, Custom);
    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
 
    // These operations have direct equivalents.
    setOperationAction(ISD::FADD, MVT::v2f64, Legal);
    setOperationAction(ISD::FNEG, MVT::v2f64, Legal);
    setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
    setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
    setOperationAction(ISD::FMA, MVT::v2f64, Legal);
    setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
    setOperationAction(ISD::FABS, MVT::v2f64, Legal);
    setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
    setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
    setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
    setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
    setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
    setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
    setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
    setOperationAction(ISD::FROUNDEVEN, MVT::v2f64, Legal);
 
    // Handle constrained floating-point operations.
    setOperationAction(ISD::STRICT_FADD, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FSUB, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FMUL, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FMA, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FDIV, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FSQRT, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FRINT, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FNEARBYINT, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FFLOOR, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FCEIL, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FTRUNC, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FROUND, MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FROUNDEVEN, MVT::v2f64, Legal);
 
    setOperationAction(ISD::SETCC, MVT::v2f64, Custom);
    setOperationAction(ISD::SETCC, MVT::v4f32, Custom);
    setOperationAction(ISD::STRICT_FSETCC, MVT::v2f64, Custom);
    setOperationAction(ISD::STRICT_FSETCC, MVT::v4f32, Custom);
    if (Subtarget.hasVectorEnhancements1()) {
      setOperationAction(ISD::STRICT_FSETCCS, MVT::v2f64, Custom);
      setOperationAction(ISD::STRICT_FSETCCS, MVT::v4f32, Custom);
    }
  }
 
  // The vector enhancements facility 1 has instructions for these.
  if (Subtarget.hasVectorEnhancements1()) {
    setOperationAction(ISD::FADD, MVT::v4f32, Legal);
    setOperationAction(ISD::FNEG, MVT::v4f32, Legal);
    setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
    setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
    setOperationAction(ISD::FMA, MVT::v4f32, Legal);
    setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
    setOperationAction(ISD::FABS, MVT::v4f32, Legal);
    setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
    setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
    setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
    setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
    setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
    setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
    setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
    setOperationAction(ISD::FROUNDEVEN, MVT::v4f32, Legal);
 
    for (MVT Type : {MVT::f64, MVT::v2f64, MVT::f32, MVT::v4f32, MVT::f128}) {
      setOperationAction(ISD::FMAXNUM, Type, Legal);
      setOperationAction(ISD::FMAXIMUM, Type, Legal);
      setOperationAction(ISD::FMAXIMUMNUM, Type, Legal);
      setOperationAction(ISD::PSEUDO_FMAX, Type, Legal);
      setOperationAction(ISD::FMINNUM, Type, Legal);
      setOperationAction(ISD::FMINIMUM, Type, Legal);
      setOperationAction(ISD::FMINIMUMNUM, Type, Legal);
      setOperationAction(ISD::PSEUDO_FMIN, Type, Legal);
    }
 
    // Handle constrained floating-point operations.
    setOperationAction(ISD::STRICT_FADD, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FSUB, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FMUL, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FMA, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FDIV, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FSQRT, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FRINT, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FNEARBYINT, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FFLOOR, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FCEIL, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FTRUNC, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FROUND, MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FROUNDEVEN, MVT::v4f32, Legal);
    for (auto VT : { MVT::f32, MVT::f64, MVT::f128,
                     MVT::v4f32, MVT::v2f64 }) {
      setOperationAction(ISD::STRICT_FMAXNUM, VT, Legal);
      setOperationAction(ISD::STRICT_FMINNUM, VT, Legal);
      setOperationAction(ISD::STRICT_FMAXIMUM, VT, Legal);
      setOperationAction(ISD::STRICT_FMINIMUM, VT, Legal);
      setOperationAction(ISD::STRICT_PSEUDO_FMAX, VT, Legal);
      setOperationAction(ISD::STRICT_PSEUDO_FMIN, VT, Legal);
    }
  }
 
  // We only have fused f128 multiply-addition on vector registers.
  if (!Subtarget.hasVectorEnhancements1()) {
    setOperationAction(ISD::FMA, MVT::f128, Expand);
    setOperationAction(ISD::STRICT_FMA, MVT::f128, Expand);
  }
 
  // We don't have a copysign instruction on vector registers.
  if (Subtarget.hasVectorEnhancements1())
    setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
 
  // Needed so that we don't try to implement f128 constant loads using
  // a load-and-extend of a f80 constant (in cases where the constant
  // would fit in an f80).
  for (MVT VT : MVT::fp_valuetypes())
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
 
  // We don't have extending load instruction on vector registers.
  if (Subtarget.hasVectorEnhancements1()) {
    setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f32, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f64, Expand);
  }
 
  // Floating-point truncation and stores need to be done separately.
  setTruncStoreAction(MVT::f64,  MVT::f32, Expand);
  setTruncStoreAction(MVT::f128, MVT::f32, Expand);
  setTruncStoreAction(MVT::f128, MVT::f64, Expand);
 
  // We have 64-bit FPR<->GPR moves, but need special handling for
  // 32-bit forms.
  if (!Subtarget.hasVector()) {
    setOperationAction(ISD::BITCAST, MVT::i32, Custom);
    setOperationAction(ISD::BITCAST, MVT::f32, Custom);
  }
 
  // VASTART and VACOPY need to deal with the SystemZ-specific varargs
  // structure, but VAEND is a no-op.
  setOperationAction(ISD::VASTART, MVT::Other, Custom);
  setOperationAction(ISD::VACOPY,  MVT::Other, Custom);
  setOperationAction(ISD::VAEND,   MVT::Other, Expand);
 
  if (Subtarget.isTargetzOS()) {
    // Handle address space casts between mixed sized pointers.
    setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom);
    setOperationAction(ISD::ADDRSPACECAST, MVT::i64, Custom);
  }
 
  setOperationAction(ISD::GET_ROUNDING, MVT::i32, Custom);
 
  // Codes for which we want to perform some z-specific combinations.
  setTargetDAGCombine({ISD::ZERO_EXTEND,
                       ISD::SIGN_EXTEND,
                       ISD::SIGN_EXTEND_INREG,
                       ISD::LOAD,
                       ISD::STORE,
                       ISD::VECTOR_SHUFFLE,
                       ISD::EXTRACT_VECTOR_ELT,
                       ISD::FP_ROUND,
                       ISD::STRICT_FP_ROUND,
                       ISD::FP_EXTEND,
                       ISD::SINT_TO_FP,
                       ISD::UINT_TO_FP,
                       ISD::STRICT_FP_EXTEND,
                       ISD::FCOPYSIGN,
                       ISD::BSWAP,
                       ISD::SETCC,
                       ISD::SRL,
                       ISD::SRA,
                       ISD::MUL,
                       ISD::SDIV,
                       ISD::UDIV,
                       ISD::SREM,
                       ISD::UREM,
                       ISD::INTRINSIC_VOID,
                       ISD::INTRINSIC_W_CHAIN});
 
  // Handle intrinsics.
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
 
  // We're not using SJLJ for exception handling, but they're implemented
  // solely to support use of __builtin_setjmp / __builtin_longjmp.
  setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
  setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
 
  // We want to use MVC in preference to even a single load/store pair.
  MaxStoresPerMemcpy = Subtarget.hasVector() ? 2 : 0;
  MaxStoresPerMemcpyOptSize = 0;
 
  // Same with memmove.
  MaxStoresPerMemmove = Subtarget.hasVector() ? 2 : 0;
  MaxStoresPerMemmoveOptSize = 0;
 
  // The main memset sequence is a byte store followed by an MVC.
  // Two STC or MV..I stores win over that, but the kind of fused stores
  // generated by target-independent code don't when the byte value is
  // variable.  E.g.  "STC <reg>;MHI <reg>,257;STH <reg>" is not better
  // than "STC;MVC".  Handle the choice in target-specific code instead.
  MaxStoresPerMemset = Subtarget.hasVector() ? 2 : 0;
  MaxStoresPerMemsetOptSize = 0;
 
  // Default to having -disable-strictnode-mutation on
  IsStrictFPEnabled = true;
}
 
bool SystemZTargetLowering::useSoftFloat() const {
  return Subtarget.hasSoftFloat();
}
 
unsigned SystemZTargetLowering::getVectorTypeBreakdownForCallingConv(
    LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
    unsigned &NumIntermediates, MVT &RegisterVT) const {
  // Pass fp16 vectors in VR(s).
  if (Subtarget.hasVector() && VT.isVectorOf(MVT::f16)) {
    IntermediateVT = RegisterVT = MVT::v8f16;
    return NumIntermediates =
               divideCeil(VT.getVectorNumElements(), SystemZ::VectorBytes / 2);
  }
  return TargetLowering::getVectorTypeBreakdownForCallingConv(
      Context, CC, VT, IntermediateVT, NumIntermediates, RegisterVT);
}
 
MVT SystemZTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
                                                         CallingConv::ID CC,
                                                         EVT VT) const {
  // 128-bit single-element vector types are passed like other vectors,
  // not like their element type.
  if (Subtarget.hasVector() && VT.isVector() && VT.getSizeInBits() == 128 &&
      VT.getVectorNumElements() == 1)
    return MVT::v16i8;
  // Pass fp16 vectors in VR(s).
  if (Subtarget.hasVector() && VT.isVectorOf(MVT::f16))
    return MVT::v8f16;
  return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
}
 
unsigned SystemZTargetLowering::getNumRegistersForCallingConv(
    LLVMContext &Context, CallingConv::ID CC, EVT VT) const {
  // Pass fp16 vectors in VR(s).
  if (Subtarget.hasVector() && VT.isVectorOf(MVT::f16))
    return divideCeil(VT.getVectorNumElements(), SystemZ::VectorBytes / 2);
  return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
}
 
EVT SystemZTargetLowering::getSetCCResultType(const DataLayout &DL,
                                              LLVMContext &, EVT VT) const {
  if (!VT.isVector())
    return MVT::i32;
  return VT.changeVectorElementTypeToInteger();
}
 
bool SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(
    const MachineFunction &MF, EVT VT) const {
  if (useSoftFloat())
    return false;
 
  VT = VT.getScalarType();
 
  if (!VT.isSimple())
    return false;
 
  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::f32:
  case MVT::f64:
    return true;
  case MVT::f128:
    return Subtarget.hasVectorEnhancements1();
  default:
    break;
  }
 
  return false;
}
 
// Return true if the constant can be generated with a vector instruction,
// such as VGM, VGMB or VREPI.
bool SystemZVectorConstantInfo::isVectorConstantLegal(
    const SystemZSubtarget &Subtarget) {
  const SystemZInstrInfo *TII = Subtarget.getInstrInfo();
  if (!Subtarget.hasVector() ||
      (isFP128 && !Subtarget.hasVectorEnhancements1()))
    return false;
 
  // Try using VECTOR GENERATE BYTE MASK.  This is the architecturally-
  // preferred way of creating all-zero and all-one vectors so give it
  // priority over other methods below.
  unsigned Mask = 0;
  unsigned I = 0;
  for (; I < SystemZ::VectorBytes; ++I) {
    uint64_t Byte = IntBits.lshr(I * 8).trunc(8).getZExtValue();
    if (Byte == 0xff)
      Mask |= 1ULL << I;
    else if (Byte != 0)
      break;
  }
  if (I == SystemZ::VectorBytes) {
    Opcode = SystemZISD::BYTE_MASK;
    OpVals.push_back(Mask);
    VecVT = MVT::getVectorVT(MVT::getIntegerVT(8), 16);
    return true;
  }
 
  if (SplatBitSize > 64)
    return false;
 
  auto TryValue = [&](uint64_t Value) -> bool {
    // Try VECTOR REPLICATE IMMEDIATE
    int64_t SignedValue = SignExtend64(Value, SplatBitSize);
    if (isInt<16>(SignedValue)) {
      OpVals.push_back(((unsigned) SignedValue));
      Opcode = SystemZISD::REPLICATE;
      VecVT = MVT::getVectorVT(MVT::getIntegerVT(SplatBitSize),
                               SystemZ::VectorBits / SplatBitSize);
      return true;
    }
    // Try VECTOR GENERATE MASK
    unsigned Start, End;
    if (TII->isRxSBGMask(Value, SplatBitSize, Start, End)) {
      // isRxSBGMask returns the bit numbers for a full 64-bit value, with 0
      // denoting 1 << 63 and 63 denoting 1.  Convert them to bit numbers for
      // an SplatBitSize value, so that 0 denotes 1 << (SplatBitSize-1).
      OpVals.push_back(Start - (64 - SplatBitSize));
      OpVals.push_back(End - (64 - SplatBitSize));
      Opcode = SystemZISD::ROTATE_MASK;
      VecVT = MVT::getVectorVT(MVT::getIntegerVT(SplatBitSize),
                               SystemZ::VectorBits / SplatBitSize);
      return true;
    }
    return false;
  };
 
  // First try assuming that any undefined bits above the highest set bit
  // and below the lowest set bit are 1s.  This increases the likelihood of
  // being able to use a sign-extended element value in VECTOR REPLICATE
  // IMMEDIATE or a wraparound mask in VECTOR GENERATE MASK.
  uint64_t SplatBitsZ = SplatBits.getZExtValue();
  uint64_t SplatUndefZ = SplatUndef.getZExtValue();
  unsigned LowerBits = llvm::countr_zero(SplatBitsZ);
  unsigned UpperBits = llvm::countl_zero(SplatBitsZ);
  uint64_t Lower = SplatUndefZ & maskTrailingOnes<uint64_t>(LowerBits);
  uint64_t Upper = SplatUndefZ & maskLeadingOnes<uint64_t>(UpperBits);
  if (TryValue(SplatBitsZ | Upper | Lower))
    return true;
 
  // Now try assuming that any undefined bits between the first and
  // last defined set bits are set.  This increases the chances of
  // using a non-wraparound mask.
  uint64_t Middle = SplatUndefZ & ~Upper & ~Lower;
  return TryValue(SplatBitsZ | Middle);
}
 
SystemZVectorConstantInfo::SystemZVectorConstantInfo(APInt IntImm) {
  if (IntImm.isSingleWord()) {
    IntBits = APInt(128, IntImm.getZExtValue());
    IntBits <<= (SystemZ::VectorBits - IntImm.getBitWidth());
  } else
    IntBits = IntImm;
  assert(IntBits.getBitWidth() == 128 && "Unsupported APInt.");
 
  // Find the smallest splat.
  SplatBits = IntImm;
  unsigned Width = SplatBits.getBitWidth();
  while (Width > 8) {
    unsigned HalfSize = Width / 2;
    APInt HighValue = SplatBits.lshr(HalfSize).trunc(HalfSize);
    APInt LowValue = SplatBits.trunc(HalfSize);
 
    // If the two halves do not match, stop here.
    if (HighValue != LowValue || 8 > HalfSize)
      break;
 
    SplatBits = HighValue;
    Width = HalfSize;
  }
  SplatUndef = 0;
  SplatBitSize = Width;
}
 
SystemZVectorConstantInfo::SystemZVectorConstantInfo(BuildVectorSDNode *BVN) {
  assert(BVN->isConstant() && "Expected a constant BUILD_VECTOR");
  bool HasAnyUndefs;
 
  // Get IntBits by finding the 128 bit splat.
  BVN->isConstantSplat(IntBits, SplatUndef, SplatBitSize, HasAnyUndefs, 128,
                       true);
 
  // Get SplatBits by finding the 8 bit or greater splat.
  BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs, 8,
                       true);
}
 
bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
                                         bool ForCodeSize) const {
  // We can load zero using LZ?R and negative zero using LZ?R;LC?BR.
  if (Imm.isZero() || Imm.isNegZero())
    return true;
 
  return SystemZVectorConstantInfo(Imm).isVectorConstantLegal(Subtarget);
}
 
MachineBasicBlock *
SystemZTargetLowering::emitEHSjLjSetJmp(MachineInstr &MI,
                                        MachineBasicBlock *MBB) const {
  DebugLoc DL = MI.getDebugLoc();
  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
  const SystemZRegisterInfo *TRI = Subtarget.getRegisterInfo();
 
  MachineFunction *MF = MBB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
 
  const BasicBlock *BB = MBB->getBasicBlock();
  MachineFunction::iterator I = ++MBB->getIterator();
 
  Register DstReg = MI.getOperand(0).getReg();
  const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
  assert(TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!");
  (void)TRI;
  Register MainDstReg = MRI.createVirtualRegister(RC);
  Register RestoreDstReg = MRI.createVirtualRegister(RC);
 
  MVT PVT = getPointerTy(MF->getDataLayout());
  assert((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!");
  // For v = setjmp(buf), we generate.
  // Algorithm:
  //
  //                 ---------
  //                | thisMBB |
  //                 ---------
  //                     |
  //         ------------------------
  //        |                        |
  //     ----------           ---------------
  //    |  mainMBB |         | restoreMBB    |
  //    |   v = 0  |         |  v = 1        |
  //     ----------           ---------------
  //        |                        |
  //         -------------------------
  //                     |
  //           -----------------------------
  //          |       sinkMBB               |
  //          | phi(v_mainMBB,v_restoreMBB) |
  //           -----------------------------
  // thisMBB:
  //  buf[FPOffset] = Frame Pointer if hasFP.
  //  buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB.
  //  buf[BCOffset] = Backchain value if building with -mbackchain.
  //  buf[SPOffset] = Stack Pointer.
  //  buf[LPOffset] = We never write this slot with  R13, gcc stores R13 always.
  //  SjLjSetup restoreMBB
  // mainMBB:
  //  v_main = 0
  // sinkMBB:
  //  v = phi(v_main, v_restore)
  // restoreMBB:
  //  v_restore = 1
 
  MachineBasicBlock *ThisMBB = MBB;
  MachineBasicBlock *MainMBB = MF->CreateMachineBasicBlock(BB);
  MachineBasicBlock *SinkMBB = MF->CreateMachineBasicBlock(BB);
  MachineBasicBlock *RestoreMBB = MF->CreateMachineBasicBlock(BB);
 
  MF->insert(I, MainMBB);
  MF->insert(I, SinkMBB);
  MF->push_back(RestoreMBB);
  RestoreMBB->setMachineBlockAddressTaken();
 
  MachineInstrBuilder MIB;
 
  // Transfer the remainder of BB and its successor edges to sinkMBB.
  SinkMBB->splice(SinkMBB->begin(), MBB,
                  std::next(MachineBasicBlock::iterator(MI)), MBB->end());
  SinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
 
  // thisMBB:
  const int64_t FPOffset = 0;                         // Slot 1.
  const int64_t LabelOffset = 1 * PVT.getStoreSize(); // Slot 2.
  const int64_t BCOffset = 2 * PVT.getStoreSize();    // Slot 3.
  const int64_t SPOffset = 3 * PVT.getStoreSize();    // Slot 4.
 
  // Buf address.
  Register BufReg = MI.getOperand(1).getReg();
 
  const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
  Register LabelReg = MRI.createVirtualRegister(PtrRC);
 
  // Prepare IP for longjmp.
  BuildMI(*ThisMBB, MI, DL, TII->get(SystemZ::LARL), LabelReg)
      .addMBB(RestoreMBB);
  // Store IP for return from jmp, slot 2, offset = 1.
  BuildMI(*ThisMBB, MI, DL, TII->get(SystemZ::STG))
      .addReg(LabelReg)
      .addReg(BufReg)
      .addImm(LabelOffset)
      .addReg(0);
 
  auto *SpecialRegs = Subtarget.getSpecialRegisters();
  bool HasFP = Subtarget.getFrameLowering()->hasFP(*MF);
  if (HasFP) {
    BuildMI(*ThisMBB, MI, DL, TII->get(SystemZ::STG))
        .addReg(SpecialRegs->getFramePointerRegister())
        .addReg(BufReg)
        .addImm(FPOffset)
        .addReg(0);
  }
 
  // Store SP.
  BuildMI(*ThisMBB, MI, DL, TII->get(SystemZ::STG))
      .addReg(SpecialRegs->getStackPointerRegister())
      .addReg(BufReg)
      .addImm(SPOffset)
      .addReg(0);
 
  // Slot 3(Offset = 2) Backchain value (if building with -mbackchain).
  bool BackChain = MF->getSubtarget<SystemZSubtarget>().hasBackChain();
  if (BackChain) {
    Register BCReg = MRI.createVirtualRegister(PtrRC);
    auto *TFL = Subtarget.getFrameLowering<SystemZFrameLowering>();
    MIB = BuildMI(*ThisMBB, MI, DL, TII->get(SystemZ::LG), BCReg)
              .addReg(SpecialRegs->getStackPointerRegister())
              .addImm(TFL->getBackchainOffset(*MF))
              .addReg(0);
 
    BuildMI(*ThisMBB, MI, DL, TII->get(SystemZ::STG))
        .addReg(BCReg)
        .addReg(BufReg)
        .addImm(BCOffset)
        .addReg(0);
  }
 
  // Setup.
  MIB = BuildMI(*ThisMBB, MI, DL, TII->get(SystemZ::EH_SjLj_Setup))
            .addMBB(RestoreMBB);
 
  const SystemZRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
  MIB.addRegMask(RegInfo->getNoPreservedMask());
 
  ThisMBB->addSuccessor(MainMBB);
  ThisMBB->addSuccessor(RestoreMBB);
 
  // mainMBB:
  BuildMI(MainMBB, DL, TII->get(SystemZ::LHI), MainDstReg).addImm(0);
  MainMBB->addSuccessor(SinkMBB);
 
  // sinkMBB:
  BuildMI(*SinkMBB, SinkMBB->begin(), DL, TII->get(SystemZ::PHI), DstReg)
      .addReg(MainDstReg)
      .addMBB(MainMBB)
      .addReg(RestoreDstReg)
      .addMBB(RestoreMBB);
 
  // restoreMBB.
  BuildMI(RestoreMBB, DL, TII->get(SystemZ::LHI), RestoreDstReg).addImm(1);
  BuildMI(RestoreMBB, DL, TII->get(SystemZ::J)).addMBB(SinkMBB);
  RestoreMBB->addSuccessor(SinkMBB);
 
  MI.eraseFromParent();
 
  return SinkMBB;
}
 
MachineBasicBlock *
SystemZTargetLowering::emitEHSjLjLongJmp(MachineInstr &MI,
                                         MachineBasicBlock *MBB) const {
 
  DebugLoc DL = MI.getDebugLoc();
  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
 
  MachineFunction *MF = MBB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
 
  MVT PVT = getPointerTy(MF->getDataLayout());
  assert((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!");
  Register BufReg = MI.getOperand(0).getReg();
  const TargetRegisterClass *RC = MRI.getRegClass(BufReg);
  auto *SpecialRegs = Subtarget.getSpecialRegisters();
 
  Register Tmp = MRI.createVirtualRegister(RC);
  Register BCReg = MRI.createVirtualRegister(RC);
 
  MachineInstrBuilder MIB;
 
  const int64_t FPOffset = 0;
  const int64_t LabelOffset = 1 * PVT.getStoreSize();
  const int64_t BCOffset = 2 * PVT.getStoreSize();
  const int64_t SPOffset = 3 * PVT.getStoreSize();
  const int64_t LPOffset = 4 * PVT.getStoreSize();
 
  MIB = BuildMI(*MBB, MI, DL, TII->get(SystemZ::LG), Tmp)
            .addReg(BufReg)
            .addImm(LabelOffset)
            .addReg(0);
 
  MIB = BuildMI(*MBB, MI, DL, TII->get(SystemZ::LG),
                SpecialRegs->getFramePointerRegister())
            .addReg(BufReg)
            .addImm(FPOffset)
            .addReg(0);
 
  // We are restoring R13 even though we never stored in setjmp from llvm,
  // as gcc always stores R13 in builtin_setjmp. We could have mixed code
  // gcc setjmp and llvm longjmp.
  MIB = BuildMI(*MBB, MI, DL, TII->get(SystemZ::LG), SystemZ::R13D)
            .addReg(BufReg)
            .addImm(LPOffset)
            .addReg(0);
 
  bool BackChain = MF->getSubtarget<SystemZSubtarget>().hasBackChain();
  if (BackChain) {
    MIB = BuildMI(*MBB, MI, DL, TII->get(SystemZ::LG), BCReg)
              .addReg(BufReg)
              .addImm(BCOffset)
              .addReg(0);
  }
 
  MIB = BuildMI(*MBB, MI, DL, TII->get(SystemZ::LG),
                SpecialRegs->getStackPointerRegister())
            .addReg(BufReg)
            .addImm(SPOffset)
            .addReg(0);
 
  if (BackChain) {
    auto *TFL = Subtarget.getFrameLowering<SystemZFrameLowering>();
    BuildMI(*MBB, MI, DL, TII->get(SystemZ::STG))
        .addReg(BCReg)
        .addReg(SpecialRegs->getStackPointerRegister())
        .addImm(TFL->getBackchainOffset(*MF))
        .addReg(0);
  }
 
  MIB = BuildMI(*MBB, MI, DL, TII->get(SystemZ::BR)).addReg(Tmp);
 
  MI.eraseFromParent();
  return MBB;
}
 
/// Returns true if stack probing through inline assembly is requested.
bool SystemZTargetLowering::hasInlineStackProbe(const MachineFunction &MF) const {
  // If the function specifically requests inline stack probes, emit them.
  if (MF.getFunction().hasFnAttribute("probe-stack"))
    return MF.getFunction().getFnAttribute("probe-stack").getValueAsString() ==
           "inline-asm";
  return false;
}
 
TargetLowering::AtomicExpansionKind
SystemZTargetLowering::shouldCastAtomicLoadInIR(LoadInst *LI) const {
  return AtomicExpansionKind::None;
}
 
TargetLowering::AtomicExpansionKind
SystemZTargetLowering::shouldCastAtomicStoreInIR(StoreInst *SI) const {
  return AtomicExpansionKind::None;
}
 
TargetLowering::AtomicExpansionKind
SystemZTargetLowering::shouldExpandAtomicRMWInIR(
    const AtomicRMWInst *RMW) const {
  // Don't expand subword operations as they require special treatment.
  if (RMW->getType()->isIntegerTy(8) || RMW->getType()->isIntegerTy(16))
    return AtomicExpansionKind::None;
 
  // Don't expand if there is a target instruction available.
  if (Subtarget.hasInterlockedAccess1() &&
      (RMW->getType()->isIntegerTy(32) || RMW->getType()->isIntegerTy(64)) &&
      (RMW->getOperation() == AtomicRMWInst::BinOp::Add ||
       RMW->getOperation() == AtomicRMWInst::BinOp::Sub ||
       RMW->getOperation() == AtomicRMWInst::BinOp::And ||
       RMW->getOperation() == AtomicRMWInst::BinOp::Or ||
       RMW->getOperation() == AtomicRMWInst::BinOp::Xor))
    return AtomicExpansionKind::None;
 
  return AtomicExpansionKind::CmpXChg;
}
 
bool SystemZTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
  // We can use CGFI or CLGFI.
  return isInt<32>(Imm) || isUInt<32>(Imm);
}
 
bool SystemZTargetLowering::isLegalAddImmediate(int64_t Imm) const {
  // We can use ALGFI or SLGFI.
  return isUInt<32>(Imm) || isUInt<32>(-Imm);
}
 
bool SystemZTargetLowering::allowsMisalignedMemoryAccesses(
    EVT VT, unsigned, Align, MachineMemOperand::Flags, unsigned *Fast) const {
  // Unaligned accesses should never be slower than the expanded version.
  // We check specifically for aligned accesses in the few cases where
  // they are required.
  if (Fast)
    *Fast = 1;
  return true;
}
 
bool SystemZTargetLowering::hasAndNot(SDValue Y) const {
  EVT VT = Y.getValueType();
 
  // We can use NC(G)RK for types in GPRs ...
  if (VT == MVT::i32 || VT == MVT::i64)
    return Subtarget.hasMiscellaneousExtensions3();
 
  // ... or VNC for types in VRs.
  if (VT.isVector() || VT == MVT::i128)
    return Subtarget.hasVector();
 
  return false;
}
 
// Information about the addressing mode for a memory access.
struct AddressingMode {
  // True if a long displacement is supported.
  bool LongDisplacement;
 
  // True if use of index register is supported.
  bool IndexReg;
 
  AddressingMode(bool LongDispl, bool IdxReg) :
    LongDisplacement(LongDispl), IndexReg(IdxReg) {}
};
 
// Return the desired addressing mode for a Load which has only one use (in
// the same block) which is a Store.
static AddressingMode getLoadStoreAddrMode(bool HasVector,
                                          Type *Ty) {
  // With vector support a Load->Store combination may be combined to either
  // an MVC or vector operations and it seems to work best to allow the
  // vector addressing mode.
  if (HasVector)
    return AddressingMode(false/*LongDispl*/, true/*IdxReg*/);
 
  // Otherwise only the MVC case is special.
  bool MVC = Ty->isIntegerTy(8);
  return AddressingMode(!MVC/*LongDispl*/, !MVC/*IdxReg*/);
}
 
// Return the addressing mode which seems most desirable given an LLVM
// Instruction pointer.
static AddressingMode
supportedAddressingMode(Instruction *I, bool HasVector) {
  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
    switch (II->getIntrinsicID()) {
    default: break;
    case Intrinsic::memset:
    case Intrinsic::memmove:
    case Intrinsic::memcpy:
      return AddressingMode(false/*LongDispl*/, false/*IdxReg*/);
    }
  }
 
  if (isa<LoadInst>(I) && I->hasOneUse()) {
    auto *SingleUser = cast<Instruction>(*I->user_begin());
    if (SingleUser->getParent() == I->getParent()) {
      if (isa<ICmpInst>(SingleUser)) {
        if (auto *C = dyn_cast<ConstantInt>(SingleUser->getOperand(1)))
          if (C->getBitWidth() <= 64 &&
              (isInt<16>(C->getSExtValue()) || isUInt<16>(C->getZExtValue())))
            // Comparison of memory with 16 bit signed / unsigned immediate
            return AddressingMode(false/*LongDispl*/, false/*IdxReg*/);
      } else if (isa<StoreInst>(SingleUser))
        // Load->Store
        return getLoadStoreAddrMode(HasVector, I->getType());
    }
  } else if (auto *StoreI = dyn_cast<StoreInst>(I)) {
    if (auto *LoadI = dyn_cast<LoadInst>(StoreI->getValueOperand()))
      if (LoadI->hasOneUse() && LoadI->getParent() == I->getParent())
        // Load->Store
        return getLoadStoreAddrMode(HasVector, LoadI->getType());
  }
 
  if (HasVector && (isa<LoadInst>(I) || isa<StoreInst>(I))) {
 
    // * Use LDE instead of LE/LEY for z13 to avoid partial register
    //   dependencies (LDE only supports small offsets).
    // * Utilize the vector registers to hold floating point
    //   values (vector load / store instructions only support small
    //   offsets).
 
    Type *MemAccessTy = (isa<LoadInst>(I) ? I->getType() :
                         I->getOperand(0)->getType());
    bool IsFPAccess = MemAccessTy->isFloatingPointTy();
    bool IsVectorAccess = MemAccessTy->isVectorTy();
 
    // A store of an extracted vector element will be combined into a VSTE type
    // instruction.
    if (!IsVectorAccess && isa<StoreInst>(I)) {
      Value *DataOp = I->getOperand(0);
      if (isa<ExtractElementInst>(DataOp))
        IsVectorAccess = true;
    }
 
    // A load which gets inserted into a vector element will be combined into a
    // VLE type instruction.
    if (!IsVectorAccess && isa<LoadInst>(I) && I->hasOneUse()) {
      User *LoadUser = *I->user_begin();
      if (isa<InsertElementInst>(LoadUser))
        IsVectorAccess = true;
    }
 
    if (IsFPAccess || IsVectorAccess)
      return AddressingMode(false/*LongDispl*/, true/*IdxReg*/);
  }
 
  return AddressingMode(true/*LongDispl*/, true/*IdxReg*/);
}
 
bool SystemZTargetLowering::isLegalAddressingMode(const DataLayout &DL,
       const AddrMode &AM, Type *Ty, unsigned AS, Instruction *I) const {
  // Punt on globals for now, although they can be used in limited
  // RELATIVE LONG cases.
  if (AM.BaseGV)
    return false;
 
  // Require a 20-bit signed offset.
  if (!isInt<20>(AM.BaseOffs))
    return false;
 
  bool RequireD12 =
      Subtarget.hasVector() && (Ty->isVectorTy() || Ty->isIntegerTy(128));
  AddressingMode SupportedAM(!RequireD12, true);
  if (I != nullptr)
    SupportedAM = supportedAddressingMode(I, Subtarget.hasVector());
 
  if (!SupportedAM.LongDisplacement && !isUInt<12>(AM.BaseOffs))
    return false;
 
  if (!SupportedAM.IndexReg)
    // No indexing allowed.
    return AM.Scale == 0;
  else
    // Indexing is OK but no scale factor can be applied.
    return AM.Scale == 0 || AM.Scale == 1;
}
 
bool SystemZTargetLowering::findOptimalMemOpLowering(
    LLVMContext &Context, std::vector<EVT> &MemOps, unsigned Limit,
    const MemOp &Op, unsigned DstAS, unsigned SrcAS,
    const AttributeList &FuncAttributes, EVT *LargestVT) const {
 
  assert(Limit != ~0U &&
         "Expected EmitTargetCodeForMemXXX() to handle AlwaysInline cases.");
 
  if (Op.isZeroMemset())
    return false; // Memset zero: Use XC.
 
  const int MVCFastLen = 16;
  // Use MVC up to 16 bytes for memcpy.  Small memset uses STC/MVI for first
  // byte.
  if (Op.isMemcpy() && Op.size() <= MVCFastLen)
    return false;
  if (Op.isMemset() && Op.size() - 1 <= MVCFastLen)
    return false;
 
  // Avoid unaligned VL/VST:s.
  if ((Op.size() >= 16 && !Op.isAligned(Align(8))) ||
      (Op.size() >= 25 && Op.size() <= 31))
    return false;
 
  return TargetLowering::findOptimalMemOpLowering(
      Context, MemOps, Limit, Op, DstAS, SrcAS, FuncAttributes, LargestVT);
}
 
EVT SystemZTargetLowering::getOptimalMemOpType(
    LLVMContext &Context, const MemOp &Op,
    const AttributeList &FuncAttributes) const {
  return Subtarget.hasVector() ? MVT::v2i64 : MVT::Other;
}
 
bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const {
  if (!FromType->isIntegerTy() || !ToType->isIntegerTy())
    return false;
  unsigned FromBits = FromType->getPrimitiveSizeInBits().getFixedValue();
  unsigned ToBits = ToType->getPrimitiveSizeInBits().getFixedValue();
  return FromBits > ToBits;
}
 
bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const {
  if (!FromVT.isInteger() || !ToVT.isInteger())
    return false;
  unsigned FromBits = FromVT.getFixedSizeInBits();
  unsigned ToBits = ToVT.getFixedSizeInBits();
  return FromBits > ToBits;
}
 
//===----------------------------------------------------------------------===//
// Inline asm support
//===----------------------------------------------------------------------===//
 
TargetLowering::ConstraintType
SystemZTargetLowering::getConstraintType(StringRef Constraint) const {
  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    case 'a': // Address register
    case 'd': // Data register (equivalent to 'r')
    case 'f': // Floating-point register
    case 'h': // High-part register
    case 'r': // General-purpose register
    case 'v': // Vector register
      return C_RegisterClass;
 
    case 'Q': // Memory with base and unsigned 12-bit displacement
    case 'R': // Likewise, plus an index
    case 'S': // Memory with base and signed 20-bit displacement
    case 'T': // Likewise, plus an index
    case 'm': // Equivalent to 'T'.
      return C_Memory;
 
    case 'I': // Unsigned 8-bit constant
    case 'J': // Unsigned 12-bit constant
    case 'K': // Signed 16-bit constant
    case 'L': // Signed 20-bit displacement (on all targets we support)
    case 'M': // 0x7fffffff
      return C_Immediate;
 
    default:
      break;
    }
  } else if (Constraint.size() == 2 && Constraint[0] == 'Z') {
    switch (Constraint[1]) {
    case 'Q': // Address with base and unsigned 12-bit displacement
    case 'R': // Likewise, plus an index
    case 'S': // Address with base and signed 20-bit displacement
    case 'T': // Likewise, plus an index
      return C_Address;
 
    default:
      break;
    }
  } else if (Constraint.size() == 5 && Constraint.starts_with("{")) {
    if (StringRef("{@cc}").compare(Constraint) == 0)
      return C_Other;
  }
  return TargetLowering::getConstraintType(Constraint);
}
 
TargetLowering::ConstraintWeight
SystemZTargetLowering::getSingleConstraintMatchWeight(
    AsmOperandInfo &Info, const char *Constraint) const {
  ConstraintWeight Weight = CW_Invalid;
  Value *CallOperandVal = Info.CallOperandVal;
  // If we don't have a value, we can't do a match,
  // but allow it at the lowest weight.
  if (!CallOperandVal)
    return CW_Default;
  Type *type = CallOperandVal->getType();
  // Look at the constraint type.
  switch (*Constraint) {
  default:
    Weight = TargetLowering::getSingleConstraintMatchWeight(Info, Constraint);
    break;
 
  case 'a': // Address register
  case 'd': // Data register (equivalent to 'r')
  case 'h': // High-part register
  case 'r': // General-purpose register
    Weight =
        CallOperandVal->getType()->isIntegerTy() ? CW_Register : CW_Default;
    break;
 
  case 'f': // Floating-point register
    if (!useSoftFloat())
      Weight = type->isFloatingPointTy() ? CW_Register : CW_Default;
    break;
 
  case 'v': // Vector register
    if (Subtarget.hasVector())
      Weight = (type->isVectorTy() || type->isFloatingPointTy()) ? CW_Register
                                                                 : CW_Default;
    break;
 
  case 'I': // Unsigned 8-bit constant
    if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
      if (isUInt<8>(C->getZExtValue()))
        Weight = CW_Constant;
    break;
 
  case 'J': // Unsigned 12-bit constant
    if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
      if (isUInt<12>(C->getZExtValue()))
        Weight = CW_Constant;
    break;
 
  case 'K': // Signed 16-bit constant
    if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
      if (isInt<16>(C->getSExtValue()))
        Weight = CW_Constant;
    break;
 
  case 'L': // Signed 20-bit displacement (on all targets we support)
    if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
      if (isInt<20>(C->getSExtValue()))
        Weight = CW_Constant;
    break;
 
  case 'M': // 0x7fffffff
    if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
      if (C->getZExtValue() == 0x7fffffff)
        Weight = CW_Constant;
    break;
  }
  return Weight;
}
 
