SIISelLowering.cpp

Go to the documentation of this file.

//===-- SIISelLowering.cpp - SI 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
//
//===----------------------------------------------------------------------===//
//
/// \file
/// Custom DAG lowering for SI
//
//===----------------------------------------------------------------------===//

#if defined(_MSC_VER) || defined(__MINGW32__)
// Provide M_PI.
#define _USE_MATH_DEFINES
#endif

#include "SIISelLowering.h"
#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "AMDGPUTargetMachine.h"
#include "SIDefines.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/ADT/StringSwitch.h"
#include "llvm/ADT/Twine.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/DAGCombine.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/CodeGen/TargetCallingConv.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MachineValueType.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
#include <cassert>
#include <cmath>
#include <cstdint>
#include <iterator>
#include <tuple>
#include <utility>
#include <vector>

using namespace llvm;

#define DEBUG_TYPE "si-lower"

STATISTIC(NumTailCalls, "Number of tail calls");

static cl::opt<bool> EnableVGPRIndexMode(
  "amdgpu-vgpr-index-mode",
  cl::desc("Use GPR indexing mode instead of movrel for vector indexing"),
  cl::init(false));

static cl::opt<bool> DisableLoopAlignment(
  "amdgpu-disable-loop-alignment",
  cl::desc("Do not align and prefetch loops"),
  cl::init(false));

static unsigned findFirstFreeSGPR(CCState &CCInfo) {
  unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs();
  for (unsigned Reg = 0; Reg < NumSGPRs; ++Reg) {
    if (!CCInfo.isAllocated(AMDGPU::SGPR0 + Reg)) {
      return AMDGPU::SGPR0 + Reg;
    }
  }
  llvm_unreachable("Cannot allocate sgpr");
}

SITargetLowering::SITargetLowering(const TargetMachine &TM,
                                   const GCNSubtarget &STI)
    : AMDGPUTargetLowering(TM, STI),
      Subtarget(&STI) {
  addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass);
  addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);

  addRegisterClass(MVT::i32, &AMDGPU::SReg_32_XM0RegClass);
  addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass);

  addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
  addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
  addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);

  addRegisterClass(MVT::v3i32, &AMDGPU::SGPR_96RegClass);
  addRegisterClass(MVT::v3f32, &AMDGPU::VReg_96RegClass);

  addRegisterClass(MVT::v2i64, &AMDGPU::SReg_128RegClass);
  addRegisterClass(MVT::v2f64, &AMDGPU::SReg_128RegClass);

  addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
  addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);

  addRegisterClass(MVT::v5i32, &AMDGPU::SGPR_160RegClass);
  addRegisterClass(MVT::v5f32, &AMDGPU::VReg_160RegClass);

  addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
  addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);

  addRegisterClass(MVT::v16i32, &AMDGPU::SReg_512RegClass);
  addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);

  if (Subtarget->has16BitInsts()) {
    addRegisterClass(MVT::i16, &AMDGPU::SReg_32_XM0RegClass);
    addRegisterClass(MVT::f16, &AMDGPU::SReg_32_XM0RegClass);

    // Unless there are also VOP3P operations, not operations are really legal.
    addRegisterClass(MVT::v2i16, &AMDGPU::SReg_32_XM0RegClass);
    addRegisterClass(MVT::v2f16, &AMDGPU::SReg_32_XM0RegClass);
    addRegisterClass(MVT::v4i16, &AMDGPU::SReg_64RegClass);
    addRegisterClass(MVT::v4f16, &AMDGPU::SReg_64RegClass);
  }

  if (Subtarget->hasMAIInsts()) {
    addRegisterClass(MVT::v32i32, &AMDGPU::VReg_1024RegClass);
    addRegisterClass(MVT::v32f32, &AMDGPU::VReg_1024RegClass);
  }

  computeRegisterProperties(Subtarget->getRegisterInfo());

  // We need to custom lower vector stores from local memory
  setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v3i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v5i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
  setOperationAction(ISD::LOAD, MVT::i1, Custom);
  setOperationAction(ISD::LOAD, MVT::v32i32, Custom);

  setOperationAction(ISD::STORE, MVT::v2i32, Custom);
  setOperationAction(ISD::STORE, MVT::v3i32, Custom);
  setOperationAction(ISD::STORE, MVT::v4i32, Custom);
  setOperationAction(ISD::STORE, MVT::v5i32, Custom);
  setOperationAction(ISD::STORE, MVT::v8i32, Custom);
  setOperationAction(ISD::STORE, MVT::v16i32, Custom);
  setOperationAction(ISD::STORE, MVT::i1, Custom);
  setOperationAction(ISD::STORE, MVT::v32i32, Custom);

  setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
  setTruncStoreAction(MVT::v3i32, MVT::v3i16, Expand);
  setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
  setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
  setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
  setTruncStoreAction(MVT::v32i32, MVT::v32i16, Expand);
  setTruncStoreAction(MVT::v2i32, MVT::v2i8, Expand);
  setTruncStoreAction(MVT::v4i32, MVT::v4i8, Expand);
  setTruncStoreAction(MVT::v8i32, MVT::v8i8, Expand);
  setTruncStoreAction(MVT::v16i32, MVT::v16i8, Expand);
  setTruncStoreAction(MVT::v32i32, MVT::v32i8, Expand);

  setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
  setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);

  setOperationAction(ISD::SELECT, MVT::i1, Promote);
  setOperationAction(ISD::SELECT, MVT::i64, Custom);
  setOperationAction(ISD::SELECT, MVT::f64, Promote);
  AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);

  setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);

  setOperationAction(ISD::SETCC, MVT::i1, Promote);
  setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
  setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
  AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);

  setOperationAction(ISD::TRUNCATE, MVT::v2i32, Expand);
  setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);

  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v3i16, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);

  setOperationAction(ISD::BRCOND, MVT::Other, Custom);
  setOperationAction(ISD::BR_CC, MVT::i1, Expand);
  setOperationAction(ISD::BR_CC, MVT::i32, Expand);
  setOperationAction(ISD::BR_CC, MVT::i64, Expand);
  setOperationAction(ISD::BR_CC, MVT::f32, Expand);
  setOperationAction(ISD::BR_CC, MVT::f64, Expand);

  setOperationAction(ISD::UADDO, MVT::i32, Legal);
  setOperationAction(ISD::USUBO, MVT::i32, Legal);

  setOperationAction(ISD::ADDCARRY, MVT::i32, Legal);
  setOperationAction(ISD::SUBCARRY, MVT::i32, Legal);

  setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
  setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
  setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);

#if 0
  setOperationAction(ISD::ADDCARRY, MVT::i64, Legal);
  setOperationAction(ISD::SUBCARRY, MVT::i64, Legal);
#endif

  // We only support LOAD/STORE and vector manipulation ops for vectors
  // with > 4 elements.
  for (MVT VT : { MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32,
                  MVT::v2i64, MVT::v2f64, MVT::v4i16, MVT::v4f16,
                  MVT::v32i32, MVT::v32f32 }) {
    for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
      switch (Op) {
      case ISD::LOAD:
      case ISD::STORE:
      case ISD::BUILD_VECTOR:
      case ISD::BITCAST:
      case ISD::EXTRACT_VECTOR_ELT:
      case ISD::INSERT_VECTOR_ELT:
      case ISD::INSERT_SUBVECTOR:
      case ISD::EXTRACT_SUBVECTOR:
      case ISD::SCALAR_TO_VECTOR:
        break;
      case ISD::CONCAT_VECTORS:
        setOperationAction(Op, VT, Custom);
        break;
      default:
        setOperationAction(Op, VT, Expand);
        break;
      }
    }
  }

  setOperationAction(ISD::FP_EXTEND, MVT::v4f32, Expand);

  // TODO: For dynamic 64-bit vector inserts/extracts, should emit a pseudo that
  // is expanded to avoid having two separate loops in case the index is a VGPR.

  // Most operations are naturally 32-bit vector operations. We only support
  // load and store of i64 vectors, so promote v2i64 vector operations to v4i32.
  for (MVT Vec64 : { MVT::v2i64, MVT::v2f64 }) {
    setOperationAction(ISD::BUILD_VECTOR, Vec64, Promote);
    AddPromotedToType(ISD::BUILD_VECTOR, Vec64, MVT::v4i32);

    setOperationAction(ISD::EXTRACT_VECTOR_ELT, Vec64, Promote);
    AddPromotedToType(ISD::EXTRACT_VECTOR_ELT, Vec64, MVT::v4i32);

    setOperationAction(ISD::INSERT_VECTOR_ELT, Vec64, Promote);
    AddPromotedToType(ISD::INSERT_VECTOR_ELT, Vec64, MVT::v4i32);

    setOperationAction(ISD::SCALAR_TO_VECTOR, Vec64, Promote);
    AddPromotedToType(ISD::SCALAR_TO_VECTOR, Vec64, MVT::v4i32);
  }

  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);

  setOperationAction(ISD::BUILD_VECTOR, MVT::v4f16, Custom);
  setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);

  // Avoid stack access for these.
  // TODO: Generalize to more vector types.
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i16, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f16, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f16, Custom);

  setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i16, Custom);
  setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f16, Custom);
  setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i8, Custom);
  setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i8, Custom);
  setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i8, Custom);

  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i8, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i8, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i8, Custom);

  setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i16, Custom);
  setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f16, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f16, Custom);

  // Deal with vec3 vector operations when widened to vec4.
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v3i32, Custom);
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v3f32, Custom);
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i32, Custom);
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4f32, Custom);

  // Deal with vec5 vector operations when widened to vec8.
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v5i32, Custom);
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v5f32, Custom);
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i32, Custom);
  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8f32, Custom);

  // BUFFER/FLAT_ATOMIC_CMP_SWAP on GCN GPUs needs input marshalling,
  // and output demarshalling
  setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);

  // We can't return success/failure, only the old value,
  // let LLVM add the comparison
  setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i32, Expand);
  setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i64, Expand);

  if (Subtarget->hasFlatAddressSpace()) {
    setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom);
    setOperationAction(ISD::ADDRSPACECAST, MVT::i64, Custom);
  }

  setOperationAction(ISD::BSWAP, MVT::i32, Legal);
  setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);

  // On SI this is s_memtime and s_memrealtime on VI.
  setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
  setOperationAction(ISD::TRAP, MVT::Other, Custom);
  setOperationAction(ISD::DEBUGTRAP, MVT::Other, Custom);

  if (Subtarget->has16BitInsts()) {
    setOperationAction(ISD::FLOG, MVT::f16, Custom);
    setOperationAction(ISD::FEXP, MVT::f16, Custom);
    setOperationAction(ISD::FLOG10, MVT::f16, Custom);
  }

  // v_mad_f32 does not support denormals according to some sources.
  if (!Subtarget->hasFP32Denormals())
    setOperationAction(ISD::FMAD, MVT::f32, Legal);

  if (!Subtarget->hasBFI()) {
    // fcopysign can be done in a single instruction with BFI.
    setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
    setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
  }

  if (!Subtarget->hasBCNT(32))
    setOperationAction(ISD::CTPOP, MVT::i32, Expand);

  if (!Subtarget->hasBCNT(64))
    setOperationAction(ISD::CTPOP, MVT::i64, Expand);

  if (Subtarget->hasFFBH())
    setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Custom);

  if (Subtarget->hasFFBL())
    setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Custom);

  // We only really have 32-bit BFE instructions (and 16-bit on VI).
  //
  // On SI+ there are 64-bit BFEs, but they are scalar only and there isn't any
  // effort to match them now. We want this to be false for i64 cases when the
  // extraction isn't restricted to the upper or lower half. Ideally we would
  // have some pass reduce 64-bit extracts to 32-bit if possible. Extracts that
  // span the midpoint are probably relatively rare, so don't worry about them
  // for now.
  if (Subtarget->hasBFE())
    setHasExtractBitsInsn(true);

  setOperationAction(ISD::FMINNUM, MVT::f32, Custom);
  setOperationAction(ISD::FMAXNUM, MVT::f32, Custom);
  setOperationAction(ISD::FMINNUM, MVT::f64, Custom);
  setOperationAction(ISD::FMAXNUM, MVT::f64, Custom);


  // These are really only legal for ieee_mode functions. We should be avoiding
  // them for functions that don't have ieee_mode enabled, so just say they are
  // legal.
  setOperationAction(ISD::FMINNUM_IEEE, MVT::f32, Legal);
  setOperationAction(ISD::FMAXNUM_IEEE, MVT::f32, Legal);
  setOperationAction(ISD::FMINNUM_IEEE, MVT::f64, Legal);
  setOperationAction(ISD::FMAXNUM_IEEE, MVT::f64, Legal);


  if (Subtarget->haveRoundOpsF64()) {
    setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
    setOperationAction(ISD::FCEIL, MVT::f64, Legal);
    setOperationAction(ISD::FRINT, MVT::f64, Legal);
  } else {
    setOperationAction(ISD::FCEIL, MVT::f64, Custom);
    setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
    setOperationAction(ISD::FRINT, MVT::f64, Custom);
    setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
  }

  setOperationAction(ISD::FFLOOR, MVT::f64, Legal);

  setOperationAction(ISD::FSIN, MVT::f32, Custom);
  setOperationAction(ISD::FCOS, MVT::f32, Custom);
  setOperationAction(ISD::FDIV, MVT::f32, Custom);
  setOperationAction(ISD::FDIV, MVT::f64, Custom);

  if (Subtarget->has16BitInsts()) {
    setOperationAction(ISD::Constant, MVT::i16, Legal);

    setOperationAction(ISD::SMIN, MVT::i16, Legal);
    setOperationAction(ISD::SMAX, MVT::i16, Legal);

    setOperationAction(ISD::UMIN, MVT::i16, Legal);
    setOperationAction(ISD::UMAX, MVT::i16, Legal);

    setOperationAction(ISD::SIGN_EXTEND, MVT::i16, Promote);
    AddPromotedToType(ISD::SIGN_EXTEND, MVT::i16, MVT::i32);

    setOperationAction(ISD::ROTR, MVT::i16, Promote);
    setOperationAction(ISD::ROTL, MVT::i16, Promote);

    setOperationAction(ISD::SDIV, MVT::i16, Promote);
    setOperationAction(ISD::UDIV, MVT::i16, Promote);
    setOperationAction(ISD::SREM, MVT::i16, Promote);
    setOperationAction(ISD::UREM, MVT::i16, Promote);

    setOperationAction(ISD::BSWAP, MVT::i16, Promote);
    setOperationAction(ISD::BITREVERSE, MVT::i16, Promote);

    setOperationAction(ISD::CTTZ, MVT::i16, Promote);
    setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Promote);
    setOperationAction(ISD::CTLZ, MVT::i16, Promote);
    setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Promote);
    setOperationAction(ISD::CTPOP, MVT::i16, Promote);

    setOperationAction(ISD::SELECT_CC, MVT::i16, Expand);

    setOperationAction(ISD::BR_CC, MVT::i16, Expand);

    setOperationAction(ISD::LOAD, MVT::i16, Custom);

    setTruncStoreAction(MVT::i64, MVT::i16, Expand);

    setOperationAction(ISD::FP16_TO_FP, MVT::i16, Promote);
    AddPromotedToType(ISD::FP16_TO_FP, MVT::i16, MVT::i32);
    setOperationAction(ISD::FP_TO_FP16, MVT::i16, Promote);
    AddPromotedToType(ISD::FP_TO_FP16, MVT::i16, MVT::i32);

    setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
    setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
    setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
    setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);