// Parse a "{tNNN}" register constraint for which the register type "t"
// has already been verified.  MC is the class associated with "t" and
// Map maps 0-based register numbers to LLVM register numbers.
static std::pair<unsigned, const TargetRegisterClass *>
parseRegisterNumber(StringRef Constraint, const TargetRegisterClass *RC,
                    const unsigned *Map, unsigned Size) {
  assert(*(Constraint.end()-1) == '}' && "Missing '}'");
  if (isdigit(Constraint[2])) {
    unsigned Index;
    bool Failed =
        Constraint.slice(2, Constraint.size() - 1).getAsInteger(10, Index);
    if (!Failed && Index < Size && Map[Index])
      return std::make_pair(Map[Index], RC);
  }
  return std::make_pair(0U, nullptr);
}
 
std::pair<unsigned, const TargetRegisterClass *>
SystemZTargetLowering::getRegForInlineAsmConstraint(
    const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
  if (Constraint.size() == 1) {
    // GCC Constraint Letters
    switch (Constraint[0]) {
    default: break;
    case 'd': // Data register (equivalent to 'r')
    case 'r': // General-purpose register
      if (VT.getSizeInBits() == 64)
        return std::make_pair(0U, &SystemZ::GR64BitRegClass);
      else if (VT.getSizeInBits() == 128)
        return std::make_pair(0U, &SystemZ::GR128BitRegClass);
      return std::make_pair(0U, &SystemZ::GR32BitRegClass);
 
    case 'a': // Address register
      if (VT == MVT::i64)
        return std::make_pair(0U, &SystemZ::ADDR64BitRegClass);
      else if (VT == MVT::i128)
        return std::make_pair(0U, &SystemZ::ADDR128BitRegClass);
      return std::make_pair(0U, &SystemZ::ADDR32BitRegClass);
 
    case 'h': // High-part register (an LLVM extension)
      return std::make_pair(0U, &SystemZ::GRH32BitRegClass);
 
    case 'f': // Floating-point register
      if (!useSoftFloat()) {
        if (VT.getSizeInBits() == 16)
          return std::make_pair(0U, &SystemZ::FP16BitRegClass);
        else if (VT.getSizeInBits() == 64)
          return std::make_pair(0U, &SystemZ::FP64BitRegClass);
        else if (VT.getSizeInBits() == 128)
          return std::make_pair(0U, &SystemZ::FP128BitRegClass);
        return std::make_pair(0U, &SystemZ::FP32BitRegClass);
      }
      break;
 
    case 'v': // Vector register
      if (Subtarget.hasVector()) {
        if (VT.getSizeInBits() == 16)
          return std::make_pair(0U, &SystemZ::VR16BitRegClass);
        if (VT.getSizeInBits() == 32)
          return std::make_pair(0U, &SystemZ::VR32BitRegClass);
        if (VT.getSizeInBits() == 64)
          return std::make_pair(0U, &SystemZ::VR64BitRegClass);
        return std::make_pair(0U, &SystemZ::VR128BitRegClass);
      }
      break;
    }
  }
  if (Constraint.starts_with("{")) {
 
    // A clobber constraint (e.g. ~{f0}) will have MVT::Other which is illegal
    // to check the size on.
    auto getVTSizeInBits = [&VT]() {
      return VT == MVT::Other ? 0 : VT.getSizeInBits();
    };
 
    // We need to override the default register parsing for GPRs and FPRs
    // because the interpretation depends on VT.  The internal names of
    // the registers are also different from the external names
    // (F0D and F0S instead of F0, etc.).
    if (Constraint[1] == 'r') {
      if (getVTSizeInBits() == 32)
        return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass,
                                   SystemZMC::GR32Regs, 16);
      if (getVTSizeInBits() == 128)
        return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass,
                                   SystemZMC::GR128Regs, 16);
      return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass,
                                 SystemZMC::GR64Regs, 16);
    }
    if (Constraint[1] == 'f') {
      if (useSoftFloat())
        return std::make_pair(
            0u, static_cast<const TargetRegisterClass *>(nullptr));
      if (getVTSizeInBits() == 16)
        return parseRegisterNumber(Constraint, &SystemZ::FP16BitRegClass,
                                   SystemZMC::FP16Regs, 16);
      if (getVTSizeInBits() == 32)
        return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass,
                                   SystemZMC::FP32Regs, 16);
      if (getVTSizeInBits() == 128)
        return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass,
                                   SystemZMC::FP128Regs, 16);
      return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass,
                                 SystemZMC::FP64Regs, 16);
    }
    if (Constraint[1] == 'v') {
      if (!Subtarget.hasVector())
        return std::make_pair(
            0u, static_cast<const TargetRegisterClass *>(nullptr));
      if (getVTSizeInBits() == 16)
        return parseRegisterNumber(Constraint, &SystemZ::VR16BitRegClass,
                                   SystemZMC::VR16Regs, 32);
      if (getVTSizeInBits() == 32)
        return parseRegisterNumber(Constraint, &SystemZ::VR32BitRegClass,
                                   SystemZMC::VR32Regs, 32);
      if (getVTSizeInBits() == 64)
        return parseRegisterNumber(Constraint, &SystemZ::VR64BitRegClass,
                                   SystemZMC::VR64Regs, 32);
      return parseRegisterNumber(Constraint, &SystemZ::VR128BitRegClass,
                                 SystemZMC::VR128Regs, 32);
    }
    if (Constraint[1] == '@') {
      if (StringRef("{@cc}").compare(Constraint) == 0)
        return std::make_pair(SystemZ::CC, &SystemZ::CCRRegClass);
    }
  }
  return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
}
 
// FIXME? Maybe this could be a TableGen attribute on some registers and
// this table could be generated automatically from RegInfo.
Register
SystemZTargetLowering::getRegisterByName(const char *RegName, LLT VT,
                                         const MachineFunction &MF) const {
  Register Reg =
      StringSwitch<Register>(RegName)
          .Case("r4", Subtarget.isTargetXPLINK64() ? SystemZ::R4D
                                                   : SystemZ::NoRegister)
          .Case("r15",
                Subtarget.isTargetELF() ? SystemZ::R15D : SystemZ::NoRegister)
          .Default(Register());
 
  return Reg;
}
 
Register SystemZTargetLowering::getExceptionPointerRegister(
    ExceptionHandling EH, const Constant *PersonalityFn) const {
  return Subtarget.isTargetXPLINK64() ? SystemZ::R1D : SystemZ::R6D;
}
 
Register SystemZTargetLowering::getExceptionSelectorRegister(
    ExceptionHandling EH, const Constant *PersonalityFn) const {
  return Subtarget.isTargetXPLINK64() ? SystemZ::R2D : SystemZ::R7D;
}
 
// Convert condition code in CCReg to an i32 value.
static SDValue getCCResult(SelectionDAG &DAG, SDValue CCReg) {
  SDLoc DL(CCReg);
  SDValue IPM = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, CCReg);
  return DAG.getNode(ISD::SRL, DL, MVT::i32, IPM,
                     DAG.getConstant(SystemZ::IPM_CC, DL, MVT::i32));
}
 
// Lower @cc targets via setcc.
SDValue SystemZTargetLowering::LowerAsmOutputForConstraint(
    SDValue &Chain, SDValue &Glue, const SDLoc &DL,
    const AsmOperandInfo &OpInfo, SelectionDAG &DAG) const {
  if (StringRef("{@cc}").compare(OpInfo.ConstraintCode) != 0)
    return SDValue();
 
  // Check that return type is valid.
  if (OpInfo.ConstraintVT.isVector() || !OpInfo.ConstraintVT.isInteger() ||
      OpInfo.ConstraintVT.getSizeInBits() < 8)
    report_fatal_error("Glue output operand is of invalid type");
 
  if (Glue.getNode()) {
    Glue = DAG.getCopyFromReg(Chain, DL, SystemZ::CC, MVT::i32, Glue);
    Chain = Glue.getValue(1);
  } else
    Glue = DAG.getCopyFromReg(Chain, DL, SystemZ::CC, MVT::i32);
  return getCCResult(DAG, Glue);
}
 
void SystemZTargetLowering::LowerAsmOperandForConstraint(
    SDValue Op, StringRef Constraint, std::vector<SDValue> &Ops,
    SelectionDAG &DAG) const {
  // Only support length 1 constraints for now.
  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    case 'I': // Unsigned 8-bit constant
      if (auto *C = dyn_cast<ConstantSDNode>(Op))
        if (isUInt<8>(C->getZExtValue()))
          Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
                                              Op.getValueType()));
      return;
 
    case 'J': // Unsigned 12-bit constant
      if (auto *C = dyn_cast<ConstantSDNode>(Op))
        if (isUInt<12>(C->getZExtValue()))
          Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
                                              Op.getValueType()));
      return;
 
    case 'K': // Signed 16-bit constant
      if (auto *C = dyn_cast<ConstantSDNode>(Op))
        if (isInt<16>(C->getSExtValue()))
          Ops.push_back(DAG.getSignedTargetConstant(
              C->getSExtValue(), SDLoc(Op), Op.getValueType()));
      return;
 
    case 'L': // Signed 20-bit displacement (on all targets we support)
      if (auto *C = dyn_cast<ConstantSDNode>(Op))
        if (isInt<20>(C->getSExtValue()))
          Ops.push_back(DAG.getSignedTargetConstant(
              C->getSExtValue(), SDLoc(Op), Op.getValueType()));
      return;
 
    case 'M': // 0x7fffffff
      if (auto *C = dyn_cast<ConstantSDNode>(Op))
        if (C->getZExtValue() == 0x7fffffff)
          Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
                                              Op.getValueType()));
      return;
    }
  }
  TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
}
 
//===----------------------------------------------------------------------===//
// Calling conventions
//===----------------------------------------------------------------------===//
 
#define GET_CALLING_CONV_IMPL
#include "SystemZGenCallingConv.inc"
 
const MCPhysReg *SystemZTargetLowering::getScratchRegisters(
  CallingConv::ID) const {
  static const MCPhysReg ScratchRegs[] = { SystemZ::R0D, SystemZ::R1D,
                                           SystemZ::R14D, 0 };
  return ScratchRegs;
}
 
bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType,
                                                     Type *ToType) const {
  return isTruncateFree(FromType, ToType);
}
 
bool SystemZTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
  return CI->isTailCall();
}
 
// Value is a value that has been passed to us in the location described by VA
// (and so has type VA.getLocVT()).  Convert Value to VA.getValVT(), chaining
// any loads onto Chain.
static SDValue convertLocVTToValVT(SelectionDAG &DAG, const SDLoc &DL,
                                   CCValAssign &VA, SDValue Chain,
                                   SDValue Value) {
  // If the argument has been promoted from a smaller type, insert an
  // assertion to capture this.
  if (VA.getLocInfo() == CCValAssign::SExt)
    Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value,
                        DAG.getValueType(VA.getValVT()));
  else if (VA.getLocInfo() == CCValAssign::ZExt)
    Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value,
                        DAG.getValueType(VA.getValVT()));
 
  if (VA.isExtInLoc())
    Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value);
  else if (VA.getLocInfo() == CCValAssign::BCvt) {
    // If this is a short vector argument loaded from the stack,
    // extend from i64 to full vector size and then bitcast.
    assert(VA.getLocVT() == MVT::i64);
    assert(VA.getValVT().isVector());
    Value = DAG.getBuildVector(MVT::v2i64, DL, {Value, DAG.getUNDEF(MVT::i64)});
    Value = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Value);
  } else
    assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
  return Value;
}
 
// Value is a value of type VA.getValVT() that we need to copy into
// the location described by VA.  Return a copy of Value converted to
// VA.getValVT().  The caller is responsible for handling indirect values.
static SDValue convertValVTToLocVT(SelectionDAG &DAG, const SDLoc &DL,
                                   CCValAssign &VA, SDValue Value) {
  switch (VA.getLocInfo()) {
  case CCValAssign::SExt:
    return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value);
  case CCValAssign::ZExt:
    return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
  case CCValAssign::AExt:
    return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
  case CCValAssign::BCvt: {
    assert(VA.getLocVT() == MVT::i64 || VA.getLocVT() == MVT::i128);
    assert(VA.getValVT().isVector() || VA.getValVT() == MVT::f32 ||
           VA.getValVT() == MVT::f64 || VA.getValVT() == MVT::f128);
    // For an f32 vararg we need to first promote it to an f64 and then
    // bitcast it to an i64.
    if (VA.getValVT() == MVT::f32 && VA.getLocVT() == MVT::i64)
      Value = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f64, Value);
    MVT BitCastToType = VA.getValVT().isVector() && VA.getLocVT() == MVT::i64
                            ? MVT::v2i64
                            : VA.getLocVT();
    Value = DAG.getNode(ISD::BITCAST, DL, BitCastToType, Value);
    // For ELF, this is a short vector argument to be stored to the stack,
    // bitcast to v2i64 and then extract first element.
    if (BitCastToType == MVT::v2i64)
      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VA.getLocVT(), Value,
                         DAG.getConstant(0, DL, MVT::i32));
    return Value;
  }
  case CCValAssign::Full:
    return Value;
  default:
    llvm_unreachable("Unhandled getLocInfo()");
  }
}
 
static SDValue lowerI128ToGR128(SelectionDAG &DAG, SDValue In) {
  SDLoc DL(In);
  SDValue Lo, Hi;
  if (DAG.getTargetLoweringInfo().isTypeLegal(MVT::i128)) {
    Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i64, In);
    Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i64,
                     DAG.getNode(ISD::SRL, DL, MVT::i128, In,
                                 DAG.getConstant(64, DL, MVT::i32)));
  } else {
    std::tie(Lo, Hi) = DAG.SplitScalar(In, DL, MVT::i64, MVT::i64);
  }
 
  // FIXME: If v2i64 were a legal type, we could use it instead of
  // Untyped here.  This might enable improved folding.
  SDNode *Pair = DAG.getMachineNode(SystemZ::PAIR128, DL,
                                    MVT::Untyped, Hi, Lo);
  return SDValue(Pair, 0);
}
 
static SDValue lowerGR128ToI128(SelectionDAG &DAG, SDValue In) {
  SDLoc DL(In);
  SDValue Hi = DAG.getTargetExtractSubreg(SystemZ::subreg_h64,
                                          DL, MVT::i64, In);
  SDValue Lo = DAG.getTargetExtractSubreg(SystemZ::subreg_l64,
                                          DL, MVT::i64, In);
 
  if (DAG.getTargetLoweringInfo().isTypeLegal(MVT::i128)) {
    Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i128, Lo);
    Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i128, Hi);
    Hi = DAG.getNode(ISD::SHL, DL, MVT::i128, Hi,
                     DAG.getConstant(64, DL, MVT::i32));
    return DAG.getNode(ISD::OR, DL, MVT::i128, Lo, Hi);
  } else {
    return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128, Lo, Hi);
  }
}
 
bool SystemZTargetLowering::splitValueIntoRegisterParts(
    SelectionDAG &DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
    unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) const {
  EVT ValueVT = Val.getValueType();
  if (ValueVT.getSizeInBits() == 128 && NumParts == 1 && PartVT == MVT::Untyped) {
    // Inline assembly operand.
    Parts[0] = lowerI128ToGR128(DAG, DAG.getBitcast(MVT::i128, Val));
    return true;
  }
 
  return false;
}
 
SDValue SystemZTargetLowering::joinRegisterPartsIntoValue(
    SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts, unsigned NumParts,
    MVT PartVT, EVT ValueVT, std::optional<CallingConv::ID> CC) const {
  if (ValueVT.getSizeInBits() == 128 && NumParts == 1 && PartVT == MVT::Untyped) {
    // Inline assembly operand.
    SDValue Res = lowerGR128ToI128(DAG, Parts[0]);
    return DAG.getBitcast(ValueVT, Res);
  }
 
  return SDValue();
}
 
// The first part of a split stack argument is at index I in Args (and
// ArgLocs). Return the type of a part and the number of them by reference.
template <class ArgTy>
static bool analyzeArgSplit(const SmallVectorImpl<ArgTy> &Args,
                            SmallVector<CCValAssign, 16> &ArgLocs, unsigned I,
                            MVT &PartVT, unsigned &NumParts) {
  if (!Args[I].Flags.isSplit())
    return false;
  assert(I < ArgLocs.size() && ArgLocs.size() == Args.size() &&
         "ArgLocs havoc.");
  PartVT = ArgLocs[I].getValVT();
  NumParts = 1;
  for (unsigned PartIdx = I + 1;; ++PartIdx) {
    assert(PartIdx != ArgLocs.size() && "SplitEnd not found.");
    assert(ArgLocs[PartIdx].getValVT() == PartVT && "Unsupported split.");
    ++NumParts;
    if (Args[PartIdx].Flags.isSplitEnd())
      break;
  }
  return true;
}
 
SDValue SystemZTargetLowering::LowerFormalArguments(
    SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
    const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
    SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  MachineRegisterInfo &MRI = MF.getRegInfo();
  SystemZMachineFunctionInfo *FuncInfo =
      MF.getInfo<SystemZMachineFunctionInfo>();
  auto *TFL = Subtarget.getFrameLowering<SystemZELFFrameLowering>();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
 
  // Assign locations to all of the incoming arguments.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
  CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
  FuncInfo->setSizeOfFnParams(CCInfo.getStackSize());
 
  unsigned NumFixedGPRs = 0;
  unsigned NumFixedFPRs = 0;
  for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
    SDValue ArgValue;
    CCValAssign &VA = ArgLocs[I];
    EVT LocVT = VA.getLocVT();
    if (VA.isRegLoc()) {
      // Arguments passed in registers
      const TargetRegisterClass *RC;
      switch (LocVT.getSimpleVT().SimpleTy) {
      default:
        // Integers smaller than i64 should be promoted to i64.
        llvm_unreachable("Unexpected argument type");
      case MVT::i32:
        NumFixedGPRs += 1;
        RC = &SystemZ::GR32BitRegClass;
        break;
      case MVT::i64:
        NumFixedGPRs += 1;
        RC = &SystemZ::GR64BitRegClass;
        break;
      case MVT::f16:
        NumFixedFPRs += 1;
        RC = &SystemZ::FP16BitRegClass;
        break;
      case MVT::f32:
        NumFixedFPRs += 1;
        RC = &SystemZ::FP32BitRegClass;
        break;
      case MVT::f64:
        NumFixedFPRs += 1;
        RC = &SystemZ::FP64BitRegClass;
        break;
      case MVT::f128:
        NumFixedFPRs += 2;
        RC = &SystemZ::FP128BitRegClass;
        break;
      case MVT::v16i8:
      case MVT::v8i16:
      case MVT::v4i32:
      case MVT::v2i64:
      case MVT::v8f16:
      case MVT::v4f32:
      case MVT::v2f64:
        RC = &SystemZ::VR128BitRegClass;
        break;
      }
 
      Register VReg = MRI.createVirtualRegister(RC);
      MRI.addLiveIn(VA.getLocReg(), VReg);
      ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
    } else {
      assert(VA.isMemLoc() && "Argument not register or memory");
 
      // Create the frame index object for this incoming parameter.
      // FIXME: Pre-include call frame size in the offset, should not
      // need to manually add it here.
      int64_t ArgSPOffset = VA.getLocMemOffset();
      if (Subtarget.isTargetXPLINK64()) {
        auto &XPRegs =
            Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>();
        ArgSPOffset += XPRegs.getCallFrameSize();
      }
      int FI =
          MFI.CreateFixedObject(LocVT.getSizeInBits() / 8, ArgSPOffset, true);
 
      // Create the SelectionDAG nodes corresponding to a load
      // from this parameter.  Unpromoted ints and floats are
      // passed as right-justified 8-byte values.
      SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
      if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32 ||
          VA.getLocVT() == MVT::f16) {
        unsigned SlotOffs = VA.getLocVT() == MVT::f16 ? 6 : 4;
        FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
                          DAG.getIntPtrConstant(SlotOffs, DL));
      }
      ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN,
                             MachinePointerInfo::getFixedStack(MF, FI));
    }
 
    // Convert the value of the argument register into the value that's
    // being passed.
    if (VA.getLocInfo() == CCValAssign::Indirect) {
      InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, ArgValue,
                                   MachinePointerInfo()));
      // If the original argument was split (e.g. i128), we need
      // to load all parts of it here (using the same address).
      MVT PartVT;
      unsigned NumParts;
      if (analyzeArgSplit(Ins, ArgLocs, I, PartVT, NumParts)) {
        for (unsigned PartIdx = 1; PartIdx < NumParts; ++PartIdx) {
          ++I;
          CCValAssign &PartVA = ArgLocs[I];
          unsigned PartOffset = Ins[I].PartOffset;
          SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, ArgValue,
                                        DAG.getIntPtrConstant(PartOffset, DL));
          InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address,
                                       MachinePointerInfo()));
          assert(PartOffset && "Offset should be non-zero.");
        }
      }
    } else
      InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue));
  }
 
  if (IsVarArg && Subtarget.isTargetXPLINK64()) {
    // Save the number of non-varargs registers for later use by va_start, etc.
    FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
    FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
 
    auto *Regs = static_cast<SystemZXPLINK64Registers *>(
        Subtarget.getSpecialRegisters());
 
    // Likewise the address (in the form of a frame index) of where the
    // first stack vararg would be.  The 1-byte size here is arbitrary.
    // FIXME: Pre-include call frame size in the offset, should not
    // need to manually add it here.
    int64_t VarArgOffset = CCInfo.getStackSize() + Regs->getCallFrameSize();
    int FI = MFI.CreateFixedObject(1, VarArgOffset, true);
    FuncInfo->setVarArgsFrameIndex(FI);
  }
 
  if (IsVarArg && Subtarget.isTargetELF()) {
    // Save the number of non-varargs registers for later use by va_start, etc.
    FuncInfo->setVarArgsFirstGPR(NumFixedGPRs);
    FuncInfo->setVarArgsFirstFPR(NumFixedFPRs);
 
    // Likewise the address (in the form of a frame index) of where the
    // first stack vararg would be.  The 1-byte size here is arbitrary.
    int64_t VarArgsOffset = CCInfo.getStackSize();
    FuncInfo->setVarArgsFrameIndex(
        MFI.CreateFixedObject(1, VarArgsOffset, true));
 
    // ...and a similar frame index for the caller-allocated save area
    // that will be used to store the incoming registers.
    int64_t RegSaveOffset =
      -SystemZMC::ELFCallFrameSize + TFL->getRegSpillOffset(MF, SystemZ::R2D) - 16;
    unsigned RegSaveIndex = MFI.CreateFixedObject(1, RegSaveOffset, true);
    FuncInfo->setRegSaveFrameIndex(RegSaveIndex);
 
    // Store the FPR varargs in the reserved frame slots.  (We store the
    // GPRs as part of the prologue.)
    if (NumFixedFPRs < SystemZ::ELFNumArgFPRs && !useSoftFloat()) {
      SDValue MemOps[SystemZ::ELFNumArgFPRs];
      for (unsigned I = NumFixedFPRs; I < SystemZ::ELFNumArgFPRs; ++I) {
        unsigned Offset = TFL->getRegSpillOffset(MF, SystemZ::ELFArgFPRs[I]);
        int FI =
          MFI.CreateFixedObject(8, -SystemZMC::ELFCallFrameSize + Offset, true);
        SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
        Register VReg = MF.addLiveIn(SystemZ::ELFArgFPRs[I],
                                     &SystemZ::FP64BitRegClass);
        SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64);
        MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN,
                                 MachinePointerInfo::getFixedStack(MF, FI));
      }
      // Join the stores, which are independent of one another.
      Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
                          ArrayRef(&MemOps[NumFixedFPRs],
                                   SystemZ::ELFNumArgFPRs - NumFixedFPRs));
    }
  }
 
  if (Subtarget.isTargetXPLINK64()) {
    // Create virual register  for handling incoming "ADA" special register (R5)
    const TargetRegisterClass *RC = &SystemZ::ADDR64BitRegClass;
    Register ADAvReg = MRI.createVirtualRegister(RC);
    auto *Regs = static_cast<SystemZXPLINK64Registers *>(
        Subtarget.getSpecialRegisters());
    MRI.addLiveIn(Regs->getADARegister(), ADAvReg);
    FuncInfo->setADAVirtualRegister(ADAvReg);
  }
  return Chain;
}
 
static bool canUseSiblingCall(const CCState &ArgCCInfo,
                              SmallVectorImpl<CCValAssign> &ArgLocs,
                              SmallVectorImpl<ISD::OutputArg> &Outs) {
  // Punt if there are any indirect or stack arguments, or if the call
  // needs the callee-saved argument register R6, or if the call uses
  // the callee-saved register arguments SwiftSelf and SwiftError.
  for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
    CCValAssign &VA = ArgLocs[I];
    if (VA.getLocInfo() == CCValAssign::Indirect)
      return false;
    if (!VA.isRegLoc())
      return false;
    Register Reg = VA.getLocReg();
    if (Reg == SystemZ::R6H || Reg == SystemZ::R6L || Reg == SystemZ::R6D)
      return false;
    if (Outs[I].Flags.isSwiftSelf() || Outs[I].Flags.isSwiftError())
      return false;
  }
  return true;
}
 
static SDValue getADAEntry(SelectionDAG &DAG, SDValue Val, SDLoc DL,
                           unsigned Offset, bool LoadAdr = false) {
  MachineFunction &MF = DAG.getMachineFunction();
  SystemZMachineFunctionInfo *MFI = MF.getInfo<SystemZMachineFunctionInfo>();
  Register ADAvReg = MFI->getADAVirtualRegister();
  EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
 
  SDValue Reg = DAG.getRegister(ADAvReg, PtrVT);
  SDValue Ofs = DAG.getTargetConstant(Offset, DL, PtrVT);
 
  SDValue Result = DAG.getNode(SystemZISD::ADA_ENTRY, DL, PtrVT, Val, Reg, Ofs);
  if (!LoadAdr)
    Result = DAG.getLoad(
        PtrVT, DL, DAG.getEntryNode(), Result, MachinePointerInfo(), Align(8),
        MachineMemOperand::MODereferenceable | MachineMemOperand::MOInvariant);
 
  return Result;
}
 
// ADA access using Global value
// Note: for functions, address of descriptor is returned
static SDValue getADAEntry(SelectionDAG &DAG, const GlobalValue *GV, SDLoc DL,
                           EVT PtrVT) {
  unsigned ADAtype;
  bool LoadAddr = false;
  const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV);
  bool IsFunction =
      (isa<Function>(GV)) || (GA && isa<Function>(GA->getAliaseeObject()));
  bool IsInternal = (GV->hasInternalLinkage() || GV->hasPrivateLinkage());
 
  if (IsFunction) {
    if (IsInternal) {
      ADAtype = SystemZII::MO_ADA_DIRECT_FUNC_DESC;
      LoadAddr = true;
    } else
      ADAtype = SystemZII::MO_ADA_INDIRECT_FUNC_DESC;
  } else {
    ADAtype = SystemZII::MO_ADA_DATA_SYMBOL_ADDR;
  }
  SDValue Val = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ADAtype);
 
  return getADAEntry(DAG, Val, DL, 0, LoadAddr);
}
 
static bool getzOSCalleeAndADA(SelectionDAG &DAG, SDValue &Callee, SDValue &ADA,
                               SDLoc &DL, SDValue &Chain) {
  unsigned ADADelta = 0; // ADA offset in desc.
  unsigned EPADelta = 8; // EPA offset in desc.
  MachineFunction &MF = DAG.getMachineFunction();
  EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
 
  // XPLink calling convention.
  if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
    bool IsInternal = (G->getGlobal()->hasInternalLinkage() ||
                       G->getGlobal()->hasPrivateLinkage());
    if (IsInternal) {
      SystemZMachineFunctionInfo *MFI =
          MF.getInfo<SystemZMachineFunctionInfo>();
      Register ADAvReg = MFI->getADAVirtualRegister();
      ADA = DAG.getCopyFromReg(Chain, DL, ADAvReg, PtrVT);
      Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
      Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
      return true;
    } else {
      SDValue GA = DAG.getTargetGlobalAddress(
          G->getGlobal(), DL, PtrVT, 0, SystemZII::MO_ADA_DIRECT_FUNC_DESC);
      ADA = getADAEntry(DAG, GA, DL, ADADelta);
      Callee = getADAEntry(DAG, GA, DL, EPADelta);
    }
  } else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
    SDValue ES = DAG.getTargetExternalSymbol(
        E->getSymbol(), PtrVT, SystemZII::MO_ADA_DIRECT_FUNC_DESC);
    ADA = getADAEntry(DAG, ES, DL, ADADelta);
    Callee = getADAEntry(DAG, ES, DL, EPADelta);
  } else {
    // Function pointer case
    ADA = DAG.getNode(ISD::ADD, DL, PtrVT, Callee,
                      DAG.getConstant(ADADelta, DL, PtrVT));
    ADA = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), ADA,
                      MachinePointerInfo::getGOT(DAG.getMachineFunction()));
    Callee = DAG.getNode(ISD::ADD, DL, PtrVT, Callee,
                         DAG.getConstant(EPADelta, DL, PtrVT));
    Callee = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Callee,
                         MachinePointerInfo::getGOT(DAG.getMachineFunction()));
  }
  return false;
}
 
SDValue
SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI,
                                 SmallVectorImpl<SDValue> &InVals) const {
  SelectionDAG &DAG = CLI.DAG;
  SDLoc &DL = CLI.DL;
  SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
  SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
  SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
  SDValue Chain = CLI.Chain;
  SDValue Callee = CLI.Callee;
  bool &IsTailCall = CLI.IsTailCall;
  CallingConv::ID CallConv = CLI.CallConv;
  bool IsVarArg = CLI.IsVarArg;
  MachineFunction &MF = DAG.getMachineFunction();
  EVT PtrVT = getPointerTy(MF.getDataLayout());
  LLVMContext &Ctx = *DAG.getContext();
  SystemZCallingConventionRegisters *Regs = Subtarget.getSpecialRegisters();
 
  // FIXME: z/OS support to be added in later.
  if (Subtarget.isTargetXPLINK64())
    IsTailCall = false;
 
  // Integer args <=32 bits should have an extension attribute.
  verifyNarrowIntegerArgs_Call(Outs, &MF.getFunction(), Callee);
 
  // Analyze the operands of the call, assigning locations to each operand.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, Ctx);
  ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
 
  // We don't support GuaranteedTailCallOpt, only automatically-detected
  // sibling calls.
  if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs, Outs))
    IsTailCall = false;
 
  // Get a count of how many bytes are to be pushed on the stack.
  unsigned NumBytes = ArgCCInfo.getStackSize();
 
  // Mark the start of the call.
  if (!IsTailCall)
    Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, DL);
 
  // Copy argument values to their designated locations.
  SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass;
  SmallVector<SDValue, 8> MemOpChains;
  SDValue StackPtr;
  for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
    CCValAssign &VA = ArgLocs[I];
    SDValue ArgValue = OutVals[I];
 
    if (VA.getLocInfo() == CCValAssign::Indirect) {
      // Store the argument in a stack slot and pass its address.
      EVT SlotVT;
      MVT PartVT;
      unsigned NumParts = 1;
      if (analyzeArgSplit(Outs, ArgLocs, I, PartVT, NumParts))
        SlotVT = EVT::getIntegerVT(Ctx, PartVT.getSizeInBits() * NumParts);
      else
        SlotVT = Outs[I].VT;
      SDValue SpillSlot = DAG.CreateStackTemporary(SlotVT);
      int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
 
      MachinePointerInfo StackPtrInfo =
          MachinePointerInfo::getFixedStack(MF, FI);
      MemOpChains.push_back(
          DAG.getStore(Chain, DL, ArgValue, SpillSlot, StackPtrInfo));
      // If the original argument was split (e.g. i128), we need
      // to store all parts of it here (and pass just one address).
      assert(Outs[I].PartOffset == 0);
      for (unsigned PartIdx = 1; PartIdx < NumParts; ++PartIdx) {
        ++I;
        SDValue PartValue = OutVals[I];
        unsigned PartOffset = Outs[I].PartOffset;
        SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, SpillSlot,
                                      DAG.getIntPtrConstant(PartOffset, DL));
        MemOpChains.push_back(
            DAG.getStore(Chain, DL, PartValue, Address,
                         StackPtrInfo.getWithOffset(PartOffset)));
        assert(PartOffset && "Offset should be non-zero.");
        assert((PartOffset + PartValue.getValueType().getStoreSize() <=
                SlotVT.getStoreSize()) && "Not enough space for argument part!");
      }
      ArgValue = SpillSlot;
    } else
      ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue);
 
    if (VA.isRegLoc()) {
      // In XPLINK64, for the 128-bit vararg case, ArgValue is bitcasted to a
      // MVT::i128 type. We decompose the 128-bit type to a pair of its high
      // and low values.
      if (VA.getLocVT() == MVT::i128)
        ArgValue = lowerI128ToGR128(DAG, ArgValue);
      // Queue up the argument copies and emit them at the end.
      RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
    } else {
      assert(VA.isMemLoc() && "Argument not register or memory");
 
      // Work out the address of the stack slot.  Unpromoted ints and
      // floats are passed as right-justified 8-byte values.
      if (!StackPtr.getNode())
        StackPtr = DAG.getCopyFromReg(Chain, DL,
                                      Regs->getStackPointerRegister(), PtrVT);
      unsigned Offset = Regs->getStackPointerBias() + Regs->getCallFrameSize() +
                        VA.getLocMemOffset();
      if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32)
        Offset += 4;
      else if (VA.getLocVT() == MVT::f16)
        Offset += 6;
      SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
                                    DAG.getIntPtrConstant(Offset, DL));
 
      // Emit the store.
      MemOpChains.push_back(
          DAG.getStore(Chain, DL, ArgValue, Address, MachinePointerInfo()));
 
      // Although long doubles or vectors are passed through the stack when
      // they are vararg (non-fixed arguments), if a long double or vector
      // occupies the third and fourth slot of the argument list GPR3 should
      // still shadow the third slot of the argument list.
      if (Subtarget.isTargetXPLINK64() && VA.needsCustom()) {
        SDValue ShadowArgValue =
            DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64, ArgValue,
                        DAG.getIntPtrConstant(1, DL));
        RegsToPass.push_back(std::make_pair(SystemZ::R3D, ShadowArgValue));
      }
    }
  }
 
  // Join the stores, which are independent of one another.
  if (!MemOpChains.empty())
    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
 
  // Accept direct calls by converting symbolic call addresses to the
  // associated Target* opcodes.  Force %r1 to be used for indirect
  // tail calls.
  SDValue Glue;
 
  if (Subtarget.isTargetXPLINK64()) {
    SDValue ADA;
    bool IsBRASL = getzOSCalleeAndADA(DAG, Callee, ADA, DL, Chain);
    if (!IsBRASL) {
      unsigned CalleeReg = static_cast<SystemZXPLINK64Registers *>(Regs)
                               ->getAddressOfCalleeRegister();
      Chain = DAG.getCopyToReg(Chain, DL, CalleeReg, Callee, Glue);
      Glue = Chain.getValue(1);
      Callee = DAG.getRegister(CalleeReg, Callee.getValueType());
    }
    RegsToPass.push_back(std::make_pair(
        static_cast<SystemZXPLINK64Registers *>(Regs)->getADARegister(), ADA));
  } else {
    if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
      Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
      Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
    } else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
      Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
      Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
    } else if (IsTailCall) {
      Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue);
      Glue = Chain.getValue(1);
      Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType());
    }
  }
 
  // Build a sequence of copy-to-reg nodes, chained and glued together.
  for (const auto &[Reg, N] : RegsToPass) {
    Chain = DAG.getCopyToReg(Chain, DL, Reg, N, Glue);
    Glue = Chain.getValue(1);
  }
 
  // The first call operand is the chain and the second is the target address.
  SmallVector<SDValue, 8> Ops;
  Ops.push_back(Chain);
  Ops.push_back(Callee);
 
  // Add argument registers to the end of the list so that they are
  // known live into the call.
  for (const auto &[Reg, N] : RegsToPass)
    Ops.push_back(DAG.getRegister(Reg, N.getValueType()));
 
  // Add a register mask operand representing the call-preserved registers.
  const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
  const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
  assert(Mask && "Missing call preserved mask for calling convention");
  Ops.push_back(DAG.getRegisterMask(Mask));
 
  // Glue the call to the argument copies, if any.
  if (Glue.getNode())
    Ops.push_back(Glue);
 
  // Emit the call.
  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
  if (IsTailCall) {
    SDValue Ret = DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, Ops);
    DAG.addNoMergeSiteInfo(Ret.getNode(), CLI.NoMerge);
    return Ret;
  }
  Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, Ops);
  DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge);
  Glue = Chain.getValue(1);
 
  // Mark the end of the call, which is glued to the call itself.
  Chain = DAG.getCALLSEQ_END(Chain, NumBytes, 0, Glue, DL);
  Glue = Chain.getValue(1);
 
  // Assign locations to each value returned by this call.
  SmallVector<CCValAssign, 16> RetLocs;
  CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, Ctx);
  RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
 
  // Copy all of the result registers out of their specified physreg.
  for (CCValAssign &VA : RetLocs) {
    // Copy the value out, gluing the copy to the end of the call sequence.
    SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(),
                                          VA.getLocVT(), Glue);
    Chain = RetValue.getValue(1);
    Glue = RetValue.getValue(2);
 
    // Convert the value of the return register into the value that's
    // being returned.
    InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue));
  }
 
  return Chain;
}
 
// Generate a call taking the given operands as arguments and returning a
// result of type RetVT.
std::pair<SDValue, SDValue> SystemZTargetLowering::makeExternalCall(
    SDValue Chain, SelectionDAG &DAG, const char *CalleeName, EVT RetVT,
    ArrayRef<SDValue> Ops, CallingConv::ID CallConv, bool IsSigned, SDLoc DL,
    bool DoesNotReturn, bool IsReturnValueUsed) const {
  TargetLowering::ArgListTy Args;
  Args.reserve(Ops.size());
 
  for (SDValue Op : Ops) {
    TargetLowering::ArgListEntry Entry(
        Op, Op.getValueType().getTypeForEVT(*DAG.getContext()));
    Entry.IsSExt = shouldSignExtendTypeInLibCall(Entry.Ty, IsSigned);
    Entry.IsZExt = !Entry.IsSExt;
    Args.push_back(Entry);
  }
 
  SDValue Callee =
      DAG.getExternalSymbol(CalleeName, getPointerTy(DAG.getDataLayout()));
 
  Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
  TargetLowering::CallLoweringInfo CLI(DAG);
  bool SignExtend = shouldSignExtendTypeInLibCall(RetTy, IsSigned);
  CLI.setDebugLoc(DL)
      .setChain(Chain)
      .setCallee(CallConv, RetTy, Callee, std::move(Args))
      .setNoReturn(DoesNotReturn)
      .setDiscardResult(!IsReturnValueUsed)
      .setSExtResult(SignExtend)
      .setZExtResult(!SignExtend);
  return LowerCallTo(CLI);
}
 
bool SystemZTargetLowering::CanLowerReturn(
    CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg,
    const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context,
    const Type *RetTy) const {
  // Special case that we cannot easily detect in RetCC_SystemZ since
  // i128 may not be a legal type.
  for (auto &Out : Outs)
    if (Out.ArgVT.isScalarInteger() && Out.ArgVT.getSizeInBits() > 64)
      return false;
 
  SmallVector<CCValAssign, 16> RetLocs;
  CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, Context);
  return RetCCInfo.CheckReturn(Outs, RetCC_SystemZ);
}
 
SDValue
SystemZTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
                                   bool IsVarArg,
                                   const SmallVectorImpl<ISD::OutputArg> &Outs,
                                   const SmallVectorImpl<SDValue> &OutVals,
                                   const SDLoc &DL, SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
 
  // Integer args <=32 bits should have an extension attribute.
  verifyNarrowIntegerArgs_Ret(Outs, &MF.getFunction());
 
  // Assign locations to each returned value.
  SmallVector<CCValAssign, 16> RetLocs;
  CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
  RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
 
  // Quick exit for void returns
  if (RetLocs.empty())
    return DAG.getNode(SystemZISD::RET_GLUE, DL, MVT::Other, Chain);
 
  if (CallConv == CallingConv::GHC)
    report_fatal_error("GHC functions return void only");
 
  // Copy the result values into the output registers.
  SDValue Glue;
  SmallVector<SDValue, 4> RetOps;
  RetOps.push_back(Chain);
  for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) {
    CCValAssign &VA = RetLocs[I];
    SDValue RetValue = OutVals[I];
 
    // Make the return register live on exit.
    assert(VA.isRegLoc() && "Can only return in registers!");
 
    // Promote the value as required.
    RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue);
 
    // Chain and glue the copies together.
    Register Reg = VA.getLocReg();
    Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue);
    Glue = Chain.getValue(1);
    RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT()));
  }
 
  // Update chain and glue.
  RetOps[0] = Chain;
  if (Glue.getNode())
    RetOps.push_back(Glue);
 
  return DAG.getNode(SystemZISD::RET_GLUE, DL, MVT::Other, RetOps);
}
 
// Return true if Op is an intrinsic node with chain that returns the CC value
// as its only (other) argument.  Provide the associated SystemZISD opcode and
// the mask of valid CC values if so.
static bool isIntrinsicWithCCAndChain(SDValue Op, unsigned &Opcode,
                                      unsigned &CCValid) {
  unsigned Id = Op.getConstantOperandVal(1);
  switch (Id) {
  case Intrinsic::s390_tbegin:
    Opcode = SystemZISD::TBEGIN;
    CCValid = SystemZ::CCMASK_TBEGIN;
    return true;
 
  case Intrinsic::s390_tbegin_nofloat:
    Opcode = SystemZISD::TBEGIN_NOFLOAT;
    CCValid = SystemZ::CCMASK_TBEGIN;
    return true;
 
  case Intrinsic::s390_tend:
    Opcode = SystemZISD::TEND;
    CCValid = SystemZ::CCMASK_TEND;
    return true;
 
  default:
    return false;
  }
}
 
// Return true if Op is an intrinsic node without chain that returns the
// CC value as its final argument.  Provide the associated SystemZISD
// opcode and the mask of valid CC values if so.
static bool isIntrinsicWithCC(SDValue Op, unsigned &Opcode, unsigned &CCValid) {
  unsigned Id = Op.getConstantOperandVal(0);
  switch (Id) {
  case Intrinsic::s390_vpkshs:
  case Intrinsic::s390_vpksfs:
  case Intrinsic::s390_vpksgs:
    Opcode = SystemZISD::PACKS_CC;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vpklshs:
  case Intrinsic::s390_vpklsfs:
  case Intrinsic::s390_vpklsgs:
    Opcode = SystemZISD::PACKLS_CC;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vceqbs:
  case Intrinsic::s390_vceqhs:
  case Intrinsic::s390_vceqfs:
  case Intrinsic::s390_vceqgs:
  case Intrinsic::s390_vceqqs:
    Opcode = SystemZISD::VICMPES;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vchbs:
  case Intrinsic::s390_vchhs:
  case Intrinsic::s390_vchfs:
  case Intrinsic::s390_vchgs:
  case Intrinsic::s390_vchqs:
    Opcode = SystemZISD::VICMPHS;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vchlbs:
  case Intrinsic::s390_vchlhs:
  case Intrinsic::s390_vchlfs:
  case Intrinsic::s390_vchlgs:
  case Intrinsic::s390_vchlqs:
    Opcode = SystemZISD::VICMPHLS;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vtm:
    Opcode = SystemZISD::VTM;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vfaebs:
  case Intrinsic::s390_vfaehs:
  case Intrinsic::s390_vfaefs:
    Opcode = SystemZISD::VFAE_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vfaezbs:
  case Intrinsic::s390_vfaezhs:
  case Intrinsic::s390_vfaezfs:
    Opcode = SystemZISD::VFAEZ_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vfeebs:
  case Intrinsic::s390_vfeehs:
  case Intrinsic::s390_vfeefs:
    Opcode = SystemZISD::VFEE_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vfeezbs:
  case Intrinsic::s390_vfeezhs:
  case Intrinsic::s390_vfeezfs:
    Opcode = SystemZISD::VFEEZ_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vfenebs:
  case Intrinsic::s390_vfenehs:
  case Intrinsic::s390_vfenefs:
    Opcode = SystemZISD::VFENE_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vfenezbs:
  case Intrinsic::s390_vfenezhs:
  case Intrinsic::s390_vfenezfs:
    Opcode = SystemZISD::VFENEZ_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vistrbs:
  case Intrinsic::s390_vistrhs:
  case Intrinsic::s390_vistrfs:
    Opcode = SystemZISD::VISTR_CC;
    CCValid = SystemZ::CCMASK_0 | SystemZ::CCMASK_3;
    return true;
 
  case Intrinsic::s390_vstrcbs:
  case Intrinsic::s390_vstrchs:
  case Intrinsic::s390_vstrcfs:
    Opcode = SystemZISD::VSTRC_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vstrczbs:
  case Intrinsic::s390_vstrczhs:
  case Intrinsic::s390_vstrczfs:
    Opcode = SystemZISD::VSTRCZ_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vstrsb:
  case Intrinsic::s390_vstrsh:
  case Intrinsic::s390_vstrsf:
    Opcode = SystemZISD::VSTRS_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vstrszb:
  case Intrinsic::s390_vstrszh:
  case Intrinsic::s390_vstrszf:
    Opcode = SystemZISD::VSTRSZ_CC;
    CCValid = SystemZ::CCMASK_ANY;
    return true;
 
  case Intrinsic::s390_vfcedbs:
  case Intrinsic::s390_vfcesbs:
    Opcode = SystemZISD::VFCMPES;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vfchdbs:
  case Intrinsic::s390_vfchsbs:
    Opcode = SystemZISD::VFCMPHS;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vfchedbs:
  case Intrinsic::s390_vfchesbs:
    Opcode = SystemZISD::VFCMPHES;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_vftcidb:
  case Intrinsic::s390_vftcisb:
    Opcode = SystemZISD::VFTCI;
    CCValid = SystemZ::CCMASK_VCMP;
    return true;
 
  case Intrinsic::s390_tdc:
    Opcode = SystemZISD::TDC;
    CCValid = SystemZ::CCMASK_TDC;
    return true;
 
  default:
    return false;
  }
}
 
// Emit an intrinsic with chain and an explicit CC register result.
static SDNode *emitIntrinsicWithCCAndChain(SelectionDAG &DAG, SDValue Op,
                                           unsigned Opcode) {
  // Copy all operands except the intrinsic ID.
  unsigned NumOps = Op.getNumOperands();
  SmallVector<SDValue, 6> Ops;
  Ops.reserve(NumOps - 1);
  Ops.push_back(Op.getOperand(0));
  for (unsigned I = 2; I < NumOps; ++I)
    Ops.push_back(Op.getOperand(I));
 
  assert(Op->getNumValues() == 2 && "Expected only CC result and chain");
  SDVTList RawVTs = DAG.getVTList(MVT::i32, MVT::Other);
  SDValue Intr = DAG.getNode(Opcode, SDLoc(Op), RawVTs, Ops);
  SDValue OldChain = SDValue(Op.getNode(), 1);
  SDValue NewChain = SDValue(Intr.getNode(), 1);
  DAG.ReplaceAllUsesOfValueWith(OldChain, NewChain);
  return Intr.getNode();
}
 
// Emit an intrinsic with an explicit CC register result.
static SDNode *emitIntrinsicWithCC(SelectionDAG &DAG, SDValue Op,
                                   unsigned Opcode) {
  // Copy all operands except the intrinsic ID.
  SDLoc DL(Op);
  unsigned NumOps = Op.getNumOperands();
  SmallVector<SDValue, 6> Ops;
  Ops.reserve(NumOps - 1);
  for (unsigned I = 1; I < NumOps; ++I) {
    SDValue CurrOper = Op.getOperand(I);
    if (CurrOper.getValueType() == MVT::f16) {
      assert((Op.getConstantOperandVal(0) == Intrinsic::s390_tdc && I == 1) &&
             "Unhandled intrinsic with f16 operand.");
      CurrOper = DAG.getFPExtendOrRound(CurrOper, DL, MVT::f32);
    }
    Ops.push_back(CurrOper);
  }
 
  SDValue Intr = DAG.getNode(Opcode, DL, Op->getVTList(), Ops);
  return Intr.getNode();
}
 
// CC is a comparison that will be implemented using an integer or
// floating-point comparison.  Return the condition code mask for
// a branch on true.  In the integer case, CCMASK_CMP_UO is set for
// unsigned comparisons and clear for signed ones.  In the floating-point
// case, CCMASK_CMP_UO has its normal mask meaning (unordered).
static unsigned CCMaskForCondCode(ISD::CondCode CC) {
#define CONV(X) \
  case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \
  case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \
  case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X
 
  switch (CC) {
  default:
    llvm_unreachable("Invalid integer condition!");
 
  CONV(EQ);
  CONV(NE);
  CONV(GT);
  CONV(GE);
  CONV(LT);
  CONV(LE);
 
  case ISD::SETO:  return SystemZ::CCMASK_CMP_O;
  case ISD::SETUO: return SystemZ::CCMASK_CMP_UO;
  }
#undef CONV
}
 
// If C can be converted to a comparison against zero, adjust the operands
// as necessary.
static void adjustZeroCmp(SelectionDAG &DAG, const SDLoc &DL, Comparison &C) {
  if (C.ICmpType == SystemZICMP::UnsignedOnly)
    return;
 
  auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1.getNode());
  if (!ConstOp1 || ConstOp1->getValueSizeInBits(0) > 64)
    return;
 
  int64_t Value = ConstOp1->getSExtValue();
  if ((Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_GT) ||
      (Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_LE) ||
      (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_LT) ||
      (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_GE)) {
    C.CCMask ^= SystemZ::CCMASK_CMP_EQ;
    C.Op1 = DAG.getConstant(0, DL, C.Op1.getValueType());
  }
}
 
// If a comparison described by C is suitable for CLI(Y), CHHSI or CLHHSI,
// adjust the operands as necessary.
static void adjustSubwordCmp(SelectionDAG &DAG, const SDLoc &DL,
                             Comparison &C) {
  // For us to make any changes, it must a comparison between a single-use
  // load and a constant.
  if (!C.Op0.hasOneUse() ||
      C.Op0.getOpcode() != ISD::LOAD ||
      C.Op1.getOpcode() != ISD::Constant)
    return;
 
  // We must have an 8- or 16-bit load.
  auto *Load = cast<LoadSDNode>(C.Op0);
  unsigned NumBits = Load->getMemoryVT().getSizeInBits();
  if ((NumBits != 8 && NumBits != 16) ||
      NumBits != Load->getMemoryVT().getStoreSizeInBits())
    return;
 
  // The load must be an extending one and the constant must be within the
  // range of the unextended value.
  auto *ConstOp1 = cast<ConstantSDNode>(C.Op1);
  if (!ConstOp1 || ConstOp1->getValueSizeInBits(0) > 64)
    return;
  uint64_t Value = ConstOp1->getZExtValue();
  uint64_t Mask = (1 << NumBits) - 1;
  if (Load->getExtensionType() == ISD::SEXTLOAD) {
    // Make sure that ConstOp1 is in range of C.Op0.
    int64_t SignedValue = ConstOp1->getSExtValue();
    if (uint64_t(SignedValue) + (uint64_t(1) << (NumBits - 1)) > Mask)
      return;
    if (C.ICmpType != SystemZICMP::SignedOnly) {
      // Unsigned comparison between two sign-extended values is equivalent
      // to unsigned comparison between two zero-extended values.
      Value &= Mask;
    } else if (NumBits == 8) {
      // Try to treat the comparison as unsigned, so that we can use CLI.
      // Adjust CCMask and Value as necessary.
      if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_LT)
        // Test whether the high bit of the byte is set.
        Value = 127, C.CCMask = SystemZ::CCMASK_CMP_GT;
      else if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_GE)
        // Test whether the high bit of the byte is clear.
        Value = 128, C.CCMask = SystemZ::CCMASK_CMP_LT;
      else
        // No instruction exists for this combination.
        return;
      C.ICmpType = SystemZICMP::UnsignedOnly;
    }
  } else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
    if (Value > Mask)
      return;
    // If the constant is in range, we can use any comparison.
    C.ICmpType = SystemZICMP::Any;
  } else
    return;
 
  // Make sure that the first operand is an i32 of the right extension type.
  ISD::LoadExtType ExtType = (C.ICmpType == SystemZICMP::SignedOnly ?
                              ISD::SEXTLOAD :
                              ISD::ZEXTLOAD);
  if (C.Op0.getValueType() != MVT::i32 ||
      Load->getExtensionType() != ExtType) {
    C.Op0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32, Load->getChain(),
                           Load->getBasePtr(), Load->getPointerInfo(),
                           Load->getMemoryVT(), Load->getAlign(),
                           Load->getMemOperand()->getFlags());
    // Update the chain uses.
    DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), C.Op0.getValue(1));
  }
 
  // Make sure that the second operand is an i32 with the right value.
  if (C.Op1.getValueType() != MVT::i32 ||
      Value != ConstOp1->getZExtValue())
    C.Op1 = DAG.getConstant((uint32_t)Value, DL, MVT::i32);
}
 
// Return true if Op is either an unextended load, or a load suitable
// for integer register-memory comparisons of type ICmpType.
static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) {
  auto *Load = dyn_cast<LoadSDNode>(Op.getNode());
  if (Load) {
    // There are no instructions to compare a register with a memory byte.
    if (Load->getMemoryVT() == MVT::i8)
      return false;
    // Otherwise decide on extension type.
    switch (Load->getExtensionType()) {
    case ISD::NON_EXTLOAD:
      return true;
    case ISD::SEXTLOAD:
      return ICmpType != SystemZICMP::UnsignedOnly;
    case ISD::ZEXTLOAD:
      return ICmpType != SystemZICMP::SignedOnly;
    default:
      break;
    }
  }
  return false;
}
 
// Return true if it is better to swap the operands of C.
static bool shouldSwapCmpOperands(const Comparison &C) {
  // If one side of the compare is a load of the stackguard reference value,
  // then that load should be Op1.
  if (C.Op0.isMachineOpcode() &&
      (C.Op0.getMachineOpcode() == SystemZ::LOAD_STACK_GUARD))
    return true;
 
  // Leave i128 and f128 comparisons alone, since they have no memory forms.
  if (C.Op0.getValueType() == MVT::i128)
    return false;
  if (C.Op0.getValueType() == MVT::f128)
    return false;
 
  // Always keep a floating-point constant second, since comparisons with
  // zero can use LOAD TEST and comparisons with other constants make a
  // natural memory operand.
  if (isa<ConstantFPSDNode>(C.Op1))
    return false;
 
  // Never swap comparisons with zero since there are many ways to optimize
  // those later.
  auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
  if (ConstOp1 && ConstOp1->getZExtValue() == 0)
    return false;
 
  // Also keep natural memory operands second if the loaded value is
  // only used here.  Several comparisons have memory forms.
  if (isNaturalMemoryOperand(C.Op1, C.ICmpType) && C.Op1.hasOneUse())
    return false;
 
  // Look for cases where Cmp0 is a single-use load and Cmp1 isn't.
  // In that case we generally prefer the memory to be second.
  if (isNaturalMemoryOperand(C.Op0, C.ICmpType) && C.Op0.hasOneUse()) {
    // The only exceptions are when the second operand is a constant and
    // we can use things like CHHSI.
    if (!ConstOp1)
      return true;
    // The unsigned memory-immediate instructions can handle 16-bit
    // unsigned integers.
    if (C.ICmpType != SystemZICMP::SignedOnly &&
        isUInt<16>(ConstOp1->getZExtValue()))
      return false;
    // The signed memory-immediate instructions can handle 16-bit
    // signed integers.
    if (C.ICmpType != SystemZICMP::UnsignedOnly &&
        isInt<16>(ConstOp1->getSExtValue()))
      return false;
    return true;
  }
 
  // Try to promote the use of CGFR and CLGFR.
  unsigned Opcode0 = C.Op0.getOpcode();
  if (C.ICmpType != SystemZICMP::UnsignedOnly && Opcode0 == ISD::SIGN_EXTEND)
    return true;
  if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::ZERO_EXTEND)
    return true;
  if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::AND &&
      C.Op0.getOperand(1).getOpcode() == ISD::Constant &&
      C.Op0.getConstantOperandVal(1) == 0xffffffff)
    return true;
 
  return false;
}
 
// Check whether C tests for equality between X and Y and whether X - Y
// or Y - X is also computed.  In that case it's better to compare the
// result of the subtraction against zero.
static void adjustForSubtraction(SelectionDAG &DAG, const SDLoc &DL,
                                 Comparison &C) {
  if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
      C.CCMask == SystemZ::CCMASK_CMP_NE) {
    for (SDNode *N : C.Op0->users()) {
      if (N->getOpcode() == ISD::SUB &&
          ((N->getOperand(0) == C.Op0 && N->getOperand(1) == C.Op1) ||
           (N->getOperand(0) == C.Op1 && N->getOperand(1) == C.Op0))) {
        // Disable the nsw and nuw flags: the backend needs to handle
        // overflow as well during comparison elimination.
        N->dropFlags(SDNodeFlags::NoWrap);
        C.Op0 = SDValue(N, 0);
        C.Op1 = DAG.getConstant(0, DL, N->getValueType(0));
        return;
      }
    }
  }
}
 
// Check whether C compares a floating-point value with zero and if that
// floating-point value is also negated.  In this case we can use the
// negation to set CC, so avoiding separate LOAD AND TEST and
// LOAD (NEGATIVE/COMPLEMENT) instructions.
static void adjustForFNeg(Comparison &C) {
  // This optimization is invalid for strict comparisons, since FNEG
  // does not raise any exceptions.
  if (C.Chain)
    return;
  auto *C1 = dyn_cast<ConstantFPSDNode>(C.Op1);
  if (C1 && C1->isZero()) {
    for (SDNode *N : C.Op0->users()) {
      if (N->getOpcode() == ISD::FNEG) {
        C.Op0 = SDValue(N, 0);
        C.CCMask = SystemZ::reverseCCMask(C.CCMask);
        return;
      }
    }
  }
}
 
// Check whether C compares (shl X, 32) with 0 and whether X is
// also sign-extended.  In that case it is better to test the result
// of the sign extension using LTGFR.
//
// This case is important because InstCombine transforms a comparison
// with (sext (trunc X)) into a comparison with (shl X, 32).
static void adjustForLTGFR(Comparison &C) {
  // Check for a comparison between (shl X, 32) and 0.
  if (C.Op0.getOpcode() == ISD::SHL && C.Op0.getValueType() == MVT::i64 &&
      C.Op1.getOpcode() == ISD::Constant && C.Op1->getAsZExtVal() == 0) {
    auto *C1 = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
    if (C1 && C1->getZExtValue() == 32) {
      SDValue ShlOp0 = C.Op0.getOperand(0);
      // See whether X has any SIGN_EXTEND_INREG uses.
      for (SDNode *N : ShlOp0->users()) {
        if (N->getOpcode() == ISD::SIGN_EXTEND_INREG &&
            cast<VTSDNode>(N->getOperand(1))->getVT() == MVT::i32) {
          C.Op0 = SDValue(N, 0);
          return;
        }
      }
    }
  }
}
 
// If C compares the truncation of an extending load, try to compare
// the untruncated value instead.  This exposes more opportunities to
// reuse CC.
static void adjustICmpTruncate(SelectionDAG &DAG, const SDLoc &DL,
                               Comparison &C) {
  if (C.Op0.getOpcode() == ISD::TRUNCATE &&
      C.Op0.getOperand(0).getOpcode() == ISD::LOAD &&
      C.Op1.getOpcode() == ISD::Constant &&
      cast<ConstantSDNode>(C.Op1)->getValueSizeInBits(0) <= 64 &&
      C.Op1->getAsZExtVal() == 0) {
    auto *L = cast<LoadSDNode>(C.Op0.getOperand(0));
    if (L->getMemoryVT().getStoreSizeInBits().getFixedValue() <=
        C.Op0.getValueSizeInBits().getFixedValue()) {
      unsigned Type = L->getExtensionType();
      if ((Type == ISD::ZEXTLOAD && C.ICmpType != SystemZICMP::SignedOnly) ||
          (Type == ISD::SEXTLOAD && C.ICmpType != SystemZICMP::UnsignedOnly)) {
        C.Op0 = C.Op0.getOperand(0);
        C.Op1 = DAG.getConstant(0, DL, C.Op0.getValueType());
      }
    }
  }
}
 
// Adjust if a given Compare is a check of the stack guard against a stack
// guard instance on the stack. Specifically, this checks if:
// - The operands are a load of the stack guard, and a load from a stack slot
// - The original opcode is ICMP
// - ICMPType is compatible with unsigned comparison.
static void adjustForStackGuardCompare(SelectionDAG &DAG, const SDLoc &DL,
                                       Comparison &C) {
 
  // Opcode must be ICMP.
  if (C.Opcode != SystemZISD::ICMP)
    return;
  // ICmpType must be Unsigned or Any.
  if (C.ICmpType == SystemZICMP::SignedOnly)
    return;
  // Op0 must be FrameIndex Load.
  if (!(ISD::isNormalLoad(C.Op0.getNode()) &&
        dyn_cast<FrameIndexSDNode>(C.Op0.getOperand(1))))
    return;
  // Op1 must be LOAD_STACK_GUARD.
  if (!C.Op1.isMachineOpcode() ||
      C.Op1.getMachineOpcode() != SystemZ::LOAD_STACK_GUARD)
    return;
 
  // At this point we are sure that this is a proper CMP_STACKGUARD
  // case, update the opcode to reflect this.
  C.Opcode = SystemZISD::CMP_STACKGUARD;
  C.Op1 = SDValue();
}
 
// Return true if shift operation N has an in-range constant shift value.
// Store it in ShiftVal if so.
static bool isSimpleShift(SDValue N, unsigned &ShiftVal) {
  auto *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
  if (!Shift)
    return false;
 
  uint64_t Amount = Shift->getZExtValue();
  if (Amount >= N.getValueSizeInBits())
    return false;
 
  ShiftVal = Amount;
  return true;
}
 
// Check whether an AND with Mask is suitable for a TEST UNDER MASK
// instruction and whether the CC value is descriptive enough to handle
// a comparison of type Opcode between the AND result and CmpVal.
// CCMask says which comparison result is being tested and BitSize is
// the number of bits in the operands.  If TEST UNDER MASK can be used,
// return the corresponding CC mask, otherwise return 0.
static unsigned getTestUnderMaskCond(unsigned BitSize, unsigned CCMask,
                                     uint64_t Mask, uint64_t CmpVal,
                                     unsigned ICmpType) {
  assert(Mask != 0 && "ANDs with zero should have been removed by now");
 
  // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL.
  if (!SystemZ::isImmLL(Mask) && !SystemZ::isImmLH(Mask) &&
      !SystemZ::isImmHL(Mask) && !SystemZ::isImmHH(Mask))
    return 0;
 
  // Work out the masks for the lowest and highest bits.
  uint64_t High = llvm::bit_floor(Mask);
  uint64_t Low = uint64_t(1) << llvm::countr_zero(Mask);
 
  // Signed ordered comparisons are effectively unsigned if the sign
  // bit is dropped.
  bool EffectivelyUnsigned = (ICmpType != SystemZICMP::SignedOnly);
 
  // Check for equality comparisons with 0, or the equivalent.
  if (CmpVal == 0) {
    if (CCMask == SystemZ::CCMASK_CMP_EQ)
      return SystemZ::CCMASK_TM_ALL_0;
    if (CCMask == SystemZ::CCMASK_CMP_NE)
      return SystemZ::CCMASK_TM_SOME_1;
  }
  if (EffectivelyUnsigned && CmpVal > 0 && CmpVal <= Low) {
    if (CCMask == SystemZ::CCMASK_CMP_LT)
      return SystemZ::CCMASK_TM_ALL_0;
    if (CCMask == SystemZ::CCMASK_CMP_GE)
      return SystemZ::CCMASK_TM_SOME_1;
  }
  if (EffectivelyUnsigned && CmpVal < Low) {
    if (CCMask == SystemZ::CCMASK_CMP_LE)
      return SystemZ::CCMASK_TM_ALL_0;
    if (CCMask == SystemZ::CCMASK_CMP_GT)
      return SystemZ::CCMASK_TM_SOME_1;
  }
 
  // Check for equality comparisons with the mask, or the equivalent.
  if (CmpVal == Mask) {
    if (CCMask == SystemZ::CCMASK_CMP_EQ)
      return SystemZ::CCMASK_TM_ALL_1;
    if (CCMask == SystemZ::CCMASK_CMP_NE)
      return SystemZ::CCMASK_TM_SOME_0;
  }
  if (EffectivelyUnsigned && CmpVal >= Mask - Low && CmpVal < Mask) {
    if (CCMask == SystemZ::CCMASK_CMP_GT)
      return SystemZ::CCMASK_TM_ALL_1;
    if (CCMask == SystemZ::CCMASK_CMP_LE)
      return SystemZ::CCMASK_TM_SOME_0;
  }
  if (EffectivelyUnsigned && CmpVal > Mask - Low && CmpVal <= Mask) {
    if (CCMask == SystemZ::CCMASK_CMP_GE)
      return SystemZ::CCMASK_TM_ALL_1;
    if (CCMask == SystemZ::CCMASK_CMP_LT)
      return SystemZ::CCMASK_TM_SOME_0;
  }
 
  // Check for ordered comparisons with the top bit.
  if (EffectivelyUnsigned && CmpVal >= Mask - High && CmpVal < High) {
    if (CCMask == SystemZ::CCMASK_CMP_LE)
      return SystemZ::CCMASK_TM_MSB_0;
    if (CCMask == SystemZ::CCMASK_CMP_GT)
      return SystemZ::CCMASK_TM_MSB_1;
  }
  if (EffectivelyUnsigned && CmpVal > Mask - High && CmpVal <= High) {
    if (CCMask == SystemZ::CCMASK_CMP_LT)
      return SystemZ::CCMASK_TM_MSB_0;
    if (CCMask == SystemZ::CCMASK_CMP_GE)
      return SystemZ::CCMASK_TM_MSB_1;
  }
 
  // If there are just two bits, we can do equality checks for Low and High
  // as well.
  if (Mask == Low + High) {
    if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == Low)
      return SystemZ::CCMASK_TM_MIXED_MSB_0;
    if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == Low)
      return SystemZ::CCMASK_TM_MIXED_MSB_0 ^ SystemZ::CCMASK_ANY;
    if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == High)
      return SystemZ::CCMASK_TM_MIXED_MSB_1;
    if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == High)
      return SystemZ::CCMASK_TM_MIXED_MSB_1 ^ SystemZ::CCMASK_ANY;
  }
 
  // Looks like we've exhausted our options.
  return 0;
}
 
// See whether C can be implemented as a TEST UNDER MASK instruction.
// Update the arguments with the TM version if so.
static void adjustForTestUnderMask(SelectionDAG &DAG, const SDLoc &DL,
                                   Comparison &C) {
  // Use VECTOR TEST UNDER MASK for i128 operations.
  if (C.Op0.getValueType() == MVT::i128) {
    // We can use VTM for EQ/NE comparisons of x & y against 0.
    if (C.Op0.getOpcode() == ISD::AND &&
        (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
         C.CCMask == SystemZ::CCMASK_CMP_NE)) {
      auto *Mask = dyn_cast<ConstantSDNode>(C.Op1);
      if (Mask && Mask->getAPIntValue() == 0) {
        C.Opcode = SystemZISD::VTM;
        C.Op1 = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, C.Op0.getOperand(1));
        C.Op0 = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, C.Op0.getOperand(0));
        C.CCValid = SystemZ::CCMASK_VCMP;
        if (C.CCMask == SystemZ::CCMASK_CMP_EQ)
          C.CCMask = SystemZ::CCMASK_VCMP_ALL;
        else
          C.CCMask = SystemZ::CCMASK_VCMP_ALL ^ C.CCValid;
      }
    }
    return;
  }
 
  // Check that we have a comparison with a constant.
  auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
  if (!ConstOp1)
    return;
  uint64_t CmpVal = ConstOp1->getZExtValue();
 
  // Check whether the nonconstant input is an AND with a constant mask.
  Comparison NewC(C);
  uint64_t MaskVal;
  ConstantSDNode *Mask = nullptr;
  if (C.Op0.getOpcode() == ISD::AND) {
    NewC.Op0 = C.Op0.getOperand(0);
    NewC.Op1 = C.Op0.getOperand(1);
    Mask = dyn_cast<ConstantSDNode>(NewC.Op1);
    if (!Mask)
      return;
    MaskVal = Mask->getZExtValue();
  } else {
    // There is no instruction to compare with a 64-bit immediate
    // so use TMHH instead if possible.  We need an unsigned ordered
    // comparison with an i64 immediate.
    if (NewC.Op0.getValueType() != MVT::i64 ||
        NewC.CCMask == SystemZ::CCMASK_CMP_EQ ||
        NewC.CCMask == SystemZ::CCMASK_CMP_NE ||
        NewC.ICmpType == SystemZICMP::SignedOnly)
      return;
    // Convert LE and GT comparisons into LT and GE.
    if (NewC.CCMask == SystemZ::CCMASK_CMP_LE ||
        NewC.CCMask == SystemZ::CCMASK_CMP_GT) {
      if (CmpVal == uint64_t(-1))
        return;
      CmpVal += 1;
      NewC.CCMask ^= SystemZ::CCMASK_CMP_EQ;
    }
    // If the low N bits of Op1 are zero than the low N bits of Op0 can
    // be masked off without changing the result.
    MaskVal = -(CmpVal & -CmpVal);
    NewC.ICmpType = SystemZICMP::UnsignedOnly;
  }
  if (!MaskVal)
    return;
 
  // Check whether the combination of mask, comparison value and comparison
  // type are suitable.
  unsigned BitSize = NewC.Op0.getValueSizeInBits();
  unsigned NewCCMask, ShiftVal;
  if (NewC.ICmpType != SystemZICMP::SignedOnly &&
      NewC.Op0.getOpcode() == ISD::SHL &&
      isSimpleShift(NewC.Op0, ShiftVal) &&
      (MaskVal >> ShiftVal != 0) &&
      ((CmpVal >> ShiftVal) << ShiftVal) == CmpVal &&
      (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
                                        MaskVal >> ShiftVal,
                                        CmpVal >> ShiftVal,
                                        SystemZICMP::Any))) {
    NewC.Op0 = NewC.Op0.getOperand(0);
    MaskVal >>= ShiftVal;
  } else if (NewC.ICmpType != SystemZICMP::SignedOnly &&
             NewC.Op0.getOpcode() == ISD::SRL &&
             isSimpleShift(NewC.Op0, ShiftVal) &&
             (MaskVal << ShiftVal != 0) &&
             ((CmpVal << ShiftVal) >> ShiftVal) == CmpVal &&
             (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
                                               MaskVal << ShiftVal,
                                               CmpVal << ShiftVal,
                                               SystemZICMP::UnsignedOnly))) {
    NewC.Op0 = NewC.Op0.getOperand(0);
    MaskVal <<= ShiftVal;
  } else {
    NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask, MaskVal, CmpVal,
                                     NewC.ICmpType);
    if (!NewCCMask)
      return;
  }
 
  // Go ahead and make the change.
  C.Opcode = SystemZISD::TM;
  C.Op0 = NewC.Op0;
  if (Mask && Mask->getZExtValue() == MaskVal)
    C.Op1 = SDValue(Mask, 0);
  else
    C.Op1 = DAG.getConstant(MaskVal, DL, C.Op0.getValueType());
  C.CCValid = SystemZ::CCMASK_TM;
  C.CCMask = NewCCMask;
}
 
// Implement i128 comparison in vector registers.
static void adjustICmp128(SelectionDAG &DAG, const SDLoc &DL,
                          Comparison &C) {
  if (C.Opcode != SystemZISD::ICMP)
    return;
  if (C.Op0.getValueType() != MVT::i128)
    return;
 
  // Recognize vector comparison reductions.
  if ((C.CCMask == SystemZ::CCMASK_CMP_EQ ||
       C.CCMask == SystemZ::CCMASK_CMP_NE) &&
      (isNullConstant(C.Op1) || isAllOnesConstant(C.Op1))) {
    bool CmpEq = C.CCMask == SystemZ::CCMASK_CMP_EQ;
    bool CmpNull = isNullConstant(C.Op1);
    SDValue Src = peekThroughBitcasts(C.Op0);
    if (Src.hasOneUse() && isBitwiseNot(Src)) {
      Src = Src.getOperand(0);
      CmpNull = !CmpNull;
    }
    unsigned Opcode = 0;
    if (Src.hasOneUse()) {
      switch (Src.getOpcode()) {
      case SystemZISD::VICMPE: Opcode = SystemZISD::VICMPES; break;
      case SystemZISD::VICMPH: Opcode = SystemZISD::VICMPHS; break;
      case SystemZISD::VICMPHL: Opcode = SystemZISD::VICMPHLS; break;
      case SystemZISD::VFCMPE: Opcode = SystemZISD::VFCMPES; break;
      case SystemZISD::VFCMPH: Opcode = SystemZISD::VFCMPHS; break;
      case SystemZISD::VFCMPHE: Opcode = SystemZISD::VFCMPHES; break;
      default: break;
      }
    }
    if (Opcode) {
      C.Opcode = Opcode;
      C.Op0 = Src->getOperand(0);
      C.Op1 = Src->getOperand(1);
      C.CCValid = SystemZ::CCMASK_VCMP;
      C.CCMask = CmpNull ? SystemZ::CCMASK_VCMP_NONE : SystemZ::CCMASK_VCMP_ALL;
      if (!CmpEq)
        C.CCMask ^= C.CCValid;
      return;
    }
  }
 
  // Everything below here is not useful if we have native i128 compares.
  if (DAG.getSubtarget<SystemZSubtarget>().hasVectorEnhancements3())
    return;
 
  // (In-)Equality comparisons can be implemented via VCEQGS.
  if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
      C.CCMask == SystemZ::CCMASK_CMP_NE) {
    C.Opcode = SystemZISD::VICMPES;
    C.Op0 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, C.Op0);
    C.Op1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, C.Op1);
    C.CCValid = SystemZ::CCMASK_VCMP;
    if (C.CCMask == SystemZ::CCMASK_CMP_EQ)
      C.CCMask = SystemZ::CCMASK_VCMP_ALL;
    else
      C.CCMask = SystemZ::CCMASK_VCMP_ALL ^ C.CCValid;
    return;
  }
 
  // Normalize other comparisons to GT.
  bool Swap = false, Invert = false;
  switch (C.CCMask) {
    case SystemZ::CCMASK_CMP_GT: break;
    case SystemZ::CCMASK_CMP_LT: Swap = true; break;
    case SystemZ::CCMASK_CMP_LE: Invert = true; break;
    case SystemZ::CCMASK_CMP_GE: Swap = Invert = true; break;
    default: llvm_unreachable("Invalid integer condition!");
  }
  if (Swap)
    std::swap(C.Op0, C.Op1);
 
  if (C.ICmpType == SystemZICMP::UnsignedOnly)
    C.Opcode = SystemZISD::UCMP128HI;
  else
    C.Opcode = SystemZISD::SCMP128HI;
  C.CCValid = SystemZ::CCMASK_ANY;
  C.CCMask = SystemZ::CCMASK_1;
 
  if (Invert)
    C.CCMask ^= C.CCValid;
}
 
// See whether the comparison argument contains a redundant AND
// and remove it if so.  This sometimes happens due to the generic
// BRCOND expansion.
static void adjustForRedundantAnd(SelectionDAG &DAG, const SDLoc &DL,
                                  Comparison &C) {
  if (C.Op0.getOpcode() != ISD::AND)
    return;
  auto *Mask = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
  if (!Mask || Mask->getValueSizeInBits(0) > 64)
    return;
  KnownBits Known = DAG.computeKnownBits(C.Op0.getOperand(0));
  if ((~Known.Zero).getZExtValue() & ~Mask->getZExtValue())
    return;
 
  C.Op0 = C.Op0.getOperand(0);
}
 
// Return a Comparison that tests the condition-code result of intrinsic
// node Call against constant integer CC using comparison code Cond.
// Opcode is the opcode of the SystemZISD operation for the intrinsic
// and CCValid is the set of possible condition-code results.
static Comparison getIntrinsicCmp(SelectionDAG &DAG, unsigned Opcode,
                                  SDValue Call, unsigned CCValid, uint64_t CC,
                                  ISD::CondCode Cond) {
  Comparison C(Call, SDValue(), SDValue());
  C.Opcode = Opcode;
  C.CCValid = CCValid;
  if (Cond == ISD::SETEQ)
    // bit 3 for CC==0, bit 0 for CC==3, always false for CC>3.
    C.CCMask = CC < 4 ? 1 << (3 - CC) : 0;
  else if (Cond == ISD::SETNE)
    // ...and the inverse of that.
    C.CCMask = CC < 4 ? ~(1 << (3 - CC)) : -1;
  else if (Cond == ISD::SETLT || Cond == ISD::SETULT)
    // bits above bit 3 for CC==0 (always false), bits above bit 0 for CC==3,
    // always true for CC>3.
    C.CCMask = CC < 4 ? ~0U << (4 - CC) : -1;
  else if (Cond == ISD::SETGE || Cond == ISD::SETUGE)
    // ...and the inverse of that.
    C.CCMask = CC < 4 ? ~(~0U << (4 - CC)) : 0;
  else if (Cond == ISD::SETLE || Cond == ISD::SETULE)
    // bit 3 and above for CC==0, bit 0 and above for CC==3 (always true),
    // always true for CC>3.
    C.CCMask = CC < 4 ? ~0U << (3 - CC) : -1;
  else if (Cond == ISD::SETGT || Cond == ISD::SETUGT)
    // ...and the inverse of that.
    C.CCMask = CC < 4 ? ~(~0U << (3 - CC)) : 0;
  else
    llvm_unreachable("Unexpected integer comparison type");
  C.CCMask &= CCValid;
  return C;
}
 
// Decide how to implement a comparison of type Cond between CmpOp0 with CmpOp1.
static Comparison getCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
                         ISD::CondCode Cond, const SDLoc &DL,
                         SDValue Chain = SDValue(),
                         bool IsSignaling = false) {
  if (CmpOp1.getOpcode() == ISD::Constant) {
    assert(!Chain);
    unsigned Opcode, CCValid;
    if (CmpOp0.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
        CmpOp0.getResNo() == 0 && CmpOp0->hasNUsesOfValue(1, 0) &&
        isIntrinsicWithCCAndChain(CmpOp0, Opcode, CCValid))
      return getIntrinsicCmp(DAG, Opcode, CmpOp0, CCValid,
                             CmpOp1->getAsZExtVal(), Cond);
    if (CmpOp0.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
        CmpOp0.getResNo() == CmpOp0->getNumValues() - 1 &&
        isIntrinsicWithCC(CmpOp0, Opcode, CCValid))
      return getIntrinsicCmp(DAG, Opcode, CmpOp0, CCValid,
                             CmpOp1->getAsZExtVal(), Cond);
  }
  Comparison C(CmpOp0, CmpOp1, Chain);
  C.CCMask = CCMaskForCondCode(Cond);
  if (C.Op0.getValueType().isFloatingPoint()) {
    C.CCValid = SystemZ::CCMASK_FCMP;
    if (!C.Chain)
      C.Opcode = SystemZISD::FCMP;
    else if (!IsSignaling)
      C.Opcode = SystemZISD::STRICT_FCMP;
    else
      C.Opcode = SystemZISD::STRICT_FCMPS;
    adjustForFNeg(C);
  } else {
    assert(!C.Chain);
    C.CCValid = SystemZ::CCMASK_ICMP;
    C.Opcode = SystemZISD::ICMP;
    // Choose the type of comparison.  Equality and inequality tests can
    // use either signed or unsigned comparisons.  The choice also doesn't
    // matter if both sign bits are known to be clear.  In those cases we
    // want to give the main isel code the freedom to choose whichever
    // form fits best.
    if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
        C.CCMask == SystemZ::CCMASK_CMP_NE ||
        (DAG.SignBitIsZero(C.Op0) && DAG.SignBitIsZero(C.Op1)))
      C.ICmpType = SystemZICMP::Any;
    else if (C.CCMask & SystemZ::CCMASK_CMP_UO)
      C.ICmpType = SystemZICMP::UnsignedOnly;
    else
      C.ICmpType = SystemZICMP::SignedOnly;
    C.CCMask &= ~SystemZ::CCMASK_CMP_UO;
    adjustForRedundantAnd(DAG, DL, C);
    adjustZeroCmp(DAG, DL, C);
    adjustSubwordCmp(DAG, DL, C);
    adjustForSubtraction(DAG, DL, C);
    adjustForLTGFR(C);
    adjustICmpTruncate(DAG, DL, C);
  }
 
  if (shouldSwapCmpOperands(C)) {
    std::swap(C.Op0, C.Op1);
    C.CCMask = SystemZ::reverseCCMask(C.CCMask);
  }
 
  adjustForTestUnderMask(DAG, DL, C);
  adjustICmp128(DAG, DL, C);
  adjustForStackGuardCompare(DAG, DL, C);
  return C;
}
 
// Emit the comparison instruction described by C.
static SDValue emitCmp(SelectionDAG &DAG, const SDLoc &DL, Comparison &C) {
  if (!C.Op1.getNode()) {
    if (C.Opcode == SystemZISD::CMP_STACKGUARD)
      return DAG.getNode(SystemZISD::CMP_STACKGUARD, DL, MVT::i32, C.Op0);
    SDNode *Node;
    switch (C.Op0.getOpcode()) {
    case ISD::INTRINSIC_W_CHAIN:
      Node = emitIntrinsicWithCCAndChain(DAG, C.Op0, C.Opcode);
      return SDValue(Node, 0);
    case ISD::INTRINSIC_WO_CHAIN:
      Node = emitIntrinsicWithCC(DAG, C.Op0, C.Opcode);
      return SDValue(Node, Node->getNumValues() - 1);
    default:
      llvm_unreachable("Invalid comparison operands");
    }
  }
  if (C.Opcode == SystemZISD::ICMP)
    return DAG.getNode(SystemZISD::ICMP, DL, MVT::i32, C.Op0, C.Op1,
                       DAG.getTargetConstant(C.ICmpType, DL, MVT::i32));
  if (C.Opcode == SystemZISD::TM) {
    bool RegisterOnly = (bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
                         bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
    return DAG.getNode(SystemZISD::TM, DL, MVT::i32, C.Op0, C.Op1,
                       DAG.getTargetConstant(RegisterOnly, DL, MVT::i32));
  }
  if (C.Opcode == SystemZISD::VICMPES ||
      C.Opcode == SystemZISD::VICMPHS ||
      C.Opcode == SystemZISD::VICMPHLS ||
      C.Opcode == SystemZISD::VFCMPES ||
      C.Opcode == SystemZISD::VFCMPHS ||
      C.Opcode == SystemZISD::VFCMPHES) {
    EVT IntVT = C.Op0.getValueType().changeVectorElementTypeToInteger();
    SDVTList VTs = DAG.getVTList(IntVT, MVT::i32);
    SDValue Val = DAG.getNode(C.Opcode, DL, VTs, C.Op0, C.Op1);
    return SDValue(Val.getNode(), 1);
  }
  if (C.Chain) {
    SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other);
    return DAG.getNode(C.Opcode, DL, VTs, C.Chain, C.Op0, C.Op1);
  }
  return DAG.getNode(C.Opcode, DL, MVT::i32, C.Op0, C.Op1);
}
 
// Implement a 32-bit *MUL_LOHI operation by extending both operands to
// 64 bits.  Extend is the extension type to use.  Store the high part
// in Hi and the low part in Lo.
static void lowerMUL_LOHI32(SelectionDAG &DAG, const SDLoc &DL, unsigned Extend,
                            SDValue Op0, SDValue Op1, SDValue &Hi,
                            SDValue &Lo) {
  Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0);
  Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1);
  SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1);
  Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
                   DAG.getConstant(32, DL, MVT::i64));
  Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi);
  Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);
}
 
// Lower a binary operation that produces two VT results, one in each
// half of a GR128 pair.  Op0 and Op1 are the VT operands to the operation,
// and Opcode performs the GR128 operation.  Store the even register result
// in Even and the odd register result in Odd.
static void lowerGR128Binary(SelectionDAG &DAG, const SDLoc &DL, EVT VT,
                             unsigned Opcode, SDValue Op0, SDValue Op1,
                             SDValue &Even, SDValue &Odd) {
  SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped, Op0, Op1);
  bool Is32Bit = is32Bit(VT);
  Even = DAG.getTargetExtractSubreg(SystemZ::even128(Is32Bit), DL, VT, Result);
  Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
}
 
// Return an i32 value that is 1 if the CC value produced by CCReg is
// in the mask CCMask and 0 otherwise.  CC is known to have a value
// in CCValid, so other values can be ignored.
static SDValue emitSETCC(SelectionDAG &DAG, const SDLoc &DL, SDValue CCReg,
                         unsigned CCValid, unsigned CCMask) {
  SDValue Ops[] = {DAG.getConstant(1, DL, MVT::i32),
                   DAG.getConstant(0, DL, MVT::i32),
                   DAG.getTargetConstant(CCValid, DL, MVT::i32),
                   DAG.getTargetConstant(CCMask, DL, MVT::i32), CCReg};
  return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, MVT::i32, Ops);
}
 
// Return the SystemISD vector comparison operation for CC, or 0 if it cannot
// be done directly.  Mode is CmpMode::Int for integer comparisons, CmpMode::FP
// for regular floating-point comparisons, CmpMode::StrictFP for strict (quiet)
// floating-point comparisons, and CmpMode::SignalingFP for strict signaling
// floating-point comparisons.
enum class CmpMode { Int, FP, StrictFP, SignalingFP };
static unsigned getVectorComparison(ISD::CondCode CC, CmpMode Mode) {
  switch (CC) {
  case ISD::SETOEQ:
  case ISD::SETEQ:
    switch (Mode) {
    case CmpMode::Int:         return SystemZISD::VICMPE;
    case CmpMode::FP:          return SystemZISD::VFCMPE;
    case CmpMode::StrictFP:    return SystemZISD::STRICT_VFCMPE;
    case CmpMode::SignalingFP: return SystemZISD::STRICT_VFCMPES;
    }
    llvm_unreachable("Bad mode");
 
  case ISD::SETOGE:
  case ISD::SETGE:
    switch (Mode) {
    case CmpMode::Int:         return 0;
    case CmpMode::FP:          return SystemZISD::VFCMPHE;
    case CmpMode::StrictFP:    return SystemZISD::STRICT_VFCMPHE;
    case CmpMode::SignalingFP: return SystemZISD::STRICT_VFCMPHES;
    }
    llvm_unreachable("Bad mode");
 
  case ISD::SETOGT:
  case ISD::SETGT:
    switch (Mode) {
    case CmpMode::Int:         return SystemZISD::VICMPH;
    case CmpMode::FP:          return SystemZISD::VFCMPH;
    case CmpMode::StrictFP:    return SystemZISD::STRICT_VFCMPH;
    case CmpMode::SignalingFP: return SystemZISD::STRICT_VFCMPHS;
    }
    llvm_unreachable("Bad mode");
 
  case ISD::SETUGT:
    switch (Mode) {
    case CmpMode::Int:         return SystemZISD::VICMPHL;
    case CmpMode::FP:          return 0;
    case CmpMode::StrictFP:    return 0;
    case CmpMode::SignalingFP: return 0;
    }
    llvm_unreachable("Bad mode");
 
  default:
    return 0;
  }
}
 
// Return the SystemZISD vector comparison operation for CC or its inverse,
// or 0 if neither can be done directly.  Indicate in Invert whether the
// result is for the inverse of CC.  Mode is as above.
static unsigned getVectorComparisonOrInvert(ISD::CondCode CC, CmpMode Mode,
                                            bool &Invert) {
  if (unsigned Opcode = getVectorComparison(CC, Mode)) {
    Invert = false;
    return Opcode;
  }
 
  CC = ISD::getSetCCInverse(CC, Mode == CmpMode::Int ? MVT::i32 : MVT::f32);
  if (unsigned Opcode = getVectorComparison(CC, Mode)) {
    Invert = true;
    return Opcode;
  }
 
  return 0;
}
 
// Return a v2f64 that contains the extended form of elements Start and Start+1
// of v4f32 value Op.  If Chain is nonnull, return the strict form.
static SDValue expandV4F32ToV2F64(SelectionDAG &DAG, int Start, const SDLoc &DL,
                                  SDValue Op, SDValue Chain) {
  int Mask[] = { Start, -1, Start + 1, -1 };
  Op = DAG.getVectorShuffle(MVT::v4f32, DL, Op, DAG.getUNDEF(MVT::v4f32), Mask);
  if (Chain) {
    SDVTList VTs = DAG.getVTList(MVT::v2f64, MVT::Other);
    return DAG.getNode(SystemZISD::STRICT_VEXTEND, DL, VTs, Chain, Op);
  }
  return DAG.getNode(SystemZISD::VEXTEND, DL, MVT::v2f64, Op);
}
 
// Build a comparison of vectors CmpOp0 and CmpOp1 using opcode Opcode,
// producing a result of type VT.  If Chain is nonnull, return the strict form.
SDValue SystemZTargetLowering::getVectorCmp(SelectionDAG &DAG, unsigned Opcode,
                                            const SDLoc &DL, EVT VT,
                                            SDValue CmpOp0,
                                            SDValue CmpOp1,
                                            SDValue Chain) const {
  // There is no hardware support for v4f32 (unless we have the vector
  // enhancements facility 1), so extend the vector into two v2f64s
  // and compare those.
  if (CmpOp0.getValueType() == MVT::v4f32 &&
      !Subtarget.hasVectorEnhancements1()) {
    SDValue H0 = expandV4F32ToV2F64(DAG, 0, DL, CmpOp0, Chain);
    SDValue L0 = expandV4F32ToV2F64(DAG, 2, DL, CmpOp0, Chain);
    SDValue H1 = expandV4F32ToV2F64(DAG, 0, DL, CmpOp1, Chain);
    SDValue L1 = expandV4F32ToV2F64(DAG, 2, DL, CmpOp1, Chain);
    if (Chain) {
      SDVTList VTs = DAG.getVTList(MVT::v2i64, MVT::Other);
      SDValue HRes = DAG.getNode(Opcode, DL, VTs, Chain, H0, H1);
      SDValue LRes = DAG.getNode(Opcode, DL, VTs, Chain, L0, L1);
      SDValue Res = DAG.getNode(SystemZISD::PACK, DL, VT, HRes, LRes);
      SDValue Chains[6] = { H0.getValue(1), L0.getValue(1),
                            H1.getValue(1), L1.getValue(1),
                            HRes.getValue(1), LRes.getValue(1) };
      SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
      SDValue Ops[2] = { Res, NewChain };
      return DAG.getMergeValues(Ops, DL);
    }
    SDValue HRes = DAG.getNode(Opcode, DL, MVT::v2i64, H0, H1);
    SDValue LRes = DAG.getNode(Opcode, DL, MVT::v2i64, L0, L1);
    return DAG.getNode(SystemZISD::PACK, DL, VT, HRes, LRes);
  }
  if (Chain) {
    SDVTList VTs = DAG.getVTList(VT, MVT::Other);
    return DAG.getNode(Opcode, DL, VTs, Chain, CmpOp0, CmpOp1);
  }
  return DAG.getNode(Opcode, DL, VT, CmpOp0, CmpOp1);
}
 
// Lower a vector comparison of type CC between CmpOp0 and CmpOp1, producing
// an integer mask of type VT.  If Chain is nonnull, we have a strict
// floating-point comparison.  If in addition IsSignaling is true, we have
// a strict signaling floating-point comparison.
SDValue SystemZTargetLowering::lowerVectorSETCC(SelectionDAG &DAG,
                                                const SDLoc &DL, EVT VT,
                                                ISD::CondCode CC,
                                                SDValue CmpOp0,
                                                SDValue CmpOp1,
                                                SDValue Chain,
                                                bool IsSignaling) const {
  bool IsFP = CmpOp0.getValueType().isFloatingPoint();
  assert (!Chain || IsFP);
  assert (!IsSignaling || Chain);
  CmpMode Mode = IsSignaling ? CmpMode::SignalingFP :
                 Chain ? CmpMode::StrictFP : IsFP ? CmpMode::FP : CmpMode::Int;
  bool Invert = false;
  SDValue Cmp;
  switch (CC) {
    // Handle tests for order using (or (ogt y x) (oge x y)).
  case ISD::SETUO:
    Invert = true;
    [[fallthrough]];
  case ISD::SETO: {
    assert(IsFP && "Unexpected integer comparison");
    SDValue LT = getVectorCmp(DAG, getVectorComparison(ISD::SETOGT, Mode),
                              DL, VT, CmpOp1, CmpOp0, Chain);
    SDValue GE = getVectorCmp(DAG, getVectorComparison(ISD::SETOGE, Mode),
                              DL, VT, CmpOp0, CmpOp1, Chain);
    Cmp = DAG.getNode(ISD::OR, DL, VT, LT, GE);
    if (Chain)
      Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
                          LT.getValue(1), GE.getValue(1));
    break;
  }
 
    // Handle <> tests using (or (ogt y x) (ogt x y)).
  case ISD::SETUEQ:
    Invert = true;
    [[fallthrough]];
  case ISD::SETONE: {
    assert(IsFP && "Unexpected integer comparison");
    SDValue LT = getVectorCmp(DAG, getVectorComparison(ISD::SETOGT, Mode),
                              DL, VT, CmpOp1, CmpOp0, Chain);
    SDValue GT = getVectorCmp(DAG, getVectorComparison(ISD::SETOGT, Mode),
                              DL, VT, CmpOp0, CmpOp1, Chain);
    Cmp = DAG.getNode(ISD::OR, DL, VT, LT, GT);
    if (Chain)
      Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
                          LT.getValue(1), GT.getValue(1));
    break;
  }
 
    // Otherwise a single comparison is enough.  It doesn't really
    // matter whether we try the inversion or the swap first, since
    // there are no cases where both work.
  default:
    // Optimize sign-bit comparisons to signed compares.
    if (Mode == CmpMode::Int && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
        ISD::isConstantSplatVectorAllZeros(CmpOp1.getNode())) {
      unsigned EltSize = VT.getVectorElementType().getSizeInBits();
      APInt Mask;
      if (CmpOp0.getOpcode() == ISD::AND
          && ISD::isConstantSplatVector(CmpOp0.getOperand(1).getNode(), Mask)
          && Mask == APInt::getSignMask(EltSize)) {
        CC = CC == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
        CmpOp0 = CmpOp0.getOperand(0);
      }
    }
    if (unsigned Opcode = getVectorComparisonOrInvert(CC, Mode, Invert))
      Cmp = getVectorCmp(DAG, Opcode, DL, VT, CmpOp0, CmpOp1, Chain);
    else {
      CC = ISD::getSetCCSwappedOperands(CC);
      if (unsigned Opcode = getVectorComparisonOrInvert(CC, Mode, Invert))
        Cmp = getVectorCmp(DAG, Opcode, DL, VT, CmpOp1, CmpOp0, Chain);
      else
        llvm_unreachable("Unhandled comparison");
    }
    if (Chain)
      Chain = Cmp.getValue(1);
    break;
  }
  if (Invert) {
    SDValue Mask =
        DAG.getSplatBuildVector(VT, DL, DAG.getAllOnesConstant(DL, MVT::i64));
    Cmp = DAG.getNode(ISD::XOR, DL, VT, Cmp, Mask);
  }
  if (Chain && Chain.getNode() != Cmp.getNode()) {
    SDValue Ops[2] = { Cmp, Chain };
    Cmp = DAG.getMergeValues(Ops, DL);
  }
  return Cmp;
}
 
SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
                                          SelectionDAG &DAG) const {
  SDValue CmpOp0   = Op.getOperand(0);
  SDValue CmpOp1   = Op.getOperand(1);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  if (VT.isVector())
    return lowerVectorSETCC(DAG, DL, VT, CC, CmpOp0, CmpOp1);
 
  Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC, DL));
  SDValue CCReg = emitCmp(DAG, DL, C);
  return emitSETCC(DAG, DL, CCReg, C.CCValid, C.CCMask);
}
 
SDValue SystemZTargetLowering::lowerSTRICT_FSETCC(SDValue Op,
                                                  SelectionDAG &DAG,
                                                  bool IsSignaling) const {
  SDValue Chain    = Op.getOperand(0);
  SDValue CmpOp0   = Op.getOperand(1);
  SDValue CmpOp1   = Op.getOperand(2);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(3))->get();
  SDLoc DL(Op);
  EVT VT = Op.getNode()->getValueType(0);
  if (VT.isVector()) {
    SDValue Res = lowerVectorSETCC(DAG, DL, VT, CC, CmpOp0, CmpOp1,
                                   Chain, IsSignaling);
    return Res.getValue(Op.getResNo());
  }
 
  Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC, DL, Chain, IsSignaling));
  SDValue CCReg = emitCmp(DAG, DL, C);
  CCReg->setFlags(Op->getFlags());
  SDValue Result = emitSETCC(DAG, DL, CCReg, C.CCValid, C.CCMask);
  SDValue Ops[2] = { Result, CCReg.getValue(1) };
  return DAG.getMergeValues(Ops, DL);
}
 
SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
  SDValue CmpOp0   = Op.getOperand(2);
  SDValue CmpOp1   = Op.getOperand(3);
  SDValue Dest     = Op.getOperand(4);
  SDLoc DL(Op);
 
  Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC, DL));
  SDValue CCReg = emitCmp(DAG, DL, C);
  return DAG.getNode(
      SystemZISD::BR_CCMASK, DL, Op.getValueType(), Op.getOperand(0),
      DAG.getTargetConstant(C.CCValid, DL, MVT::i32),
      DAG.getTargetConstant(C.CCMask, DL, MVT::i32), Dest, CCReg);
}
 
// Return true if Pos is CmpOp and Neg is the negative of CmpOp,
// allowing Pos and Neg to be wider than CmpOp.
static bool isAbsolute(SDValue CmpOp, SDValue Pos, SDValue Neg) {
  return (Neg.getOpcode() == ISD::SUB &&
          Neg.getOperand(0).getOpcode() == ISD::Constant &&
          Neg.getConstantOperandVal(0) == 0 && Neg.getOperand(1) == Pos &&
          (Pos == CmpOp || (Pos.getOpcode() == ISD::SIGN_EXTEND &&
                            Pos.getOperand(0) == CmpOp)));
}
 
// Return the absolute or negative absolute of Op; IsNegative decides which.
static SDValue getAbsolute(SelectionDAG &DAG, const SDLoc &DL, SDValue Op,
                           bool IsNegative) {
  Op = DAG.getNode(ISD::ABS, DL, Op.getValueType(), Op);
  if (IsNegative)
    Op = DAG.getNode(ISD::SUB, DL, Op.getValueType(),
                     DAG.getConstant(0, DL, Op.getValueType()), Op);
  return Op;
}
 
static SDValue getI128Select(SelectionDAG &DAG, const SDLoc &DL,
                             Comparison C, SDValue TrueOp, SDValue FalseOp) {
  EVT VT = MVT::i128;
  unsigned Op;
 
  if (C.CCMask == SystemZ::CCMASK_CMP_NE ||
      C.CCMask == SystemZ::CCMASK_CMP_GE ||
      C.CCMask == SystemZ::CCMASK_CMP_LE) {
    std::swap(TrueOp, FalseOp);
    C.CCMask ^= C.CCValid;
  }
  if (C.CCMask == SystemZ::CCMASK_CMP_LT) {
    std::swap(C.Op0, C.Op1);
    C.CCMask = SystemZ::CCMASK_CMP_GT;
  }
  switch (C.CCMask) {
  case SystemZ::CCMASK_CMP_EQ:
    Op = SystemZISD::VICMPE;
    break;
  case SystemZ::CCMASK_CMP_GT:
    if (C.ICmpType == SystemZICMP::UnsignedOnly)
      Op = SystemZISD::VICMPHL;
    else
      Op = SystemZISD::VICMPH;
    break;
  default:
    llvm_unreachable("Unhandled comparison");
    break;
  }
 
  SDValue Mask = DAG.getNode(Op, DL, VT, C.Op0, C.Op1);
  TrueOp = DAG.getNode(ISD::AND, DL, VT, TrueOp, Mask);
  FalseOp = DAG.getNode(ISD::AND, DL, VT, FalseOp, DAG.getNOT(DL, Mask, VT));
  return DAG.getNode(ISD::OR, DL, VT, TrueOp, FalseOp);
}
 
SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
                                              SelectionDAG &DAG) const {
  SDValue CmpOp0   = Op.getOperand(0);
  SDValue CmpOp1   = Op.getOperand(1);
  SDValue TrueOp   = Op.getOperand(2);
  SDValue FalseOp  = Op.getOperand(3);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
  SDLoc DL(Op);
 
  // SELECT_CC involving f16 will not have the cmp-ops promoted by the
  // legalizer, as it will be handled according to the type of the resulting
  // value. Extend them here if needed.
  if (CmpOp0.getSimpleValueType() == MVT::f16) {
    CmpOp0 = DAG.getFPExtendOrRound(CmpOp0, SDLoc(CmpOp0), MVT::f32);
    CmpOp1 = DAG.getFPExtendOrRound(CmpOp1, SDLoc(CmpOp1), MVT::f32);
  }
 
  Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC, DL));
 
  // Check for absolute and negative-absolute selections, including those
  // where the comparison value is sign-extended (for LPGFR and LNGFR).
  // This check supplements the one in DAGCombiner.
  if (C.Opcode == SystemZISD::ICMP && C.CCMask != SystemZ::CCMASK_CMP_EQ &&
      C.CCMask != SystemZ::CCMASK_CMP_NE &&
      C.Op1.getOpcode() == ISD::Constant &&
      cast<ConstantSDNode>(C.Op1)->getValueSizeInBits(0) <= 64 &&
      C.Op1->getAsZExtVal() == 0) {
    if (isAbsolute(C.Op0, TrueOp, FalseOp))
      return getAbsolute(DAG, DL, TrueOp, C.CCMask & SystemZ::CCMASK_CMP_LT);
    if (isAbsolute(C.Op0, FalseOp, TrueOp))
      return getAbsolute(DAG, DL, FalseOp, C.CCMask & SystemZ::CCMASK_CMP_GT);
  }
 
  if (Subtarget.hasVectorEnhancements3() &&
      C.Opcode == SystemZISD::ICMP &&
      C.Op0.getValueType() == MVT::i128 &&
      TrueOp.getValueType() == MVT::i128) {
    return getI128Select(DAG, DL, C, TrueOp, FalseOp);
  }
 
  SDValue CCReg = emitCmp(DAG, DL, C);
  SDValue Ops[] = {TrueOp, FalseOp,
                   DAG.getTargetConstant(C.CCValid, DL, MVT::i32),
                   DAG.getTargetConstant(C.CCMask, DL, MVT::i32), CCReg};
 
  return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, Op.getValueType(), Ops);
}
 
SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
                                                  SelectionDAG &DAG) const {
  SDLoc DL(Node);
  const GlobalValue *GV = Node->getGlobal();
  int64_t Offset = Node->getOffset();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  CodeModel::Model CM = DAG.getTarget().getCodeModel();
 
  SDValue Result;
  if (Subtarget.isPC32DBLSymbol(GV, CM)) {
    if (isInt<32>(Offset)) {
      // Assign anchors at 1<<12 byte boundaries.
      uint64_t Anchor = Offset & ~uint64_t(0xfff);
      Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor);
      Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
 
      // The offset can be folded into the address if it is aligned to a
      // halfword.
      Offset -= Anchor;
      if (Offset != 0 && (Offset & 1) == 0) {
        SDValue Full =
          DAG.getTargetGlobalAddress(GV, DL, PtrVT, Anchor + Offset);
        Result = DAG.getNode(SystemZISD::PCREL_OFFSET, DL, PtrVT, Full, Result);
        Offset = 0;
      }
    } else {
      // Conservatively load a constant offset greater than 32 bits into a
      // register below.
      Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT);
      Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
    }
  } else if (Subtarget.isTargetELF()) {
    Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT);
    Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
    Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
                         MachinePointerInfo::getGOT(DAG.getMachineFunction()));
  } else if (Subtarget.isTargetzOS()) {
    Result = getADAEntry(DAG, GV, DL, PtrVT);
  } else
    llvm_unreachable("Unexpected Subtarget");
 
  // If there was a non-zero offset that we didn't fold, create an explicit
  // addition for it.
  if (Offset != 0)
    Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
                         DAG.getSignedConstant(Offset, DL, PtrVT));
 
  return Result;
}
 
SDValue SystemZTargetLowering::lowerTLSGetOffset(GlobalAddressSDNode *Node,
                                                 SelectionDAG &DAG,
                                                 unsigned Opcode,
                                                 SDValue GOTOffset) const {
  SDLoc DL(Node);
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDValue Chain = DAG.getEntryNode();
  SDValue Glue;
 
  if (DAG.getMachineFunction().getFunction().getCallingConv() ==
      CallingConv::GHC)
    report_fatal_error("In GHC calling convention TLS is not supported");
 
  // __tls_get_offset takes the GOT offset in %r2 and the GOT in %r12.
  SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
  Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R12D, GOT, Glue);
  Glue = Chain.getValue(1);
  Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R2D, GOTOffset, Glue);
  Glue = Chain.getValue(1);
 
  // The first call operand is the chain and the second is the TLS symbol.
  SmallVector<SDValue, 8> Ops;
  Ops.push_back(Chain);
  Ops.push_back(DAG.getTargetGlobalAddress(Node->getGlobal(), DL,
                                           Node->getValueType(0),
                                           0, 0));
 
  // Add argument registers to the end of the list so that they are
  // known live into the call.
  Ops.push_back(DAG.getRegister(SystemZ::R2D, PtrVT));
  Ops.push_back(DAG.getRegister(SystemZ::R12D, PtrVT));
 
  // Add a register mask operand representing the call-preserved registers.
  const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
  const uint32_t *Mask =
      TRI->getCallPreservedMask(DAG.getMachineFunction(), CallingConv::C);
  assert(Mask && "Missing call preserved mask for calling convention");
  Ops.push_back(DAG.getRegisterMask(Mask));
 
  // Glue the call to the argument copies.
  Ops.push_back(Glue);
 
  // Emit the call.
  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
  Chain = DAG.getNode(Opcode, DL, NodeTys, Ops);
  Glue = Chain.getValue(1);
 
  // Copy the return value from %r2.
  return DAG.getCopyFromReg(Chain, DL, SystemZ::R2D, PtrVT, Glue);
}
 
SDValue SystemZTargetLowering::lowerThreadPointer(const SDLoc &DL,
                                                  SelectionDAG &DAG) const {
  SDValue Chain = DAG.getEntryNode();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
 
  // The high part of the thread pointer is in access register 0.
  SDValue TPHi = DAG.getCopyFromReg(Chain, DL, SystemZ::A0, MVT::i32);
  TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi);
 
  // The low part of the thread pointer is in access register 1.
  SDValue TPLo = DAG.getCopyFromReg(Chain, DL, SystemZ::A1, MVT::i32);
  TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo);
 
  // Merge them into a single 64-bit address.
  SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi,
                                    DAG.getConstant(32, DL, PtrVT));
  return DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
}
 
SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
                                                     SelectionDAG &DAG) const {
  if (DAG.getTarget().useEmulatedTLS())
    return LowerToTLSEmulatedModel(Node, DAG);
  SDLoc DL(Node);
  const GlobalValue *GV = Node->getGlobal();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
 
  if (DAG.getMachineFunction().getFunction().getCallingConv() ==
      CallingConv::GHC)
    report_fatal_error("In GHC calling convention TLS is not supported");
 
  SDValue TP = lowerThreadPointer(DL, DAG);
 
  // Get the offset of GA from the thread pointer, based on the TLS model.
  SDValue Offset;
  switch (model) {
    case TLSModel::GeneralDynamic: {
      // Load the GOT offset of the tls_index (module ID / per-symbol offset).
      SystemZConstantPoolValue *CPV =
        SystemZConstantPoolValue::Create(GV, SystemZCP::TLSGD);
 
      Offset = DAG.getConstantPool(CPV, PtrVT, Align(8));
      Offset = DAG.getLoad(
          PtrVT, DL, DAG.getEntryNode(), Offset,
          MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
 
      // Call __tls_get_offset to retrieve the offset.
      Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_GDCALL, Offset);
      break;
    }
 
    case TLSModel::LocalDynamic: {
      // Load the GOT offset of the module ID.
      SystemZConstantPoolValue *CPV =
        SystemZConstantPoolValue::Create(GV, SystemZCP::TLSLDM);
 
      Offset = DAG.getConstantPool(CPV, PtrVT, Align(8));
      Offset = DAG.getLoad(
          PtrVT, DL, DAG.getEntryNode(), Offset,
          MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
 
      // Call __tls_get_offset to retrieve the module base offset.
      Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_LDCALL, Offset);
 
      // Note: The SystemZLDCleanupPass will remove redundant computations
      // of the module base offset.  Count total number of local-dynamic
      // accesses to trigger execution of that pass.
      SystemZMachineFunctionInfo* MFI =
        DAG.getMachineFunction().getInfo<SystemZMachineFunctionInfo>();
      MFI->incNumLocalDynamicTLSAccesses();
 
      // Add the per-symbol offset.
      CPV = SystemZConstantPoolValue::Create(GV, SystemZCP::DTPOFF);
 
      SDValue DTPOffset = DAG.getConstantPool(CPV, PtrVT, Align(8));
      DTPOffset = DAG.getLoad(
          PtrVT, DL, DAG.getEntryNode(), DTPOffset,
          MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
 
      Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Offset, DTPOffset);
      break;
    }
 
    case TLSModel::InitialExec: {
      // Load the offset from the GOT.
      Offset = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
                                          SystemZII::MO_INDNTPOFF);
      Offset = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Offset);
      Offset =
          DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Offset,
                      MachinePointerInfo::getGOT(DAG.getMachineFunction()));
      break;
    }
 
    case TLSModel::LocalExec: {
      // Force the offset into the constant pool and load it from there.
      SystemZConstantPoolValue *CPV =
        SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
 
      Offset = DAG.getConstantPool(CPV, PtrVT, Align(8));
      Offset = DAG.getLoad(
          PtrVT, DL, DAG.getEntryNode(), Offset,
          MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
      break;
    }
  }
 
  // Add the base and offset together.
  return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
}
 
SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node,
                                                 SelectionDAG &DAG) const {
  SDLoc DL(Node);
  const BlockAddress *BA = Node->getBlockAddress();
  int64_t Offset = Node->getOffset();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
 
  SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset);
  Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
  return Result;
}
 
SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT,
                                              SelectionDAG &DAG) const {
  SDLoc DL(JT);
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
 
  // Use LARL to load the address of the table.
  return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
}
 
SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP,
                                                 SelectionDAG &DAG) const {
  SDLoc DL(CP);
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
 
  SDValue Result;
  if (CP->isMachineConstantPoolEntry())
    Result =
        DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, CP->getAlign());
  else
    Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlign(),
                                       CP->getOffset());
 
  // Use LARL to load the address of the constant pool entry.
  return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result);
}
 
SDValue SystemZTargetLowering::lowerFRAMEADDR(SDValue Op,
                                              SelectionDAG &DAG) const {
  auto *TFL = Subtarget.getFrameLowering<SystemZFrameLowering>();
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  MFI.setFrameAddressIsTaken(true);
 
  SDLoc DL(Op);
  unsigned Depth = Op.getConstantOperandVal(0);
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
 
  // By definition, the frame address is the address of the back chain.  (In
  // the case of packed stack without backchain, return the address where the
  // backchain would have been stored. This will either be an unused space or
  // contain a saved register).
  int BackChainIdx = TFL->getOrCreateFramePointerSaveIndex(MF);
  SDValue BackChain = DAG.getFrameIndex(BackChainIdx, PtrVT);
 
  if (Depth > 0) {
    // FIXME The frontend should detect this case.
    if (!MF.getSubtarget<SystemZSubtarget>().hasBackChain())
      report_fatal_error("Unsupported stack frame traversal count");
 
    SDValue Offset = DAG.getConstant(TFL->getBackchainOffset(MF), DL, PtrVT);
    while (Depth--) {
      BackChain = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), BackChain,
                              MachinePointerInfo());
      BackChain = DAG.getNode(ISD::ADD, DL, PtrVT, BackChain, Offset);
    }
  }
 
  return BackChain;
}
 
SDValue SystemZTargetLowering::lowerRETURNADDR(SDValue Op,
                                               SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  MFI.setReturnAddressIsTaken(true);
 
  SDLoc DL(Op);
  unsigned Depth = Op.getConstantOperandVal(0);
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
 
  if (Depth > 0) {
    // FIXME The frontend should detect this case.
    if (!MF.getSubtarget<SystemZSubtarget>().hasBackChain())
      report_fatal_error("Unsupported stack frame traversal count");
 
    SDValue FrameAddr = lowerFRAMEADDR(Op, DAG);
    const auto *TFL = Subtarget.getFrameLowering<SystemZFrameLowering>();
    int Offset = TFL->getReturnAddressOffset(MF);
    SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, FrameAddr,
                              DAG.getSignedConstant(Offset, DL, PtrVT));
    return DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr,
                       MachinePointerInfo());
  }
 
  // Return R14D (Elf) / R7D (XPLINK), which has the return address. Mark it an
  // implicit live-in.
  SystemZCallingConventionRegisters *CCR = Subtarget.getSpecialRegisters();
  Register LinkReg = MF.addLiveIn(CCR->getReturnFunctionAddressRegister(),
                                  &SystemZ::GR64BitRegClass);
  return DAG.getCopyFromReg(DAG.getEntryNode(), DL, LinkReg, PtrVT);
}
 
SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op,
                                            SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue In = Op.getOperand(0);
  EVT InVT = In.getValueType();
  EVT ResVT = Op.getValueType();
 
  // Convert loads directly.  This is normally done by DAGCombiner,
  // but we need this case for bitcasts that are created during lowering
  // and which are then lowered themselves.
  if (auto *LoadN = dyn_cast<LoadSDNode>(In))
    if (ISD::isNormalLoad(LoadN)) {
      SDValue NewLoad = DAG.getLoad(ResVT, DL, LoadN->getChain(),
                                    LoadN->getBasePtr(), LoadN->getMemOperand());
      // Update the chain uses.
      DAG.ReplaceAllUsesOfValueWith(SDValue(LoadN, 1), NewLoad.getValue(1));
      return NewLoad;
    }
 
  if (InVT == MVT::i32 && ResVT == MVT::f32) {
    SDValue In64;
    if (Subtarget.hasHighWord()) {
      SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL,
                                       MVT::i64);
      In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
                                       MVT::i64, SDValue(U64, 0), In);
    } else {
      In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In);
      In64 = DAG.getNode(ISD::SHL, DL, MVT::i64, In64,
                         DAG.getConstant(32, DL, MVT::i64));
    }
    SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, In64);
    return DAG.getTargetExtractSubreg(SystemZ::subreg_h32,
                                      DL, MVT::f32, Out64);
  }
  if (InVT == MVT::f32 && ResVT == MVT::i32) {
    SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
    SDValue In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h32, DL,
                                             MVT::f64, SDValue(U64, 0), In);
    SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, In64);
    if (Subtarget.hasHighWord())
      return DAG.getTargetExtractSubreg(SystemZ::subreg_h32, DL,
                                        MVT::i32, Out64);
    SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64,
                                DAG.getConstant(32, DL, MVT::i64));
    return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift);
  }
  llvm_unreachable("Unexpected bitcast combination");
}
 