    // F16 - Constant Actions.
    setOperationAction(ISD::ConstantFP, MVT::f16, Legal);

    // F16 - Load/Store Actions.
    setOperationAction(ISD::LOAD, MVT::f16, Promote);
    AddPromotedToType(ISD::LOAD, MVT::f16, MVT::i16);
    setOperationAction(ISD::STORE, MVT::f16, Promote);
    AddPromotedToType(ISD::STORE, MVT::f16, MVT::i16);

    // F16 - VOP1 Actions.
    setOperationAction(ISD::FP_ROUND, MVT::f16, Custom);
    setOperationAction(ISD::FCOS, MVT::f16, Promote);
    setOperationAction(ISD::FSIN, MVT::f16, Promote);
    setOperationAction(ISD::FP_TO_SINT, MVT::f16, Promote);
    setOperationAction(ISD::FP_TO_UINT, MVT::f16, Promote);
    setOperationAction(ISD::SINT_TO_FP, MVT::f16, Promote);
    setOperationAction(ISD::UINT_TO_FP, MVT::f16, Promote);
    setOperationAction(ISD::FROUND, MVT::f16, Custom);

    // F16 - VOP2 Actions.
    setOperationAction(ISD::BR_CC, MVT::f16, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);

    setOperationAction(ISD::FDIV, MVT::f16, Custom);

    // F16 - VOP3 Actions.
    setOperationAction(ISD::FMA, MVT::f16, Legal);
    if (!Subtarget->hasFP16Denormals() && STI.hasMadF16())
      setOperationAction(ISD::FMAD, MVT::f16, Legal);

    for (MVT VT : {MVT::v2i16, MVT::v2f16, MVT::v4i16, MVT::v4f16}) {
      for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
        switch (Op) {
        case ISD::LOAD:
        case ISD::STORE:
        case ISD::BUILD_VECTOR:
        case ISD::BITCAST:
        case ISD::EXTRACT_VECTOR_ELT:
        case ISD::INSERT_VECTOR_ELT:
        case ISD::INSERT_SUBVECTOR:
        case ISD::EXTRACT_SUBVECTOR:
        case ISD::SCALAR_TO_VECTOR:
          break;
        case ISD::CONCAT_VECTORS:
          setOperationAction(Op, VT, Custom);
          break;
        default:
          setOperationAction(Op, VT, Expand);
          break;
        }
      }
    }

    // XXX - Do these do anything? Vector constants turn into build_vector.
    setOperationAction(ISD::Constant, MVT::v2i16, Legal);
    setOperationAction(ISD::ConstantFP, MVT::v2f16, Legal);

    setOperationAction(ISD::UNDEF, MVT::v2i16, Legal);
    setOperationAction(ISD::UNDEF, MVT::v2f16, Legal);

    setOperationAction(ISD::STORE, MVT::v2i16, Promote);
    AddPromotedToType(ISD::STORE, MVT::v2i16, MVT::i32);
    setOperationAction(ISD::STORE, MVT::v2f16, Promote);
    AddPromotedToType(ISD::STORE, MVT::v2f16, MVT::i32);

    setOperationAction(ISD::LOAD, MVT::v2i16, Promote);
    AddPromotedToType(ISD::LOAD, MVT::v2i16, MVT::i32);
    setOperationAction(ISD::LOAD, MVT::v2f16, Promote);
    AddPromotedToType(ISD::LOAD, MVT::v2f16, MVT::i32);

    setOperationAction(ISD::AND, MVT::v2i16, Promote);
    AddPromotedToType(ISD::AND, MVT::v2i16, MVT::i32);
    setOperationAction(ISD::OR, MVT::v2i16, Promote);
    AddPromotedToType(ISD::OR, MVT::v2i16, MVT::i32);
    setOperationAction(ISD::XOR, MVT::v2i16, Promote);
    AddPromotedToType(ISD::XOR, MVT::v2i16, MVT::i32);

    setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
    AddPromotedToType(ISD::LOAD, MVT::v4i16, MVT::v2i32);
    setOperationAction(ISD::LOAD, MVT::v4f16, Promote);
    AddPromotedToType(ISD::LOAD, MVT::v4f16, MVT::v2i32);

    setOperationAction(ISD::STORE, MVT::v4i16, Promote);
    AddPromotedToType(ISD::STORE, MVT::v4i16, MVT::v2i32);
    setOperationAction(ISD::STORE, MVT::v4f16, Promote);
    AddPromotedToType(ISD::STORE, MVT::v4f16, MVT::v2i32);

    setOperationAction(ISD::ANY_EXTEND, MVT::v2i32, Expand);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v2i32, Expand);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v2i32, Expand);
    setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Expand);

    setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Expand);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v4i32, Expand);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Expand);

    if (!Subtarget->hasVOP3PInsts()) {
      setOperationAction(ISD::BUILD_VECTOR, MVT::v2i16, Custom);
      setOperationAction(ISD::BUILD_VECTOR, MVT::v2f16, Custom);
    }

    setOperationAction(ISD::FNEG, MVT::v2f16, Legal);
    // This isn't really legal, but this avoids the legalizer unrolling it (and
    // allows matching fneg (fabs x) patterns)
    setOperationAction(ISD::FABS, MVT::v2f16, Legal);

    setOperationAction(ISD::FMAXNUM, MVT::f16, Custom);
    setOperationAction(ISD::FMINNUM, MVT::f16, Custom);
    setOperationAction(ISD::FMAXNUM_IEEE, MVT::f16, Legal);
    setOperationAction(ISD::FMINNUM_IEEE, MVT::f16, Legal);

    setOperationAction(ISD::FMINNUM_IEEE, MVT::v4f16, Custom);
    setOperationAction(ISD::FMAXNUM_IEEE, MVT::v4f16, Custom);

    setOperationAction(ISD::FMINNUM, MVT::v4f16, Expand);
    setOperationAction(ISD::FMAXNUM, MVT::v4f16, Expand);
  }

  if (Subtarget->hasVOP3PInsts()) {
    setOperationAction(ISD::ADD, MVT::v2i16, Legal);
    setOperationAction(ISD::SUB, MVT::v2i16, Legal);
    setOperationAction(ISD::MUL, MVT::v2i16, Legal);
    setOperationAction(ISD::SHL, MVT::v2i16, Legal);
    setOperationAction(ISD::SRL, MVT::v2i16, Legal);
    setOperationAction(ISD::SRA, MVT::v2i16, Legal);
    setOperationAction(ISD::SMIN, MVT::v2i16, Legal);
    setOperationAction(ISD::UMIN, MVT::v2i16, Legal);
    setOperationAction(ISD::SMAX, MVT::v2i16, Legal);
    setOperationAction(ISD::UMAX, MVT::v2i16, Legal);

    setOperationAction(ISD::FADD, MVT::v2f16, Legal);
    setOperationAction(ISD::FMUL, MVT::v2f16, Legal);
    setOperationAction(ISD::FMA, MVT::v2f16, Legal);

    setOperationAction(ISD::FMINNUM_IEEE, MVT::v2f16, Legal);
    setOperationAction(ISD::FMAXNUM_IEEE, MVT::v2f16, Legal);

    setOperationAction(ISD::FCANONICALIZE, MVT::v2f16, Legal);

    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i16, Custom);
    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f16, Custom);

    setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f16, Custom);
    setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);

    setOperationAction(ISD::SHL, MVT::v4i16, Custom);
    setOperationAction(ISD::SRA, MVT::v4i16, Custom);
    setOperationAction(ISD::SRL, MVT::v4i16, Custom);
    setOperationAction(ISD::ADD, MVT::v4i16, Custom);
    setOperationAction(ISD::SUB, MVT::v4i16, Custom);
    setOperationAction(ISD::MUL, MVT::v4i16, Custom);

    setOperationAction(ISD::SMIN, MVT::v4i16, Custom);
    setOperationAction(ISD::SMAX, MVT::v4i16, Custom);
    setOperationAction(ISD::UMIN, MVT::v4i16, Custom);
    setOperationAction(ISD::UMAX, MVT::v4i16, Custom);

    setOperationAction(ISD::FADD, MVT::v4f16, Custom);
    setOperationAction(ISD::FMUL, MVT::v4f16, Custom);
    setOperationAction(ISD::FMA, MVT::v4f16, Custom);

    setOperationAction(ISD::FMAXNUM, MVT::v2f16, Custom);
    setOperationAction(ISD::FMINNUM, MVT::v2f16, Custom);

    setOperationAction(ISD::FMINNUM, MVT::v4f16, Custom);
    setOperationAction(ISD::FMAXNUM, MVT::v4f16, Custom);
    setOperationAction(ISD::FCANONICALIZE, MVT::v4f16, Custom);

    setOperationAction(ISD::FEXP, MVT::v2f16, Custom);
    setOperationAction(ISD::SELECT, MVT::v4i16, Custom);
    setOperationAction(ISD::SELECT, MVT::v4f16, Custom);
  }

  setOperationAction(ISD::FNEG, MVT::v4f16, Custom);
  setOperationAction(ISD::FABS, MVT::v4f16, Custom);

  if (Subtarget->has16BitInsts()) {
    setOperationAction(ISD::SELECT, MVT::v2i16, Promote);
    AddPromotedToType(ISD::SELECT, MVT::v2i16, MVT::i32);
    setOperationAction(ISD::SELECT, MVT::v2f16, Promote);
    AddPromotedToType(ISD::SELECT, MVT::v2f16, MVT::i32);
  } else {
    // Legalization hack.
    setOperationAction(ISD::SELECT, MVT::v2i16, Custom);
    setOperationAction(ISD::SELECT, MVT::v2f16, Custom);

    setOperationAction(ISD::FNEG, MVT::v2f16, Custom);
    setOperationAction(ISD::FABS, MVT::v2f16, Custom);
  }

  for (MVT VT : { MVT::v4i16, MVT::v4f16, MVT::v2i8, MVT::v4i8, MVT::v8i8 }) {
    setOperationAction(ISD::SELECT, VT, Custom);
  }

  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i16, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f16, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v2i16, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v2f16, Custom);

  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::v2f16, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::v2i16, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::v4f16, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::v4i16, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::v8f16, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::f16, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i16, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);

  setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::v2i16, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::v2f16, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::v4f16, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::v4i16, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::f16, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::i16, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::i8, Custom);

  setTargetDAGCombine(ISD::ADD);
  setTargetDAGCombine(ISD::ADDCARRY);
  setTargetDAGCombine(ISD::SUB);
  setTargetDAGCombine(ISD::SUBCARRY);
  setTargetDAGCombine(ISD::FADD);
  setTargetDAGCombine(ISD::FSUB);
  setTargetDAGCombine(ISD::FMINNUM);
  setTargetDAGCombine(ISD::FMAXNUM);
  setTargetDAGCombine(ISD::FMINNUM_IEEE);
  setTargetDAGCombine(ISD::FMAXNUM_IEEE);
  setTargetDAGCombine(ISD::FMA);
  setTargetDAGCombine(ISD::SMIN);
  setTargetDAGCombine(ISD::SMAX);
  setTargetDAGCombine(ISD::UMIN);
  setTargetDAGCombine(ISD::UMAX);
  setTargetDAGCombine(ISD::SETCC);
  setTargetDAGCombine(ISD::AND);
  setTargetDAGCombine(ISD::OR);
  setTargetDAGCombine(ISD::XOR);
  setTargetDAGCombine(ISD::SINT_TO_FP);
  setTargetDAGCombine(ISD::UINT_TO_FP);
  setTargetDAGCombine(ISD::FCANONICALIZE);
  setTargetDAGCombine(ISD::SCALAR_TO_VECTOR);
  setTargetDAGCombine(ISD::ZERO_EXTEND);
  setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
  setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
  setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);

  // All memory operations. Some folding on the pointer operand is done to help
  // matching the constant offsets in the addressing modes.
  setTargetDAGCombine(ISD::LOAD);
  setTargetDAGCombine(ISD::STORE);
  setTargetDAGCombine(ISD::ATOMIC_LOAD);
  setTargetDAGCombine(ISD::ATOMIC_STORE);
  setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
  setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
  setTargetDAGCombine(ISD::ATOMIC_SWAP);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
  setTargetDAGCombine(ISD::ATOMIC_LOAD_FADD);

  setSchedulingPreference(Sched::RegPressure);
}

const GCNSubtarget *SITargetLowering::getSubtarget() const {
  return Subtarget;
}

//===----------------------------------------------------------------------===//
// TargetLowering queries
//===----------------------------------------------------------------------===//

// v_mad_mix* support a conversion from f16 to f32.
//
// There is only one special case when denormals are enabled we don't currently,
// where this is OK to use.
bool SITargetLowering::isFPExtFoldable(unsigned Opcode,
                                           EVT DestVT, EVT SrcVT) const {
  return ((Opcode == ISD::FMAD && Subtarget->hasMadMixInsts()) ||
          (Opcode == ISD::FMA && Subtarget->hasFmaMixInsts())) &&
         DestVT.getScalarType() == MVT::f32 && !Subtarget->hasFP32Denormals() &&
         SrcVT.getScalarType() == MVT::f16;
}

bool SITargetLowering::isShuffleMaskLegal(ArrayRef<int>, EVT) const {
  // SI has some legal vector types, but no legal vector operations. Say no
  // shuffles are legal in order to prefer scalarizing some vector operations.
  return false;
}

MVT SITargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
                                                    CallingConv::ID CC,
                                                    EVT VT) const {
  if (CC == CallingConv::AMDGPU_KERNEL)
    return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);

  if (VT.isVector()) {
    EVT ScalarVT = VT.getScalarType();
    unsigned Size = ScalarVT.getSizeInBits();
    if (Size == 32)
      return ScalarVT.getSimpleVT();

    if (Size > 32)
      return MVT::i32;

    if (Size == 16 && Subtarget->has16BitInsts())
      return VT.isInteger() ? MVT::v2i16 : MVT::v2f16;
  } else if (VT.getSizeInBits() > 32)
    return MVT::i32;

  return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
}

unsigned SITargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
                                                         CallingConv::ID CC,
                                                         EVT VT) const {
  if (CC == CallingConv::AMDGPU_KERNEL)
    return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);

  if (VT.isVector()) {
    unsigned NumElts = VT.getVectorNumElements();
    EVT ScalarVT = VT.getScalarType();
    unsigned Size = ScalarVT.getSizeInBits();

    if (Size == 32)
      return NumElts;

    if (Size > 32)
      return NumElts * ((Size + 31) / 32);

    if (Size == 16 && Subtarget->has16BitInsts())
      return (NumElts + 1) / 2;
  } else if (VT.getSizeInBits() > 32)
    return (VT.getSizeInBits() + 31) / 32;

  return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
}

unsigned SITargetLowering::getVectorTypeBreakdownForCallingConv(
  LLVMContext &Context, CallingConv::ID CC,
  EVT VT, EVT &IntermediateVT,
  unsigned &NumIntermediates, MVT &RegisterVT) const {
  if (CC != CallingConv::AMDGPU_KERNEL && VT.isVector()) {
    unsigned NumElts = VT.getVectorNumElements();
    EVT ScalarVT = VT.getScalarType();
    unsigned Size = ScalarVT.getSizeInBits();
    if (Size == 32) {
      RegisterVT = ScalarVT.getSimpleVT();
      IntermediateVT = RegisterVT;
      NumIntermediates = NumElts;
      return NumIntermediates;
    }

    if (Size > 32) {
      RegisterVT = MVT::i32;
      IntermediateVT = RegisterVT;
      NumIntermediates = NumElts * ((Size + 31) / 32);
      return NumIntermediates;
    }