SDValue SystemZTargetLowering::lowerVASTART(SDValue Op,
                                            SelectionDAG &DAG) const {
 
  if (Subtarget.isTargetXPLINK64())
    return lowerVASTART_XPLINK(Op, DAG);
  else
    return lowerVASTART_ELF(Op, DAG);
}
 
SDValue SystemZTargetLowering::lowerVASTART_XPLINK(SDValue Op,
                                                   SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  SystemZMachineFunctionInfo *FuncInfo =
      MF.getInfo<SystemZMachineFunctionInfo>();
 
  SDLoc DL(Op);
 
  // vastart just stores the address of the VarArgsFrameIndex slot into the
  // memory location argument.
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
                      MachinePointerInfo(SV));
}
 
SDValue SystemZTargetLowering::lowerVASTART_ELF(SDValue Op,
                                                SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  SystemZMachineFunctionInfo *FuncInfo =
    MF.getInfo<SystemZMachineFunctionInfo>();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
 
  SDValue Chain   = Op.getOperand(0);
  SDValue Addr    = Op.getOperand(1);
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  SDLoc DL(Op);
 
  // The initial values of each field.
  const unsigned NumFields = 4;
  SDValue Fields[NumFields] = {
    DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), DL, PtrVT),
    DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), DL, PtrVT),
    DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT),
    DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT)
  };
 
  // Store each field into its respective slot.
  SDValue MemOps[NumFields];
  unsigned Offset = 0;
  for (unsigned I = 0; I < NumFields; ++I) {
    SDValue FieldAddr = Addr;
    if (Offset != 0)
      FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr,
                              DAG.getIntPtrConstant(Offset, DL));
    MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr,
                             MachinePointerInfo(SV, Offset));
    Offset += 8;
  }
  return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
}
 
SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
                                           SelectionDAG &DAG) const {
  SDValue Chain      = Op.getOperand(0);
  SDValue DstPtr     = Op.getOperand(1);
  SDValue SrcPtr     = Op.getOperand(2);
  const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
  const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
  SDLoc DL(Op);
 
  uint32_t Sz =
      Subtarget.isTargetXPLINK64() ? getTargetMachine().getPointerSize(0) : 32;
  return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(Sz, DL),
                       Align(8), Align(8), /*isVolatile*/ false,
                       /*AlwaysInline*/ false,
                       /*CI=*/nullptr, std::nullopt, MachinePointerInfo(DstSV),
                       MachinePointerInfo(SrcSV));
}
 
SDValue
SystemZTargetLowering::lowerDYNAMIC_STACKALLOC(SDValue Op,
                                               SelectionDAG &DAG) const {
  if (Subtarget.isTargetXPLINK64())
    return lowerDYNAMIC_STACKALLOC_XPLINK(Op, DAG);
  else
    return lowerDYNAMIC_STACKALLOC_ELF(Op, DAG);
}
 
SDValue
SystemZTargetLowering::lowerDYNAMIC_STACKALLOC_XPLINK(SDValue Op,
                                                      SelectionDAG &DAG) const {
  const TargetFrameLowering *TFI = Subtarget.getFrameLowering();
  MachineFunction &MF = DAG.getMachineFunction();
  bool RealignOpt = !MF.getFunction().hasFnAttribute("no-realign-stack");
  SDValue Chain = Op.getOperand(0);
  SDValue Size = Op.getOperand(1);
  SDValue Align = Op.getOperand(2);
  SDLoc DL(Op);
 
  // If user has set the no alignment function attribute, ignore
  // alloca alignments.
  uint64_t AlignVal = (RealignOpt ? Align->getAsZExtVal() : 0);
 
  uint64_t StackAlign = TFI->getStackAlignment();
  uint64_t RequiredAlign = std::max(AlignVal, StackAlign);
  uint64_t ExtraAlignSpace = RequiredAlign - StackAlign;
 
  SDValue NeededSpace = Size;
 
  // Add extra space for alignment if needed.
  EVT PtrVT = getPointerTy(MF.getDataLayout());
  if (ExtraAlignSpace)
    NeededSpace = DAG.getNode(ISD::ADD, DL, PtrVT, NeededSpace,
                              DAG.getConstant(ExtraAlignSpace, DL, PtrVT));
 
  bool IsSigned = false;
  bool DoesNotReturn = false;
  bool IsReturnValueUsed = false;
  EVT VT = Op.getValueType();
  SDValue AllocaCall =
      makeExternalCall(Chain, DAG, "@@ALCAXP", VT, ArrayRef(NeededSpace),
                       CallingConv::C, IsSigned, DL, DoesNotReturn,
                       IsReturnValueUsed)
          .first;
 
  // Perform a CopyFromReg from %GPR4 (stack pointer register). Chain and Glue
  // to end of call in order to ensure it isn't broken up from the call
  // sequence.
  auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>();
  Register SPReg = Regs.getStackPointerRegister();
  Chain = AllocaCall.getValue(1);
  SDValue Glue = AllocaCall.getValue(2);
  SDValue NewSPRegNode = DAG.getCopyFromReg(Chain, DL, SPReg, PtrVT, Glue);
  Chain = NewSPRegNode.getValue(1);
 
  MVT PtrMVT = getPointerMemTy(MF.getDataLayout());
  SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, PtrMVT);
  SDValue Result = DAG.getNode(ISD::ADD, DL, PtrMVT, NewSPRegNode, ArgAdjust);
 
  // Dynamically realign if needed.
  if (ExtraAlignSpace) {
    Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result,
                         DAG.getConstant(ExtraAlignSpace, DL, PtrVT));
    Result = DAG.getNode(ISD::AND, DL, PtrVT, Result,
                         DAG.getConstant(~(RequiredAlign - 1), DL, PtrVT));
  }
 
  SDValue Ops[2] = {Result, Chain};
  return DAG.getMergeValues(Ops, DL);
}
 
SDValue
SystemZTargetLowering::lowerDYNAMIC_STACKALLOC_ELF(SDValue Op,
                                                   SelectionDAG &DAG) const {
  const TargetFrameLowering *TFI = Subtarget.getFrameLowering();
  MachineFunction &MF = DAG.getMachineFunction();
  bool RealignOpt = !MF.getFunction().hasFnAttribute("no-realign-stack");
  bool StoreBackchain = MF.getSubtarget<SystemZSubtarget>().hasBackChain();
 
  SDValue Chain = Op.getOperand(0);
  SDValue Size  = Op.getOperand(1);
  SDValue Align = Op.getOperand(2);
  SDLoc DL(Op);
 
  // If user has set the no alignment function attribute, ignore
  // alloca alignments.
  uint64_t AlignVal = (RealignOpt ? Align->getAsZExtVal() : 0);
 
  uint64_t StackAlign = TFI->getStackAlignment();
  uint64_t RequiredAlign = std::max(AlignVal, StackAlign);
  uint64_t ExtraAlignSpace = RequiredAlign - StackAlign;
 
  Register SPReg = getStackPointerRegisterToSaveRestore();
  SDValue NeededSpace = Size;
 
  // Get a reference to the stack pointer.
  SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64);
 
  // If we need a backchain, save it now.
  SDValue Backchain;
  if (StoreBackchain)
    Backchain = DAG.getLoad(MVT::i64, DL, Chain, getBackchainAddress(OldSP, DAG),
                            MachinePointerInfo());
 
  // Add extra space for alignment if needed.
  if (ExtraAlignSpace)
    NeededSpace = DAG.getNode(ISD::ADD, DL, MVT::i64, NeededSpace,
                              DAG.getConstant(ExtraAlignSpace, DL, MVT::i64));
 
  // Get the new stack pointer value.
  SDValue NewSP;
  if (hasInlineStackProbe(MF)) {
    NewSP = DAG.getNode(SystemZISD::PROBED_ALLOCA, DL,
                DAG.getVTList(MVT::i64, MVT::Other), Chain, OldSP, NeededSpace);
    Chain = NewSP.getValue(1);
  }
  else {
    NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, NeededSpace);
    // Copy the new stack pointer back.
    Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP);
  }
 
  // The allocated data lives above the 160 bytes allocated for the standard
  // frame, plus any outgoing stack arguments.  We don't know how much that
  // amounts to yet, so emit a special ADJDYNALLOC placeholder.
  SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
  SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
 
  // Dynamically realign if needed.
  if (RequiredAlign > StackAlign) {
    Result =
      DAG.getNode(ISD::ADD, DL, MVT::i64, Result,
                  DAG.getConstant(ExtraAlignSpace, DL, MVT::i64));
    Result =
      DAG.getNode(ISD::AND, DL, MVT::i64, Result,
                  DAG.getConstant(~(RequiredAlign - 1), DL, MVT::i64));
  }
 
  if (StoreBackchain)
    Chain = DAG.getStore(Chain, DL, Backchain, getBackchainAddress(NewSP, DAG),
                         MachinePointerInfo());
 
  SDValue Ops[2] = { Result, Chain };
  return DAG.getMergeValues(Ops, DL);
}
 
SDValue SystemZTargetLowering::lowerGET_DYNAMIC_AREA_OFFSET(
    SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
 
  return DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64);
}
 
SDValue SystemZTargetLowering::lowerMULH(SDValue Op,
                                         SelectionDAG &DAG,
                                         unsigned Opcode) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue Even, Odd;
 
  // This custom expander is only used on z17 and later for 64-bit types.
  assert(!is32Bit(VT));
  assert(Subtarget.hasMiscellaneousExtensions2());
 
  // SystemZISD::xMUL_LOHI returns the low result in the odd register and
  // the high result in the even register.  Return the latter.
  lowerGR128Binary(DAG, DL, VT, Opcode,
                   Op.getOperand(0), Op.getOperand(1), Even, Odd);
  return Even;
}
 
SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
                                              SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue Ops[2];
  if (is32Bit(VT))
    // Just do a normal 64-bit multiplication and extract the results.
    // We define this so that it can be used for constant division.
    lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0),
                    Op.getOperand(1), Ops[1], Ops[0]);
  else if (Subtarget.hasMiscellaneousExtensions2())
    // SystemZISD::SMUL_LOHI returns the low result in the odd register and
    // the high result in the even register.  ISD::SMUL_LOHI is defined to
    // return the low half first, so the results are in reverse order.
    lowerGR128Binary(DAG, DL, VT, SystemZISD::SMUL_LOHI,
                     Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
  else {
    // Do a full 128-bit multiplication based on SystemZISD::UMUL_LOHI:
    //
    //   (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64)
    //
    // but using the fact that the upper halves are either all zeros
    // or all ones:
    //
    //   (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64)
    //
    // and grouping the right terms together since they are quicker than the
    // multiplication:
    //
    //   (ll * rl) - (((lh & rl) + (ll & rh)) << 64)
    SDValue C63 = DAG.getConstant(63, DL, MVT::i64);
    SDValue LL = Op.getOperand(0);
    SDValue RL = Op.getOperand(1);
    SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63);
    SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63);
    // SystemZISD::UMUL_LOHI returns the low result in the odd register and
    // the high result in the even register.  ISD::SMUL_LOHI is defined to
    // return the low half first, so the results are in reverse order.
    lowerGR128Binary(DAG, DL, VT, SystemZISD::UMUL_LOHI,
                     LL, RL, Ops[1], Ops[0]);
    SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH);
    SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL);
    SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
    Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
  }
  return DAG.getMergeValues(Ops, DL);
}
 
SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
                                              SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue Ops[2];
  if (is32Bit(VT))
    // Just do a normal 64-bit multiplication and extract the results.
    // We define this so that it can be used for constant division.
    lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0),
                    Op.getOperand(1), Ops[1], Ops[0]);
  else
    // SystemZISD::UMUL_LOHI returns the low result in the odd register and
    // the high result in the even register.  ISD::UMUL_LOHI is defined to
    // return the low half first, so the results are in reverse order.
    lowerGR128Binary(DAG, DL, VT, SystemZISD::UMUL_LOHI,
                     Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
  return DAG.getMergeValues(Ops, DL);
}
 
SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
                                            SelectionDAG &DAG) const {
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
 
  // We use DSGF for 32-bit division.  This means the first operand must
  // always be 64-bit, and the second operand should be 32-bit whenever
  // that is possible, to improve performance.
  if (is32Bit(VT))
    Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0);
  else if (DAG.ComputeNumSignBits(Op1) > 32)
    Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1);
 
  // DSG(F) returns the remainder in the even register and the
  // quotient in the odd register.
  SDValue Ops[2];
  lowerGR128Binary(DAG, DL, VT, SystemZISD::SDIVREM, Op0, Op1, Ops[1], Ops[0]);
  return DAG.getMergeValues(Ops, DL);
}
 
SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
                                            SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
 
  // DL(G) returns the remainder in the even register and the
  // quotient in the odd register.
  SDValue Ops[2];
  lowerGR128Binary(DAG, DL, VT, SystemZISD::UDIVREM,
                   Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
  return DAG.getMergeValues(Ops, DL);
}
 
SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
  assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation");
 
  // Get the known-zero masks for each operand.
  SDValue Ops[] = {Op.getOperand(0), Op.getOperand(1)};
  KnownBits Known[2] = {DAG.computeKnownBits(Ops[0]),
                        DAG.computeKnownBits(Ops[1])};
 
  // See if the upper 32 bits of one operand and the lower 32 bits of the
  // other are known zero.  They are the low and high operands respectively.
  uint64_t Masks[] = { Known[0].Zero.getZExtValue(),
                       Known[1].Zero.getZExtValue() };
  unsigned High, Low;
  if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff)
    High = 1, Low = 0;
  else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff)
    High = 0, Low = 1;
  else
    return Op;
 
  SDValue LowOp = Ops[Low];
  SDValue HighOp = Ops[High];
 
  // If the high part is a constant, we're better off using IILH.
  if (HighOp.getOpcode() == ISD::Constant)
    return Op;
 
  // If the low part is a constant that is outside the range of LHI,
  // then we're better off using IILF.
  if (LowOp.getOpcode() == ISD::Constant) {
    int64_t Value = int32_t(LowOp->getAsZExtVal());
    if (!isInt<16>(Value))
      return Op;
  }
 
  // Check whether the high part is an AND that doesn't change the
  // high 32 bits and just masks out low bits.  We can skip it if so.
  if (HighOp.getOpcode() == ISD::AND &&
      HighOp.getOperand(1).getOpcode() == ISD::Constant) {
    SDValue HighOp0 = HighOp.getOperand(0);
    uint64_t Mask = HighOp.getConstantOperandVal(1);
    if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
      HighOp = HighOp0;
  }
 
  // Take advantage of the fact that all GR32 operations only change the
  // low 32 bits by truncating Low to an i32 and inserting it directly
  // using a subreg.  The interesting cases are those where the truncation
  // can be folded.
  SDLoc DL(Op);
  SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp);
  return DAG.getTargetInsertSubreg(SystemZ::subreg_l32, DL,
                                   MVT::i64, HighOp, Low32);
}
 
// Lower SADDO/SSUBO/UADDO/USUBO nodes.
SDValue SystemZTargetLowering::lowerXALUO(SDValue Op,
                                          SelectionDAG &DAG) const {
  SDNode *N = Op.getNode();
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  SDLoc DL(N);
 
  if (N->getValueType(0) == MVT::i128) {
    unsigned BaseOp = 0;
    unsigned FlagOp = 0;
    bool IsBorrow = false;
    switch (Op.getOpcode()) {
    default: llvm_unreachable("Unknown instruction!");
    case ISD::UADDO:
      BaseOp = ISD::ADD;
      FlagOp = SystemZISD::VACC;
      break;
    case ISD::USUBO:
      BaseOp = ISD::SUB;
      FlagOp = SystemZISD::VSCBI;
      IsBorrow = true;
      break;
    }
    SDValue Result = DAG.getNode(BaseOp, DL, MVT::i128, LHS, RHS);
    SDValue Flag = DAG.getNode(FlagOp, DL, MVT::i128, LHS, RHS);
    Flag = DAG.getNode(ISD::AssertZext, DL, MVT::i128, Flag,
                       DAG.getValueType(MVT::i1));
    Flag = DAG.getZExtOrTrunc(Flag, DL, N->getValueType(1));
    if (IsBorrow)
      Flag = DAG.getNode(ISD::XOR, DL, Flag.getValueType(),
                         Flag, DAG.getConstant(1, DL, Flag.getValueType()));
    return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Result, Flag);
  }
 
  unsigned BaseOp = 0;
  unsigned CCValid = 0;
  unsigned CCMask = 0;
 
  switch (Op.getOpcode()) {
  default: llvm_unreachable("Unknown instruction!");
  case ISD::SADDO:
    BaseOp = SystemZISD::SADDO;
    CCValid = SystemZ::CCMASK_ARITH;
    CCMask = SystemZ::CCMASK_ARITH_OVERFLOW;
    break;
  case ISD::SSUBO:
    BaseOp = SystemZISD::SSUBO;
    CCValid = SystemZ::CCMASK_ARITH;
    CCMask = SystemZ::CCMASK_ARITH_OVERFLOW;
    break;
  case ISD::UADDO:
    BaseOp = SystemZISD::UADDO;
    CCValid = SystemZ::CCMASK_LOGICAL;
    CCMask = SystemZ::CCMASK_LOGICAL_CARRY;
    break;
  case ISD::USUBO:
    BaseOp = SystemZISD::USUBO;
    CCValid = SystemZ::CCMASK_LOGICAL;
    CCMask = SystemZ::CCMASK_LOGICAL_BORROW;
    break;
  }
 
  SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
  SDValue Result = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
 
  SDValue SetCC = emitSETCC(DAG, DL, Result.getValue(1), CCValid, CCMask);
  if (N->getValueType(1) == MVT::i1)
    SetCC = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SetCC);
 
  return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Result, SetCC);
}
 
static bool isAddCarryChain(SDValue Carry) {
  while (Carry.getOpcode() == ISD::UADDO_CARRY &&
         Carry->getValueType(0) != MVT::i128)
    Carry = Carry.getOperand(2);
  return Carry.getOpcode() == ISD::UADDO &&
         Carry->getValueType(0) != MVT::i128;
}
 
static bool isSubBorrowChain(SDValue Carry) {
  while (Carry.getOpcode() == ISD::USUBO_CARRY &&
         Carry->getValueType(0) != MVT::i128)
    Carry = Carry.getOperand(2);
  return Carry.getOpcode() == ISD::USUBO &&
         Carry->getValueType(0) != MVT::i128;
}
 
// Lower UADDO_CARRY/USUBO_CARRY nodes.
SDValue SystemZTargetLowering::lowerUADDSUBO_CARRY(SDValue Op,
                                                   SelectionDAG &DAG) const {
 
  SDNode *N = Op.getNode();
  MVT VT = N->getSimpleValueType(0);
 
  // Let legalize expand this if it isn't a legal type yet.
  if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
    return SDValue();
 
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  SDValue Carry = Op.getOperand(2);
  SDLoc DL(N);
 
  if (VT == MVT::i128) {
    unsigned BaseOp = 0;
    unsigned FlagOp = 0;
    bool IsBorrow = false;
    switch (Op.getOpcode()) {
    default: llvm_unreachable("Unknown instruction!");
    case ISD::UADDO_CARRY:
      BaseOp = SystemZISD::VAC;
      FlagOp = SystemZISD::VACCC;
      break;
    case ISD::USUBO_CARRY:
      BaseOp = SystemZISD::VSBI;
      FlagOp = SystemZISD::VSBCBI;
      IsBorrow = true;
      break;
    }
    if (IsBorrow)
      Carry = DAG.getNode(ISD::XOR, DL, Carry.getValueType(),
                          Carry, DAG.getConstant(1, DL, Carry.getValueType()));
    Carry = DAG.getZExtOrTrunc(Carry, DL, MVT::i128);
    SDValue Result = DAG.getNode(BaseOp, DL, MVT::i128, LHS, RHS, Carry);
    SDValue Flag = DAG.getNode(FlagOp, DL, MVT::i128, LHS, RHS, Carry);
    Flag = DAG.getNode(ISD::AssertZext, DL, MVT::i128, Flag,
                       DAG.getValueType(MVT::i1));
    Flag = DAG.getZExtOrTrunc(Flag, DL, N->getValueType(1));
    if (IsBorrow)
      Flag = DAG.getNode(ISD::XOR, DL, Flag.getValueType(),
                         Flag, DAG.getConstant(1, DL, Flag.getValueType()));
    return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Result, Flag);
  }
 
  unsigned BaseOp = 0;
  unsigned CCValid = 0;
  unsigned CCMask = 0;
 
  switch (Op.getOpcode()) {
  default: llvm_unreachable("Unknown instruction!");
  case ISD::UADDO_CARRY:
    if (!isAddCarryChain(Carry))
      return SDValue();
 
    BaseOp = SystemZISD::ADDCARRY;
    CCValid = SystemZ::CCMASK_LOGICAL;
    CCMask = SystemZ::CCMASK_LOGICAL_CARRY;
    break;
  case ISD::USUBO_CARRY:
    if (!isSubBorrowChain(Carry))
      return SDValue();
 
    BaseOp = SystemZISD::SUBCARRY;
    CCValid = SystemZ::CCMASK_LOGICAL;
    CCMask = SystemZ::CCMASK_LOGICAL_BORROW;
    break;
  }
 
  // Set the condition code from the carry flag.
  Carry = DAG.getNode(SystemZISD::GET_CCMASK, DL, MVT::i32, Carry,
                      DAG.getConstant(CCValid, DL, MVT::i32),
                      DAG.getConstant(CCMask, DL, MVT::i32));
 
  SDVTList VTs = DAG.getVTList(VT, MVT::i32);
  SDValue Result = DAG.getNode(BaseOp, DL, VTs, LHS, RHS, Carry);
 
  SDValue SetCC = emitSETCC(DAG, DL, Result.getValue(1), CCValid, CCMask);
  if (N->getValueType(1) == MVT::i1)
    SetCC = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SetCC);
 
  return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Result, SetCC);
}
 
SDValue SystemZTargetLowering::lowerCTPOP(SDValue Op,
                                          SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  Op = Op.getOperand(0);
 
  if (VT.getScalarSizeInBits() == 128) {
    Op = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Op);
    Op = DAG.getNode(ISD::CTPOP, DL, MVT::v2i64, Op);
    SDValue Tmp = DAG.getSplatBuildVector(MVT::v2i64, DL,
                                          DAG.getConstant(0, DL, MVT::i64));
    Op = DAG.getNode(SystemZISD::VSUM, DL, VT, Op, Tmp);
    return Op;
  }
 
  // Handle vector types via VPOPCT.
  if (VT.isVector()) {
    Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Op);
    Op = DAG.getNode(SystemZISD::POPCNT, DL, MVT::v16i8, Op);
    switch (VT.getScalarSizeInBits()) {
    case 8:
      break;
    case 16: {
      Op = DAG.getNode(ISD::BITCAST, DL, VT, Op);
      SDValue Shift = DAG.getConstant(8, DL, MVT::i32);
      SDValue Tmp = DAG.getNode(SystemZISD::VSHL_BY_SCALAR, DL, VT, Op, Shift);
      Op = DAG.getNode(ISD::ADD, DL, VT, Op, Tmp);
      Op = DAG.getNode(SystemZISD::VSRL_BY_SCALAR, DL, VT, Op, Shift);
      break;
    }
    case 32: {
      SDValue Tmp = DAG.getSplatBuildVector(MVT::v16i8, DL,
                                            DAG.getConstant(0, DL, MVT::i32));
      Op = DAG.getNode(SystemZISD::VSUM, DL, VT, Op, Tmp);
      break;
    }
    case 64: {
      SDValue Tmp = DAG.getSplatBuildVector(MVT::v16i8, DL,
                                            DAG.getConstant(0, DL, MVT::i32));
      Op = DAG.getNode(SystemZISD::VSUM, DL, MVT::v4i32, Op, Tmp);
      Op = DAG.getNode(SystemZISD::VSUM, DL, VT, Op, Tmp);
      break;
    }
    default:
      llvm_unreachable("Unexpected type");
    }
    return Op;
  }
 
  // Get the known-zero mask for the operand.
  KnownBits Known = DAG.computeKnownBits(Op);
  unsigned NumSignificantBits = Known.getMaxValue().getActiveBits();
  if (NumSignificantBits == 0)
    return DAG.getConstant(0, DL, VT);
 
  // Skip known-zero high parts of the operand.
  int64_t OrigBitSize = VT.getSizeInBits();
  int64_t BitSize = llvm::bit_ceil(NumSignificantBits);
  BitSize = std::min(BitSize, OrigBitSize);
 
  // The POPCNT instruction counts the number of bits in each byte.
  Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op);
  Op = DAG.getNode(SystemZISD::POPCNT, DL, MVT::i64, Op);
  Op = DAG.getNode(ISD::TRUNCATE, DL, VT, Op);
 
  // Add up per-byte counts in a binary tree.  All bits of Op at
  // position larger than BitSize remain zero throughout.
  for (int64_t I = BitSize / 2; I >= 8; I = I / 2) {
    SDValue Tmp = DAG.getNode(ISD::SHL, DL, VT, Op, DAG.getConstant(I, DL, VT));
    if (BitSize != OrigBitSize)
      Tmp = DAG.getNode(ISD::AND, DL, VT, Tmp,
                        DAG.getConstant(((uint64_t)1 << BitSize) - 1, DL, VT));
    Op = DAG.getNode(ISD::ADD, DL, VT, Op, Tmp);
  }
 
  // Extract overall result from high byte.
  if (BitSize > 8)
    Op = DAG.getNode(ISD::SRL, DL, VT, Op,
                     DAG.getConstant(BitSize - 8, DL, VT));
 
  return Op;
}
 
SDValue SystemZTargetLowering::lowerATOMIC_FENCE(SDValue Op,
                                                 SelectionDAG &DAG) const {
  SDLoc DL(Op);
  AtomicOrdering FenceOrdering =
      static_cast<AtomicOrdering>(Op.getConstantOperandVal(1));
  SyncScope::ID FenceSSID =
      static_cast<SyncScope::ID>(Op.getConstantOperandVal(2));
 
  // The only fence that needs an instruction is a sequentially-consistent
  // cross-thread fence.
  if (FenceOrdering == AtomicOrdering::SequentiallyConsistent &&
      FenceSSID == SyncScope::System) {
    return SDValue(DAG.getMachineNode(SystemZ::Serialize, DL, MVT::Other,
                                      Op.getOperand(0)),
                   0);
  }
 
  // MEMBARRIER is a compiler barrier; it codegens to a no-op.
  return DAG.getNode(ISD::MEMBARRIER, DL, MVT::Other, Op.getOperand(0));
}
 
SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
                                                SelectionDAG &DAG) const {
  EVT RegVT = Op.getValueType();
  if (RegVT.getSizeInBits() == 128)
    return lowerATOMIC_LDST_I128(Op, DAG);
  return lowerLoadF16(Op, DAG);
}
 
SDValue SystemZTargetLowering::lowerATOMIC_STORE(SDValue Op,
                                                 SelectionDAG &DAG) const {
  auto *Node = cast<AtomicSDNode>(Op.getNode());
  if (Node->getMemoryVT().getSizeInBits() == 128)
    return lowerATOMIC_LDST_I128(Op, DAG);
  return lowerStoreF16(Op, DAG);
}
 
SDValue SystemZTargetLowering::lowerATOMIC_LDST_I128(SDValue Op,
                                                     SelectionDAG &DAG) const {
  auto *Node = cast<AtomicSDNode>(Op.getNode());
  assert(
      (Node->getMemoryVT() == MVT::i128 || Node->getMemoryVT() == MVT::f128) &&
      "Only custom lowering i128 or f128.");
  // Use same code to handle both legal and non-legal i128 types.
  SmallVector<SDValue, 2> Results;
  LowerOperationWrapper(Node, Results, DAG);
  return DAG.getMergeValues(Results, SDLoc(Op));
}
 
// Prepare for a Compare And Swap for a subword operation. This needs to be
// done in memory with 4 bytes at natural alignment.
static void getCSAddressAndShifts(SDValue Addr, SelectionDAG &DAG, SDLoc DL,
                                  SDValue &AlignedAddr, SDValue &BitShift,
                                  SDValue &NegBitShift) {
  EVT PtrVT = Addr.getValueType();
  EVT WideVT = MVT::i32;
 
  // Get the address of the containing word.
  AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr,
                            DAG.getSignedConstant(-4, DL, PtrVT));
 
  // Get the number of bits that the word must be rotated left in order
  // to bring the field to the top bits of a GR32.
  BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr,
                         DAG.getConstant(3, DL, PtrVT));
  BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift);
 
  // Get the complementing shift amount, for rotating a field in the top
  // bits back to its proper position.
  NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT,
                            DAG.getConstant(0, DL, WideVT), BitShift);
 
}
 
// Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation.  Lower the first
// two into the fullword ATOMIC_LOADW_* operation given by Opcode.
SDValue SystemZTargetLowering::lowerATOMIC_LOAD_OP(SDValue Op,
                                                   SelectionDAG &DAG,
                                                   unsigned Opcode) const {
  auto *Node = cast<AtomicSDNode>(Op.getNode());
 
  // 32-bit operations need no special handling.
  EVT NarrowVT = Node->getMemoryVT();
  EVT WideVT = MVT::i32;
  if (NarrowVT == WideVT)
    return Op;
 
  int64_t BitSize = NarrowVT.getSizeInBits();
  SDValue ChainIn = Node->getChain();
  SDValue Addr = Node->getBasePtr();
  SDValue Src2 = Node->getVal();
  MachineMemOperand *MMO = Node->getMemOperand();
  SDLoc DL(Node);
 
  // Convert atomic subtracts of constants into additions.
  if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
    if (auto *Const = dyn_cast<ConstantSDNode>(Src2)) {
      Opcode = SystemZISD::ATOMIC_LOADW_ADD;
      Src2 = DAG.getSignedConstant(-Const->getSExtValue(), DL,
                                   Src2.getValueType());
    }
 
  SDValue AlignedAddr, BitShift, NegBitShift;
  getCSAddressAndShifts(Addr, DAG, DL, AlignedAddr, BitShift, NegBitShift);
 
  // Extend the source operand to 32 bits and prepare it for the inner loop.
  // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other
  // operations require the source to be shifted in advance.  (This shift
  // can be folded if the source is constant.)  For AND and NAND, the lower
  // bits must be set, while for other opcodes they should be left clear.
  if (Opcode != SystemZISD::ATOMIC_SWAPW)
    Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2,
                       DAG.getConstant(32 - BitSize, DL, WideVT));
  if (Opcode == SystemZISD::ATOMIC_LOADW_AND ||
      Opcode == SystemZISD::ATOMIC_LOADW_NAND)
    Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2,
                       DAG.getConstant(uint32_t(-1) >> BitSize, DL, WideVT));
 
  // Construct the ATOMIC_LOADW_* node.
  SDVTList VTList = DAG.getVTList(WideVT, MVT::Other);
  SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
                    DAG.getConstant(BitSize, DL, WideVT) };
  SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
                                             NarrowVT, MMO);
 
  // Rotate the result of the final CS so that the field is in the lower
  // bits of a GR32, then truncate it.
  SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift,
                                    DAG.getConstant(BitSize, DL, WideVT));
  SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
 
  SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
  return DAG.getMergeValues(RetOps, DL);
}
 
// Op is an ATOMIC_LOAD_SUB operation.  Lower 8- and 16-bit operations into
// ATOMIC_LOADW_SUBs and convert 32- and 64-bit operations into additions.
SDValue SystemZTargetLowering::lowerATOMIC_LOAD_SUB(SDValue Op,
                                                    SelectionDAG &DAG) const {
  auto *Node = cast<AtomicSDNode>(Op.getNode());
  EVT MemVT = Node->getMemoryVT();
  if (MemVT == MVT::i32 || MemVT == MVT::i64) {
    // A full-width operation: negate and use LAA(G).
    assert(Op.getValueType() == MemVT && "Mismatched VTs");
    assert(Subtarget.hasInterlockedAccess1() &&
           "Should have been expanded by AtomicExpand pass.");
    SDValue Src2 = Node->getVal();
    SDLoc DL(Src2);
    SDValue NegSrc2 =
      DAG.getNode(ISD::SUB, DL, MemVT, DAG.getConstant(0, DL, MemVT), Src2);
    return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, MemVT,
                         Node->getChain(), Node->getBasePtr(), NegSrc2,
                         Node->getMemOperand());
  }
 
  return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
}
 
// Lower 8/16/32/64-bit ATOMIC_CMP_SWAP_WITH_SUCCESS node.
SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
                                                    SelectionDAG &DAG) const {
  auto *Node = cast<AtomicSDNode>(Op.getNode());
  SDValue ChainIn = Node->getOperand(0);
  SDValue Addr = Node->getOperand(1);
  SDValue CmpVal = Node->getOperand(2);
  SDValue SwapVal = Node->getOperand(3);
  MachineMemOperand *MMO = Node->getMemOperand();
  SDLoc DL(Node);
 
  if (Node->getMemoryVT() == MVT::i128) {
    // Use same code to handle both legal and non-legal i128 types.
    SmallVector<SDValue, 3> Results;
    LowerOperationWrapper(Node, Results, DAG);
    return DAG.getMergeValues(Results, DL);
  }
 
  // We have native support for 32-bit and 64-bit compare and swap, but we
  // still need to expand extracting the "success" result from the CC.
  EVT NarrowVT = Node->getMemoryVT();
  EVT WideVT = NarrowVT == MVT::i64 ? MVT::i64 : MVT::i32;
  if (NarrowVT == WideVT) {
    SDVTList Tys = DAG.getVTList(WideVT, MVT::i32, MVT::Other);
    SDValue Ops[] = { ChainIn, Addr, CmpVal, SwapVal };
    SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAP,
                                               DL, Tys, Ops, NarrowVT, MMO);
    SDValue Success = emitSETCC(DAG, DL, AtomicOp.getValue(1),
                                SystemZ::CCMASK_CS, SystemZ::CCMASK_CS_EQ);
 
    DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), AtomicOp.getValue(0));
    DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
    DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), AtomicOp.getValue(2));
    return SDValue();
  }
 
  // Convert 8-bit and 16-bit compare and swap to a loop, implemented
  // via a fullword ATOMIC_CMP_SWAPW operation.
  int64_t BitSize = NarrowVT.getSizeInBits();
 
  SDValue AlignedAddr, BitShift, NegBitShift;
  getCSAddressAndShifts(Addr, DAG, DL, AlignedAddr, BitShift, NegBitShift);
 
  // Construct the ATOMIC_CMP_SWAPW node.
  SDVTList VTList = DAG.getVTList(WideVT, MVT::i32, MVT::Other);
  SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
                    NegBitShift, DAG.getConstant(BitSize, DL, WideVT) };
  SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
                                             VTList, Ops, NarrowVT, MMO);
  SDValue Success = emitSETCC(DAG, DL, AtomicOp.getValue(1),
                              SystemZ::CCMASK_ICMP, SystemZ::CCMASK_CMP_EQ);
 
  // emitAtomicCmpSwapW() will zero extend the result (original value).
  SDValue OrigVal = DAG.getNode(ISD::AssertZext, DL, WideVT, AtomicOp.getValue(0),
                                DAG.getValueType(NarrowVT));
  DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), OrigVal);
  DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
  DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), AtomicOp.getValue(2));
  return SDValue();
}
 
MachineMemOperand::Flags
SystemZTargetLowering::getTargetMMOFlags(const Instruction &I) const {
  // Because of how we convert atomic_load and atomic_store to normal loads and
  // stores in the DAG, we need to ensure that the MMOs are marked volatile
  // since DAGCombine hasn't been updated to account for atomic, but non
  // volatile loads.  (See D57601)
  if (auto *SI = dyn_cast<StoreInst>(&I))
    if (SI->isAtomic())
      return MachineMemOperand::MOVolatile;
  if (auto *LI = dyn_cast<LoadInst>(&I))
    if (LI->isAtomic())
      return MachineMemOperand::MOVolatile;
  if (auto *AI = dyn_cast<AtomicRMWInst>(&I))
    if (AI->isAtomic())
      return MachineMemOperand::MOVolatile;
  if (auto *AI = dyn_cast<AtomicCmpXchgInst>(&I))
    if (AI->isAtomic())
      return MachineMemOperand::MOVolatile;
  return MachineMemOperand::MONone;
}
 
SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op,
                                              SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  auto *Regs = Subtarget.getSpecialRegisters();
  if (MF.getFunction().getCallingConv() == CallingConv::GHC)
    report_fatal_error("Variable-sized stack allocations are not supported "
                       "in GHC calling convention");
  return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op),
                            Regs->getStackPointerRegister(), Op.getValueType());
}
 
SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op,
                                                 SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  auto *Regs = Subtarget.getSpecialRegisters();
  bool StoreBackchain = MF.getSubtarget<SystemZSubtarget>().hasBackChain();
 
  if (MF.getFunction().getCallingConv() == CallingConv::GHC)
    report_fatal_error("Variable-sized stack allocations are not supported "
                       "in GHC calling convention");
 
  SDValue Chain = Op.getOperand(0);
  SDValue NewSP = Op.getOperand(1);
  SDValue Backchain;
  SDLoc DL(Op);
 
  if (StoreBackchain) {
    SDValue OldSP = DAG.getCopyFromReg(
        Chain, DL, Regs->getStackPointerRegister(), MVT::i64);
    Backchain = DAG.getLoad(MVT::i64, DL, Chain, getBackchainAddress(OldSP, DAG),
                            MachinePointerInfo());
  }
 
  Chain = DAG.getCopyToReg(Chain, DL, Regs->getStackPointerRegister(), NewSP);
 
  if (StoreBackchain)
    Chain = DAG.getStore(Chain, DL, Backchain, getBackchainAddress(NewSP, DAG),
                         MachinePointerInfo());
 
  return Chain;
}
 
SDValue SystemZTargetLowering::lowerPREFETCH(SDValue Op,
                                             SelectionDAG &DAG) const {
  bool IsData = Op.getConstantOperandVal(4);
  if (!IsData)
    // Just preserve the chain.
    return Op.getOperand(0);
 
  SDLoc DL(Op);
  bool IsWrite = Op.getConstantOperandVal(2);
  unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ;
  auto *Node = cast<MemIntrinsicSDNode>(Op.getNode());
  SDValue Ops[] = {Op.getOperand(0), DAG.getTargetConstant(Code, DL, MVT::i32),
                   Op.getOperand(1)};
  return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, DL,
                                 Node->getVTList(), Ops,
                                 Node->getMemoryVT(), Node->getMemOperand());
}
 