    // FIXME: We should fix the ABI to be the same on targets without 16-bit
    // support, but unless we can properly handle 3-vectors, it will be still be
    // inconsistent.
    if (Size == 16 && Subtarget->has16BitInsts()) {
      RegisterVT = VT.isInteger() ? MVT::v2i16 : MVT::v2f16;
      IntermediateVT = RegisterVT;
      NumIntermediates = (NumElts + 1) / 2;
      return NumIntermediates;
    }
  }

  return TargetLowering::getVectorTypeBreakdownForCallingConv(
    Context, CC, VT, IntermediateVT, NumIntermediates, RegisterVT);
}

static MVT memVTFromAggregate(Type *Ty) {
  // Only limited forms of aggregate type currently expected.
  assert(Ty->isStructTy() && "Expected struct type");


  Type *ElementType = nullptr;
  unsigned NumElts;
  if (Ty->getContainedType(0)->isVectorTy()) {
    VectorType *VecComponent = cast<VectorType>(Ty->getContainedType(0));
    ElementType = VecComponent->getElementType();
    NumElts = VecComponent->getNumElements();
  } else {
    ElementType = Ty->getContainedType(0);
    NumElts = 1;
  }

  assert((Ty->getContainedType(1) && Ty->getContainedType(1)->isIntegerTy(32)) && "Expected int32 type");

  // Calculate the size of the memVT type from the aggregate
  unsigned Pow2Elts = 0;
  unsigned ElementSize;
  switch (ElementType->getTypeID()) {
    default:
      llvm_unreachable("Unknown type!");
    case Type::IntegerTyID:
      ElementSize = cast<IntegerType>(ElementType)->getBitWidth();
      break;
    case Type::HalfTyID:
      ElementSize = 16;
      break;
    case Type::FloatTyID:
      ElementSize = 32;
      break;
  }
  unsigned AdditionalElts = ElementSize == 16 ? 2 : 1;
  Pow2Elts = 1 << Log2_32_Ceil(NumElts + AdditionalElts);

  return MVT::getVectorVT(MVT::getVT(ElementType, false),
                          Pow2Elts);
}

bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
                                          const CallInst &CI,
                                          MachineFunction &MF,
                                          unsigned IntrID) const {
  if (const AMDGPU::RsrcIntrinsic *RsrcIntr =
          AMDGPU::lookupRsrcIntrinsic(IntrID)) {
    AttributeList Attr = Intrinsic::getAttributes(CI.getContext(),
                                                  (Intrinsic::ID)IntrID);
    if (Attr.hasFnAttribute(Attribute::ReadNone))
      return false;

    SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

    if (RsrcIntr->IsImage) {
      Info.ptrVal = MFI->getImagePSV(
        *MF.getSubtarget<GCNSubtarget>().getInstrInfo(),
        CI.getArgOperand(RsrcIntr->RsrcArg));
      Info.align.reset();
    } else {
      Info.ptrVal = MFI->getBufferPSV(
        *MF.getSubtarget<GCNSubtarget>().getInstrInfo(),
        CI.getArgOperand(RsrcIntr->RsrcArg));
    }

    Info.flags = MachineMemOperand::MODereferenceable;
    if (Attr.hasFnAttribute(Attribute::ReadOnly)) {
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.memVT = MVT::getVT(CI.getType(), true);
      if (Info.memVT == MVT::Other) {
        // Some intrinsics return an aggregate type - special case to work out
        // the correct memVT
        Info.memVT = memVTFromAggregate(CI.getType());
      }
      Info.flags |= MachineMemOperand::MOLoad;
    } else if (Attr.hasFnAttribute(Attribute::WriteOnly)) {
      Info.opc = ISD::INTRINSIC_VOID;
      Info.memVT = MVT::getVT(CI.getArgOperand(0)->getType());
      Info.flags |= MachineMemOperand::MOStore;
    } else {
      // Atomic
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.memVT = MVT::getVT(CI.getType());
      Info.flags = MachineMemOperand::MOLoad |
                   MachineMemOperand::MOStore |
                   MachineMemOperand::MODereferenceable;

      // XXX - Should this be volatile without known ordering?
      Info.flags |= MachineMemOperand::MOVolatile;
    }
    return true;
  }

  switch (IntrID) {
  case Intrinsic::amdgcn_atomic_inc:
  case Intrinsic::amdgcn_atomic_dec:
  case Intrinsic::amdgcn_ds_ordered_add:
  case Intrinsic::amdgcn_ds_ordered_swap:
  case Intrinsic::amdgcn_ds_fadd:
  case Intrinsic::amdgcn_ds_fmin:
  case Intrinsic::amdgcn_ds_fmax: {
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(CI.getType());
    Info.ptrVal = CI.getOperand(0);
    Info.align.reset();
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;

    const ConstantInt *Vol = cast<ConstantInt>(CI.getOperand(4));
    if (!Vol->isZero())
      Info.flags |= MachineMemOperand::MOVolatile;

    return true;
  }
  case Intrinsic::amdgcn_buffer_atomic_fadd: {
    SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

    Info.opc = ISD::INTRINSIC_VOID;
    Info.memVT = MVT::getVT(CI.getOperand(0)->getType());
    Info.ptrVal = MFI->getBufferPSV(
      *MF.getSubtarget<GCNSubtarget>().getInstrInfo(),
      CI.getArgOperand(1));
    Info.align.reset();
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;

    const ConstantInt *Vol = dyn_cast<ConstantInt>(CI.getOperand(4));
    if (!Vol || !Vol->isZero())
      Info.flags |= MachineMemOperand::MOVolatile;

    return true;
  }
  case Intrinsic::amdgcn_global_atomic_fadd: {
    Info.opc = ISD::INTRINSIC_VOID;
    Info.memVT = MVT::getVT(CI.getOperand(0)->getType()
                            ->getPointerElementType());
    Info.ptrVal = CI.getOperand(0);
    Info.align.reset();
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;

    return true;
  }
  case Intrinsic::amdgcn_ds_append:
  case Intrinsic::amdgcn_ds_consume: {
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(CI.getType());
    Info.ptrVal = CI.getOperand(0);
    Info.align.reset();
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;

    const ConstantInt *Vol = cast<ConstantInt>(CI.getOperand(1));
    if (!Vol->isZero())
      Info.flags |= MachineMemOperand::MOVolatile;

    return true;
  }
  case Intrinsic::amdgcn_ds_gws_init:
  case Intrinsic::amdgcn_ds_gws_barrier:
  case Intrinsic::amdgcn_ds_gws_sema_v:
  case Intrinsic::amdgcn_ds_gws_sema_br:
  case Intrinsic::amdgcn_ds_gws_sema_p:
  case Intrinsic::amdgcn_ds_gws_sema_release_all: {
    Info.opc = ISD::INTRINSIC_VOID;

    SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
    Info.ptrVal =
        MFI->getGWSPSV(*MF.getSubtarget<GCNSubtarget>().getInstrInfo());

    // This is an abstract access, but we need to specify a type and size.
    Info.memVT = MVT::i32;
    Info.size = 4;
    Info.align = Align(4);

    Info.flags = MachineMemOperand::MOStore;
    if (IntrID == Intrinsic::amdgcn_ds_gws_barrier)
      Info.flags = MachineMemOperand::MOLoad;
    return true;
  }
  default:
    return false;
  }
}

bool SITargetLowering::getAddrModeArguments(IntrinsicInst *II,
                                            SmallVectorImpl<Value*> &Ops,
                                            Type *&AccessTy) const {
  switch (II->getIntrinsicID()) {
  case Intrinsic::amdgcn_atomic_inc:
  case Intrinsic::amdgcn_atomic_dec:
  case Intrinsic::amdgcn_ds_ordered_add:
  case Intrinsic::amdgcn_ds_ordered_swap:
  case Intrinsic::amdgcn_ds_fadd:
  case Intrinsic::amdgcn_ds_fmin:
  case Intrinsic::amdgcn_ds_fmax: {
    Value *Ptr = II->getArgOperand(0);
    AccessTy = II->getType();
    Ops.push_back(Ptr);
    return true;
  }
  default:
    return false;
  }
}

bool SITargetLowering::isLegalFlatAddressingMode(const AddrMode &AM) const {
  if (!Subtarget->hasFlatInstOffsets()) {
    // Flat instructions do not have offsets, and only have the register
    // address.
    return AM.BaseOffs == 0 && AM.Scale == 0;
  }

  // GFX9 added a 13-bit signed offset. When using regular flat instructions,
  // the sign bit is ignored and is treated as a 12-bit unsigned offset.

  // GFX10 shrinked signed offset to 12 bits. When using regular flat
  // instructions, the sign bit is also ignored and is treated as 11-bit
  // unsigned offset.

  if (Subtarget->getGeneration() >= AMDGPUSubtarget::GFX10)
    return isUInt<11>(AM.BaseOffs) && AM.Scale == 0;

  // Just r + i
  return isUInt<12>(AM.BaseOffs) && AM.Scale == 0;
}

bool SITargetLowering::isLegalGlobalAddressingMode(const AddrMode &AM) const {
  if (Subtarget->hasFlatGlobalInsts())
    return isInt<13>(AM.BaseOffs) && AM.Scale == 0;

  if (!Subtarget->hasAddr64() || Subtarget->useFlatForGlobal()) {
      // Assume the we will use FLAT for all global memory accesses
      // on VI.
      // FIXME: This assumption is currently wrong.  On VI we still use
      // MUBUF instructions for the r + i addressing mode.  As currently
      // implemented, the MUBUF instructions only work on buffer < 4GB.
      // It may be possible to support > 4GB buffers with MUBUF instructions,
      // by setting the stride value in the resource descriptor which would
      // increase the size limit to (stride * 4GB).  However, this is risky,
      // because it has never been validated.
    return isLegalFlatAddressingMode(AM);
  }

  return isLegalMUBUFAddressingMode(AM);
}

bool SITargetLowering::isLegalMUBUFAddressingMode(const AddrMode &AM) const {
  // MUBUF / MTBUF instructions have a 12-bit unsigned byte offset, and
  // additionally can do r + r + i with addr64. 32-bit has more addressing
  // mode options. Depending on the resource constant, it can also do
  // (i64 r0) + (i32 r1) * (i14 i).
  //
  // Private arrays end up using a scratch buffer most of the time, so also
  // assume those use MUBUF instructions. Scratch loads / stores are currently
  // implemented as mubuf instructions with offen bit set, so slightly
  // different than the normal addr64.
  if (!isUInt<12>(AM.BaseOffs))
    return false;

  // FIXME: Since we can split immediate into soffset and immediate offset,
  // would it make sense to allow any immediate?

  switch (AM.Scale) {
  case 0: // r + i or just i, depending on HasBaseReg.
    return true;
  case 1:
    return true; // We have r + r or r + i.
  case 2:
    if (AM.HasBaseReg) {
      // Reject 2 * r + r.
      return false;
    }

    // Allow 2 * r as r + r
    // Or  2 * r + i is allowed as r + r + i.
    return true;
  default: // Don't allow n * r
    return false;
  }
}

bool SITargetLowering::isLegalAddressingMode(const DataLayout &DL,
                                             const AddrMode &AM, Type *Ty,
                                             unsigned AS, Instruction *I) const {
  // No global is ever allowed as a base.
  if (AM.BaseGV)
    return false;

  if (AS == AMDGPUAS::GLOBAL_ADDRESS)
    return isLegalGlobalAddressingMode(AM);

  if (AS == AMDGPUAS::CONSTANT_ADDRESS ||
      AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
      AS == AMDGPUAS::BUFFER_FAT_POINTER) {
    // If the offset isn't a multiple of 4, it probably isn't going to be
    // correctly aligned.
    // FIXME: Can we get the real alignment here?
    if (AM.BaseOffs % 4 != 0)
      return isLegalMUBUFAddressingMode(AM);

    // There are no SMRD extloads, so if we have to do a small type access we
    // will use a MUBUF load.
    // FIXME?: We also need to do this if unaligned, but we don't know the
    // alignment here.
    if (Ty->isSized() && DL.getTypeStoreSize(Ty) < 4)
      return isLegalGlobalAddressingMode(AM);

    if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) {
      // SMRD instructions have an 8-bit, dword offset on SI.
      if (!isUInt<8>(AM.BaseOffs / 4))
        return false;
    } else if (Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS) {
      // On CI+, this can also be a 32-bit literal constant offset. If it fits
      // in 8-bits, it can use a smaller encoding.
      if (!isUInt<32>(AM.BaseOffs / 4))
        return false;
    } else if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
      // On VI, these use the SMEM format and the offset is 20-bit in bytes.
      if (!isUInt<20>(AM.BaseOffs))
        return false;
    } else
      llvm_unreachable("unhandled generation");

    if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
      return true;

    if (AM.Scale == 1 && AM.HasBaseReg)
      return true;

    return false;

  } else if (AS == AMDGPUAS::PRIVATE_ADDRESS) {
    return isLegalMUBUFAddressingMode(AM);
  } else if (AS == AMDGPUAS::LOCAL_ADDRESS ||
             AS == AMDGPUAS::REGION_ADDRESS) {
    // Basic, single offset DS instructions allow a 16-bit unsigned immediate
    // field.
    // XXX - If doing a 4-byte aligned 8-byte type access, we effectively have
    // an 8-bit dword offset but we don't know the alignment here.
    if (!isUInt<16>(AM.BaseOffs))
      return false;

    if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
      return true;

    if (AM.Scale == 1 && AM.HasBaseReg)
      return true;

    return false;
  } else if (AS == AMDGPUAS::FLAT_ADDRESS ||
             AS == AMDGPUAS::UNKNOWN_ADDRESS_SPACE) {
    // For an unknown address space, this usually means that this is for some
    // reason being used for pure arithmetic, and not based on some addressing
    // computation. We don't have instructions that compute pointers with any
    // addressing modes, so treat them as having no offset like flat
    // instructions.
    return isLegalFlatAddressingMode(AM);
  } else {
    llvm_unreachable("unhandled address space");
  }
}

bool SITargetLowering::canMergeStoresTo(unsigned AS, EVT MemVT,
                                        const SelectionDAG &DAG) const {
  if (AS == AMDGPUAS::GLOBAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS) {
    return (MemVT.getSizeInBits() <= 4 * 32);
  } else if (AS == AMDGPUAS::PRIVATE_ADDRESS) {
    unsigned MaxPrivateBits = 8 * getSubtarget()->getMaxPrivateElementSize();
    return (MemVT.getSizeInBits() <= MaxPrivateBits);
  } else if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) {
    return (MemVT.getSizeInBits() <= 2 * 32);
  }
  return true;
}

bool SITargetLowering::allowsMisalignedMemoryAccessesImpl(
    unsigned Size, unsigned AddrSpace, unsigned Align,
    MachineMemOperand::Flags Flags, bool *IsFast) const {
  if (IsFast)
    *IsFast = false;

  if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS ||
      AddrSpace == AMDGPUAS::REGION_ADDRESS) {
    // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
    // aligned, 8 byte access in a single operation using ds_read2/write2_b32
    // with adjacent offsets.
    bool AlignedBy4 = (Align % 4 == 0);
    if (IsFast)
      *IsFast = AlignedBy4;

    return AlignedBy4;
  }

  // FIXME: We have to be conservative here and assume that flat operations
  // will access scratch.  If we had access to the IR function, then we
  // could determine if any private memory was used in the function.
  if (!Subtarget->hasUnalignedScratchAccess() &&
      (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS ||
       AddrSpace == AMDGPUAS::FLAT_ADDRESS)) {
    bool AlignedBy4 = Align >= 4;
    if (IsFast)
      *IsFast = AlignedBy4;

    return AlignedBy4;
  }

  if (Subtarget->hasUnalignedBufferAccess()) {
    // If we have an uniform constant load, it still requires using a slow
    // buffer instruction if unaligned.
    if (IsFast) {
      *IsFast = (AddrSpace == AMDGPUAS::CONSTANT_ADDRESS ||
                 AddrSpace == AMDGPUAS::CONSTANT_ADDRESS_32BIT) ?
        (Align % 4 == 0) : true;
    }

    return true;
  }

  // Smaller than dword value must be aligned.
  if (Size < 32)
    return false;

  // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
  // byte-address are ignored, thus forcing Dword alignment.
  // This applies to private, global, and constant memory.
  if (IsFast)
    *IsFast = true;

  return Size >= 32 && Align >= 4;
}

bool SITargetLowering::allowsMisalignedMemoryAccesses(
    EVT VT, unsigned AddrSpace, unsigned Align, MachineMemOperand::Flags Flags,
    bool *IsFast) const {
  if (IsFast)
    *IsFast = false;

  // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
  // which isn't a simple VT.
  // Until MVT is extended to handle this, simply check for the size and
  // rely on the condition below: allow accesses if the size is a multiple of 4.
  if (VT == MVT::Other || (VT != MVT::Other && VT.getSizeInBits() > 1024 &&
                           VT.getStoreSize() > 16)) {
    return false;
  }

  return allowsMisalignedMemoryAccessesImpl(VT.getSizeInBits(), AddrSpace,
                                            Align, Flags, IsFast);
}

EVT SITargetLowering::getOptimalMemOpType(
    uint64_t Size, unsigned DstAlign, unsigned SrcAlign, bool IsMemset,
    bool ZeroMemset, bool MemcpyStrSrc,
    const AttributeList &FuncAttributes) const {
  // FIXME: Should account for address space here.