SDValue
SystemZTargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
                                              SelectionDAG &DAG) const {
  unsigned Opcode, CCValid;
  if (isIntrinsicWithCCAndChain(Op, Opcode, CCValid)) {
    assert(Op->getNumValues() == 2 && "Expected only CC result and chain");
    SDNode *Node = emitIntrinsicWithCCAndChain(DAG, Op, Opcode);
    SDValue CC = getCCResult(DAG, SDValue(Node, 0));
    DAG.ReplaceAllUsesOfValueWith(SDValue(Op.getNode(), 0), CC);
    return SDValue();
  }
 
  return SDValue();
}
 
SDValue
SystemZTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
                                               SelectionDAG &DAG) const {
  unsigned Opcode, CCValid;
  if (isIntrinsicWithCC(Op, Opcode, CCValid)) {
    SDNode *Node = emitIntrinsicWithCC(DAG, Op, Opcode);
    if (Op->getNumValues() == 1)
      return getCCResult(DAG, SDValue(Node, 0));
    assert(Op->getNumValues() == 2 && "Expected a CC and non-CC result");
    return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), Op->getVTList(),
                       SDValue(Node, 0), getCCResult(DAG, SDValue(Node, 1)));
  }
 
  unsigned Id = Op.getConstantOperandVal(0);
  switch (Id) {
  case Intrinsic::thread_pointer:
    return lowerThreadPointer(SDLoc(Op), DAG);
 
  case Intrinsic::s390_vpdi:
    return DAG.getNode(SystemZISD::PERMUTE_DWORDS, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
  case Intrinsic::s390_vperm:
    return DAG.getNode(SystemZISD::PERMUTE, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
  case Intrinsic::s390_vuphb:
  case Intrinsic::s390_vuphh:
  case Intrinsic::s390_vuphf:
  case Intrinsic::s390_vuphg:
    return DAG.getNode(SystemZISD::UNPACK_HIGH, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1));
 
  case Intrinsic::s390_vuplhb:
  case Intrinsic::s390_vuplhh:
  case Intrinsic::s390_vuplhf:
  case Intrinsic::s390_vuplhg:
    return DAG.getNode(SystemZISD::UNPACKL_HIGH, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1));
 
  case Intrinsic::s390_vuplb:
  case Intrinsic::s390_vuplhw:
  case Intrinsic::s390_vuplf:
  case Intrinsic::s390_vuplg:
    return DAG.getNode(SystemZISD::UNPACK_LOW, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1));
 
  case Intrinsic::s390_vupllb:
  case Intrinsic::s390_vupllh:
  case Intrinsic::s390_vupllf:
  case Intrinsic::s390_vupllg:
    return DAG.getNode(SystemZISD::UNPACKL_LOW, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1));
 
  case Intrinsic::s390_vsumb:
  case Intrinsic::s390_vsumh:
  case Intrinsic::s390_vsumgh:
  case Intrinsic::s390_vsumgf:
  case Intrinsic::s390_vsumqf:
  case Intrinsic::s390_vsumqg:
    return DAG.getNode(SystemZISD::VSUM, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
 
  case Intrinsic::s390_vaq:
    return DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vaccb:
  case Intrinsic::s390_vacch:
  case Intrinsic::s390_vaccf:
  case Intrinsic::s390_vaccg:
  case Intrinsic::s390_vaccq:
    return DAG.getNode(SystemZISD::VACC, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vacq:
    return DAG.getNode(SystemZISD::VAC, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
  case Intrinsic::s390_vacccq:
    return DAG.getNode(SystemZISD::VACCC, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
  case Intrinsic::s390_vsq:
    return DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vscbib:
  case Intrinsic::s390_vscbih:
  case Intrinsic::s390_vscbif:
  case Intrinsic::s390_vscbig:
  case Intrinsic::s390_vscbiq:
    return DAG.getNode(SystemZISD::VSCBI, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vsbiq:
    return DAG.getNode(SystemZISD::VSBI, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
  case Intrinsic::s390_vsbcbiq:
    return DAG.getNode(SystemZISD::VSBCBI, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
  case Intrinsic::s390_vmhb:
  case Intrinsic::s390_vmhh:
  case Intrinsic::s390_vmhf:
  case Intrinsic::s390_vmhg:
  case Intrinsic::s390_vmhq:
    return DAG.getNode(ISD::MULHS, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vmlhb:
  case Intrinsic::s390_vmlhh:
  case Intrinsic::s390_vmlhf:
  case Intrinsic::s390_vmlhg:
  case Intrinsic::s390_vmlhq:
    return DAG.getNode(ISD::MULHU, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
 
  case Intrinsic::s390_vmahb:
  case Intrinsic::s390_vmahh:
  case Intrinsic::s390_vmahf:
  case Intrinsic::s390_vmahg:
  case Intrinsic::s390_vmahq:
    return DAG.getNode(SystemZISD::VMAH, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
  case Intrinsic::s390_vmalhb:
  case Intrinsic::s390_vmalhh:
  case Intrinsic::s390_vmalhf:
  case Intrinsic::s390_vmalhg:
  case Intrinsic::s390_vmalhq:
    return DAG.getNode(SystemZISD::VMALH, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
  case Intrinsic::s390_vmeb:
  case Intrinsic::s390_vmeh:
  case Intrinsic::s390_vmef:
  case Intrinsic::s390_vmeg:
    return DAG.getNode(SystemZISD::VME, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vmleb:
  case Intrinsic::s390_vmleh:
  case Intrinsic::s390_vmlef:
  case Intrinsic::s390_vmleg:
    return DAG.getNode(SystemZISD::VMLE, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vmob:
  case Intrinsic::s390_vmoh:
  case Intrinsic::s390_vmof:
  case Intrinsic::s390_vmog:
    return DAG.getNode(SystemZISD::VMO, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::s390_vmlob:
  case Intrinsic::s390_vmloh:
  case Intrinsic::s390_vmlof:
  case Intrinsic::s390_vmlog:
    return DAG.getNode(SystemZISD::VMLO, SDLoc(Op), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
 
  case Intrinsic::s390_vmaeb:
  case Intrinsic::s390_vmaeh:
  case Intrinsic::s390_vmaef:
  case Intrinsic::s390_vmaeg:
    return DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(),
                       DAG.getNode(SystemZISD::VME, SDLoc(Op), Op.getValueType(),
                                   Op.getOperand(1), Op.getOperand(2)),
                       Op.getOperand(3));
  case Intrinsic::s390_vmaleb:
  case Intrinsic::s390_vmaleh:
  case Intrinsic::s390_vmalef:
  case Intrinsic::s390_vmaleg:
    return DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(),
                       DAG.getNode(SystemZISD::VMLE, SDLoc(Op), Op.getValueType(),
                                   Op.getOperand(1), Op.getOperand(2)),
                       Op.getOperand(3));
  case Intrinsic::s390_vmaob:
  case Intrinsic::s390_vmaoh:
  case Intrinsic::s390_vmaof:
  case Intrinsic::s390_vmaog:
    return DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(),
                       DAG.getNode(SystemZISD::VMO, SDLoc(Op), Op.getValueType(),
                                   Op.getOperand(1), Op.getOperand(2)),
                       Op.getOperand(3));
  case Intrinsic::s390_vmalob:
  case Intrinsic::s390_vmaloh:
  case Intrinsic::s390_vmalof:
  case Intrinsic::s390_vmalog:
    return DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(),
                       DAG.getNode(SystemZISD::VMLO, SDLoc(Op), Op.getValueType(),
                                   Op.getOperand(1), Op.getOperand(2)),
                       Op.getOperand(3));
  }
 
  return SDValue();
}
 
namespace {
// Says that SystemZISD operation Opcode can be used to perform the equivalent
// of a VPERM with permute vector Bytes.  If Opcode takes three operands,
// Operand is the constant third operand, otherwise it is the number of
// bytes in each element of the result.
struct Permute {
  unsigned Opcode;
  unsigned Operand;
  unsigned char Bytes[SystemZ::VectorBytes];
};
}
 
static const Permute PermuteForms[] = {
  // VMRHG
  { SystemZISD::MERGE_HIGH, 8,
    { 0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23 } },
  // VMRHF
  { SystemZISD::MERGE_HIGH, 4,
    { 0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23 } },
  // VMRHH
  { SystemZISD::MERGE_HIGH, 2,
    { 0, 1, 16, 17, 2, 3, 18, 19, 4, 5, 20, 21, 6, 7, 22, 23 } },
  // VMRHB
  { SystemZISD::MERGE_HIGH, 1,
    { 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23 } },
  // VMRLG
  { SystemZISD::MERGE_LOW, 8,
    { 8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31 } },
  // VMRLF
  { SystemZISD::MERGE_LOW, 4,
    { 8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31 } },
  // VMRLH
  { SystemZISD::MERGE_LOW, 2,
    { 8, 9, 24, 25, 10, 11, 26, 27, 12, 13, 28, 29, 14, 15, 30, 31 } },
  // VMRLB
  { SystemZISD::MERGE_LOW, 1,
    { 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31 } },
  // VPKG
  { SystemZISD::PACK, 4,
    { 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 } },
  // VPKF
  { SystemZISD::PACK, 2,
    { 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 } },
  // VPKH
  { SystemZISD::PACK, 1,
    { 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31 } },
  // VPDI V1, V2, 4  (low half of V1, high half of V2)
  { SystemZISD::PERMUTE_DWORDS, 4,
    { 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 } },
  // VPDI V1, V2, 1  (high half of V1, low half of V2)
  { SystemZISD::PERMUTE_DWORDS, 1,
    { 0, 1, 2, 3, 4, 5, 6, 7, 24, 25, 26, 27, 28, 29, 30, 31 } }
};
 
// Called after matching a vector shuffle against a particular pattern.
// Both the original shuffle and the pattern have two vector operands.
// OpNos[0] is the operand of the original shuffle that should be used for
// operand 0 of the pattern, or -1 if operand 0 of the pattern can be anything.
// OpNos[1] is the same for operand 1 of the pattern.  Resolve these -1s and
// set OpNo0 and OpNo1 to the shuffle operands that should actually be used
// for operands 0 and 1 of the pattern.
static bool chooseShuffleOpNos(int *OpNos, unsigned &OpNo0, unsigned &OpNo1) {
  if (OpNos[0] < 0) {
    if (OpNos[1] < 0)
      return false;
    OpNo0 = OpNo1 = OpNos[1];
  } else if (OpNos[1] < 0) {
    OpNo0 = OpNo1 = OpNos[0];
  } else {
    OpNo0 = OpNos[0];
    OpNo1 = OpNos[1];
  }
  return true;
}
 
// Bytes is a VPERM-like permute vector, except that -1 is used for
// undefined bytes.  Return true if the VPERM can be implemented using P.
// When returning true set OpNo0 to the VPERM operand that should be
// used for operand 0 of P and likewise OpNo1 for operand 1 of P.
//
// For example, if swapping the VPERM operands allows P to match, OpNo0
// will be 1 and OpNo1 will be 0.  If instead Bytes only refers to one
// operand, but rewriting it to use two duplicated operands allows it to
// match P, then OpNo0 and OpNo1 will be the same.
static bool matchPermute(const SmallVectorImpl<int> &Bytes, const Permute &P,
                         unsigned &OpNo0, unsigned &OpNo1) {
  int OpNos[] = { -1, -1 };
  for (unsigned I = 0; I < SystemZ::VectorBytes; ++I) {
    int Elt = Bytes[I];
    if (Elt >= 0) {
      // Make sure that the two permute vectors use the same suboperand
      // byte number.  Only the operand numbers (the high bits) are
      // allowed to differ.
      if ((Elt ^ P.Bytes[I]) & (SystemZ::VectorBytes - 1))
        return false;
      int ModelOpNo = P.Bytes[I] / SystemZ::VectorBytes;
      int RealOpNo = unsigned(Elt) / SystemZ::VectorBytes;
      // Make sure that the operand mappings are consistent with previous
      // elements.
      if (OpNos[ModelOpNo] == 1 - RealOpNo)
        return false;
      OpNos[ModelOpNo] = RealOpNo;
    }
  }
  return chooseShuffleOpNos(OpNos, OpNo0, OpNo1);
}
 
// As above, but search for a matching permute.
static const Permute *matchPermute(const SmallVectorImpl<int> &Bytes,
                                   unsigned &OpNo0, unsigned &OpNo1) {
  for (auto &P : PermuteForms)
    if (matchPermute(Bytes, P, OpNo0, OpNo1))
      return &P;
  return nullptr;
}
 
// Bytes is a VPERM-like permute vector, except that -1 is used for
// undefined bytes.  This permute is an operand of an outer permute.
// See whether redistributing the -1 bytes gives a shuffle that can be
// implemented using P.  If so, set Transform to a VPERM-like permute vector
// that, when applied to the result of P, gives the original permute in Bytes.
static bool matchDoublePermute(const SmallVectorImpl<int> &Bytes,
                               const Permute &P,
                               SmallVectorImpl<int> &Transform) {
  unsigned To = 0;
  for (unsigned From = 0; From < SystemZ::VectorBytes; ++From) {
    int Elt = Bytes[From];
    if (Elt < 0)
      // Byte number From of the result is undefined.
      Transform[From] = -1;
    else {
      while (P.Bytes[To] != Elt) {
        To += 1;
        if (To == SystemZ::VectorBytes)
          return false;
      }
      Transform[From] = To;
    }
  }
  return true;
}
 
// As above, but search for a matching permute.
static const Permute *matchDoublePermute(const SmallVectorImpl<int> &Bytes,
                                         SmallVectorImpl<int> &Transform) {
  for (auto &P : PermuteForms)
    if (matchDoublePermute(Bytes, P, Transform))
      return &P;
  return nullptr;
}
 
// Convert the mask of the given shuffle op into a byte-level mask,
// as if it had type vNi8.
static bool getVPermMask(SDValue ShuffleOp,
                         SmallVectorImpl<int> &Bytes) {
  EVT VT = ShuffleOp.getValueType();
  unsigned NumElements = VT.getVectorNumElements();
  unsigned BytesPerElement = VT.getVectorElementType().getStoreSize();
 
  if (auto *VSN = dyn_cast<ShuffleVectorSDNode>(ShuffleOp)) {
    Bytes.resize(NumElements * BytesPerElement, -1);
    for (unsigned I = 0; I < NumElements; ++I) {
      int Index = VSN->getMaskElt(I);
      if (Index >= 0)
        for (unsigned J = 0; J < BytesPerElement; ++J)
          Bytes[I * BytesPerElement + J] = Index * BytesPerElement + J;
    }
    return true;
  }
  if (SystemZISD::SPLAT == ShuffleOp.getOpcode() &&
      isa<ConstantSDNode>(ShuffleOp.getOperand(1))) {
    unsigned Index = ShuffleOp.getConstantOperandVal(1);
    Bytes.resize(NumElements * BytesPerElement, -1);
    for (unsigned I = 0; I < NumElements; ++I)
      for (unsigned J = 0; J < BytesPerElement; ++J)
        Bytes[I * BytesPerElement + J] = Index * BytesPerElement + J;
    return true;
  }
  return false;
}
 
// Bytes is a VPERM-like permute vector, except that -1 is used for
// undefined bytes.  See whether bytes [Start, Start + BytesPerElement) of
// the result come from a contiguous sequence of bytes from one input.
// Set Base to the selector for the first byte if so.
static bool getShuffleInput(const SmallVectorImpl<int> &Bytes, unsigned Start,
                            unsigned BytesPerElement, int &Base) {
  Base = -1;
  for (unsigned I = 0; I < BytesPerElement; ++I) {
    if (Bytes[Start + I] >= 0) {
      unsigned Elem = Bytes[Start + I];
      if (Base < 0) {
        Base = Elem - I;
        // Make sure the bytes would come from one input operand.
        if (unsigned(Base) % Bytes.size() + BytesPerElement > Bytes.size())
          return false;
      } else if (unsigned(Base) != Elem - I)
        return false;
    }
  }
  return true;
}
 
// Bytes is a VPERM-like permute vector, except that -1 is used for
// undefined bytes.  Return true if it can be performed using VSLDB.
// When returning true, set StartIndex to the shift amount and OpNo0
// and OpNo1 to the VPERM operands that should be used as the first
// and second shift operand respectively.
static bool isShlDoublePermute(const SmallVectorImpl<int> &Bytes,
                               unsigned &StartIndex, unsigned &OpNo0,
                               unsigned &OpNo1) {
  int OpNos[] = { -1, -1 };
  int Shift = -1;
  for (unsigned I = 0; I < 16; ++I) {
    int Index = Bytes[I];
    if (Index >= 0) {
      int ExpectedShift = (Index - I) % SystemZ::VectorBytes;
      int ModelOpNo = unsigned(ExpectedShift + I) / SystemZ::VectorBytes;
      int RealOpNo = unsigned(Index) / SystemZ::VectorBytes;
      if (Shift < 0)
        Shift = ExpectedShift;
      else if (Shift != ExpectedShift)
        return false;
      // Make sure that the operand mappings are consistent with previous
      // elements.
      if (OpNos[ModelOpNo] == 1 - RealOpNo)
        return false;
      OpNos[ModelOpNo] = RealOpNo;
    }
  }
  StartIndex = Shift;
  return chooseShuffleOpNos(OpNos, OpNo0, OpNo1);
}
 
// Create a node that performs P on operands Op0 and Op1, casting the
// operands to the appropriate type.  The type of the result is determined by P.
static SDValue getPermuteNode(SelectionDAG &DAG, const SDLoc &DL,
                              const Permute &P, SDValue Op0, SDValue Op1) {
  // VPDI (PERMUTE_DWORDS) always operates on v2i64s.  The input
  // elements of a PACK are twice as wide as the outputs.
  unsigned InBytes = (P.Opcode == SystemZISD::PERMUTE_DWORDS ? 8 :
                      P.Opcode == SystemZISD::PACK ? P.Operand * 2 :
                      P.Operand);
  // Cast both operands to the appropriate type.
  MVT InVT = MVT::getVectorVT(MVT::getIntegerVT(InBytes * 8),
                              SystemZ::VectorBytes / InBytes);
  Op0 = DAG.getNode(ISD::BITCAST, DL, InVT, Op0);
  Op1 = DAG.getNode(ISD::BITCAST, DL, InVT, Op1);
  SDValue Op;
  if (P.Opcode == SystemZISD::PERMUTE_DWORDS) {
    SDValue Op2 = DAG.getTargetConstant(P.Operand, DL, MVT::i32);
    Op = DAG.getNode(SystemZISD::PERMUTE_DWORDS, DL, InVT, Op0, Op1, Op2);
  } else if (P.Opcode == SystemZISD::PACK) {
    MVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(P.Operand * 8),
                                 SystemZ::VectorBytes / P.Operand);
    Op = DAG.getNode(SystemZISD::PACK, DL, OutVT, Op0, Op1);
  } else {
    Op = DAG.getNode(P.Opcode, DL, InVT, Op0, Op1);
  }
  return Op;
}
 
static bool isZeroVector(SDValue N) {
  if (N->getOpcode() == ISD::BITCAST)
    N = N->getOperand(0);
  if (N->getOpcode() == ISD::SPLAT_VECTOR)
    if (auto *Op = dyn_cast<ConstantSDNode>(N->getOperand(0)))
      return Op->getZExtValue() == 0;
  return ISD::isBuildVectorAllZeros(N.getNode());
}
 
// Return the index of the zero/undef vector, or UINT32_MAX if not found.
static uint32_t findZeroVectorIdx(SDValue *Ops, unsigned Num) {
  for (unsigned I = 0; I < Num ; I++)
    if (isZeroVector(Ops[I]))
      return I;
  return UINT32_MAX;
}
 
// Bytes is a VPERM-like permute vector, except that -1 is used for
// undefined bytes.  Implement it on operands Ops[0] and Ops[1] using
// VSLDB or VPERM.
static SDValue getGeneralPermuteNode(SelectionDAG &DAG, const SDLoc &DL,
                                     SDValue *Ops,
                                     const SmallVectorImpl<int> &Bytes) {
  for (unsigned I = 0; I < 2; ++I)
    Ops[I] = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Ops[I]);
 
  // First see whether VSLDB can be used.
  unsigned StartIndex, OpNo0, OpNo1;
  if (isShlDoublePermute(Bytes, StartIndex, OpNo0, OpNo1))
    return DAG.getNode(SystemZISD::SHL_DOUBLE, DL, MVT::v16i8, Ops[OpNo0],
                       Ops[OpNo1],
                       DAG.getTargetConstant(StartIndex, DL, MVT::i32));
 
  // Fall back on VPERM.  Construct an SDNode for the permute vector.  Try to
  // eliminate a zero vector by reusing any zero index in the permute vector.
  unsigned ZeroVecIdx = findZeroVectorIdx(&Ops[0], 2);
  if (ZeroVecIdx != UINT32_MAX) {
    bool MaskFirst = true;
    int ZeroIdx = -1;
    for (unsigned I = 0; I < SystemZ::VectorBytes; ++I) {
      unsigned OpNo = unsigned(Bytes[I]) / SystemZ::VectorBytes;
      unsigned Byte = unsigned(Bytes[I]) % SystemZ::VectorBytes;
      if (OpNo == ZeroVecIdx && I == 0) {
        // If the first byte is zero, use mask as first operand.
        ZeroIdx = 0;
        break;
      }
      if (OpNo != ZeroVecIdx && Byte == 0) {
        // If mask contains a zero, use it by placing that vector first.
        ZeroIdx = I + SystemZ::VectorBytes;
        MaskFirst = false;
        break;
      }
    }
    if (ZeroIdx != -1) {
      SDValue IndexNodes[SystemZ::VectorBytes];
      for (unsigned I = 0; I < SystemZ::VectorBytes; ++I) {
        if (Bytes[I] >= 0) {
          unsigned OpNo = unsigned(Bytes[I]) / SystemZ::VectorBytes;
          unsigned Byte = unsigned(Bytes[I]) % SystemZ::VectorBytes;
          if (OpNo == ZeroVecIdx)
            IndexNodes[I] = DAG.getConstant(ZeroIdx, DL, MVT::i32);
          else {
            unsigned BIdx = MaskFirst ? Byte + SystemZ::VectorBytes : Byte;
            IndexNodes[I] = DAG.getConstant(BIdx, DL, MVT::i32);
          }
        } else
          IndexNodes[I] = DAG.getUNDEF(MVT::i32);
      }
      SDValue Mask = DAG.getBuildVector(MVT::v16i8, DL, IndexNodes);
      SDValue Src = ZeroVecIdx == 0 ? Ops[1] : Ops[0];
      if (MaskFirst)
        return DAG.getNode(SystemZISD::PERMUTE, DL, MVT::v16i8, Mask, Src,
                           Mask);
      else
        return DAG.getNode(SystemZISD::PERMUTE, DL, MVT::v16i8, Src, Mask,
                           Mask);
    }
  }
 
  SDValue IndexNodes[SystemZ::VectorBytes];
  for (unsigned I = 0; I < SystemZ::VectorBytes; ++I)
    if (Bytes[I] >= 0)
      IndexNodes[I] = DAG.getConstant(Bytes[I], DL, MVT::i32);
    else
      IndexNodes[I] = DAG.getUNDEF(MVT::i32);
  SDValue Op2 = DAG.getBuildVector(MVT::v16i8, DL, IndexNodes);
  return DAG.getNode(SystemZISD::PERMUTE, DL, MVT::v16i8, Ops[0],
                     (!Ops[1].isUndef() ? Ops[1] : Ops[0]), Op2);
}
 
namespace {
// Describes a general N-operand vector shuffle.
struct GeneralShuffle {
  GeneralShuffle(EVT vt)
      : VT(vt), UnpackFromEltSize(UINT_MAX), UnpackLow(false) {}
  void addUndef();
  bool add(SDValue, unsigned);
  SDValue getNode(SelectionDAG &, const SDLoc &);
  void tryPrepareForUnpack();
  bool unpackWasPrepared() { return UnpackFromEltSize <= 4; }
  SDValue insertUnpackIfPrepared(SelectionDAG &DAG, const SDLoc &DL, SDValue Op);
 
  // The operands of the shuffle.
  SmallVector<SDValue, SystemZ::VectorBytes> Ops;
 
  // Index I is -1 if byte I of the result is undefined.  Otherwise the
  // result comes from byte Bytes[I] % SystemZ::VectorBytes of operand
  // Bytes[I] / SystemZ::VectorBytes.
  SmallVector<int, SystemZ::VectorBytes> Bytes;
 
  // The type of the shuffle result.
  EVT VT;
 
  // Holds a value of 1, 2 or 4 if a final unpack has been prepared for.
  unsigned UnpackFromEltSize;
  // True if the final unpack uses the low half.
  bool UnpackLow;
};
} // namespace
 
// Add an extra undefined element to the shuffle.
void GeneralShuffle::addUndef() {
  unsigned BytesPerElement = VT.getVectorElementType().getStoreSize();
  for (unsigned I = 0; I < BytesPerElement; ++I)
    Bytes.push_back(-1);
}
 
// Add an extra element to the shuffle, taking it from element Elem of Op.
// A null Op indicates a vector input whose value will be calculated later;
// there is at most one such input per shuffle and it always has the same
// type as the result. Aborts and returns false if the source vector elements
// of an EXTRACT_VECTOR_ELT are smaller than the destination elements. Per
// LLVM they become implicitly extended, but this is rare and not optimized.
bool GeneralShuffle::add(SDValue Op, unsigned Elem) {
  unsigned BytesPerElement = VT.getVectorElementType().getStoreSize();
 
  // The source vector can have wider elements than the result,
  // either through an explicit TRUNCATE or because of type legalization.
  // We want the least significant part.
  EVT FromVT = Op.getNode() ? Op.getValueType() : VT;
  unsigned FromBytesPerElement = FromVT.getVectorElementType().getStoreSize();
 
  // Return false if the source elements are smaller than their destination
  // elements.
  if (FromBytesPerElement < BytesPerElement)
    return false;
 
  unsigned Byte = ((Elem * FromBytesPerElement) % SystemZ::VectorBytes +
                   (FromBytesPerElement - BytesPerElement));
 
  // Look through things like shuffles and bitcasts.
  while (Op.getNode()) {
    if (Op.getOpcode() == ISD::BITCAST)
      Op = Op.getOperand(0);
    else if (Op.getOpcode() == ISD::VECTOR_SHUFFLE && Op.hasOneUse()) {
      // See whether the bytes we need come from a contiguous part of one
      // operand.
      SmallVector<int, SystemZ::VectorBytes> OpBytes;
      if (!getVPermMask(Op, OpBytes))
        break;
      int NewByte;
      if (!getShuffleInput(OpBytes, Byte, BytesPerElement, NewByte))
        break;
      if (NewByte < 0) {
        addUndef();
        return true;
      }
      Op = Op.getOperand(unsigned(NewByte) / SystemZ::VectorBytes);
      Byte = unsigned(NewByte) % SystemZ::VectorBytes;
    } else if (Op.isUndef()) {
      addUndef();
      return true;
    } else
      break;
  }
 
  // Make sure that the source of the extraction is in Ops.
  unsigned OpNo = 0;
  for (; OpNo < Ops.size(); ++OpNo)
    if (Ops[OpNo] == Op)
      break;
  if (OpNo == Ops.size())
    Ops.push_back(Op);
 
  // Add the element to Bytes.
  unsigned Base = OpNo * SystemZ::VectorBytes + Byte;
  for (unsigned I = 0; I < BytesPerElement; ++I)
    Bytes.push_back(Base + I);
 
  return true;
}
 
// Return SDNodes for the completed shuffle.
SDValue GeneralShuffle::getNode(SelectionDAG &DAG, const SDLoc &DL) {
  assert(Bytes.size() == SystemZ::VectorBytes && "Incomplete vector");
 
  if (Ops.size() == 0)
    return DAG.getUNDEF(VT);
 
  // Use a single unpack if possible as the last operation.
  tryPrepareForUnpack();
 
  // Make sure that there are at least two shuffle operands.
  if (Ops.size() == 1)
    Ops.push_back(DAG.getUNDEF(MVT::v16i8));
 
  // Create a tree of shuffles, deferring root node until after the loop.
  // Try to redistribute the undefined elements of non-root nodes so that
  // the non-root shuffles match something like a pack or merge, then adjust
  // the parent node's permute vector to compensate for the new order.
  // Among other things, this copes with vectors like <2 x i16> that were
  // padded with undefined elements during type legalization.
  //
  // In the best case this redistribution will lead to the whole tree
  // using packs and merges.  It should rarely be a loss in other cases.
  unsigned Stride = 1;
  for (; Stride * 2 < Ops.size(); Stride *= 2) {
    for (unsigned I = 0; I < Ops.size() - Stride; I += Stride * 2) {
      SDValue SubOps[] = { Ops[I], Ops[I + Stride] };
 
      // Create a mask for just these two operands.
      SmallVector<int, SystemZ::VectorBytes> NewBytes(SystemZ::VectorBytes);
      for (unsigned J = 0; J < SystemZ::VectorBytes; ++J) {
        unsigned OpNo = unsigned(Bytes[J]) / SystemZ::VectorBytes;
        unsigned Byte = unsigned(Bytes[J]) % SystemZ::VectorBytes;
        if (OpNo == I)
          NewBytes[J] = Byte;
        else if (OpNo == I + Stride)
          NewBytes[J] = SystemZ::VectorBytes + Byte;
        else
          NewBytes[J] = -1;
      }
      // See if it would be better to reorganize NewMask to avoid using VPERM.
      SmallVector<int, SystemZ::VectorBytes> NewBytesMap(SystemZ::VectorBytes);
      if (const Permute *P = matchDoublePermute(NewBytes, NewBytesMap)) {
        Ops[I] = getPermuteNode(DAG, DL, *P, SubOps[0], SubOps[1]);
        // Applying NewBytesMap to Ops[I] gets back to NewBytes.
        for (unsigned J = 0; J < SystemZ::VectorBytes; ++J) {
          if (NewBytes[J] >= 0) {
            assert(unsigned(NewBytesMap[J]) < SystemZ::VectorBytes &&
                   "Invalid double permute");
            Bytes[J] = I * SystemZ::VectorBytes + NewBytesMap[J];
          } else
            assert(NewBytesMap[J] < 0 && "Invalid double permute");
        }
      } else {
        // Just use NewBytes on the operands.
        Ops[I] = getGeneralPermuteNode(DAG, DL, SubOps, NewBytes);
        for (unsigned J = 0; J < SystemZ::VectorBytes; ++J)
          if (NewBytes[J] >= 0)
            Bytes[J] = I * SystemZ::VectorBytes + J;
      }
    }
  }
 
  // Now we just have 2 inputs.  Put the second operand in Ops[1].
  if (Stride > 1) {
    Ops[1] = Ops[Stride];
    for (unsigned I = 0; I < SystemZ::VectorBytes; ++I)
      if (Bytes[I] >= int(SystemZ::VectorBytes))
        Bytes[I] -= (Stride - 1) * SystemZ::VectorBytes;
  }
 
  // Look for an instruction that can do the permute without resorting
  // to VPERM.
  unsigned OpNo0, OpNo1;
  SDValue Op;
  if (unpackWasPrepared() && Ops[1].isUndef())
    Op = Ops[0];
  else if (const Permute *P = matchPermute(Bytes, OpNo0, OpNo1))
    Op = getPermuteNode(DAG, DL, *P, Ops[OpNo0], Ops[OpNo1]);
  else
    Op = getGeneralPermuteNode(DAG, DL, &Ops[0], Bytes);
 
  Op = insertUnpackIfPrepared(DAG, DL, Op);
 
  return DAG.getNode(ISD::BITCAST, DL, VT, Op);
}
 
#ifndef NDEBUG
static void dumpBytes(const SmallVectorImpl<int> &Bytes, std::string Msg) {
  dbgs() << Msg.c_str() << " { ";
  for (unsigned I = 0; I < Bytes.size(); I++)
    dbgs() << Bytes[I] << " ";
  dbgs() << "}\n";
}
#endif
 
// If the Bytes vector matches an unpack operation, prepare to do the unpack
// after all else by removing the zero vector and the effect of the unpack on
// Bytes.
void GeneralShuffle::tryPrepareForUnpack() {
  uint32_t ZeroVecOpNo = findZeroVectorIdx(&Ops[0], Ops.size());
  if (ZeroVecOpNo == UINT32_MAX || Ops.size() == 1)
    return;
 
  // Only do this if removing the zero vector reduces the depth, otherwise
  // the critical path will increase with the final unpack.
  if (Ops.size() > 2 &&
      Log2_32_Ceil(Ops.size()) == Log2_32_Ceil(Ops.size() - 1))
    return;
 
  // Find an unpack that would allow removing the zero vector from Ops.
  UnpackFromEltSize = 1;
  for (; UnpackFromEltSize <= 4; UnpackFromEltSize *= 2) {
    bool MatchUnpack = true;
    SmallVector<int, SystemZ::VectorBytes> SrcBytes;
    for (unsigned Elt = 0; Elt < SystemZ::VectorBytes; Elt++) {
      unsigned ToEltSize = UnpackFromEltSize * 2;
      bool IsZextByte = (Elt % ToEltSize) < UnpackFromEltSize;
      if (!IsZextByte)
        SrcBytes.push_back(Bytes[Elt]);
      if (Bytes[Elt] != -1) {
        unsigned OpNo = unsigned(Bytes[Elt]) / SystemZ::VectorBytes;
        if (IsZextByte != (OpNo == ZeroVecOpNo)) {
          MatchUnpack = false;
          break;
        }
      }
    }
    if (MatchUnpack) {
      if (Ops.size() == 2) {
        // Don't use unpack if a single source operand needs rearrangement.
        bool CanUseUnpackLow = true, CanUseUnpackHigh = true;
        for (unsigned i = 0; i < SystemZ::VectorBytes / 2; i++) {
          if (SrcBytes[i] == -1)
            continue;
          if (SrcBytes[i] % 16 != int(i))
            CanUseUnpackHigh = false;
          if (SrcBytes[i] % 16 != int(i + SystemZ::VectorBytes / 2))
            CanUseUnpackLow = false;
          if (!CanUseUnpackLow && !CanUseUnpackHigh) {
            UnpackFromEltSize = UINT_MAX;
            return;
          }
        }
        if (!CanUseUnpackHigh)
          UnpackLow = true;
      }
      break;
    }
  }
  if (UnpackFromEltSize > 4)
    return;
 
  LLVM_DEBUG(dbgs() << "Preparing for final unpack of element size "
             << UnpackFromEltSize << ". Zero vector is Op#" << ZeroVecOpNo
             << ".\n";
             dumpBytes(Bytes, "Original Bytes vector:"););
 
  // Apply the unpack in reverse to the Bytes array.
  unsigned B = 0;
  if (UnpackLow) {
    while (B < SystemZ::VectorBytes / 2)
      Bytes[B++] = -1;
  }
  for (unsigned Elt = 0; Elt < SystemZ::VectorBytes;) {
    Elt += UnpackFromEltSize;
    for (unsigned i = 0; i < UnpackFromEltSize; i++, Elt++, B++)
      Bytes[B] = Bytes[Elt];
  }
  if (!UnpackLow) {
    while (B < SystemZ::VectorBytes)
      Bytes[B++] = -1;
  }
 
  // Remove the zero vector from Ops
  Ops.erase(&Ops[ZeroVecOpNo]);
  for (unsigned I = 0; I < SystemZ::VectorBytes; ++I)
    if (Bytes[I] >= 0) {
      unsigned OpNo = unsigned(Bytes[I]) / SystemZ::VectorBytes;
      if (OpNo > ZeroVecOpNo)
        Bytes[I] -= SystemZ::VectorBytes;
    }
 
  LLVM_DEBUG(dumpBytes(Bytes, "Resulting Bytes vector, zero vector removed:");
             dbgs() << "\n";);
}
 
SDValue GeneralShuffle::insertUnpackIfPrepared(SelectionDAG &DAG,
                                               const SDLoc &DL,
                                               SDValue Op) {
  if (!unpackWasPrepared())
    return Op;
  unsigned InBits = UnpackFromEltSize * 8;
  EVT InVT = MVT::getVectorVT(MVT::getIntegerVT(InBits),
                                SystemZ::VectorBits / InBits);
  SDValue PackedOp = DAG.getNode(ISD::BITCAST, DL, InVT, Op);
  unsigned OutBits = InBits * 2;
  EVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(OutBits),
                               SystemZ::VectorBits / OutBits);
  return DAG.getNode(UnpackLow ? SystemZISD::UNPACKL_LOW
                               : SystemZISD::UNPACKL_HIGH,
                     DL, OutVT, PackedOp);
}
 
// Return true if the given BUILD_VECTOR is a scalar-to-vector conversion.
static bool isScalarToVector(SDValue Op) {
  for (unsigned I = 1, E = Op.getNumOperands(); I != E; ++I)
    if (!Op.getOperand(I).isUndef())
      return false;
  return true;
}
 
// Return a vector of type VT that contains Value in the first element.
// The other elements don't matter.
static SDValue buildScalarToVector(SelectionDAG &DAG, const SDLoc &DL, EVT VT,
                                   SDValue Value) {
  // If we have a constant, replicate it to all elements and let the
  // BUILD_VECTOR lowering take care of it.
  if (Value.getOpcode() == ISD::Constant ||
      Value.getOpcode() == ISD::ConstantFP) {
    SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Value);
    return DAG.getBuildVector(VT, DL, Ops);
  }
  if (Value.isUndef())
    return DAG.getUNDEF(VT);
  return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
}
 
// Return a vector of type VT in which Op0 is in element 0 and Op1 is in
// element 1.  Used for cases in which replication is cheap.
static SDValue buildMergeScalars(SelectionDAG &DAG, const SDLoc &DL, EVT VT,
                                 SDValue Op0, SDValue Op1) {
  if (Op0.isUndef()) {
    if (Op1.isUndef())
      return DAG.getUNDEF(VT);
    return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op1);
  }
  if (Op1.isUndef())
    return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op0);
  return DAG.getNode(SystemZISD::MERGE_HIGH, DL, VT,
                     buildScalarToVector(DAG, DL, VT, Op0),
                     buildScalarToVector(DAG, DL, VT, Op1));
}
 
// Extend GPR scalars Op0 and Op1 to doublewords and return a v2i64
// vector for them.
static SDValue joinDwords(SelectionDAG &DAG, const SDLoc &DL, SDValue Op0,
                          SDValue Op1) {
  if (Op0.isUndef() && Op1.isUndef())
    return DAG.getUNDEF(MVT::v2i64);
  // If one of the two inputs is undefined then replicate the other one,
  // in order to avoid using another register unnecessarily.
  if (Op0.isUndef())
    Op0 = Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op1);
  else if (Op1.isUndef())
    Op0 = Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
  else {
    Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
    Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op1);
  }
  return DAG.getNode(SystemZISD::JOIN_DWORDS, DL, MVT::v2i64, Op0, Op1);
}
 