  // The default fallback uses the private pointer size as a guess for a type to
  // use. Make sure we switch these to 64-bit accesses.

  if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global
    return MVT::v4i32;

  if (Size >= 8 && DstAlign >= 4)
    return MVT::v2i32;

  // Use the default.
  return MVT::Other;
}

static bool isFlatGlobalAddrSpace(unsigned AS) {
  return AS == AMDGPUAS::GLOBAL_ADDRESS ||
         AS == AMDGPUAS::FLAT_ADDRESS ||
         AS == AMDGPUAS::CONSTANT_ADDRESS ||
         AS > AMDGPUAS::MAX_AMDGPU_ADDRESS;
}

bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
                                           unsigned DestAS) const {
  return isFlatGlobalAddrSpace(SrcAS) && isFlatGlobalAddrSpace(DestAS);
}

bool SITargetLowering::isMemOpHasNoClobberedMemOperand(const SDNode *N) const {
  const MemSDNode *MemNode = cast<MemSDNode>(N);
  const Value *Ptr = MemNode->getMemOperand()->getValue();
  const Instruction *I = dyn_cast_or_null<Instruction>(Ptr);
  return I && I->getMetadata("amdgpu.noclobber");
}

bool SITargetLowering::isFreeAddrSpaceCast(unsigned SrcAS,
                                           unsigned DestAS) const {
  // Flat -> private/local is a simple truncate.
  // Flat -> global is no-op
  if (SrcAS == AMDGPUAS::FLAT_ADDRESS)
    return true;

  return isNoopAddrSpaceCast(SrcAS, DestAS);
}

bool SITargetLowering::isMemOpUniform(const SDNode *N) const {
  const MemSDNode *MemNode = cast<MemSDNode>(N);

  return AMDGPUInstrInfo::isUniformMMO(MemNode->getMemOperand());
}

TargetLoweringBase::LegalizeTypeAction
SITargetLowering::getPreferredVectorAction(MVT VT) const {
  int NumElts = VT.getVectorNumElements();
  if (NumElts != 1 && VT.getScalarType().bitsLE(MVT::i16))
    return VT.isPow2VectorType() ? TypeSplitVector : TypeWidenVector;
  return TargetLoweringBase::getPreferredVectorAction(VT);
}

bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                                         Type *Ty) const {
  // FIXME: Could be smarter if called for vector constants.
  return true;
}

bool SITargetLowering::isTypeDesirableForOp(unsigned Op, EVT VT) const {
  if (Subtarget->has16BitInsts() && VT == MVT::i16) {
    switch (Op) {
    case ISD::LOAD:
    case ISD::STORE:

    // These operations are done with 32-bit instructions anyway.
    case ISD::AND:
    case ISD::OR:
    case ISD::XOR:
    case ISD::SELECT:
      // TODO: Extensions?
      return true;
    default:
      return false;
    }
  }

  // SimplifySetCC uses this function to determine whether or not it should
  // create setcc with i1 operands.  We don't have instructions for i1 setcc.
  if (VT == MVT::i1 && Op == ISD::SETCC)
    return false;

  return TargetLowering::isTypeDesirableForOp(Op, VT);
}

SDValue SITargetLowering::lowerKernArgParameterPtr(SelectionDAG &DAG,
                                                   const SDLoc &SL,
                                                   SDValue Chain,
                                                   uint64_t Offset) const {
  const DataLayout &DL = DAG.getDataLayout();
  MachineFunction &MF = DAG.getMachineFunction();
  const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();

  const ArgDescriptor *InputPtrReg;
  const TargetRegisterClass *RC;

  std::tie(InputPtrReg, RC)
    = Info->getPreloadedValue(AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);

  MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
  MVT PtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
  SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
    MRI.getLiveInVirtReg(InputPtrReg->getRegister()), PtrVT);

  return DAG.getObjectPtrOffset(SL, BasePtr, Offset);
}

SDValue SITargetLowering::getImplicitArgPtr(SelectionDAG &DAG,
                                            const SDLoc &SL) const {
  uint64_t Offset = getImplicitParameterOffset(DAG.getMachineFunction(),
                                               FIRST_IMPLICIT);
  return lowerKernArgParameterPtr(DAG, SL, DAG.getEntryNode(), Offset);
}

SDValue SITargetLowering::convertArgType(SelectionDAG &DAG, EVT VT, EVT MemVT,
                                         const SDLoc &SL, SDValue Val,
                                         bool Signed,
                                         const ISD::InputArg *Arg) const {
  // First, if it is a widened vector, narrow it.
  if (VT.isVector() &&
      VT.getVectorNumElements() != MemVT.getVectorNumElements()) {
    EVT NarrowedVT =
        EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(),
                         VT.getVectorNumElements());
    Val = DAG.getNode(ISD::EXTRACT_SUBVECTOR, SL, NarrowedVT, Val,
                      DAG.getConstant(0, SL, MVT::i32));
  }

  // Then convert the vector elements or scalar value.
  if (Arg && (Arg->Flags.isSExt() || Arg->Flags.isZExt()) &&
      VT.bitsLT(MemVT)) {
    unsigned Opc = Arg->Flags.isZExt() ? ISD::AssertZext : ISD::AssertSext;
    Val = DAG.getNode(Opc, SL, MemVT, Val, DAG.getValueType(VT));
  }

  if (MemVT.isFloatingPoint())
    Val = getFPExtOrFPTrunc(DAG, Val, SL, VT);
  else if (Signed)
    Val = DAG.getSExtOrTrunc(Val, SL, VT);
  else
    Val = DAG.getZExtOrTrunc(Val, SL, VT);

  return Val;
}

SDValue SITargetLowering::lowerKernargMemParameter(
  SelectionDAG &DAG, EVT VT, EVT MemVT,
  const SDLoc &SL, SDValue Chain,
  uint64_t Offset, unsigned Align, bool Signed,
  const ISD::InputArg *Arg) const {
  Type *Ty = MemVT.getTypeForEVT(*DAG.getContext());
  PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
  MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));

  // Try to avoid using an extload by loading earlier than the argument address,
  // and extracting the relevant bits. The load should hopefully be merged with
  // the previous argument.
  if (MemVT.getStoreSize() < 4 && Align < 4) {
    // TODO: Handle align < 4 and size >= 4 (can happen with packed structs).
    int64_t AlignDownOffset = alignDown(Offset, 4);
    int64_t OffsetDiff = Offset - AlignDownOffset;

    EVT IntVT = MemVT.changeTypeToInteger();

    // TODO: If we passed in the base kernel offset we could have a better
    // alignment than 4, but we don't really need it.
    SDValue Ptr = lowerKernArgParameterPtr(DAG, SL, Chain, AlignDownOffset);
    SDValue Load = DAG.getLoad(MVT::i32, SL, Chain, Ptr, PtrInfo, 4,
                               MachineMemOperand::MODereferenceable |
                               MachineMemOperand::MOInvariant);

    SDValue ShiftAmt = DAG.getConstant(OffsetDiff * 8, SL, MVT::i32);
    SDValue Extract = DAG.getNode(ISD::SRL, SL, MVT::i32, Load, ShiftAmt);

    SDValue ArgVal = DAG.getNode(ISD::TRUNCATE, SL, IntVT, Extract);
    ArgVal = DAG.getNode(ISD::BITCAST, SL, MemVT, ArgVal);
    ArgVal = convertArgType(DAG, VT, MemVT, SL, ArgVal, Signed, Arg);


    return DAG.getMergeValues({ ArgVal, Load.getValue(1) }, SL);
  }

  SDValue Ptr = lowerKernArgParameterPtr(DAG, SL, Chain, Offset);
  SDValue Load = DAG.getLoad(MemVT, SL, Chain, Ptr, PtrInfo, Align,
                             MachineMemOperand::MODereferenceable |
                             MachineMemOperand::MOInvariant);

  SDValue Val = convertArgType(DAG, VT, MemVT, SL, Load, Signed, Arg);
  return DAG.getMergeValues({ Val, Load.getValue(1) }, SL);
}

SDValue SITargetLowering::lowerStackParameter(SelectionDAG &DAG, CCValAssign &VA,
                                              const SDLoc &SL, SDValue Chain,
                                              const ISD::InputArg &Arg) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();

  if (Arg.Flags.isByVal()) {
    unsigned Size = Arg.Flags.getByValSize();
    int FrameIdx = MFI.CreateFixedObject(Size, VA.getLocMemOffset(), false);
    return DAG.getFrameIndex(FrameIdx, MVT::i32);
  }

  unsigned ArgOffset = VA.getLocMemOffset();
  unsigned ArgSize = VA.getValVT().getStoreSize();

  int FI = MFI.CreateFixedObject(ArgSize, ArgOffset, true);

  // Create load nodes to retrieve arguments from the stack.
  SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
  SDValue ArgValue;

  // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
  ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
  MVT MemVT = VA.getValVT();

  switch (VA.getLocInfo()) {
  default:
    break;
  case CCValAssign::BCvt:
    MemVT = VA.getLocVT();
    break;
  case CCValAssign::SExt:
    ExtType = ISD::SEXTLOAD;
    break;
  case CCValAssign::ZExt:
    ExtType = ISD::ZEXTLOAD;
    break;
  case CCValAssign::AExt:
    ExtType = ISD::EXTLOAD;
    break;
  }

  ArgValue = DAG.getExtLoad(
    ExtType, SL, VA.getLocVT(), Chain, FIN,
    MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
    MemVT);
  return ArgValue;
}

SDValue SITargetLowering::getPreloadedValue(SelectionDAG &DAG,
  const SIMachineFunctionInfo &MFI,
  EVT VT,
  AMDGPUFunctionArgInfo::PreloadedValue PVID) const {
  const ArgDescriptor *Reg;
  const TargetRegisterClass *RC;

  std::tie(Reg, RC) = MFI.getPreloadedValue(PVID);
  return CreateLiveInRegister(DAG, RC, Reg->getRegister(), VT);
}

static void processShaderInputArgs(SmallVectorImpl<ISD::InputArg> &Splits,
                                   CallingConv::ID CallConv,
                                   ArrayRef<ISD::InputArg> Ins,
                                   BitVector &Skipped,
                                   FunctionType *FType,
                                   SIMachineFunctionInfo *Info) {
  for (unsigned I = 0, E = Ins.size(), PSInputNum = 0; I != E; ++I) {
    const ISD::InputArg *Arg = &Ins[I];

    assert((!Arg->VT.isVector() || Arg->VT.getScalarSizeInBits() == 16) &&
           "vector type argument should have been split");

    // First check if it's a PS input addr.
    if (CallConv == CallingConv::AMDGPU_PS &&
        !Arg->Flags.isInReg() && PSInputNum <= 15) {
      bool SkipArg = !Arg->Used && !Info->isPSInputAllocated(PSInputNum);

      // Inconveniently only the first part of the split is marked as isSplit,
      // so skip to the end. We only want to increment PSInputNum once for the
      // entire split argument.
      if (Arg->Flags.isSplit()) {
        while (!Arg->Flags.isSplitEnd()) {
          assert((!Arg->VT.isVector() ||
                  Arg->VT.getScalarSizeInBits() == 16) &&
                 "unexpected vector split in ps argument type");
          if (!SkipArg)
            Splits.push_back(*Arg);
          Arg = &Ins[++I];
        }
      }

      if (SkipArg) {
        // We can safely skip PS inputs.
        Skipped.set(Arg->getOrigArgIndex());
        ++PSInputNum;
        continue;
      }

      Info->markPSInputAllocated(PSInputNum);
      if (Arg->Used)
        Info->markPSInputEnabled(PSInputNum);

      ++PSInputNum;
    }

    Splits.push_back(*Arg);
  }
}

// Allocate special inputs passed in VGPRs.
void SITargetLowering::allocateSpecialEntryInputVGPRs(CCState &CCInfo,
                                                      MachineFunction &MF,
                                                      const SIRegisterInfo &TRI,
                                                      SIMachineFunctionInfo &Info) const {
  const LLT S32 = LLT::scalar(32);
  MachineRegisterInfo &MRI = MF.getRegInfo();

  if (Info.hasWorkItemIDX()) {
    Register Reg = AMDGPU::VGPR0;
    MRI.setType(MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass), S32);

    CCInfo.AllocateReg(Reg);
    Info.setWorkItemIDX(ArgDescriptor::createRegister(Reg));
  }

  if (Info.hasWorkItemIDY()) {
    Register Reg = AMDGPU::VGPR1;
    MRI.setType(MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass), S32);

    CCInfo.AllocateReg(Reg);
    Info.setWorkItemIDY(ArgDescriptor::createRegister(Reg));
  }

  if (Info.hasWorkItemIDZ()) {
    Register Reg = AMDGPU::VGPR2;
    MRI.setType(MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass), S32);

    CCInfo.AllocateReg(Reg);
    Info.setWorkItemIDZ(ArgDescriptor::createRegister(Reg));
  }
}

// Try to allocate a VGPR at the end of the argument list, or if no argument
// VGPRs are left allocating a stack slot.
// If \p Mask is is given it indicates bitfield position in the register.
// If \p Arg is given use it with new ]p Mask instead of allocating new.
static ArgDescriptor allocateVGPR32Input(CCState &CCInfo, unsigned Mask = ~0u,
                                         ArgDescriptor Arg = ArgDescriptor()) {
  if (Arg.isSet())
    return ArgDescriptor::createArg(Arg, Mask);