// If a BUILD_VECTOR contains some EXTRACT_VECTOR_ELTs, it's usually
// better to use VECTOR_SHUFFLEs on them, only using BUILD_VECTOR for
// the non-EXTRACT_VECTOR_ELT elements.  See if the given BUILD_VECTOR
// would benefit from this representation and return it if so.
static SDValue tryBuildVectorShuffle(SelectionDAG &DAG,
                                     BuildVectorSDNode *BVN) {
  EVT VT = BVN->getValueType(0);
  unsigned NumElements = VT.getVectorNumElements();
 
  // Represent the BUILD_VECTOR as an N-operand VECTOR_SHUFFLE-like operation
  // on byte vectors.  If there are non-EXTRACT_VECTOR_ELT elements that still
  // need a BUILD_VECTOR, add an additional placeholder operand for that
  // BUILD_VECTOR and store its operands in ResidueOps.
  GeneralShuffle GS(VT);
  SmallVector<SDValue, SystemZ::VectorBytes> ResidueOps;
  bool FoundOne = false;
  for (unsigned I = 0; I < NumElements; ++I) {
    SDValue Op = BVN->getOperand(I);
    if (Op.getOpcode() == ISD::TRUNCATE)
      Op = Op.getOperand(0);
    if (Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
        Op.getOperand(1).getOpcode() == ISD::Constant) {
      unsigned Elem = Op.getConstantOperandVal(1);
      if (!GS.add(Op.getOperand(0), Elem))
        return SDValue();
      FoundOne = true;
    } else if (Op.isUndef()) {
      GS.addUndef();
    } else {
      if (!GS.add(SDValue(), ResidueOps.size()))
        return SDValue();
      ResidueOps.push_back(BVN->getOperand(I));
    }
  }
 
  // Nothing to do if there are no EXTRACT_VECTOR_ELTs.
  if (!FoundOne)
    return SDValue();
 
  // Create the BUILD_VECTOR for the remaining elements, if any.
  if (!ResidueOps.empty()) {
    while (ResidueOps.size() < NumElements)
      ResidueOps.push_back(DAG.getUNDEF(ResidueOps[0].getValueType()));
    for (auto &Op : GS.Ops) {
      if (!Op.getNode()) {
        Op = DAG.getBuildVector(VT, SDLoc(BVN), ResidueOps);
        break;
      }
    }
  }
  return GS.getNode(DAG, SDLoc(BVN));
}
 
bool SystemZTargetLowering::isVectorElementLoad(SDValue Op) const {
  if (Op.getOpcode() == ISD::LOAD && cast<LoadSDNode>(Op)->isUnindexed())
    return true;
  if (auto *AL = dyn_cast<AtomicSDNode>(Op))
    if (AL->getOpcode() == ISD::ATOMIC_LOAD)
      return true;
  if (Subtarget.hasVectorEnhancements2() && Op.getOpcode() == SystemZISD::LRV)
    return true;
  return false;
}
 
static SDValue mergeHighParts(SelectionDAG &DAG, const SDLoc &DL,
                              unsigned MergedBits, EVT VT, SDValue Op0,
                              SDValue Op1) {
  MVT IntVecVT = MVT::getVectorVT(MVT::getIntegerVT(MergedBits),
                                  SystemZ::VectorBits / MergedBits);
  assert(VT.getSizeInBits() == 128 && IntVecVT.getSizeInBits() == 128 &&
         "Handling full vectors only.");
  Op0 = DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op0);
  Op1 = DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op1);
  SDValue Op = DAG.getNode(SystemZISD::MERGE_HIGH, DL, IntVecVT, Op0, Op1);
  return DAG.getNode(ISD::BITCAST, DL, VT, Op);
}
 
static SDValue buildFPVecFromScalars4(SelectionDAG &DAG, const SDLoc &DL,
                                      EVT VT, SmallVectorImpl<SDValue> &Elems,
                                      unsigned Pos) {
  SDValue Op01 = buildMergeScalars(DAG, DL, VT, Elems[Pos + 0], Elems[Pos + 1]);
  SDValue Op23 = buildMergeScalars(DAG, DL, VT, Elems[Pos + 2], Elems[Pos + 3]);
  // Avoid unnecessary undefs by reusing the other operand.
  if (Op01.isUndef()) {
    if (Op23.isUndef())
      return Op01;
    Op01 = Op23;
  } else if (Op23.isUndef())
    Op23 = Op01;
  // Merging identical replications is a no-op.
  if (Op01.getOpcode() == SystemZISD::REPLICATE && Op01 == Op23)
    return Op01;
  unsigned MergedBits = VT.getSimpleVT().getScalarSizeInBits() * 2;
  return mergeHighParts(DAG, DL, MergedBits, VT, Op01, Op23);
}
 
// Combine GPR scalar values Elems into a vector of type VT.
SDValue
SystemZTargetLowering::buildVector(SelectionDAG &DAG, const SDLoc &DL, EVT VT,
                                   SmallVectorImpl<SDValue> &Elems) const {
  // See whether there is a single replicated value.
  SDValue Single;
  unsigned int NumElements = Elems.size();
  unsigned int Count = 0;
  for (auto Elem : Elems) {
    if (!Elem.isUndef()) {
      if (!Single.getNode())
        Single = Elem;
      else if (Elem != Single) {
        Single = SDValue();
        break;
      }
      Count += 1;
    }
  }
  // There are three cases here:
  //
  // - if the only defined element is a loaded one, the best sequence
  //   is a replicating load.
  //
  // - otherwise, if the only defined element is an i64 value, we will
  //   end up with the same VLVGP sequence regardless of whether we short-cut
  //   for replication or fall through to the later code.
  //
  // - otherwise, if the only defined element is an i32 or smaller value,
  //   we would need 2 instructions to replicate it: VLVGP followed by VREPx.
  //   This is only a win if the single defined element is used more than once.
  //   In other cases we're better off using a single VLVGx.
  if (Single.getNode() && (Count > 1 || isVectorElementLoad(Single)))
    return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Single);
 
  // If all elements are loads, use VLREP/VLEs (below).
  bool AllLoads = true;
  for (auto Elem : Elems)
    if (!isVectorElementLoad(Elem)) {
      AllLoads = false;
      break;
    }
 
  // The best way of building a v2i64 from two i64s is to use VLVGP.
  if (VT == MVT::v2i64 && !AllLoads)
    return joinDwords(DAG, DL, Elems[0], Elems[1]);
 
  // Use a 64-bit merge high to combine two doubles.
  if (VT == MVT::v2f64 && !AllLoads)
    return buildMergeScalars(DAG, DL, VT, Elems[0], Elems[1]);
 
  // Build v4f32 values directly from the FPRs:
  //
  //   <Axxx> <Bxxx> <Cxxxx> <Dxxx>
  //         V              V         VMRHF
  //      <ABxx>         <CDxx>
  //                V                 VMRHG
  //              <ABCD>
  if (VT == MVT::v4f32 && !AllLoads)
    return buildFPVecFromScalars4(DAG, DL, VT, Elems, 0);
 
  // Same for v8f16.
  if (VT == MVT::v8f16 && !AllLoads) {
    SDValue Op0123 = buildFPVecFromScalars4(DAG, DL, VT, Elems, 0);
    SDValue Op4567 = buildFPVecFromScalars4(DAG, DL, VT, Elems, 4);
    // Avoid unnecessary undefs by reusing the other operand.
    if (Op0123.isUndef())
      Op0123 = Op4567;
    else if (Op4567.isUndef())
      Op4567 = Op0123;
    // Merging identical replications is a no-op.
    if (Op0123.getOpcode() == SystemZISD::REPLICATE && Op0123 == Op4567)
      return Op0123;
    return mergeHighParts(DAG, DL, 64, VT, Op0123, Op4567);
  }
 
  // Collect the constant terms.
  SmallVector<SDValue, SystemZ::VectorBytes> Constants(NumElements, SDValue());
  SmallVector<bool, SystemZ::VectorBytes> Done(NumElements, false);
 
  unsigned NumConstants = 0;
  for (unsigned I = 0; I < NumElements; ++I) {
    SDValue Elem = Elems[I];
    if (Elem.getOpcode() == ISD::Constant ||
        Elem.getOpcode() == ISD::ConstantFP) {
      NumConstants += 1;
      Constants[I] = Elem;
      Done[I] = true;
    }
  }
  // If there was at least one constant, fill in the other elements of
  // Constants with undefs to get a full vector constant and use that
  // as the starting point.
  SDValue Result;
  SDValue ReplicatedVal;
  if (NumConstants > 0) {
    for (unsigned I = 0; I < NumElements; ++I)
      if (!Constants[I].getNode())
        Constants[I] = DAG.getUNDEF(Elems[I].getValueType());
    Result = DAG.getBuildVector(VT, DL, Constants);
  } else {
    // Otherwise try to use VLREP or VLVGP to start the sequence in order to
    // avoid a false dependency on any previous contents of the vector
    // register.
 
    // Use a VLREP if at least one element is a load. Make sure to replicate
    // the load with the most elements having its value.
    std::map<const SDNode*, unsigned> UseCounts;
    SDNode *LoadMaxUses = nullptr;
    for (unsigned I = 0; I < NumElements; ++I)
      if (isVectorElementLoad(Elems[I])) {
        SDNode *Ld = Elems[I].getNode();
        unsigned Count = ++UseCounts[Ld];
        if (LoadMaxUses == nullptr || UseCounts[LoadMaxUses] < Count)
          LoadMaxUses = Ld;
      }
    if (LoadMaxUses != nullptr) {
      ReplicatedVal = SDValue(LoadMaxUses, 0);
      Result = DAG.getNode(SystemZISD::REPLICATE, DL, VT, ReplicatedVal);
    } else {
      // Try to use VLVGP.
      unsigned I1 = NumElements / 2 - 1;
      unsigned I2 = NumElements - 1;
      bool Def1 = !Elems[I1].isUndef();
      bool Def2 = !Elems[I2].isUndef();
      if (Def1 || Def2) {
        SDValue Elem1 = Elems[Def1 ? I1 : I2];
        SDValue Elem2 = Elems[Def2 ? I2 : I1];
        Result = DAG.getNode(ISD::BITCAST, DL, VT,
                             joinDwords(DAG, DL, Elem1, Elem2));
        Done[I1] = true;
        Done[I2] = true;
      } else
        Result = DAG.getUNDEF(VT);
    }
  }
 
  // Use VLVGx to insert the other elements.
  for (unsigned I = 0; I < NumElements; ++I)
    if (!Done[I] && !Elems[I].isUndef() && Elems[I] != ReplicatedVal)
      Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Result, Elems[I],
                           DAG.getConstant(I, DL, MVT::i32));
  return Result;
}
 
SDValue SystemZTargetLowering::lowerBUILD_VECTOR(SDValue Op,
                                                 SelectionDAG &DAG) const {
  auto *BVN = cast<BuildVectorSDNode>(Op.getNode());
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
 
  if (BVN->isConstant()) {
    if (SystemZVectorConstantInfo(BVN).isVectorConstantLegal(Subtarget))
      return Op;
 
    // Fall back to loading it from memory.
    return SDValue();
  }
 
  // See if we should use shuffles to construct the vector from other vectors.
  if (SDValue Res = tryBuildVectorShuffle(DAG, BVN))
    return Res;
 
  // Detect SCALAR_TO_VECTOR conversions.
  if (isOperationLegal(ISD::SCALAR_TO_VECTOR, VT) && isScalarToVector(Op))
    return buildScalarToVector(DAG, DL, VT, Op.getOperand(0));
 
  // Otherwise use buildVector to build the vector up from GPRs.
  unsigned NumElements = Op.getNumOperands();
  SmallVector<SDValue, SystemZ::VectorBytes> Ops(NumElements);
  for (unsigned I = 0; I < NumElements; ++I)
    Ops[I] = Op.getOperand(I);
  return buildVector(DAG, DL, VT, Ops);
}
 
SDValue SystemZTargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
                                                   SelectionDAG &DAG) const {
  auto *VSN = cast<ShuffleVectorSDNode>(Op.getNode());
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  unsigned NumElements = VT.getVectorNumElements();
 
  if (VSN->isSplat()) {
    SDValue Op0 = Op.getOperand(0);
    unsigned Index = VSN->getSplatIndex();
    assert(Index < VT.getVectorNumElements() &&
           "Splat index should be defined and in first operand");
    // See whether the value we're splatting is directly available as a scalar.
    if ((Index == 0 && Op0.getOpcode() == ISD::SCALAR_TO_VECTOR) ||
        Op0.getOpcode() == ISD::BUILD_VECTOR)
      return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op0.getOperand(Index));
    // Otherwise keep it as a vector-to-vector operation.
    return DAG.getNode(SystemZISD::SPLAT, DL, VT, Op.getOperand(0),
                       DAG.getTargetConstant(Index, DL, MVT::i32));
  }
 
  GeneralShuffle GS(VT);
  for (unsigned I = 0; I < NumElements; ++I) {
    int Elt = VSN->getMaskElt(I);
    if (Elt < 0)
      GS.addUndef();
    else if (!GS.add(Op.getOperand(unsigned(Elt) / NumElements),
                     unsigned(Elt) % NumElements))
      return SDValue();
  }
  return GS.getNode(DAG, SDLoc(VSN));
}
 
SDValue SystemZTargetLowering::lowerSCALAR_TO_VECTOR(SDValue Op,
                                                     SelectionDAG &DAG) const {
  SDLoc DL(Op);
  // Just insert the scalar into element 0 of an undefined vector.
  return DAG.getNode(ISD::INSERT_VECTOR_ELT, DL,
                     Op.getValueType(), DAG.getUNDEF(Op.getValueType()),
                     Op.getOperand(0), DAG.getConstant(0, DL, MVT::i32));
}
 
// Shift the lower 2 bytes of Op to the left in order to insert into the
// upper 2 bytes of the FP register.
static SDValue convertToF16(SDValue Op, SelectionDAG &DAG) {
  assert(Op.getSimpleValueType() == MVT::i64 &&
         "Expexted to convert i64 to f16.");
  SDLoc DL(Op);
  SDValue Shft = DAG.getNode(ISD::SHL, DL, MVT::i64, Op,
                             DAG.getConstant(48, DL, MVT::i64));
  SDValue BCast = DAG.getNode(ISD::BITCAST, DL, MVT::f64, Shft);
  SDValue F16Val =
      DAG.getTargetExtractSubreg(SystemZ::subreg_h16, DL, MVT::f16, BCast);
  return F16Val;
}
 
// Extract Op into GPR and shift the 2 f16 bytes to the right.
static SDValue convertFromF16(SDValue Op, SDLoc DL, SelectionDAG &DAG) {
  assert(Op.getSimpleValueType() == MVT::f16 &&
         "Expected to convert f16 to i64.");
  SDNode *U32 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64);
  SDValue In64 = DAG.getTargetInsertSubreg(SystemZ::subreg_h16, DL, MVT::f64,
                                           SDValue(U32, 0), Op);
  SDValue BCast = DAG.getNode(ISD::BITCAST, DL, MVT::i64, In64);
  SDValue Shft = DAG.getNode(ISD::SRL, DL, MVT::i64, BCast,
                             DAG.getConstant(48, DL, MVT::i32));
  return Shft;
}
 
SDValue SystemZTargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op,
                                                      SelectionDAG &DAG) const {
  // Handle insertions of floating-point values.
  SDLoc DL(Op);
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  SDValue Op2 = Op.getOperand(2);
  EVT VT = Op.getValueType();
 
  // Insertions into constant indices of a v2f64 can be done using VPDI.
  // However, if the inserted value is a bitcast or a constant then it's
  // better to use GPRs, as below.
  if (VT == MVT::v2f64 &&
      Op1.getOpcode() != ISD::BITCAST &&
      Op1.getOpcode() != ISD::ConstantFP &&
      Op2.getOpcode() == ISD::Constant) {
    uint64_t Index = Op2->getAsZExtVal();
    unsigned Mask = VT.getVectorNumElements() - 1;
    if (Index <= Mask)
      return Op;
  }
 
  // Otherwise bitcast to the equivalent integer form and insert via a GPR.
  MVT IntVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
  MVT IntVecVT = MVT::getVectorVT(IntVT, VT.getVectorNumElements());
  SDValue IntOp1 =
      VT == MVT::v8f16
          ? DAG.getZExtOrTrunc(convertFromF16(Op1, DL, DAG), DL, MVT::i32)
          : DAG.getNode(ISD::BITCAST, DL, IntVT, Op1);
  SDValue Res =
      DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, IntVecVT,
                  DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op0), IntOp1, Op2);
  return DAG.getNode(ISD::BITCAST, DL, VT, Res);
}
 
SDValue
SystemZTargetLowering::lowerEXTRACT_VECTOR_ELT(SDValue Op,
                                               SelectionDAG &DAG) const {
  // Handle extractions of floating-point values.
  SDLoc DL(Op);
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  EVT VT = Op.getValueType();
  EVT VecVT = Op0.getValueType();
 
  // Extractions of constant indices can be done directly.
  if (auto *CIndexN = dyn_cast<ConstantSDNode>(Op1)) {
    uint64_t Index = CIndexN->getZExtValue();
    unsigned Mask = VecVT.getVectorNumElements() - 1;
    if (Index <= Mask)
      return Op;
  }
 
  // Otherwise bitcast to the equivalent integer form and extract via a GPR.
  MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits());
  MVT IntVecVT = MVT::getVectorVT(IntVT, VecVT.getVectorNumElements());
  MVT ExtrVT = IntVT == MVT::i16 ? MVT::i32 : IntVT;
  SDValue Extr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrVT,
                             DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op0), Op1);
  if (VT == MVT::f16)
    return convertToF16(DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Extr), DAG);
  return DAG.getNode(ISD::BITCAST, DL, VT, Extr);
}
 
SDValue SystemZTargetLowering::
lowerSIGN_EXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const {
  SDValue PackedOp = Op.getOperand(0);
  EVT OutVT = Op.getValueType();
  EVT InVT = PackedOp.getValueType();
  unsigned ToBits = OutVT.getScalarSizeInBits();
  unsigned FromBits = InVT.getScalarSizeInBits();
  unsigned StartOffset = 0;
 
  // If the input is a VECTOR_SHUFFLE, there are a number of important
  // cases where we can directly implement the sign-extension of the
  // original input lanes of the shuffle.
  if (PackedOp.getOpcode() == ISD::VECTOR_SHUFFLE) {
    ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(PackedOp.getNode());
    ArrayRef<int> ShuffleMask = SVN->getMask();
    int OutNumElts = OutVT.getVectorNumElements();
 
    // Recognize the special case where the sign-extension can be done
    // by the VSEG instruction.  Handled via the default expander.
    if (ToBits == 64 && OutNumElts == 2) {
      int NumElem = ToBits / FromBits;
      if (ShuffleMask[0] == NumElem - 1 && ShuffleMask[1] == 2 * NumElem - 1)
        return SDValue();
    }
 
    // Recognize the special case where we can fold the shuffle by
    // replacing some of the UNPACK_HIGH with UNPACK_LOW.
    int StartOffsetCandidate = -1;
    for (int Elt = 0; Elt < OutNumElts; Elt++) {
      if (ShuffleMask[Elt] == -1)
        continue;
      if (ShuffleMask[Elt] % OutNumElts == Elt) {
        if (StartOffsetCandidate == -1)
          StartOffsetCandidate = ShuffleMask[Elt] - Elt;
        if (StartOffsetCandidate == ShuffleMask[Elt] - Elt)
          continue;
      }
      StartOffsetCandidate = -1;
      break;
    }
    if (StartOffsetCandidate != -1) {
      StartOffset = StartOffsetCandidate;
      PackedOp = PackedOp.getOperand(0);
    }
  }
 
  do {
    FromBits *= 2;
    unsigned OutNumElts = SystemZ::VectorBits / FromBits;
    EVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(FromBits), OutNumElts);
    unsigned Opcode = SystemZISD::UNPACK_HIGH;
    if (StartOffset >= OutNumElts) {
      Opcode = SystemZISD::UNPACK_LOW;
      StartOffset -= OutNumElts;
    }
    PackedOp = DAG.getNode(Opcode, SDLoc(PackedOp), OutVT, PackedOp);
  } while (FromBits != ToBits);
  return PackedOp;
}
 
// Lower a ZERO_EXTEND_VECTOR_INREG to a vector shuffle with a zero vector.
SDValue SystemZTargetLowering::
lowerZERO_EXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const {
  SDValue PackedOp = Op.getOperand(0);
  SDLoc DL(Op);
  EVT OutVT = Op.getValueType();
  EVT InVT = PackedOp.getValueType();
  unsigned InNumElts = InVT.getVectorNumElements();
  unsigned OutNumElts = OutVT.getVectorNumElements();
  unsigned NumInPerOut = InNumElts / OutNumElts;
 
  SDValue ZeroVec =
    DAG.getSplatVector(InVT, DL, DAG.getConstant(0, DL, InVT.getScalarType()));
 
  SmallVector<int, 16> Mask(InNumElts);
  unsigned ZeroVecElt = InNumElts;
  for (unsigned PackedElt = 0; PackedElt < OutNumElts; PackedElt++) {
    unsigned MaskElt = PackedElt * NumInPerOut;
    unsigned End = MaskElt + NumInPerOut - 1;
    for (; MaskElt < End; MaskElt++)
      Mask[MaskElt] = ZeroVecElt++;
    Mask[MaskElt] = PackedElt;
  }
  SDValue Shuf = DAG.getVectorShuffle(InVT, DL, PackedOp, ZeroVec, Mask);
  return DAG.getNode(ISD::BITCAST, DL, OutVT, Shuf);
}
 
SDValue SystemZTargetLowering::lowerShift(SDValue Op, SelectionDAG &DAG,
                                          unsigned ByScalar) const {
  // Look for cases where a vector shift can use the *_BY_SCALAR form.
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  unsigned ElemBitSize = VT.getScalarSizeInBits();
 
  // See whether the shift vector is a splat represented as BUILD_VECTOR.
  if (auto *BVN = dyn_cast<BuildVectorSDNode>(Op1)) {
    APInt SplatBits, SplatUndef;
    unsigned SplatBitSize;
    bool HasAnyUndefs;
    // Check for constant splats.  Use ElemBitSize as the minimum element
    // width and reject splats that need wider elements.
    if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs,
                             ElemBitSize, true) &&
        SplatBitSize == ElemBitSize) {
      SDValue Shift = DAG.getConstant(SplatBits.getZExtValue() & 0xfff,
                                      DL, MVT::i32);
      return DAG.getNode(ByScalar, DL, VT, Op0, Shift);
    }
    // Check for variable splats.
    BitVector UndefElements;
    SDValue Splat = BVN->getSplatValue(&UndefElements);
    if (Splat) {
      // Since i32 is the smallest legal type, we either need a no-op
      // or a truncation.
      SDValue Shift = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Splat);
      return DAG.getNode(ByScalar, DL, VT, Op0, Shift);
    }
  }
 
  // See whether the shift vector is a splat represented as SHUFFLE_VECTOR,
  // and the shift amount is directly available in a GPR.
  if (auto *VSN = dyn_cast<ShuffleVectorSDNode>(Op1)) {
    if (VSN->isSplat()) {
      SDValue VSNOp0 = VSN->getOperand(0);
      unsigned Index = VSN->getSplatIndex();
      assert(Index < VT.getVectorNumElements() &&
             "Splat index should be defined and in first operand");
      if ((Index == 0 && VSNOp0.getOpcode() == ISD::SCALAR_TO_VECTOR) ||
          VSNOp0.getOpcode() == ISD::BUILD_VECTOR) {
        // Since i32 is the smallest legal type, we either need a no-op
        // or a truncation.
        SDValue Shift = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32,
                                    VSNOp0.getOperand(Index));
        return DAG.getNode(ByScalar, DL, VT, Op0, Shift);
      }
    }
  }
 
  // Otherwise just treat the current form as legal.
  return Op;
}
 
SDValue SystemZTargetLowering::lowerFSHL(SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
 
  // i128 FSHL with a constant amount that is a multiple of 8 can be
  // implemented via VECTOR_SHUFFLE.  If we have the vector-enhancements-2
  // facility, FSHL with a constant amount less than 8 can be implemented
  // via SHL_DOUBLE_BIT, and FSHL with other constant amounts by a
  // combination of the two.
  if (auto *ShiftAmtNode = dyn_cast<ConstantSDNode>(Op.getOperand(2))) {
    uint64_t ShiftAmt = ShiftAmtNode->getZExtValue() & 127;
    if ((ShiftAmt & 7) == 0 || Subtarget.hasVectorEnhancements2()) {
      SDValue Op0 = DAG.getBitcast(MVT::v16i8, Op.getOperand(0));
      SDValue Op1 = DAG.getBitcast(MVT::v16i8, Op.getOperand(1));
      if (ShiftAmt > 120) {
        // For N in 121..128, fshl N == fshr (128 - N), and for 1 <= N < 8
        // SHR_DOUBLE_BIT emits fewer instructions.
        SDValue Val =
            DAG.getNode(SystemZISD::SHR_DOUBLE_BIT, DL, MVT::v16i8, Op0, Op1,
                        DAG.getTargetConstant(128 - ShiftAmt, DL, MVT::i32));
        return DAG.getBitcast(MVT::i128, Val);
      }
      SmallVector<int, 16> Mask(16);
      for (unsigned Elt = 0; Elt < 16; Elt++)
        Mask[Elt] = (ShiftAmt >> 3) + Elt;
      SDValue Shuf1 = DAG.getVectorShuffle(MVT::v16i8, DL, Op0, Op1, Mask);
      if ((ShiftAmt & 7) == 0)
        return DAG.getBitcast(MVT::i128, Shuf1);
      SDValue Shuf2 = DAG.getVectorShuffle(MVT::v16i8, DL, Op1, Op1, Mask);
      SDValue Val =
          DAG.getNode(SystemZISD::SHL_DOUBLE_BIT, DL, MVT::v16i8, Shuf1, Shuf2,
                      DAG.getTargetConstant(ShiftAmt & 7, DL, MVT::i32));
      return DAG.getBitcast(MVT::i128, Val);
    }
  }
 
  return SDValue();
}
 
SDValue SystemZTargetLowering::lowerFSHR(SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
 
  // i128 FSHR with a constant amount that is a multiple of 8 can be
  // implemented via VECTOR_SHUFFLE.  If we have the vector-enhancements-2
  // facility, FSHR with a constant amount less than 8 can be implemented
  // via SHR_DOUBLE_BIT, and FSHR with other constant amounts by a
  // combination of the two.
  if (auto *ShiftAmtNode = dyn_cast<ConstantSDNode>(Op.getOperand(2))) {
    uint64_t ShiftAmt = ShiftAmtNode->getZExtValue() & 127;
    if ((ShiftAmt & 7) == 0 || Subtarget.hasVectorEnhancements2()) {
      SDValue Op0 = DAG.getBitcast(MVT::v16i8, Op.getOperand(0));
      SDValue Op1 = DAG.getBitcast(MVT::v16i8, Op.getOperand(1));
      if (ShiftAmt > 120) {
        // For N in 121..128, fshr N == fshl (128 - N), and for 1 <= N < 8
        // SHL_DOUBLE_BIT emits fewer instructions.
        SDValue Val =
            DAG.getNode(SystemZISD::SHL_DOUBLE_BIT, DL, MVT::v16i8, Op0, Op1,
                        DAG.getTargetConstant(128 - ShiftAmt, DL, MVT::i32));
        return DAG.getBitcast(MVT::i128, Val);
      }
      SmallVector<int, 16> Mask(16);
      for (unsigned Elt = 0; Elt < 16; Elt++)
        Mask[Elt] = 16 - (ShiftAmt >> 3) + Elt;
      SDValue Shuf1 = DAG.getVectorShuffle(MVT::v16i8, DL, Op0, Op1, Mask);
      if ((ShiftAmt & 7) == 0)
        return DAG.getBitcast(MVT::i128, Shuf1);
      SDValue Shuf2 = DAG.getVectorShuffle(MVT::v16i8, DL, Op0, Op0, Mask);
      SDValue Val =
          DAG.getNode(SystemZISD::SHR_DOUBLE_BIT, DL, MVT::v16i8, Shuf2, Shuf1,
                      DAG.getTargetConstant(ShiftAmt & 7, DL, MVT::i32));
      return DAG.getBitcast(MVT::i128, Val);
    }
  }
 
  return SDValue();
}
 
static SDValue lowerAddrSpaceCast(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  SDValue Src = Op.getOperand(0);
  MVT DstVT = Op.getSimpleValueType();
 
  AddrSpaceCastSDNode *N = cast<AddrSpaceCastSDNode>(Op.getNode());
  unsigned SrcAS = N->getSrcAddressSpace();
 
  assert(SrcAS != N->getDestAddressSpace() &&
         "addrspacecast must be between different address spaces");
 
  // addrspacecast [0 <- 1] : Assinging a ptr32 value to a 64-bit pointer.
  // addrspacecast [1 <- 0] : Assigining a 64-bit pointer to a ptr32 value.
  if (SrcAS == SYSTEMZAS::PTR32 && DstVT == MVT::i64) {
    Op = DAG.getNode(ISD::AND, DL, MVT::i32, Src,
                     DAG.getConstant(0x7fffffff, DL, MVT::i32));
    Op = DAG.getNode(ISD::ZERO_EXTEND, DL, DstVT, Op);
  } else if (DstVT == MVT::i32) {
    Op = DAG.getNode(ISD::TRUNCATE, DL, DstVT, Src);
    Op = DAG.getNode(ISD::AND, DL, MVT::i32, Op,
                     DAG.getConstant(0x7fffffff, DL, MVT::i32));
    Op = DAG.getNode(ISD::ZERO_EXTEND, DL, DstVT, Op);
  } else {
    report_fatal_error("Bad address space in addrspacecast");
  }
  return Op;
}
 
SDValue SystemZTargetLowering::lowerFP_EXTEND(SDValue Op,
                                              SelectionDAG &DAG) const {
  SDValue In = Op.getOperand(Op->isStrictFPOpcode() ? 1 : 0);
  if (In.getSimpleValueType() != MVT::f16)
    return Op;      // Legal
  return SDValue(); // Let legalizer emit the libcall.
}
 
SDValue SystemZTargetLowering::useLibCall(SelectionDAG &DAG, RTLIB::Libcall LC,
                                          MVT VT, SDValue Arg, SDLoc DL,
                                          SDValue Chain, bool IsStrict) const {
  assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected request for libcall!");
  MakeLibCallOptions CallOptions;
  SDValue Result;
  std::tie(Result, Chain) =
      makeLibCall(DAG, LC, VT, Arg, CallOptions, DL, Chain);
  return IsStrict ? DAG.getMergeValues({Result, Chain}, DL) : Result;
}
 
SDValue SystemZTargetLowering::lower_FP_TO_INT(SDValue Op,
                                               SelectionDAG &DAG) const {
  bool IsSigned = (Op->getOpcode() == ISD::FP_TO_SINT ||
                   Op->getOpcode() == ISD::STRICT_FP_TO_SINT);
  bool IsStrict = Op->isStrictFPOpcode();
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  SDValue InOp = Op.getOperand(IsStrict ? 1 : 0);
  SDValue Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode();
  EVT InVT = InOp.getValueType();
 
  // FP to unsigned is not directly supported on z10.  Promoting an i32
  // result to (signed) i64 doesn't generate an inexact condition (fp
  // exception) for values that are outside the i32 range but in the i64
  // range, so use the default expansion.
  if (!Subtarget.hasFPExtension() && !IsSigned)
    // Expand i32/i64. F16 values will be recognized to fit and extended.
    return SDValue();
 
  // Conversion from f16 is done via f32.
  if (InOp.getSimpleValueType() == MVT::f16) {
    SmallVector<SDValue, 2> Results;
    LowerOperationWrapper(Op.getNode(), Results, DAG);
    return DAG.getMergeValues(Results, DL);
  }
 
  if (VT == MVT::i128) {
    RTLIB::Libcall LC =
        IsSigned ? RTLIB::getFPTOSINT(InVT, VT) : RTLIB::getFPTOUINT(InVT, VT);
    return useLibCall(DAG, LC, VT, InOp, DL, Chain, IsStrict);
  }
 
  return Op; // Legal
}
 
SDValue SystemZTargetLowering::lower_INT_TO_FP(SDValue Op,
                                               SelectionDAG &DAG) const {
  bool IsSigned = (Op->getOpcode() == ISD::SINT_TO_FP ||
                   Op->getOpcode() == ISD::STRICT_SINT_TO_FP);
  bool IsStrict = Op->isStrictFPOpcode();
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  SDValue InOp = Op.getOperand(IsStrict ? 1 : 0);
  SDValue Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode();
  EVT InVT = InOp.getValueType();
 
  // Conversion to f16 is done via f32.
  if (VT == MVT::f16) {
    SmallVector<SDValue, 2> Results;
    LowerOperationWrapper(Op.getNode(), Results, DAG);
    return DAG.getMergeValues(Results, DL);
  }
 
  // Unsigned to fp is not directly supported on z10.
  if (!Subtarget.hasFPExtension() && !IsSigned)
    return SDValue(); // Expand i64.
 
  if (InVT == MVT::i128) {
    RTLIB::Libcall LC =
        IsSigned ? RTLIB::getSINTTOFP(InVT, VT) : RTLIB::getUINTTOFP(InVT, VT);
    return useLibCall(DAG, LC, VT, InOp, DL, Chain, IsStrict);
  }
 
  return Op; // Legal
}
 
// Lower an f16 LOAD in case of no vector support.
SDValue SystemZTargetLowering::lowerLoadF16(SDValue Op,
                                            SelectionDAG &DAG) const {
  EVT RegVT = Op.getValueType();
  assert(RegVT == MVT::f16 && "Expected to lower an f16 load.");
  (void)RegVT;
 
  // Load as integer.
  SDLoc DL(Op);
  SDValue NewLd;
  if (auto *AtomicLd = dyn_cast<AtomicSDNode>(Op.getNode())) {
    assert(EVT(RegVT) == AtomicLd->getMemoryVT() && "Unhandled f16 load");
    NewLd = DAG.getAtomicLoad(ISD::EXTLOAD, DL, MVT::i16, MVT::i64,
                              AtomicLd->getChain(), AtomicLd->getBasePtr(),
                              AtomicLd->getMemOperand());
  } else {
    LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
    assert(EVT(RegVT) == Ld->getMemoryVT() && "Unhandled f16 load");
    NewLd = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i64, Ld->getChain(),
                           Ld->getBasePtr(), Ld->getPointerInfo(), MVT::i16,
                           Ld->getBaseAlign(), Ld->getMemOperand()->getFlags());
  }
  SDValue F16Val = convertToF16(NewLd, DAG);
  return DAG.getMergeValues({F16Val, NewLd.getValue(1)}, DL);
}
 
// Lower an f16 STORE in case of no vector support.
SDValue SystemZTargetLowering::lowerStoreF16(SDValue Op,
                                             SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Shft = convertFromF16(Op->getOperand(1), DL, DAG);
 
  if (auto *AtomicSt = dyn_cast<AtomicSDNode>(Op.getNode()))
    return DAG.getAtomic(ISD::ATOMIC_STORE, DL, MVT::i16, AtomicSt->getChain(),
                         Shft, AtomicSt->getBasePtr(),
                         AtomicSt->getMemOperand());
 
  StoreSDNode *St = cast<StoreSDNode>(Op.getNode());
  return DAG.getTruncStore(St->getChain(), DL, Shft, St->getBasePtr(), MVT::i16,
                           St->getMemOperand());
}
 
SDValue SystemZTargetLowering::lowerIS_FPCLASS(SDValue Op,
                                               SelectionDAG &DAG) const {
  SDLoc DL(Op);
  MVT ResultVT = Op.getSimpleValueType();
  SDValue Arg = Op.getOperand(0);
  unsigned Check = Op.getConstantOperandVal(1);
 
  unsigned TDCMask = 0;
  if (Check & fcSNan)
    TDCMask |= SystemZ::TDCMASK_SNAN_PLUS | SystemZ::TDCMASK_SNAN_MINUS;
  if (Check & fcQNan)
    TDCMask |= SystemZ::TDCMASK_QNAN_PLUS | SystemZ::TDCMASK_QNAN_MINUS;
  if (Check & fcPosInf)
    TDCMask |= SystemZ::TDCMASK_INFINITY_PLUS;
  if (Check & fcNegInf)
    TDCMask |= SystemZ::TDCMASK_INFINITY_MINUS;
  if (Check & fcPosNormal)
    TDCMask |= SystemZ::TDCMASK_NORMAL_PLUS;
  if (Check & fcNegNormal)
    TDCMask |= SystemZ::TDCMASK_NORMAL_MINUS;
  if (Check & fcPosSubnormal)
    TDCMask |= SystemZ::TDCMASK_SUBNORMAL_PLUS;
  if (Check & fcNegSubnormal)
    TDCMask |= SystemZ::TDCMASK_SUBNORMAL_MINUS;
  if (Check & fcPosZero)
    TDCMask |= SystemZ::TDCMASK_ZERO_PLUS;
  if (Check & fcNegZero)
    TDCMask |= SystemZ::TDCMASK_ZERO_MINUS;
  SDValue TDCMaskV = DAG.getConstant(TDCMask, DL, MVT::i64);
 
  SDValue Intr = DAG.getNode(SystemZISD::TDC, DL, ResultVT, Arg, TDCMaskV);
  return getCCResult(DAG, Intr);
}
 
SDValue SystemZTargetLowering::lowerREADCYCLECOUNTER(SDValue Op,
                                                     SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Chain = Op.getOperand(0);
 
  // STCKF only supports a memory operand, so we have to use a temporary.
  SDValue StackPtr = DAG.CreateStackTemporary(MVT::i64);
  int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
  MachinePointerInfo MPI =
    MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI);
 
  // Use STCFK to store the TOD clock into the temporary.
  SDValue StoreOps[] = {Chain, StackPtr};
  Chain = DAG.getMemIntrinsicNode(
    SystemZISD::STCKF, DL, DAG.getVTList(MVT::Other), StoreOps, MVT::i64,
    MPI, MaybeAlign(), MachineMemOperand::MOStore);
 
  // And read it back from there.
  return DAG.getLoad(MVT::i64, DL, Chain, StackPtr, MPI);
}
 
SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
                                              SelectionDAG &DAG) const {
  switch (Op.getOpcode()) {
  case ISD::FRAMEADDR:
    return lowerFRAMEADDR(Op, DAG);
  case ISD::RETURNADDR:
    return lowerRETURNADDR(Op, DAG);
  case ISD::BR_CC:
    return lowerBR_CC(Op, DAG);
  case ISD::SELECT_CC:
    return lowerSELECT_CC(Op, DAG);
  case ISD::SETCC:
    return lowerSETCC(Op, DAG);
  case ISD::STRICT_FSETCC:
    return lowerSTRICT_FSETCC(Op, DAG, false);
  case ISD::STRICT_FSETCCS:
    return lowerSTRICT_FSETCC(Op, DAG, true);
  case ISD::GlobalAddress:
    return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
  case ISD::GlobalTLSAddress:
    return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG);
  case ISD::BlockAddress:
    return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG);
  case ISD::JumpTable:
    return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG);
  case ISD::ConstantPool:
    return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG);
  case ISD::BITCAST:
    return lowerBITCAST(Op, DAG);
  case ISD::VASTART:
    return lowerVASTART(Op, DAG);
  case ISD::VACOPY:
    return lowerVACOPY(Op, DAG);
  case ISD::DYNAMIC_STACKALLOC:
    return lowerDYNAMIC_STACKALLOC(Op, DAG);
  case ISD::GET_DYNAMIC_AREA_OFFSET:
    return lowerGET_DYNAMIC_AREA_OFFSET(Op, DAG);
  case ISD::MULHS:
    return lowerMULH(Op, DAG, SystemZISD::SMUL_LOHI);
  case ISD::MULHU:
    return lowerMULH(Op, DAG, SystemZISD::UMUL_LOHI);
  case ISD::SMUL_LOHI:
    return lowerSMUL_LOHI(Op, DAG);
  case ISD::UMUL_LOHI:
    return lowerUMUL_LOHI(Op, DAG);
  case ISD::SDIVREM:
    return lowerSDIVREM(Op, DAG);
  case ISD::UDIVREM:
    return lowerUDIVREM(Op, DAG);
  case ISD::SADDO:
  case ISD::SSUBO:
  case ISD::UADDO:
  case ISD::USUBO:
    return lowerXALUO(Op, DAG);
  case ISD::UADDO_CARRY:
  case ISD::USUBO_CARRY:
    return lowerUADDSUBO_CARRY(Op, DAG);
  case ISD::OR:
    return lowerOR(Op, DAG);
  case ISD::CTPOP:
    return lowerCTPOP(Op, DAG);
  case ISD::VECREDUCE_ADD:
    return lowerVECREDUCE_ADD(Op, DAG);
  case ISD::ATOMIC_FENCE:
    return lowerATOMIC_FENCE(Op, DAG);
  case ISD::ATOMIC_SWAP:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_SWAPW);
  case ISD::ATOMIC_STORE:
    return lowerATOMIC_STORE(Op, DAG);
  case ISD::ATOMIC_LOAD:
    return lowerATOMIC_LOAD(Op, DAG);
  case ISD::ATOMIC_LOAD_ADD:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
  case ISD::ATOMIC_LOAD_SUB:
    return lowerATOMIC_LOAD_SUB(Op, DAG);
  case ISD::ATOMIC_LOAD_AND:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
  case ISD::ATOMIC_LOAD_OR:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
  case ISD::ATOMIC_LOAD_XOR:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
  case ISD::ATOMIC_LOAD_NAND:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
  case ISD::ATOMIC_LOAD_MIN:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
  case ISD::ATOMIC_LOAD_MAX:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
  case ISD::ATOMIC_LOAD_UMIN:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
  case ISD::ATOMIC_LOAD_UMAX:
    return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
  case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
    return lowerATOMIC_CMP_SWAP(Op, DAG);
  case ISD::STACKSAVE:
    return lowerSTACKSAVE(Op, DAG);
  case ISD::STACKRESTORE:
    return lowerSTACKRESTORE(Op, DAG);
  case ISD::PREFETCH:
    return lowerPREFETCH(Op, DAG);
  case ISD::INTRINSIC_W_CHAIN:
    return lowerINTRINSIC_W_CHAIN(Op, DAG);
  case ISD::INTRINSIC_WO_CHAIN:
    return lowerINTRINSIC_WO_CHAIN(Op, DAG);
  case ISD::BUILD_VECTOR:
    return lowerBUILD_VECTOR(Op, DAG);
  case ISD::VECTOR_SHUFFLE:
    return lowerVECTOR_SHUFFLE(Op, DAG);
  case ISD::SCALAR_TO_VECTOR:
    return lowerSCALAR_TO_VECTOR(Op, DAG);
  case ISD::INSERT_VECTOR_ELT:
    return lowerINSERT_VECTOR_ELT(Op, DAG);
  case ISD::EXTRACT_VECTOR_ELT:
    return lowerEXTRACT_VECTOR_ELT(Op, DAG);
  case ISD::SIGN_EXTEND_VECTOR_INREG:
    return lowerSIGN_EXTEND_VECTOR_INREG(Op, DAG);
  case ISD::ZERO_EXTEND_VECTOR_INREG:
    return lowerZERO_EXTEND_VECTOR_INREG(Op, DAG);
  case ISD::SHL:
    return lowerShift(Op, DAG, SystemZISD::VSHL_BY_SCALAR);
  case ISD::SRL:
    return lowerShift(Op, DAG, SystemZISD::VSRL_BY_SCALAR);
  case ISD::SRA:
    return lowerShift(Op, DAG, SystemZISD::VSRA_BY_SCALAR);
  case ISD::ADDRSPACECAST:
    return lowerAddrSpaceCast(Op, DAG);
  case ISD::ROTL:
    return lowerShift(Op, DAG, SystemZISD::VROTL_BY_SCALAR);
  case ISD::FSHL:
    return lowerFSHL(Op, DAG);
  case ISD::FSHR:
    return lowerFSHR(Op, DAG);
  case ISD::FP_EXTEND:
  case ISD::STRICT_FP_EXTEND:
    return lowerFP_EXTEND(Op, DAG);
  case ISD::FP_TO_UINT:
  case ISD::FP_TO_SINT:
  case ISD::STRICT_FP_TO_UINT:
  case ISD::STRICT_FP_TO_SINT:
    return lower_FP_TO_INT(Op, DAG);
  case ISD::UINT_TO_FP:
  case ISD::SINT_TO_FP:
  case ISD::STRICT_UINT_TO_FP:
  case ISD::STRICT_SINT_TO_FP:
    return lower_INT_TO_FP(Op, DAG);
  case ISD::LOAD:
    return lowerLoadF16(Op, DAG);
  case ISD::STORE:
    return lowerStoreF16(Op, DAG);
  case ISD::IS_FPCLASS:
    return lowerIS_FPCLASS(Op, DAG);
  case ISD::GET_ROUNDING:
    return lowerGET_ROUNDING(Op, DAG);
  case ISD::READCYCLECOUNTER:
    return lowerREADCYCLECOUNTER(Op, DAG);
  case ISD::EH_SJLJ_SETJMP:
  case ISD::EH_SJLJ_LONGJMP:
    // These operations are legal on our platform, but we cannot actually
    // set the operation action to Legal as common code would treat this
    // as equivalent to Expand. Instead, we keep the operation action to
    // Custom and just leave them unchanged here.
    return Op;
 
  default:
    llvm_unreachable("Unexpected node to lower");
  }
}
 
static SDValue expandBitCastI128ToF128(SelectionDAG &DAG, SDValue Src,
                                       const SDLoc &SL) {
  // If i128 is legal, just use a normal bitcast.
  if (DAG.getTargetLoweringInfo().isTypeLegal(MVT::i128))
    return DAG.getBitcast(MVT::f128, Src);
 
  // Otherwise, f128 must live in FP128, so do a partwise move.
  assert(DAG.getTargetLoweringInfo().getRepRegClassFor(MVT::f128) ==
         &SystemZ::FP128BitRegClass);
 
  SDValue Hi, Lo;
  std::tie(Lo, Hi) = DAG.SplitScalar(Src, SL, MVT::i64, MVT::i64);
 
  Hi = DAG.getBitcast(MVT::f64, Hi);
  Lo = DAG.getBitcast(MVT::f64, Lo);
 
  SDNode *Pair = DAG.getMachineNode(
      SystemZ::REG_SEQUENCE, SL, MVT::f128,
      {DAG.getTargetConstant(SystemZ::FP128BitRegClassID, SL, MVT::i32), Lo,
       DAG.getTargetConstant(SystemZ::subreg_l64, SL, MVT::i32), Hi,
       DAG.getTargetConstant(SystemZ::subreg_h64, SL, MVT::i32)});
  return SDValue(Pair, 0);
}
 
static SDValue expandBitCastF128ToI128(SelectionDAG &DAG, SDValue Src,
                                       const SDLoc &SL) {
  // If i128 is legal, just use a normal bitcast.
  if (DAG.getTargetLoweringInfo().isTypeLegal(MVT::i128))
    return DAG.getBitcast(MVT::i128, Src);
 
  // Otherwise, f128 must live in FP128, so do a partwise move.
  assert(DAG.getTargetLoweringInfo().getRepRegClassFor(MVT::f128) ==
         &SystemZ::FP128BitRegClass);
 
  SDValue LoFP =
      DAG.getTargetExtractSubreg(SystemZ::subreg_l64, SL, MVT::f64, Src);
  SDValue HiFP =
      DAG.getTargetExtractSubreg(SystemZ::subreg_h64, SL, MVT::f64, Src);
  SDValue Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i64, LoFP);
  SDValue Hi = DAG.getNode(ISD::BITCAST, SL, MVT::i64, HiFP);
 
  return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i128, Lo, Hi);
}
 
// Lower operations with invalid operand or result types.
void
SystemZTargetLowering::LowerOperationWrapper(SDNode *N,
                                             SmallVectorImpl<SDValue> &Results,
                                             SelectionDAG &DAG) const {
  switch (N->getOpcode()) {
  case ISD::ATOMIC_LOAD: {
    SDLoc DL(N);
    SDVTList Tys = DAG.getVTList(MVT::Untyped, MVT::Other);
    SDValue Ops[] = { N->getOperand(0), N->getOperand(1) };
    MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
    SDValue Res = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_LOAD_128,
                                          DL, Tys, Ops, MVT::i128, MMO);
 
    SDValue Lowered = lowerGR128ToI128(DAG, Res);
    if (N->getValueType(0) == MVT::f128)
      Lowered = expandBitCastI128ToF128(DAG, Lowered, DL);
    Results.push_back(Lowered);
    Results.push_back(Res.getValue(1));
    break;
  }
  case ISD::ATOMIC_STORE: {
    SDLoc DL(N);
    SDVTList Tys = DAG.getVTList(MVT::Other);
    SDValue Val = N->getOperand(1);
    if (Val.getValueType() == MVT::f128)
      Val = expandBitCastF128ToI128(DAG, Val, DL);
    Val = lowerI128ToGR128(DAG, Val);
 
    SDValue Ops[] = {N->getOperand(0), Val, N->getOperand(2)};
    MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
    SDValue Res = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_STORE_128,
                                          DL, Tys, Ops, MVT::i128, MMO);
    // We have to enforce sequential consistency by performing a
    // serialization operation after the store.
    if (cast<AtomicSDNode>(N)->getSuccessOrdering() ==
        AtomicOrdering::SequentiallyConsistent)
      Res = SDValue(DAG.getMachineNode(SystemZ::Serialize, DL,
                                       MVT::Other, Res), 0);
    Results.push_back(Res);
    break;
  }
  case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
    SDLoc DL(N);
    SDVTList Tys = DAG.getVTList(MVT::Untyped, MVT::i32, MVT::Other);
    SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
                      lowerI128ToGR128(DAG, N->getOperand(2)),
                      lowerI128ToGR128(DAG, N->getOperand(3)) };
    MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
    SDValue Res = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAP_128,
                                          DL, Tys, Ops, MVT::i128, MMO);
    SDValue Success = emitSETCC(DAG, DL, Res.getValue(1),
                                SystemZ::CCMASK_CS, SystemZ::CCMASK_CS_EQ);
    Success = DAG.getZExtOrTrunc(Success, DL, N->getValueType(1));
    Results.push_back(lowerGR128ToI128(DAG, Res));
    Results.push_back(Success);
    Results.push_back(Res.getValue(2));
    break;
  }
  case ISD::BITCAST: {
    if (useSoftFloat())
      return;
    SDLoc DL(N);
    SDValue Src = N->getOperand(0);
    EVT SrcVT = Src.getValueType();
    EVT ResVT = N->getValueType(0);
    if (ResVT == MVT::i128 && SrcVT == MVT::f128)
      Results.push_back(expandBitCastF128ToI128(DAG, Src, DL));
    else if (SrcVT == MVT::i16 && ResVT == MVT::f16) {
      if (Subtarget.hasVector()) {
        SDValue In32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Src);
        Results.push_back(SDValue(
            DAG.getMachineNode(SystemZ::LEFR_16, DL, MVT::f16, In32), 0));
      } else {
        SDValue In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Src);
        Results.push_back(convertToF16(In64, DAG));
      }
    } else if (SrcVT == MVT::f16 && ResVT == MVT::i16) {
      SDValue ExtractedVal =
          Subtarget.hasVector()
              ? SDValue(DAG.getMachineNode(SystemZ::LFER_16, DL, MVT::i32, Src),
                        0)
              : convertFromF16(Src, DL, DAG);
      Results.push_back(DAG.getZExtOrTrunc(ExtractedVal, DL, ResVT));
    }
    break;
  }
  case ISD::UINT_TO_FP:
  case ISD::SINT_TO_FP:
  case ISD::STRICT_UINT_TO_FP:
  case ISD::STRICT_SINT_TO_FP: {
    if (useSoftFloat())
      return;
    bool IsStrict = N->isStrictFPOpcode();
    SDLoc DL(N);
    SDValue InOp = N->getOperand(IsStrict ? 1 : 0);
    EVT ResVT = N->getValueType(0);
    SDValue Chain = IsStrict ? N->getOperand(0) : DAG.getEntryNode();
    if (ResVT == MVT::f16) {
      if (!IsStrict) {
        SDValue OpF32 = DAG.getNode(N->getOpcode(), DL, MVT::f32, InOp);
        Results.push_back(DAG.getFPExtendOrRound(OpF32, DL, MVT::f16));
      } else {
        SDValue OpF32 =
            DAG.getNode(N->getOpcode(), DL, DAG.getVTList(MVT::f32, MVT::Other),
                        {Chain, InOp});
        SDValue F16Res;
        std::tie(F16Res, Chain) = DAG.getStrictFPExtendOrRound(
            OpF32, OpF32.getValue(1), DL, MVT::f16);
        Results.push_back(F16Res);
        Results.push_back(Chain);
      }
    }
    break;
  }
  case ISD::FP_TO_UINT:
  case ISD::FP_TO_SINT:
  case ISD::STRICT_FP_TO_UINT:
  case ISD::STRICT_FP_TO_SINT: {
    if (useSoftFloat())
      return;
    bool IsStrict = N->isStrictFPOpcode();
    SDLoc DL(N);
    EVT ResVT = N->getValueType(0);
    SDValue InOp = N->getOperand(IsStrict ? 1 : 0);
    EVT InVT = InOp->getValueType(0);
    SDValue Chain = IsStrict ? N->getOperand(0) : DAG.getEntryNode();
    if (InVT == MVT::f16) {
      if (!IsStrict) {
        SDValue InF32 = DAG.getFPExtendOrRound(InOp, DL, MVT::f32);
        Results.push_back(DAG.getNode(N->getOpcode(), DL, ResVT, InF32));
      } else {
        SDValue InF32;
        std::tie(InF32, Chain) =
            DAG.getStrictFPExtendOrRound(InOp, Chain, DL, MVT::f32);
        SDValue OpF32 =
            DAG.getNode(N->getOpcode(), DL, DAG.getVTList(ResVT, MVT::Other),
                        {Chain, InF32});
        Results.push_back(OpF32);
        Results.push_back(OpF32.getValue(1));
      }
    }
    break;
  }
  default:
    llvm_unreachable("Unexpected node to lower");
  }
}
 
void
SystemZTargetLowering::ReplaceNodeResults(SDNode *N,
                                          SmallVectorImpl<SDValue> &Results,
                                          SelectionDAG &DAG) const {
  return LowerOperationWrapper(N, Results, DAG);
}
 
// Return true if VT is a vector whose elements are a whole number of bytes
// in width. Also check for presence of vector support.
bool SystemZTargetLowering::canTreatAsByteVector(EVT VT) const {
  if (!Subtarget.hasVector())
    return false;
 
  return VT.isVector() && VT.getScalarSizeInBits() % 8 == 0 && VT.isSimple();
}
 
// Try to simplify an EXTRACT_VECTOR_ELT from a vector of type VecVT
// producing a result of type ResVT.  Op is a possibly bitcast version
// of the input vector and Index is the index (based on type VecVT) that
// should be extracted.  Return the new extraction if a simplification
// was possible or if Force is true.
SDValue SystemZTargetLowering::combineExtract(const SDLoc &DL, EVT ResVT,
                                              EVT VecVT, SDValue Op,
                                              unsigned Index,
                                              DAGCombinerInfo &DCI,
                                              bool Force) const {
  SelectionDAG &DAG = DCI.DAG;
 
  // The number of bytes being extracted.
  unsigned BytesPerElement = VecVT.getVectorElementType().getStoreSize();
 
  for (;;) {
    unsigned Opcode = Op.getOpcode();
    if (Opcode == ISD::BITCAST)
      // Look through bitcasts.
      Op = Op.getOperand(0);
    else if ((Opcode == ISD::VECTOR_SHUFFLE || Opcode == SystemZISD::SPLAT) &&
             canTreatAsByteVector(Op.getValueType())) {
      // Get a VPERM-like permute mask and see whether the bytes covered
      // by the extracted element are a contiguous sequence from one
      // source operand.
      SmallVector<int, SystemZ::VectorBytes> Bytes;
      if (!getVPermMask(Op, Bytes))
        break;
      int First;
      if (!getShuffleInput(Bytes, Index * BytesPerElement,
                           BytesPerElement, First))
        break;
      if (First < 0)
        return DAG.getUNDEF(ResVT);
      // Make sure the contiguous sequence starts at a multiple of the
      // original element size.
      unsigned Byte = unsigned(First) % Bytes.size();
      if (Byte % BytesPerElement != 0)
        break;
      // We can get the extracted value directly from an input.
      Index = Byte / BytesPerElement;
      Op = Op.getOperand(unsigned(First) / Bytes.size());
      Force = true;
    } else if (Opcode == ISD::BUILD_VECTOR &&
               canTreatAsByteVector(Op.getValueType())) {
      // We can only optimize this case if the BUILD_VECTOR elements are
      // at least as wide as the extracted value.
      EVT OpVT = Op.getValueType();
      unsigned OpBytesPerElement = OpVT.getVectorElementType().getStoreSize();
      if (OpBytesPerElement < BytesPerElement)
        break;
      // Make sure that the least-significant bit of the extracted value
      // is the least significant bit of an input.
      unsigned End = (Index + 1) * BytesPerElement;
      if (End % OpBytesPerElement != 0)
        break;
      // We're extracting the low part of one operand of the BUILD_VECTOR.
      Op = Op.getOperand(End / OpBytesPerElement - 1);
      EVT ResIntVT = MVT::getIntegerVT(ResVT.getSizeInBits());
      if (!isTypeLegal(ResIntVT))
        break;
      if (!Op.getValueType().isInteger()) {
        EVT OpIntVT = MVT::getIntegerVT(Op.getValueSizeInBits());
        if (!isTypeLegal(OpIntVT))
          break;
        Op = DAG.getNode(ISD::BITCAST, DL, OpIntVT, Op);
        DCI.AddToWorklist(Op.getNode());
      }
      Op = DAG.getNode(ISD::TRUNCATE, DL, ResIntVT, Op);
      if (ResIntVT != ResVT) {
        DCI.AddToWorklist(Op.getNode());
        Op = DAG.getNode(ISD::BITCAST, DL, ResVT, Op);
      }
      return Op;
    } else if ((Opcode == ISD::SIGN_EXTEND_VECTOR_INREG ||
                Opcode == ISD::ZERO_EXTEND_VECTOR_INREG ||
                Opcode == ISD::ANY_EXTEND_VECTOR_INREG) &&
               canTreatAsByteVector(Op.getValueType()) &&
               canTreatAsByteVector(Op.getOperand(0).getValueType())) {
      // Make sure that only the unextended bits are significant.
      EVT ExtVT = Op.getValueType();
      EVT OpVT = Op.getOperand(0).getValueType();
      unsigned ExtBytesPerElement = ExtVT.getVectorElementType().getStoreSize();
      unsigned OpBytesPerElement = OpVT.getVectorElementType().getStoreSize();
      unsigned Byte = Index * BytesPerElement;
      unsigned SubByte = Byte % ExtBytesPerElement;
      unsigned MinSubByte = ExtBytesPerElement - OpBytesPerElement;
      if (SubByte < MinSubByte ||
          SubByte + BytesPerElement > ExtBytesPerElement)
        break;
      // Get the byte offset of the unextended element
      Byte = Byte / ExtBytesPerElement * OpBytesPerElement;
      // ...then add the byte offset relative to that element.
      Byte += SubByte - MinSubByte;
      if (Byte % BytesPerElement != 0)
        break;
      Op = Op.getOperand(0);
      Index = Byte / BytesPerElement;
      Force = true;
    } else
      break;
  }
  if (Force) {
    if (Op.getValueType() != VecVT) {
      Op = DAG.getNode(ISD::BITCAST, DL, VecVT, Op);
      DCI.AddToWorklist(Op.getNode());
    }
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, Op,
                       DAG.getConstant(Index, DL, MVT::i32));
  }
  return SDValue();
}
 
// Optimize vector operations in scalar value Op on the basis that Op
// is truncated to TruncVT.
SDValue SystemZTargetLowering::combineTruncateExtract(
    const SDLoc &DL, EVT TruncVT, SDValue Op, DAGCombinerInfo &DCI) const {
  // If we have (trunc (extract_vector_elt X, Y)), try to turn it into
  // (extract_vector_elt (bitcast X), Y'), where (bitcast X) has elements
  // of type TruncVT.
  if (Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
      TruncVT.getSizeInBits() % 8 == 0) {
    SDValue Vec = Op.getOperand(0);
    EVT VecVT = Vec.getValueType();
    if (canTreatAsByteVector(VecVT)) {
      if (auto *IndexN = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
        unsigned BytesPerElement = VecVT.getVectorElementType().getStoreSize();
        unsigned TruncBytes = TruncVT.getStoreSize();
        if (BytesPerElement % TruncBytes == 0) {
          // Calculate the value of Y' in the above description.  We are
          // splitting the original elements into Scale equal-sized pieces
          // and for truncation purposes want the last (least-significant)
          // of these pieces for IndexN.  This is easiest to do by calculating
          // the start index of the following element and then subtracting 1.
          unsigned Scale = BytesPerElement / TruncBytes;
          unsigned NewIndex = (IndexN->getZExtValue() + 1) * Scale - 1;
 
          // Defer the creation of the bitcast from X to combineExtract,
          // which might be able to optimize the extraction.
          VecVT = EVT::getVectorVT(*DCI.DAG.getContext(),
                                   MVT::getIntegerVT(TruncBytes * 8),
                                   VecVT.getStoreSize() / TruncBytes);
          EVT ResVT = (TruncBytes < 4 ? MVT::i32 : TruncVT);
          return combineExtract(DL, ResVT, VecVT, Vec, NewIndex, DCI, true);
        }
      }
    }
  }
  return SDValue();
}
 
SDValue SystemZTargetLowering::combineZERO_EXTEND(
    SDNode *N, DAGCombinerInfo &DCI) const {
  // Convert (zext (select_ccmask C1, C2)) into (select_ccmask C1', C2')
  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0);
  EVT VT = N->getValueType(0);
  if (N0.getOpcode() == SystemZISD::SELECT_CCMASK) {
    auto *TrueOp = dyn_cast<ConstantSDNode>(N0.getOperand(0));
    auto *FalseOp = dyn_cast<ConstantSDNode>(N0.getOperand(1));
    if (TrueOp && FalseOp) {
      SDLoc DL(N0);
      SDValue Ops[] = { DAG.getConstant(TrueOp->getZExtValue(), DL, VT),
                        DAG.getConstant(FalseOp->getZExtValue(), DL, VT),
                        N0.getOperand(2), N0.getOperand(3), N0.getOperand(4) };
      SDValue NewSelect = DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VT, Ops);
      // If N0 has multiple uses, change other uses as well.
      if (!N0.hasOneUse()) {
        SDValue TruncSelect =
          DAG.getNode(ISD::TRUNCATE, DL, N0.getValueType(), NewSelect);
        DCI.CombineTo(N0.getNode(), TruncSelect);
      }
      return NewSelect;
    }
  }
  // Convert (zext (xor (trunc X), C)) into (xor (trunc X), C') if the size
  // of the result is smaller than the size of X and all the truncated bits
  // of X are already zero.
  if (N0.getOpcode() == ISD::XOR &&
      N0.hasOneUse() && N0.getOperand(0).hasOneUse() &&
      N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
      N0.getOperand(1).getOpcode() == ISD::Constant) {
    SDValue X = N0.getOperand(0).getOperand(0);
    if (VT.isScalarInteger() && VT.getSizeInBits() < X.getValueSizeInBits()) {
      KnownBits Known = DAG.computeKnownBits(X);
      APInt TruncatedBits = APInt::getBitsSet(X.getValueSizeInBits(),
                                              N0.getValueSizeInBits(),
                                              VT.getSizeInBits());
      if (TruncatedBits.isSubsetOf(Known.Zero)) {
        X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
        APInt Mask = N0.getConstantOperandAPInt(1).zext(VT.getSizeInBits());
        return DAG.getNode(ISD::XOR, SDLoc(N0), VT,
                           X, DAG.getConstant(Mask, SDLoc(N0), VT));
      }
    }
  }
  // Recognize patterns for VECTOR SUBTRACT COMPUTE BORROW INDICATION
  // and VECTOR ADD COMPUTE CARRY for i128:
  //   (zext (setcc_uge X Y)) --> (VSCBI X Y)
  //   (zext (setcc_ule Y X)) --> (VSCBI X Y)
  //   (zext (setcc_ult (add X Y) X/Y) -> (VACC X Y)
  //   (zext (setcc_ugt X/Y (add X Y)) -> (VACC X Y)
  // For vector types, these patterns are recognized in the .td file.
  if (N0.getOpcode() == ISD::SETCC && isTypeLegal(VT) && VT == MVT::i128 &&
      N0.getOperand(0).getValueType() == VT) {
    SDValue Op0 = N0.getOperand(0);
    SDValue Op1 = N0.getOperand(1);
    const ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
    switch (CC) {
    case ISD::SETULE:
      std::swap(Op0, Op1);
      [[fallthrough]];
    case ISD::SETUGE:
      return DAG.getNode(SystemZISD::VSCBI, SDLoc(N0), VT, Op0, Op1);
    case ISD::SETUGT:
      std::swap(Op0, Op1);
      [[fallthrough]];
    case ISD::SETULT:
      if (Op0->hasOneUse() && Op0->getOpcode() == ISD::ADD &&
          (Op0->getOperand(0) == Op1 || Op0->getOperand(1) == Op1))
        return DAG.getNode(SystemZISD::VACC, SDLoc(N0), VT, Op0->getOperand(0),
                           Op0->getOperand(1));
      break;
    default:
      break;
    }
  }
 
  return SDValue();
}
 
SDValue SystemZTargetLowering::combineSIGN_EXTEND_INREG(
    SDNode *N, DAGCombinerInfo &DCI) const {
  // Convert (sext_in_reg (setcc LHS, RHS, COND), i1)
  // and (sext_in_reg (any_extend (setcc LHS, RHS, COND)), i1)
  // into (select_cc LHS, RHS, -1, 0, COND)
  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0);
  EVT VT = N->getValueType(0);
  EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
  if (N0.hasOneUse() && N0.getOpcode() == ISD::ANY_EXTEND)
    N0 = N0.getOperand(0);
  if (EVT == MVT::i1 && N0.hasOneUse() && N0.getOpcode() == ISD::SETCC) {
    SDLoc DL(N0);
    SDValue Ops[] = { N0.getOperand(0), N0.getOperand(1),
                      DAG.getAllOnesConstant(DL, VT),
                      DAG.getConstant(0, DL, VT), N0.getOperand(2) };
    return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
  }
  return SDValue();
}
 
SDValue SystemZTargetLowering::combineSIGN_EXTEND(
    SDNode *N, DAGCombinerInfo &DCI) const {
  // Convert (sext (ashr (shl X, C1), C2)) to
  // (ashr (shl (anyext X), C1'), C2')), since wider shifts are as
  // cheap as narrower ones.
  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0);
  EVT VT = N->getValueType(0);
  if (N0.hasOneUse() && N0.getOpcode() == ISD::SRA) {
    auto *SraAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1));
    SDValue Inner = N0.getOperand(0);
    if (SraAmt && Inner.hasOneUse() && Inner.getOpcode() == ISD::SHL) {
      if (auto *ShlAmt = dyn_cast<ConstantSDNode>(Inner.getOperand(1))) {
        unsigned Extra = (VT.getSizeInBits() - N0.getValueSizeInBits());
        unsigned NewShlAmt = ShlAmt->getZExtValue() + Extra;
        unsigned NewSraAmt = SraAmt->getZExtValue() + Extra;
        EVT ShiftVT = N0.getOperand(1).getValueType();
        SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, SDLoc(Inner), VT,
                                  Inner.getOperand(0));
        SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(Inner), VT, Ext,
                                  DAG.getConstant(NewShlAmt, SDLoc(Inner),
                                                  ShiftVT));
        return DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl,
                           DAG.getConstant(NewSraAmt, SDLoc(N0), ShiftVT));
      }
    }
  }
 
  return SDValue();
}
 
SDValue SystemZTargetLowering::combineMERGE(
    SDNode *N, DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  unsigned Opcode = N->getOpcode();
  SDValue Op0 = N->getOperand(0);
  SDValue Op1 = N->getOperand(1);
  if (Op0.getOpcode() == ISD::BITCAST)
    Op0 = Op0.getOperand(0);
  if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
    // (z_merge_* 0, 0) -> 0.  This is mostly useful for using VLLEZF
    // for v4f32.
    if (Op1 == N->getOperand(0))
      return Op1;
    // (z_merge_? 0, X) -> (z_unpackl_? 0, X).
    EVT VT = Op1.getValueType();
    unsigned ElemBytes = VT.getVectorElementType().getStoreSize();
    if (ElemBytes <= 4) {
      Opcode = (Opcode == SystemZISD::MERGE_HIGH ?
                SystemZISD::UNPACKL_HIGH : SystemZISD::UNPACKL_LOW);
      EVT InVT = VT.changeVectorElementTypeToInteger();
      EVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(ElemBytes * 16),
                                   SystemZ::VectorBytes / ElemBytes / 2);
      if (VT != InVT) {
        Op1 = DAG.getNode(ISD::BITCAST, SDLoc(N), InVT, Op1);
        DCI.AddToWorklist(Op1.getNode());
      }
      SDValue Op = DAG.getNode(Opcode, SDLoc(N), OutVT, Op1);
      DCI.AddToWorklist(Op.getNode());
      return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
    }
  }
  return SDValue();
}
 
static bool isI128MovedToParts(LoadSDNode *LD, SDNode *&LoPart,
                               SDNode *&HiPart) {
  LoPart = HiPart = nullptr;
 
  // Scan through all users.
  for (SDUse &Use : LD->uses()) {
    // Skip the uses of the chain.
    if (Use.getResNo() != 0)
      continue;
 
    // Verify every user is a TRUNCATE to i64 of the low or high half.
    SDNode *User = Use.getUser();
    bool IsLoPart = true;
    if (User->getOpcode() == ISD::SRL &&
        User->getOperand(1).getOpcode() == ISD::Constant &&
        User->getConstantOperandVal(1) == 64 && User->hasOneUse()) {
      User = *User->user_begin();
      IsLoPart = false;
    }
    if (User->getOpcode() != ISD::TRUNCATE || User->getValueType(0) != MVT::i64)
      return false;
 
    if (IsLoPart) {
      if (LoPart)
        return false;
      LoPart = User;
    } else {
      if (HiPart)
        return false;
      HiPart = User;
    }
  }
  return true;
}
 
static bool isF128MovedToParts(LoadSDNode *LD, SDNode *&LoPart,
                               SDNode *&HiPart) {
  LoPart = HiPart = nullptr;
 
  // Scan through all users.
  for (SDUse &Use : LD->uses()) {
    // Skip the uses of the chain.
    if (Use.getResNo() != 0)
      continue;
 
    // Verify every user is an EXTRACT_SUBREG of the low or high half.
    SDNode *User = Use.getUser();
    if (!User->hasOneUse() || !User->isMachineOpcode() ||
        User->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
      return false;
 
    switch (User->getConstantOperandVal(1)) {
    case SystemZ::subreg_l64:
      if (LoPart)
        return false;
      LoPart = User;
      break;
    case SystemZ::subreg_h64:
      if (HiPart)
        return false;
      HiPart = User;
      break;
    default:
      return false;
    }
  }
  return true;
}
 
SDValue SystemZTargetLowering::combineLOAD(
    SDNode *N, DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  EVT LdVT = N->getValueType(0);
  if (auto *LN = dyn_cast<LoadSDNode>(N)) {
    if (LN->getAddressSpace() == SYSTEMZAS::PTR32) {
      MVT PtrVT = getPointerTy(DAG.getDataLayout());
      MVT LoadNodeVT = LN->getBasePtr().getSimpleValueType();
      if (PtrVT != LoadNodeVT) {
        SDLoc DL(LN);
        SDValue AddrSpaceCast = DAG.getAddrSpaceCast(
            DL, PtrVT, LN->getBasePtr(), SYSTEMZAS::PTR32, 0);
        return DAG.getExtLoad(LN->getExtensionType(), DL, LN->getValueType(0),
                              LN->getChain(), AddrSpaceCast, LN->getMemoryVT(),
                              LN->getMemOperand());
      }
    }
  }
  SDLoc DL(N);
 
  // Replace a 128-bit load that is used solely to move its value into GPRs
  // by separate loads of both halves.
  LoadSDNode *LD = cast<LoadSDNode>(N);
  if (LD->isSimple() && ISD::isNormalLoad(LD)) {
    SDNode *LoPart, *HiPart;
    if ((LdVT == MVT::i128 && isI128MovedToParts(LD, LoPart, HiPart)) ||
        (LdVT == MVT::f128 && isF128MovedToParts(LD, LoPart, HiPart))) {
      // Rewrite each extraction as an independent load.
      SmallVector<SDValue, 2> ArgChains;
      if (HiPart) {
        SDValue EltLoad = DAG.getLoad(
            HiPart->getValueType(0), DL, LD->getChain(), LD->getBasePtr(),
            LD->getPointerInfo(), LD->getBaseAlign(),
            LD->getMemOperand()->getFlags(), LD->getAAInfo());
 
        DCI.CombineTo(HiPart, EltLoad, true);
        ArgChains.push_back(EltLoad.getValue(1));
      }
      if (LoPart) {
        SDValue EltLoad = DAG.getLoad(
            LoPart->getValueType(0), DL, LD->getChain(),
            DAG.getObjectPtrOffset(DL, LD->getBasePtr(), TypeSize::getFixed(8)),
            LD->getPointerInfo().getWithOffset(8), LD->getBaseAlign(),
            LD->getMemOperand()->getFlags(), LD->getAAInfo());
 
        DCI.CombineTo(LoPart, EltLoad, true);
        ArgChains.push_back(EltLoad.getValue(1));
      }
 
      // Collect all chains via TokenFactor.
      SDValue Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, ArgChains);
      DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
      DCI.AddToWorklist(Chain.getNode());
      return SDValue(N, 0);
    }
  }
 
  if (LdVT.isVector() || LdVT.isInteger())
    return SDValue();
  // Transform a scalar load that is REPLICATEd as well as having other
  // use(s) to the form where the other use(s) use the first element of the
  // REPLICATE instead of the load. Otherwise instruction selection will not
  // produce a VLREP. Avoid extracting to a GPR, so only do this for floating
  // point loads.
 
  SDValue Replicate;
  SmallVector<SDNode*, 8> OtherUses;
  for (SDUse &Use : N->uses()) {
    if (Use.getUser()->getOpcode() == SystemZISD::REPLICATE) {
      if (Replicate)
        return SDValue(); // Should never happen
      Replicate = SDValue(Use.getUser(), 0);
    } else if (Use.getResNo() == 0)
      OtherUses.push_back(Use.getUser());
  }
  if (!Replicate || OtherUses.empty())
    return SDValue();
 
  SDValue Extract0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, LdVT,
                              Replicate, DAG.getConstant(0, DL, MVT::i32));
  // Update uses of the loaded Value while preserving old chains.
  for (SDNode *U : OtherUses) {
    SmallVector<SDValue, 8> Ops;
    for (SDValue Op : U->ops())
      Ops.push_back((Op.getNode() == N && Op.getResNo() == 0) ? Extract0 : Op);
    DAG.UpdateNodeOperands(U, Ops);
  }
  return SDValue(N, 0);
}
 
bool SystemZTargetLowering::canLoadStoreByteSwapped(EVT VT) const {
  if (VT == MVT::i16 || VT == MVT::i32 || VT == MVT::i64)
    return true;
  if (Subtarget.hasVectorEnhancements2())
    if (VT == MVT::v8i16 || VT == MVT::v4i32 || VT == MVT::v2i64 || VT == MVT::i128)
      return true;
  return false;
}
 
static bool isVectorElementSwap(ArrayRef<int> M, EVT VT) {
  if (!VT.isVector() || !VT.isSimple() ||
      VT.getSizeInBits() != 128 ||
      VT.getScalarSizeInBits() % 8 != 0)
    return false;
 
  unsigned NumElts = VT.getVectorNumElements();
  for (unsigned i = 0; i < NumElts; ++i) {
    if (M[i] < 0) continue; // ignore UNDEF indices
    if ((unsigned) M[i] != NumElts - 1 - i)
      return false;
  }
 
  return true;
}
 
static bool isOnlyUsedByStores(SDValue StoredVal, SelectionDAG &DAG) {
  for (auto *U : StoredVal->users()) {
    if (StoreSDNode *ST = dyn_cast<StoreSDNode>(U)) {
      EVT CurrMemVT = ST->getMemoryVT().getScalarType();
      if (CurrMemVT.isRound() && CurrMemVT.getStoreSize() <= 16)
        continue;
    } else if (isa<BuildVectorSDNode>(U)) {
      SDValue BuildVector = SDValue(U, 0);
      if (DAG.isSplatValue(BuildVector, true/*AllowUndefs*/) &&
          isOnlyUsedByStores(BuildVector, DAG))
        continue;
    }
    return false;
  }
  return true;
}
 
static bool isI128MovedFromParts(SDValue Val, SDValue &LoPart,
                                 SDValue &HiPart) {
  if (Val.