  ArrayRef<MCPhysReg> ArgVGPRs
    = makeArrayRef(AMDGPU::VGPR_32RegClass.begin(), 32);
  unsigned RegIdx = CCInfo.getFirstUnallocated(ArgVGPRs);
  if (RegIdx == ArgVGPRs.size()) {
    // Spill to stack required.
    int64_t Offset = CCInfo.AllocateStack(4, 4);

    return ArgDescriptor::createStack(Offset, Mask);
  }

  unsigned Reg = ArgVGPRs[RegIdx];
  Reg = CCInfo.AllocateReg(Reg);
  assert(Reg != AMDGPU::NoRegister);

  MachineFunction &MF = CCInfo.getMachineFunction();
  MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass);
  return ArgDescriptor::createRegister(Reg, Mask);
}

static ArgDescriptor allocateSGPR32InputImpl(CCState &CCInfo,
                                             const TargetRegisterClass *RC,
                                             unsigned NumArgRegs) {
  ArrayRef<MCPhysReg> ArgSGPRs = makeArrayRef(RC->begin(), 32);
  unsigned RegIdx = CCInfo.getFirstUnallocated(ArgSGPRs);
  if (RegIdx == ArgSGPRs.size())
    report_fatal_error("ran out of SGPRs for arguments");

  unsigned Reg = ArgSGPRs[RegIdx];
  Reg = CCInfo.AllocateReg(Reg);
  assert(Reg != AMDGPU::NoRegister);

  MachineFunction &MF = CCInfo.getMachineFunction();
  MF.addLiveIn(Reg, RC);
  return ArgDescriptor::createRegister(Reg);
}

static ArgDescriptor allocateSGPR32Input(CCState &CCInfo) {
  return allocateSGPR32InputImpl(CCInfo, &AMDGPU::SGPR_32RegClass, 32);
}

static ArgDescriptor allocateSGPR64Input(CCState &CCInfo) {
  return allocateSGPR32InputImpl(CCInfo, &AMDGPU::SGPR_64RegClass, 16);
}

void SITargetLowering::allocateSpecialInputVGPRs(CCState &CCInfo,
                                                 MachineFunction &MF,
                                                 const SIRegisterInfo &TRI,
                                                 SIMachineFunctionInfo &Info) const {
  const unsigned Mask = 0x3ff;
  ArgDescriptor Arg;

  if (Info.hasWorkItemIDX()) {
    Arg = allocateVGPR32Input(CCInfo, Mask);
    Info.setWorkItemIDX(Arg);
  }

  if (Info.hasWorkItemIDY()) {
    Arg = allocateVGPR32Input(CCInfo, Mask << 10, Arg);
    Info.setWorkItemIDY(Arg);
  }

  if (Info.hasWorkItemIDZ())
    Info.setWorkItemIDZ(allocateVGPR32Input(CCInfo, Mask << 20, Arg));
}

void SITargetLowering::allocateSpecialInputSGPRs(
  CCState &CCInfo,
  MachineFunction &MF,
  const SIRegisterInfo &TRI,
  SIMachineFunctionInfo &Info) const {
  auto &ArgInfo = Info.getArgInfo();

  // TODO: Unify handling with private memory pointers.

  if (Info.hasDispatchPtr())
    ArgInfo.DispatchPtr = allocateSGPR64Input(CCInfo);

  if (Info.hasQueuePtr())
    ArgInfo.QueuePtr = allocateSGPR64Input(CCInfo);

  if (Info.hasKernargSegmentPtr())
    ArgInfo.KernargSegmentPtr = allocateSGPR64Input(CCInfo);

  if (Info.hasDispatchID())
    ArgInfo.DispatchID = allocateSGPR64Input(CCInfo);

  // flat_scratch_init is not applicable for non-kernel functions.

  if (Info.hasWorkGroupIDX())
    ArgInfo.WorkGroupIDX = allocateSGPR32Input(CCInfo);

  if (Info.hasWorkGroupIDY())
    ArgInfo.WorkGroupIDY = allocateSGPR32Input(CCInfo);

  if (Info.hasWorkGroupIDZ())
    ArgInfo.WorkGroupIDZ = allocateSGPR32Input(CCInfo);

  if (Info.hasImplicitArgPtr())
    ArgInfo.ImplicitArgPtr = allocateSGPR64Input(CCInfo);
}

// Allocate special inputs passed in user SGPRs.
void SITargetLowering::allocateHSAUserSGPRs(CCState &CCInfo,
                                            MachineFunction &MF,
                                            const SIRegisterInfo &TRI,
                                            SIMachineFunctionInfo &Info) const {
  if (Info.hasImplicitBufferPtr()) {
    unsigned ImplicitBufferPtrReg = Info.addImplicitBufferPtr(TRI);
    MF.addLiveIn(ImplicitBufferPtrReg, &AMDGPU::SGPR_64RegClass);
    CCInfo.AllocateReg(ImplicitBufferPtrReg);
  }

  // FIXME: How should these inputs interact with inreg / custom SGPR inputs?
  if (Info.hasPrivateSegmentBuffer()) {
    unsigned PrivateSegmentBufferReg = Info.addPrivateSegmentBuffer(TRI);
    MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SGPR_128RegClass);
    CCInfo.AllocateReg(PrivateSegmentBufferReg);
  }

  if (Info.hasDispatchPtr()) {
    unsigned DispatchPtrReg = Info.addDispatchPtr(TRI);
    MF.addLiveIn(DispatchPtrReg, &AMDGPU::SGPR_64RegClass);
    CCInfo.AllocateReg(DispatchPtrReg);
  }

  if (Info.hasQueuePtr()) {
    unsigned QueuePtrReg = Info.addQueuePtr(TRI);
    MF.addLiveIn(QueuePtrReg, &AMDGPU::SGPR_64RegClass);
    CCInfo.AllocateReg(QueuePtrReg);
  }

  if (Info.hasKernargSegmentPtr()) {
    MachineRegisterInfo &MRI = MF.getRegInfo();
    Register InputPtrReg = Info.addKernargSegmentPtr(TRI);
    CCInfo.AllocateReg(InputPtrReg);

    Register VReg = MF.addLiveIn(InputPtrReg, &AMDGPU::SGPR_64RegClass);
    MRI.setType(VReg, LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64));
  }

  if (Info.hasDispatchID()) {
    unsigned DispatchIDReg = Info.addDispatchID(TRI);
    MF.addLiveIn(DispatchIDReg, &AMDGPU::SGPR_64RegClass);
    CCInfo.AllocateReg(DispatchIDReg);
  }

  if (Info.hasFlatScratchInit()) {
    unsigned FlatScratchInitReg = Info.addFlatScratchInit(TRI);
    MF.addLiveIn(FlatScratchInitReg, &AMDGPU::SGPR_64RegClass);
    CCInfo.AllocateReg(FlatScratchInitReg);
  }

  // TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read
  // these from the dispatch pointer.
}

// Allocate special input registers that are initialized per-wave.
void SITargetLowering::allocateSystemSGPRs(CCState &CCInfo,
                                           MachineFunction &MF,
                                           SIMachineFunctionInfo &Info,
                                           CallingConv::ID CallConv,
                                           bool IsShader) const {
  if (Info.hasWorkGroupIDX()) {
    unsigned Reg = Info.addWorkGroupIDX();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info.hasWorkGroupIDY()) {
    unsigned Reg = Info.addWorkGroupIDY();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info.hasWorkGroupIDZ()) {
    unsigned Reg = Info.addWorkGroupIDZ();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info.hasWorkGroupInfo()) {
    unsigned Reg = Info.addWorkGroupInfo();
    MF.addLiveIn(Reg, &AMDGPU::SReg_32_XM0RegClass);
    CCInfo.AllocateReg(Reg);
  }

  if (Info.hasPrivateSegmentWaveByteOffset()) {
    // Scratch wave offset passed in system SGPR.
    unsigned PrivateSegmentWaveByteOffsetReg;

    if (IsShader) {
      PrivateSegmentWaveByteOffsetReg =
        Info.getPrivateSegmentWaveByteOffsetSystemSGPR();

      // This is true if the scratch wave byte offset doesn't have a fixed
      // location.
      if (PrivateSegmentWaveByteOffsetReg == AMDGPU::NoRegister) {
        PrivateSegmentWaveByteOffsetReg = findFirstFreeSGPR(CCInfo);
        Info.setPrivateSegmentWaveByteOffset(PrivateSegmentWaveByteOffsetReg);
      }
    } else
      PrivateSegmentWaveByteOffsetReg = Info.addPrivateSegmentWaveByteOffset();

    MF.addLiveIn(PrivateSegmentWaveByteOffsetReg, &AMDGPU::SGPR_32RegClass);
    CCInfo.AllocateReg(PrivateSegmentWaveByteOffsetReg);
  }
}

static void reservePrivateMemoryRegs(const TargetMachine &TM,
                                     MachineFunction &MF,
                                     const SIRegisterInfo &TRI,
                                     SIMachineFunctionInfo &Info) {
  // Now that we've figured out where the scratch register inputs are, see if
  // should reserve the arguments and use them directly.
  MachineFrameInfo &MFI = MF.getFrameInfo();
  bool HasStackObjects = MFI.hasStackObjects();
  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();

  // Record that we know we have non-spill stack objects so we don't need to
  // check all stack objects later.
  if (HasStackObjects)
    Info.setHasNonSpillStackObjects(true);

  // Everything live out of a block is spilled with fast regalloc, so it's
  // almost certain that spilling will be required.
  if (TM.getOptLevel() == CodeGenOpt::None)
    HasStackObjects = true;

  // For now assume stack access is needed in any callee functions, so we need
  // the scratch registers to pass in.
  bool RequiresStackAccess = HasStackObjects || MFI.hasCalls();

  if (RequiresStackAccess && ST.isAmdHsaOrMesa(MF.getFunction())) {
    // If we have stack objects, we unquestionably need the private buffer
    // resource. For the Code Object V2 ABI, this will be the first 4 user
    // SGPR inputs. We can reserve those and use them directly.

    Register PrivateSegmentBufferReg =
        Info.getPreloadedReg(AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_BUFFER);
    Info.setScratchRSrcReg(PrivateSegmentBufferReg);
  } else {
    unsigned ReservedBufferReg = TRI.reservedPrivateSegmentBufferReg(MF);
    // We tentatively reserve the last registers (skipping the last registers
    // which may contain VCC, FLAT_SCR, and XNACK). After register allocation,
    // we'll replace these with the ones immediately after those which were
    // really allocated. In the prologue copies will be inserted from the
    // argument to these reserved registers.

    // Without HSA, relocations are used for the scratch pointer and the
    // buffer resource setup is always inserted in the prologue. Scratch wave
    // offset is still in an input SGPR.
    Info.setScratchRSrcReg(ReservedBufferReg);
  }

  // hasFP should be accurate for kernels even before the frame is finalized.
  if (ST.getFrameLowering()->hasFP(MF)) {
    MachineRegisterInfo &MRI = MF.getRegInfo();

    // Try to use s32 as the SP, but move it if it would interfere with input
    // arguments. This won't work with calls though.
    //
    // FIXME: Move SP to avoid any possible inputs, or find a way to spill input
    // registers.
    if (!MRI.isLiveIn(AMDGPU::SGPR32)) {
      Info.setStackPtrOffsetReg(AMDGPU::SGPR32);
    } else {
      assert(AMDGPU::isShader(MF.getFunction().getCallingConv()));

      if (MFI.hasCalls())
        report_fatal_error("call in graphics shader with too many input SGPRs");

      for (unsigned Reg : AMDGPU::SGPR_32RegClass) {
        if (!MRI.isLiveIn(Reg)) {
          Info.setStackPtrOffsetReg(Reg);
          break;
        }
      }

      if (Info.getStackPtrOffsetReg() == AMDGPU::SP_REG)
        report_fatal_error("failed to find register for SP");
    }

    if (MFI.hasCalls()) {
      Info.setScratchWaveOffsetReg(AMDGPU::SGPR33);
      Info.setFrameOffsetReg(AMDGPU::SGPR33);
    } else {
      unsigned ReservedOffsetReg =
        TRI.reservedPrivateSegmentWaveByteOffsetReg(MF);
      Info.setScratchWaveOffsetReg(ReservedOffsetReg);
      Info.setFrameOffsetReg(ReservedOffsetReg);
    }
  } else if (RequiresStackAccess) {
    assert(!MFI.hasCalls());
    // We know there are accesses and they will be done relative to SP, so just
    // pin it to the input.
    //
    // FIXME: Should not do this if inline asm is reading/writing these
    // registers.
    Register PreloadedSP = Info.getPreloadedReg(
        AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);

    Info.setStackPtrOffsetReg(PreloadedSP);
    Info.setScratchWaveOffsetReg(PreloadedSP);
    Info.setFrameOffsetReg(PreloadedSP);
  } else {
    assert(!MFI.hasCalls());

    // There may not be stack access at all. There may still be spills, or
    // access of a constant pointer (in which cases an extra copy will be
    // emitted in the prolog).
    unsigned ReservedOffsetReg
      = TRI.reservedPrivateSegmentWaveByteOffsetReg(MF);
    Info.setStackPtrOffsetReg(ReservedOffsetReg);
    Info.setScratchWaveOffsetReg(ReservedOffsetReg);
    Info.setFrameOffsetReg(ReservedOffsetReg);
  }
}

bool SITargetLowering::supportSplitCSR(MachineFunction *MF) const {
  const SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
  return !Info->isEntryFunction();
}

void SITargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const {

}

void SITargetLowering::insertCopiesSplitCSR(
  MachineBasicBlock *Entry,
  const SmallVectorImpl<MachineBasicBlock *> &Exits) const {
  const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();

  const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent());
  if (!IStart)
    return;

  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo();
  MachineBasicBlock::iterator MBBI = Entry->begin();
  for (const MCPhysReg *I = IStart; *I; ++I) {
    const TargetRegisterClass *RC = nullptr;
    if (AMDGPU::SReg_64RegClass.contains(*I))
      RC = &AMDGPU::SGPR_64RegClass;
    else if (AMDGPU::SReg_32RegClass.contains(*I))
      RC = &AMDGPU::SGPR_32RegClass;
    else
      llvm_unreachable("Unexpected register class in CSRsViaCopy!");

    Register NewVR = MRI->createVirtualRegister(RC);
    // Create copy from CSR to a virtual register.
    Entry->addLiveIn(*I);
    BuildMI(*Entry, MBBI, DebugLoc(), TII->get(TargetOpcode::COPY), NewVR)
      .addReg(*I);

    // Insert the copy-back instructions right before the terminator.
    for (auto *Exit : Exits)
      BuildMI(*Exit, Exit->getFirstTerminator(), DebugLoc(),
              TII->get(TargetOpcode::COPY), *I)
        .addReg(NewVR);
  }
}

SDValue SITargetLowering::LowerFormalArguments(
    SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
    const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
    SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
  const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();

  MachineFunction &MF = DAG.getMachineFunction();
  const Function &Fn = MF.getFunction();
  FunctionType *FType = MF.getFunction().getFunctionType();
  SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();

  if (Subtarget->isAmdHsaOS() && AMDGPU::isShader(CallConv)) {
    DiagnosticInfoUnsupported NoGraphicsHSA(
        Fn, "unsupported non-compute shaders with HSA", DL.getDebugLoc());
    DAG.getContext()->diagnose(NoGraphicsHSA);
    return DAG.getEntryNode();
  }

  SmallVector<ISD::InputArg, 16> Splits;
  SmallVector<CCValAssign, 16> ArgLocs;
  BitVector Skipped(Ins.size());
  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
                 *DAG.getContext());

  bool IsShader = AMDGPU::isShader(CallConv);
  bool IsKernel = AMDGPU::isKernel(CallConv);
  bool IsEntryFunc = AMDGPU::isEntryFunctionCC(CallConv);

  if (IsShader) {
    processShaderInputArgs(Splits, CallConv, Ins, Skipped, FType, Info);

    // At least one interpolation mode must be enabled or else the GPU will
    // hang.
    //
    // Check PSInputAddr instead of PSInputEnable. The idea is that if the user
    // set PSInputAddr, the user wants to enable some bits after the compilation
    // based on run-time states. Since we can't know what the final PSInputEna
    // will look like, so we shouldn't do anything here and the user should take
    // responsibility for the correct programming.
    //
    // Otherwise, the following restrictions apply:
    // - At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled.
    // - If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be
    //   enabled too.
    if (CallConv == CallingConv::AMDGPU_PS) {
      if ((Info->getPSInputAddr() & 0x7F) == 0 ||
           ((Info->getPSInputAddr() & 0xF) == 0 &&
            Info->isPSInputAllocated(11))) {
        CCInfo.AllocateReg(AMDGPU::VGPR0);
        CCInfo.AllocateReg(AMDGPU::VGPR1);
        Info->markPSInputAllocated(0);
        Info->markPSInputEnabled(0);
      }
      if (Subtarget->isAmdPalOS()) {
        // For isAmdPalOS, the user does not enable some bits after compilation
        // based on run-time states; the register values being generated here are
        // the final ones set in hardware. Therefore we need to apply the
        // workaround to PSInputAddr and PSInputEnable together.  (The case where
        // a bit is set in PSInputAddr but not PSInputEnable is where the
        // frontend set up an input arg for a particular interpolation mode, but
        // nothing uses that input arg. Really we should have an earlier pass
        // that removes such an arg.)
        unsigned PsInputBits = Info->getPSInputAddr() & Info->getPSInputEnable();
        if ((PsInputBits & 0x7F) == 0 ||
            ((PsInputBits & 0xF) == 0 &&
             (PsInputBits >> 11 & 1)))
          Info->markPSInputEnabled(
              countTrailingZeros(Info->getPSInputAddr(), ZB_Undefined));
      }
    }

    assert(!Info->hasDispatchPtr() &&
           !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() &&
           !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() &&
           !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() &&
           !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() &&
           !Info->hasWorkItemIDZ());
  } else if (IsKernel) {
    assert(Info->hasWorkGroupIDX() && Info->hasWorkItemIDX());
  } else {
    Splits.append(Ins.begin(), Ins.end());
  }

  if (IsEntryFunc) {
    allocateSpecialEntryInputVGPRs(CCInfo, MF, *TRI, *Info);
    allocateHSAUserSGPRs(CCInfo, MF, *TRI, *Info);
  }

  if (IsKernel) {
    analyzeFormalArgumentsCompute(CCInfo, Ins);
  } else {
    CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, isVarArg);
    CCInfo.AnalyzeFormalArguments(Splits, AssignFn);
  }

  SmallVector<SDValue, 16> Chains;

  // FIXME: This is the minimum kernel argument alignment. We should improve
  // this to the maximum alignment of the arguments.
  //
  // FIXME: Alignment of explicit arguments totally broken with non-0 explicit
  // kern arg offset.
  const unsigned KernelArgBaseAlign = 16;

   for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
    const ISD::InputArg &Arg = Ins[i];
    if (Arg.isOrigArg() && Skipped[Arg.getOrigArgIndex()]) {
      InVals.push_back(DAG.getUNDEF(Arg.VT));
      continue;
    }

    CCValAssign &VA = ArgLocs[ArgIdx++];
    MVT VT = VA.getLocVT();

    if (IsEntryFunc && VA.isMemLoc()) {
      VT = Ins[i].VT;
      EVT MemVT = VA.getLocVT();

      const uint64_t Offset = VA.getLocMemOffset();
      unsigned Align = MinAlign(KernelArgBaseAlign, Offset);

      SDValue Arg = lowerKernargMemParameter(
        DAG, VT, MemVT, DL, Chain, Offset, Align, Ins[i].Flags.isSExt(), &Ins[i]);
      Chains.push_back(Arg.getValue(1));

      auto *ParamTy =
        dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex()));
      if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS &&
          ParamTy && (ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
                      ParamTy->getAddressSpace() == AMDGPUAS::REGION_ADDRESS)) {
        // On SI local pointers are just offsets into LDS, so they are always
        // less than 16-bits.  On CI and newer they could potentially be
        // real pointers, so we can't guarantee their size.
        Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg,
                          DAG.getValueType(MVT::i16));
      }

      InVals.push_back(Arg);
      continue;
    } else if (!IsEntryFunc && VA.isMemLoc()) {
      SDValue Val = lowerStackParameter(DAG, VA, DL, Chain, Arg);
      InVals.push_back(Val);
      if (!Arg.Flags.isByVal())
        Chains.push_back(Val.getValue(1));
      continue;
    }

    assert(VA.isRegLoc() && "Parameter must be in a register!");

    Register Reg = VA.getLocReg();
    const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
    EVT ValVT = VA.getValVT();

    Reg = MF.addLiveIn(Reg, RC);
    SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);

    if (Arg.Flags.isSRet()) {
      // The return object should be reasonably addressable.

      // FIXME: This helps when the return is a real sret. If it is a
      // automatically inserted sret (i.e. CanLowerReturn returns false), an
      // extra copy is inserted in SelectionDAGBuilder which obscures this.
      unsigned NumBits
        = 32 - getSubtarget()->getKnownHighZeroBitsForFrameIndex();
      Val = DAG.getNode(ISD::AssertZext, DL, VT, Val,
        DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), NumBits)));
    }

    // If this is an 8 or 16-bit value, it is really passed promoted
    // to 32 bits. Insert an assert[sz]ext to capture this, then
    // truncate to the right size.
    switch (VA.getLocInfo()) {
    case CCValAssign::Full:
      break;
    case CCValAssign::BCvt:
      Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val);
      break;
    case CCValAssign::SExt:
      Val = DAG.getNode(ISD::AssertSext, DL, VT, Val,
                        DAG.getValueType(ValVT));
      Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
      break;
    case CCValAssign::ZExt:
      Val = DAG.getNode(ISD::AssertZext, DL, VT, Val,
                        DAG.getValueType(ValVT));
      Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
      break;
    case CCValAssign::AExt:
      Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
      break;
    default:
      llvm_unreachable("Unknown loc info!");
    }

    InVals.push_back(Val);
  }

  if (!IsEntryFunc) {
    // Special inputs come after user arguments.
    allocateSpecialInputVGPRs(CCInfo, MF, *TRI, *Info);
  }

  // Start adding system SGPRs.
  if (IsEntryFunc) {
    allocateSystemSGPRs(CCInfo, MF, *Info, CallConv, IsShader);
  } else {
    CCInfo.AllocateReg(Info->getScratchRSrcReg());
    CCInfo.AllocateReg(Info->getScratchWaveOffsetReg());
    CCInfo.AllocateReg(Info->getFrameOffsetReg());
    allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
  }

  auto &ArgUsageInfo =
    DAG.getPass()->getAnalysis<AMDGPUArgumentUsageInfo>();
  ArgUsageInfo.setFuncArgInfo(Fn, Info->getArgInfo());

  unsigned StackArgSize = CCInfo.getNextStackOffset();
  Info->setBytesInStackArgArea(StackArgSize);

  return Chains.empty() ? Chain :
    DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
}

// TODO: If return values can't fit in registers, we should return as many as
// possible in registers before passing on stack.
bool SITargetLowering::CanLowerReturn(
  CallingConv::ID CallConv,
  MachineFunction &MF, bool IsVarArg,
  const SmallVectorImpl<ISD::OutputArg> &Outs,
  LLVMContext &Context) const {
  // Replacing returns with sret/stack usage doesn't make sense for shaders.
  // FIXME: Also sort of a workaround for custom vector splitting in LowerReturn
  // for shaders. Vector types should be explicitly handled by CC.
  if (AMDGPU::isEntryFunctionCC(CallConv))
    return true;

  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
  return CCInfo.CheckReturn(Outs, CCAssignFnForReturn(CallConv, IsVarArg));
}

SDValue
SITargetLowering::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();
  SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();

  if (AMDGPU::isKernel(CallConv)) {
    return AMDGPUTargetLowering::LowerReturn(Chain, CallConv, isVarArg, Outs,
                                             OutVals, DL, DAG);
  }

  bool IsShader = AMDGPU::isShader(CallConv);

  Info->setIfReturnsVoid(Outs.empty());
  bool IsWaveEnd = Info->returnsVoid() && IsShader;

  // CCValAssign - represent the assignment of the return value to a location.
  SmallVector<CCValAssign, 48> RVLocs;
  SmallVector<ISD::OutputArg, 48> Splits;

  // CCState - Info about the registers and stack slots.
  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
                 *DAG.getContext());

  // Analyze outgoing return values.
  CCInfo.AnalyzeReturn(Outs, CCAssignFnForReturn(CallConv, isVarArg));

  SDValue Flag;
  SmallVector<SDValue, 48> RetOps;
  RetOps.push_back(Chain); // Operand #0 = Chain (updated below)

  // Add return address for callable functions.
  if (!Info->isEntryFunction()) {
    const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
    SDValue ReturnAddrReg = CreateLiveInRegister(
      DAG, &AMDGPU::SReg_64RegClass, TRI->getReturnAddressReg(MF), MVT::i64);

    SDValue ReturnAddrVirtualReg = DAG.getRegister(
        MF.getRegInfo().createVirtualRegister(&AMDGPU::CCR_SGPR_64RegClass),
        MVT::i64);
    Chain =
        DAG.getCopyToReg(Chain, DL, ReturnAddrVirtualReg, ReturnAddrReg, Flag);
    Flag = Chain.getValue(1);
    RetOps.push_back(ReturnAddrVirtualReg);
  }

  // Copy the result values into the output registers.
  for (unsigned I = 0, RealRVLocIdx = 0, E = RVLocs.size(); I != E;
       ++I, ++RealRVLocIdx) {
    CCValAssign &VA = RVLocs[I];
    assert(VA.isRegLoc() && "Can only return in registers!");
    // TODO: Partially return in registers if return values don't fit.
    SDValue Arg = OutVals[RealRVLocIdx];

    // Copied from other backends.
    switch (VA.getLocInfo()) {
    case CCValAssign::Full:
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::SExt:
      Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::ZExt:
      Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::AExt:
      Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    default:
      llvm_unreachable("Unknown loc info!");
    }

    Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
    Flag = Chain.getValue(1);
    RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
  }

  // FIXME: Does sret work properly?
  if (!Info->isEntryFunction()) {
    const SIRegisterInfo *TRI = Subtarget->getRegisterInfo();
    const MCPhysReg *I =
      TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction());
    if (I) {
      for (; *I; ++I) {
        if (AMDGPU::SReg_64RegClass.contains(*I))
          RetOps.push_back(DAG.getRegister(*I, MVT::i64));
        else if (AMDGPU::SReg_32RegClass.contains(*I))
          RetOps.push_back(DAG.getRegister(*I, MVT::i32));
        else
          llvm_unreachable("Unexpected register class in CSRsViaCopy!");
      }
    }
  }

  // Update chain and glue.
  RetOps[0] = Chain;
  if (Flag.getNode())
    RetOps.push_back(Flag);

  unsigned Opc = AMDGPUISD::ENDPGM;
  if (!IsWaveEnd)
    Opc = IsShader ? AMDGPUISD::RETURN_TO_EPILOG : AMDGPUISD::RET_FLAG;
  return DAG.getNode(Opc, DL, MVT::Other, RetOps);
}

SDValue SITargetLowering::LowerCallResult(
    SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
    const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
    SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, bool IsThisReturn,
    SDValue ThisVal) const {
  CCAssignFn *RetCC = CCAssignFnForReturn(CallConv, IsVarArg);

  // Assign locations to each value returned by this call.
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
                 *DAG.getContext());
  CCInfo.AnalyzeCallResult(Ins, RetCC);

  // Copy all of the result registers out of their specified physreg.
  for (unsigned i = 0; i != RVLocs.size(); ++i) {
    CCValAssign VA = RVLocs[i];
    SDValue Val;

    if (VA.isRegLoc()) {
      Val = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
      Chain = Val.getValue(1);
      InFlag = Val.getValue(2);
    } else if (VA.isMemLoc()) {
      report_fatal_error("TODO: return values in memory");
    } else
      llvm_unreachable("unknown argument location type");

    switch (VA.getLocInfo()) {
    case CCValAssign::Full:
      break;
    case CCValAssign::BCvt:
      Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
      break;
    case CCValAssign::ZExt:
      Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val,
                        DAG.getValueType(VA.getValVT()));
      Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
      break;
    case CCValAssign::SExt:
      Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val,
                        DAG.getValueType(VA.getValVT()));
      Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
      break;
    case CCValAssign::AExt:
      Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
      break;
    default:
      llvm_unreachable("Unknown loc info!");
    }

    InVals.push_back(Val);
  }

  return Chain;
}

// Add code to pass special inputs required depending on used features separate
// from the explicit user arguments present in the IR.
void SITargetLowering::passSpecialInputs(
    CallLoweringInfo &CLI,
    CCState &CCInfo,
    const SIMachineFunctionInfo &Info,
    SmallVectorImpl<std::pair<unsigned, SDValue>> &RegsToPass,
    SmallVectorImpl<SDValue> &MemOpChains,
    SDValue Chain) const {
  // If we don't have a call site, this was a call inserted by
  // legalization. These can never use special inputs.
  if (!CLI.CS)
    return;

  const Function *CalleeFunc = CLI.CS.getCalledFunction();
  assert(CalleeFunc);

  SelectionDAG &DAG = CLI.DAG;
  const SDLoc &DL = CLI.DL;

  const SIRegisterInfo *TRI = Subtarget->getRegisterInfo();

  auto &ArgUsageInfo =
    DAG.getPass()->getAnalysis<AMDGPUArgumentUsageInfo>();
  const AMDGPUFunctionArgInfo &CalleeArgInfo
    = ArgUsageInfo.lookupFuncArgInfo(*CalleeFunc);

  const AMDGPUFunctionArgInfo &CallerArgInfo = Info.getArgInfo();

  // TODO: Unify with private memory register handling. This is complicated by
  // the fact that at least in kernels, the input argument is not necessarily
  // in the same location as the input.
  AMDGPUFunctionArgInfo::PreloadedValue InputRegs[] = {
    AMDGPUFunctionArgInfo::DISPATCH_PTR,
    AMDGPUFunctionArgInfo::QUEUE_PTR,
    AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR,
    AMDGPUFunctionArgInfo::DISPATCH_ID,
    AMDGPUFunctionArgInfo::WORKGROUP_ID_X,
    AMDGPUFunctionArgInfo::WORKGROUP_ID_Y,
    AMDGPUFunctionArgInfo::WORKGROUP_ID_Z,
    AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR
  };

  for (auto InputID : InputRegs) {
    const ArgDescriptor *OutgoingArg;
    const TargetRegisterClass *ArgRC;

    std::tie(OutgoingArg, ArgRC) = CalleeArgInfo.getPreloadedValue(InputID);
    if (!OutgoingArg)
      continue;

    const ArgDescriptor *IncomingArg;
    const TargetRegisterClass *IncomingArgRC;
    std::tie(IncomingArg, IncomingArgRC)
      = CallerArgInfo.getPreloadedValue(InputID);
    assert(IncomingArgRC == ArgRC);

    // All special arguments are ints for now.
    EVT ArgVT = TRI->getSpillSize(*ArgRC) == 8 ? MVT::i64 : MVT::i32;
    SDValue InputReg;

    if (IncomingArg) {
      InputReg = loadInputValue(DAG, ArgRC, ArgVT, DL, *IncomingArg);
    } else {
      // The implicit arg ptr is special because it doesn't have a corresponding
      // input for kernels, and is computed from the kernarg segment pointer.
      assert(InputID == AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR);
      InputReg = getImplicitArgPtr(DAG, DL);
    }

    if (OutgoingArg->isRegister()) {
      RegsToPass.emplace_back(OutgoingArg->getRegister(), InputReg);
    } else {
      unsigned SpecialArgOffset = CCInfo.AllocateStack(ArgVT.getStoreSize(), 4);
      SDValue ArgStore = storeStackInputValue(DAG, DL, Chain, InputReg,
                                              SpecialArgOffset);
      MemOpChains.push_back(ArgStore);
    }
  }

  // Pack workitem IDs into a single register or pass it as is if already
  // packed.
  const ArgDescriptor *OutgoingArg;
  const TargetRegisterClass *ArgRC;

  std::tie(OutgoingArg, ArgRC) =
    CalleeArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_X);
  if (!OutgoingArg)
    std::tie(OutgoingArg, ArgRC) =
      CalleeArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Y);
  if (!OutgoingArg)
    std::tie(OutgoingArg, ArgRC) =
      CalleeArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Z);
  if (!OutgoingArg)
    return;

  const ArgDescriptor *IncomingArgX
    = CallerArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_X).first;
  const ArgDescriptor *IncomingArgY
    = CallerArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Y).first;
  const ArgDescriptor *IncomingArgZ
    = CallerArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Z).first;

  SDValue InputReg;
  SDLoc SL;

  // If incoming ids are not packed we need to pack them.
  if (IncomingArgX && !IncomingArgX->isMasked() && CalleeArgInfo.WorkItemIDX)
    InputReg = loadInputValue(DAG, ArgRC, MVT::i32, DL, *IncomingArgX);

  if (IncomingArgY && !IncomingArgY->isMasked() && CalleeArgInfo.WorkItemIDY) {
    SDValue Y = loadInputValue(DAG, ArgRC, MVT::i32, DL, *IncomingArgY);
    Y = DAG.getNode(ISD::SHL, SL, MVT::i32, Y,
                    DAG.getShiftAmountConstant(10, MVT::i32, SL));
    InputReg = InputReg.getNode() ?
                 DAG.getNode(ISD::OR, SL, MVT::i32, InputReg, Y) : Y;
  }

  if (IncomingArgZ && !IncomingArgZ->isMasked() && CalleeArgInfo.WorkItemIDZ) {
    SDValue Z = loadInputValue(DAG, ArgRC, MVT::i32, DL, *IncomingArgZ);
    Z = DAG.getNode(ISD::SHL, SL, MVT::i32, Z,
                    DAG.getShiftAmountConstant(20, MVT::i32, SL));
    InputReg = InputReg.getNode() ?
                 DAG.getNode(ISD::OR, SL, MVT::i32, InputReg, Z) : Z;
  }

  if (!InputReg.getNode()) {
    // Workitem ids are already packed, any of present incoming arguments
    // will carry all required fields.
    ArgDescriptor IncomingArg = ArgDescriptor::createArg(
      IncomingArgX ? *IncomingArgX :
      IncomingArgY ? *IncomingArgY :
                     *IncomingArgZ, ~0u);
    InputReg = loadInputValue(DAG, ArgRC, MVT::i32, DL, IncomingArg);
  }

  if (OutgoingArg->isRegister()) {
    RegsToPass.emplace_back(OutgoingArg->getRegister(), InputReg);
  } else {
    unsigned SpecialArgOffset = CCInfo.AllocateStack(4, 4);
    SDValue ArgStore = storeStackInputValue(DAG, DL, Chain, InputReg,
                                            SpecialArgOffset);
    MemOpChains.push_back(ArgStore);
  }
}

static bool canGuaranteeTCO(CallingConv::ID CC) {
  return CC == CallingConv::Fast;
}

/// Return true if we might ever do TCO for calls with this calling convention.
static bool mayTailCallThisCC(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::C:
    return true;
  default:
    return canGuaranteeTCO(CC);
  }
}

bool SITargetLowering::isEligibleForTailCallOptimization(
    SDValue Callee, CallingConv::ID CalleeCC, bool IsVarArg,
    const SmallVectorImpl<ISD::OutputArg> &Outs,
    const SmallVectorImpl<SDValue> &OutVals,
    const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
  if (!mayTailCallThisCC(CalleeCC))
    return false;

  MachineFunction &MF = DAG.getMachineFunction();
  const Function &CallerF = MF.getFunction();
  CallingConv::ID CallerCC = CallerF.getCallingConv();
  const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
  const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);

  // Kernels aren't callable, and don't have a live in return address so it
  // doesn't make sense to do a tail call with entry functions.
  if (!CallerPreserved)
    return false;

  bool CCMatch = CallerCC == CalleeCC;

  if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
    if (canGuaranteeTCO(CalleeCC) && CCMatch)
      return true;
    return false;
  }

  // TODO: Can we handle var args?
  if (IsVarArg)
    return false;

  for (const Argument &Arg : CallerF.args()) {
    if (Arg.hasByValAttr())
      return false;
  }

  LLVMContext &Ctx = *DAG.getContext();

  // Check that the call results are passed in the same way.
  if (!CCState::resultsCompatible(CalleeCC, CallerCC, MF, Ctx, Ins,
                                  CCAssignFnForCall(CalleeCC, IsVarArg),
                                  CCAssignFnForCall(CallerCC, IsVarArg)))
    return false;

  // The callee has to preserve all registers the caller needs to preserve.
  if (!CCMatch) {
    const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC);
    if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved))
      return false;
  }

  // Nothing more to check if the callee is taking no arguments.
  if (Outs.empty())
    return true;

  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CalleeCC, IsVarArg, MF, ArgLocs, Ctx);

  CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, IsVarArg));

  const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
  // If the stack arguments for this call do not fit into our own save area then
  // the call cannot be made tail.
  // TODO: Is this really necessary?
  if (CCInfo.getNextStackOffset() > FuncInfo->getBytesInStackArgArea())
    return false;

  const MachineRegisterInfo &MRI = MF.getRegInfo();
  return parametersInCSRMatch(MRI, CallerPreserved, ArgLocs, OutVals);
}

bool SITargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
  if (!CI->isTailCall())
    return false;

  const Function *ParentFn = CI->getParent()->getParent();
  if (AMDGPU::isEntryFunctionCC(ParentFn->getCallingConv()))
    return false;

  auto Attr = ParentFn->getFnAttribute("disable-tail-calls");
  return (Attr.getValueAsString() != "true");
}

// The wave scratch offset register is used as the global base pointer.
SDValue SITargetLowering::LowerCall(CallLoweringInfo &CLI,
                                    SmallVectorImpl<SDValue> &InVals) const {
  SelectionDAG &DAG = CLI.DAG;
  const SDLoc &DL = CLI.DL;
  SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
  SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
  SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
  SDValue Chain = CLI.Chain;
  SDValue Callee = CLI.Callee;
  bool &IsTailCall = CLI.IsTailCall;
  CallingConv::ID CallConv = CLI.CallConv;
  bool IsVarArg = CLI.IsVarArg;
  bool IsSibCall = false;
  bool IsThisReturn = false;
  MachineFunction &MF = DAG.getMachineFunction();

  if (IsVarArg) {
    return lowerUnhandledCall(CLI, InVals,
                              "unsupported call to variadic function ");
  }

  if (!CLI.CS.getInstruction())
    report_fatal_error("unsupported libcall legalization");

  if (!CLI.CS.getCalledFunction()) {
    return lowerUnhandledCall(CLI, InVals,
                              "unsupported indirect call to function ");
  }

  if (IsTailCall && MF.getTarget().Options.GuaranteedTailCallOpt) {
    return lowerUnhandledCall(CLI, InVals,
                              "unsupported required tail call to function ");
  }

  if (AMDGPU::isShader(MF.getFunction().getCallingConv())) {
    // Note the issue is with the CC of the calling function, not of the call
    // itself.
    return lowerUnhandledCall(CLI, InVals,
                          "unsupported call from graphics shader of function ");
  }

  if (IsTailCall) {
    IsTailCall = isEligibleForTailCallOptimization(
      Callee, CallConv, IsVarArg, Outs, OutVals, Ins, DAG);
    if (!IsTailCall && CLI.CS && CLI.CS.isMustTailCall()) {
      report_fatal_error("failed to perform tail call elimination on a call "
                         "site marked musttail");
    }

    bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;

    // A sibling call is one where we're under the usual C ABI and not planning
    // to change that but can still do a tail call:
    if (!TailCallOpt && IsTailCall)
      IsSibCall = true;

    if (IsTailCall)
      ++NumTailCalls;
  }

  const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();

  // Analyze operands of the call, assigning locations to each operand.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
  CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, IsVarArg);

  CCInfo.AnalyzeCallOperands(Outs, AssignFn);

  // Get a count of how many bytes are to be pushed on the stack.
  unsigned NumBytes = CCInfo.getNextStackOffset();

  if (IsSibCall) {
    // Since we're not changing the ABI to make this a tail call, the memory
    // operands are already available in the caller's incoming argument space.
    NumBytes = 0;
  }

  // FPDiff is the byte offset of the call's argument area from the callee's.
  // Stores to callee stack arguments will be placed in FixedStackSlots offset
  // by this amount for a tail call. In a sibling call it must be 0 because the
  // caller will deallocate the entire stack and the callee still expects its
  // arguments to begin at SP+0. Completely unused for non-tail calls.
  int32_t FPDiff = 0;
  MachineFrameInfo &MFI = MF.getFrameInfo();
  SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;

  // Adjust the stack pointer for the new arguments...
  // These operations are automatically eliminated by the prolog/epilog pass
  if (!IsSibCall) {
    Chain = DAG.getCALLSEQ_START(Chain, 0, 0, DL);

    SmallVector<SDValue, 4> CopyFromChains;

    // In the HSA case, this should be an identity copy.
    SDValue ScratchRSrcReg
      = DAG.getCopyFromReg(Chain, DL, Info->getScratchRSrcReg(), MVT::v4i32);
    RegsToPass.emplace_back(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, ScratchRSrcReg);
    CopyFromChains.push_back(ScratchRSrcReg.getValue(1));
    Chain = DAG.getTokenFactor(DL, CopyFromChains);
  }

  SmallVector<SDValue, 8> MemOpChains;
  MVT PtrVT = MVT::i32;

  // Walk the register/memloc assignments, inserting copies/loads.
  for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
       ++i, ++realArgIdx) {
    CCValAssign &VA = ArgLocs[i];
    SDValue Arg = OutVals[realArgIdx];

    // Promote the value if needed.
    switch (VA.getLocInfo()) {
    case CCValAssign::Full:
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::ZExt:
      Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::SExt:
      Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::AExt:
      Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::FPExt:
      Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    default:
      llvm_unreachable("Unknown loc info!");
    }

    if (VA.isRegLoc()) {
      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
    } else {
      assert(VA.isMemLoc());

      SDValue DstAddr;
      MachinePointerInfo DstInfo;

      unsigned LocMemOffset = VA.getLocMemOffset();
      int32_t Offset = LocMemOffset;

      SDValue PtrOff = DAG.getConstant(Offset, DL, PtrVT);
      unsigned Align = 0;

      if (IsTailCall) {
        ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
        unsigned OpSize = Flags.isByVal() ?
          Flags.getByValSize() : VA.getValVT().getStoreSize();

        // FIXME: We can have better than the minimum byval required alignment.
        Align = Flags.isByVal() ? Flags.getByValAlign() :
          MinAlign(Subtarget->getStackAlignment(), Offset);

        Offset = Offset + FPDiff;
        int FI = MFI.CreateFixedObject(OpSize, Offset, true);

        DstAddr = DAG.getFrameIndex(FI, PtrVT);
        DstInfo = MachinePointerInfo::getFixedStack(MF, FI);

        // Make sure any stack arguments overlapping with where we're storing
        // are loaded before this eventual operation. Otherwise they'll be
        // clobbered.

        // FIXME: Why is this really necessary? This seems to just result in a
        // lot of code to copy the stack and write them back to the same
        // locations, which are supposed to be immutable?
        Chain = addTokenForArgument(Chain, DAG, MFI, FI);
      } else {
        DstAddr = PtrOff;
        DstInfo = MachinePointerInfo::getStack(MF, LocMemOffset);
        Align = MinAlign(Subtarget->getStackAlignment(), LocMemOffset);
      }

      if (Outs[i].Flags.isByVal()) {
        SDValue SizeNode =
            DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i32);
        SDValue Cpy = DAG.getMemcpy(
            Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
            /*isVol = */ false, /*AlwaysInline = */ true,
            /*isTailCall = */ false, DstInfo,
            MachinePointerInfo(UndefValue::get(Type::getInt8PtrTy(
                *DAG.getContext(), AMDGPUAS::PRIVATE_ADDRESS))));

        MemOpChains.push_back(Cpy);
      } else {
        SDValue Store = DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, Align);
        MemOpChains.push_back(Store);
      }
    }
  }

  // Copy special input registers after user input arguments.
  passSpecialInputs(CLI, CCInfo, *Info, RegsToPass, MemOpChains, Chain);

  if (!MemOpChains.empty())
    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);

  // Build a sequence of copy-to-reg nodes chained together with token chain
  // and flag operands which copy the outgoing args into the appropriate regs.
  SDValue InFlag;
  for (auto &RegToPass : RegsToPass) {
    Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
                             RegToPass.second, InFlag);
    InFlag = Chain.getValue(1);
  }


  SDValue PhysReturnAddrReg;
  if (IsTailCall) {
    // Since the return is being combined with the call, we need to pass on the
    // return address.

    const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
    SDValue ReturnAddrReg = CreateLiveInRegister(
      DAG, &AMDGPU::SReg_64RegClass, TRI->getReturnAddressReg(MF), MVT::i64);

    PhysReturnAddrReg = DAG.getRegister(TRI->getReturnAddressReg(MF),
                                        MVT::i64);
    Chain = DAG.getCopyToReg(Chain, DL, PhysReturnAddrReg, ReturnAddrReg, InFlag);
    InFlag = Chain.getValue(1);
  }

  // We don't usually want to end the call-sequence here because we would tidy
  // the frame up *after* the call, however in the ABI-changing tail-call case
  // we've carefully laid out the parameters so that when sp is reset they'll be
  // in the correct location.
  if (IsTailCall && !IsSibCall) {
    Chain = DAG.getCALLSEQ_END(Chain,
                               DAG.getTargetConstant(NumBytes, DL, MVT::i32),
                               DAG.getTargetConstant(0, DL, MVT::i32),
                               InFlag, DL);
    InFlag = Chain.getValue(1);
  }

  std::vector<SDValue> Ops;
  Ops.push_back(Chain);
  Ops.push_back(Callee);
  // Add a redundant copy of the callee global which will not be legalized, as
  // we need direct access to the callee later.
  GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Callee);
  const GlobalValue *GV = GSD->getGlobal();
  Ops.push_back(DAG.getTargetGlobalAddress(GV, DL, MVT::i64));

  if (IsTailCall) {
    // Each tail call may have to adjust the stack by a different amount, so
    // this information must travel along with the operation for eventual
    // consumption by emitEpilogue.
    Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));

    Ops.push_back(PhysReturnAddrReg);
  }

  // Add argument registers to the end of the list so that they are known live
  // into the call.
  for (auto &RegToPass : RegsToPass) {
    Ops.push_back(DAG.getRegister(RegToPass.first,
                                  RegToPass.second.getValueType()));
  }

  // Add a register mask operand representing the call-preserved registers.

  auto *TRI = static_cast<const SIRegisterInfo*>(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));

  if (InFlag.getNode())
    Ops.push_back(InFlag);

  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);

  // If we're doing a tall call, use a TC_RETURN here rather than an
  // actual call instruction.
  if (IsTailCall) {
    MFI.setHasTailCall();
    return DAG.getNode(AMDGPUISD::TC_RETURN, DL, NodeTys, Ops);
  }

  // Returns a chain and a flag for retval copy to use.
  SDValue Call = DAG.getNode(AMDGPUISD::CALL, DL, NodeTys, Ops);
  Chain = Call.getValue(0);
  InFlag = Call.getValue(1);

  uint64_t CalleePopBytes = NumBytes;
  Chain = DAG.getCALLSEQ_END(Chain, DAG.getTargetConstant(0, DL, MVT::i32),
                             DAG.getTargetConstant(CalleePopBytes, DL, MVT::i32),
                             InFlag, DL);
  if (!Ins.empty())
    InFlag = Chain.getValue(1);

  // Handle result values, copying them out of physregs into vregs that we
  // return.
  return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
                         InVals, IsThisReturn,
                         IsThisReturn ? OutVals[0] : SDValue());
}

unsigned SITargetLowering::getRegisterByName(const char* RegName, EVT VT,
                                             SelectionDAG &DAG) const {
  unsigned Reg = StringSwitch<unsigned>(RegName)
    .Case("m0", AMDGPU::M0)
    .Case("exec", AMDGPU::EXEC)
    .Case("exec_lo", AMDGPU::EXEC_LO)
    .Case("exec_hi", AMDGPU::EXEC_HI)
    .Case("flat_scratch", AMDGPU::FLAT_SCR)
    .Case("flat_scratch_lo", AMDGPU::FLAT_SCR_LO)
    .Case("flat_scratch_hi", AMDGPU::FLAT_SCR_HI)
    .Default(AMDGPU::NoRegister);

  if (Reg == AMDGPU::NoRegister) {
    report_fatal_error(Twine("invalid register name \""
                             + StringRef(RegName)  + "\"."));

  }

  if (!Subtarget->hasFlatScrRegister() &&
       Subtarget->getRegisterInfo()->regsOverlap(Reg, AMDGPU::FLAT_SCR)) {
    report_fatal_error(Twine("invalid register \""
                             + StringRef(RegName)  + "\" for subtarget."));
  }

  switch (Reg) {
  case AMDGPU::M0:
  case AMDGPU::EXEC_LO:
  case AMDGPU::EXEC_HI:
  case AMDGPU::FLAT_SCR_LO:
  case AMDGPU::FLAT_SCR_HI:
    if (VT.getSizeInBits() == 32)
      return Reg;
    break;
  case AMDGPU::EXEC:
  case AMDGPU::FLAT_SCR:
    if (VT.getSizeInBits() == 64)
      return Reg;
    break;
  default:
    llvm_unreachable("missing register type checking");
  }

  report_fatal_error(Twine("invalid type for register \""
                           + StringRef(RegName) + "\"."));
}

// If kill is not the last instruction, split the block so kill is always a
// proper terminator.
MachineBasicBlock *SITargetLowering::splitKillBlock(MachineInstr &MI,
                                                    MachineBasicBlock *BB) const {
  const SIInstrInfo *TII = getSubtarget()->getInstrInfo();

  MachineBasicBlock::iterator SplitPoint(&MI);
  ++SplitPoint;

  if (SplitPoint == BB->end()) {
    // Don't bother with a new block.
    MI.setDesc(TII->getKillTerminatorFromPseudo(MI.getOpcode()));
    return BB;
  }

  MachineFunction *MF = BB->getParent();
  MachineBasicBlock *SplitBB
    = MF->CreateMachineBasicBlock(BB->getBasicBlock());

  MF->insert(++MachineFunction::iterator(BB), SplitBB);
  SplitBB->splice(SplitBB->begin(), BB, SplitPoint, BB->end());

  SplitBB->transferSuccessorsAndUpdatePHIs(BB);
  BB->addSuccessor(SplitBB);

  MI.setDesc(TII->getKillTerminatorFromPseudo(MI.getOpcode()));
  return SplitBB;
}

// Split block \p MBB at \p MI, as to insert a loop. If \p InstInLoop is true,
// \p MI will be the only instruction in the loop body block. Otherwise, it will
// be the first instruction in the remainder block.
//
/// \returns { LoopBody, Remainder }
static std::pair<MachineBasicBlock *, MachineBasicBlock *>
splitBlockForLoop(MachineInstr &MI, MachineBasicBlock &MBB, bool InstInLoop) {
  MachineFunction *MF = MBB.getParent();
  MachineBasicBlock::iterator I(&MI);

  // To insert the loop we need to split the block. Move everything after this
  // point to a new block, and insert a new empty block between the two.
  MachineBasicBlock *LoopBB = MF->CreateMachineBasicBlock();
  MachineBasicBlock *RemainderBB = MF->CreateMachineBasicBlock();
  MachineFunction::iterator MBBI(MBB);
  ++MBBI;

  MF->insert(MBBI, LoopBB);
  MF->insert(MBBI, RemainderBB);

  LoopBB->addSuccessor(LoopBB);
  LoopBB->addSuccessor(RemainderBB);

  // Move the rest of the block into a new block.
  RemainderBB->transferSuccessorsAndUpdatePHIs(&MBB);

  if (InstInLoop) {
    auto Next = std::next(I);

    // Move instruction to loop body.
    LoopBB->splice(LoopBB->begin(), &MBB, I, Next);

    // Move the rest of the block.
    RemainderBB->splice(RemainderBB->begin(), &MBB, Next, MBB.end());
  } else {
    RemainderBB->splice(RemainderBB->begin(), &MBB, I, MBB.end());
  }

  MBB.addSuccessor(LoopBB);

  return std::make_pair(LoopBB, RemainderBB);
}

/// Insert \p MI into a BUNDLE with an S_WAITCNT 0 immediately following it.
void SITargetLowering::bundleInstWithWaitcnt(MachineInstr &MI) const {
  MachineBasicBlock *MBB = MI.getParent();
  const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
  auto I = MI.getIterator();
  auto E = std::next(I);

  BuildMI(*MBB, E, MI.getDebugLoc(), TII->get(AMDGPU::S_WAITCNT))
    .addImm(0);

  MIBundleBuilder Bundler(*MBB, I, E);
  finalizeBundle(*MBB, Bundler.begin());
}

MachineBasicBlock *
SITargetLowering::emitGWSMemViolTestLoop(MachineInstr &MI,
                                         MachineBasicBlock *BB) const {
  const DebugLoc &DL = MI.getDebugLoc();

  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();

  MachineBasicBlock *LoopBB;
  MachineBasicBlock *RemainderBB;
  const SIInstrInfo *TII = getSubtarget()->getInstrInfo();

  // Apparently kill flags are only valid if the def is in the same block?
  if (MachineOperand *Src = TII->getNamedOperand(MI, AMDGPU::OpName::data0))
    Src->setIsKill(false);

  std::tie(LoopBB, RemainderBB) = splitBlockForLoop(MI, *BB, true);

  MachineBasicBlock::iterator I = LoopBB->end();

  const unsigned EncodedReg = AMDGPU::Hwreg::encodeHwreg(
    AMDGPU::Hwreg::ID_TRAPSTS, AMDGPU::Hwreg::OFFSET_MEM_VIOL, 1);

  // Clear TRAP_STS.MEM_VIOL
  BuildMI(*LoopBB, LoopBB->begin(), DL, TII->get(AMDGPU::S_SETREG_IMM32_B32))
    .addImm(0)
    .addImm(EncodedReg);

  bundleInstWithWaitcnt(MI);

  Register Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);

  // Load and check TRAP_STS.MEM_VIOL
  BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::S_GETREG_B32), Reg)
    .addImm(EncodedReg);

  // FIXME: Do we need to use an isel pseudo that may clobber scc?
  BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::S_CMP_LG_U32))
    .addReg(Reg, RegState::Kill)
    .addImm(0);
  BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::S_CBRANCH_SCC1))
    .addMBB(LoopBB);

  return RemainderBB;
}

// Do a v_movrels_b32 or v_movreld_b32 for each unique value of \p IdxReg in the
// wavefront. If the value is uniform and just happens to be in a VGPR, this
// will only do one iteration. In the worst case, this will loop 64 times.
//
// TODO: Just use v_readlane_b32 if we know the VGPR has a uniform value.
static MachineBasicBlock::iterator emitLoadM0FromVGPRLoop(
  const SIInstrInfo *TII,
  MachineRegisterInfo &MRI,
  MachineBasicBlock &OrigBB,
  MachineBasicBlock &LoopBB,
  const DebugLoc &DL,
  const MachineOperand &IdxReg,
  unsigned InitReg,
  unsigned ResultReg,
  unsigned PhiReg,
  unsigned InitSaveExecReg,
  int Offset,
  bool UseGPRIdxMode,
  bool IsIndirectSrc) {
  MachineFunction *MF = OrigBB.getParent();
  const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
  const SIRegisterInfo *TRI = ST.getRegisterInfo();
  MachineBasicBlock::iterator I = LoopBB.begin();

  const TargetRegisterClass *BoolRC = TRI->getBoolRC();
  Register PhiExec = MRI.createVirtualRegister(BoolRC);
  Register NewExec = MRI.createVirtualRegister(BoolRC);
  Register CurrentIdxReg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
  Register CondReg = MRI.createVirtualRegister(BoolRC);

  BuildMI(LoopBB, I, DL, TII->get(TargetOpcode::PHI), PhiReg)
    .addReg(InitReg)
    .addMBB(&OrigBB)
    .addReg(ResultReg)
    .addMBB(&LoopBB);

  BuildMI(LoopBB, I, DL, TII->get(TargetOpcode::PHI), PhiExec)
    .addReg(InitSaveExecReg)
    .addMBB(&OrigBB)
    .addReg(NewExec)
    .addMBB(&LoopBB);

  // Read the next variant <- also loop target.
  BuildMI(LoopBB, I, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32), CurrentIdxReg)
    .addReg(IdxReg.getReg(), getUndefRegState(IdxReg.isUndef()));

  // Compare the just read M0 value to all possible Idx values.
  BuildMI(LoopBB, I, DL, TII->get(AMDGPU::V_CMP_EQ_U32_e64), CondReg)
    .addReg(CurrentIdxReg)
    .addReg(IdxReg.getReg(), 0, IdxReg.getSubReg());

  // Update EXEC, save the original EXEC value to VCC.
  BuildMI(LoopBB, I, DL, TII->get(ST.isWave32() ? AMDGPU::S_AND_SAVEEXEC_B32
                                                : AMDGPU::S_AND_SAVEEXEC_B64),
          NewExec)
    .addReg(CondReg, RegState::Kill);

  MRI.setSimpleHint(NewExec, CondReg);

  if (UseGPRIdxMode) {
    unsigned IdxReg;
    if (Offset == 0) {
      IdxReg = CurrentIdxReg;
    } else {
      IdxReg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
      BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_ADD_I32), IdxReg)
        .addReg(CurrentIdxReg, RegState::Kill)
        .addImm(Offset);
    }
    unsigned IdxMode = IsIndirectSrc ?
      AMDGPU::VGPRIndexMode::SRC0_ENABLE : AMDGPU::VGPRIndexMode::DST_ENABLE;
    MachineInstr *SetOn =
      BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_ON))
      .addReg(IdxReg, RegState::Kill)
      .addImm(IdxMode);
    SetOn->getOperand(3).setIsUndef();
  } else {
    // Move index from VCC into M0
    if (Offset == 0) {
      BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
        .addReg(CurrentIdxReg, RegState::Kill);
    } else {
      BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_ADD_I32), AMDGPU::M0)
        .addReg(CurrentIdxReg, RegState::Kill)
        .addImm(Offset);
    }
  }

  // Update EXEC, switch all done bits to 0 and all todo bits to 1.
  unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
  MachineInstr *InsertPt =
    BuildMI(LoopBB, I, DL, TII->get(ST.isWave32() ? AMDGPU::S_XOR_B32_term
                                                  : AMDGPU::S_XOR_B64_term), Exec)
      .addReg(Exec)
      .addReg(NewExec);

  // XXX - s_xor_b64 sets scc to 1 if the result is nonzero, so can we use
  // s_cbranch_scc0?

  // Loop back to V_READFIRSTLANE_B32 if there are still variants to cover.
  BuildMI(LoopBB, I, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
    .addMBB(&LoopBB);

  return InsertPt->getIterator();
}

// This has slightly sub-optimal regalloc when the source vector is killed by
// the read. The register allocator does not understand that the kill is
// per-workitem, so is kept alive for the whole loop so we end up not re-using a
// subregister from it, using 1 more VGPR than necessary. This was saved when
// this was expanded after register allocation.
static MachineBasicBlock::iterator loadM0FromVGPR(const SIInstrInfo *TII,
                                                  MachineBasicBlock &MBB,
                                                  MachineInstr &MI,
                                                  unsigned InitResultReg,
                                                  unsigned PhiReg,
                                                  int Offset,
                                                  bool UseGPRIdxMode,
                                                  bool IsIndirectSrc) {
  MachineFunction *MF = MBB.getParent();
  const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
  const SIRegisterInfo *TRI = ST.getRegisterInfo();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  const DebugLoc &DL = MI.getDebugLoc();
  MachineBasicBlock::iterator I(&MI);

  const auto *BoolXExecRC = TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
  Register DstReg = MI.getOperand(0).getReg();
  Register SaveExec = MRI.createVirtualRegister(BoolXExecRC);
  Register TmpExec = MRI.createVirtualRegister(BoolXExecRC);
  unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
  unsigned MovExecOpc = ST.isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;

  BuildMI(MBB, I, DL, TII->get(TargetOpcode::IMPLICIT_DEF), TmpExec);

  // Save the EXEC mask
  BuildMI(MBB, I, DL, TII->get(MovExecOpc), SaveExec)
    .addReg(Exec);

  MachineBasicBlock *LoopBB;
  MachineBasicBlock *RemainderBB;
  std::tie(LoopBB, RemainderBB) = splitBlockForLoop(MI, MBB, false);

  const MachineOperand *Idx = TII->getNamedOperand(MI, AMDGPU::OpName::idx);

  auto InsPt = emitLoadM0FromVGPRLoop(TII, MRI, MBB, *LoopBB, DL, *Idx,
                                      InitResultReg, DstReg, PhiReg, TmpExec,
                                      Offset, UseGPRIdxMode, IsIndirectSrc);

  MachineBasicBlock::iterator First = RemainderBB->begin();
  BuildMI(*RemainderBB, First, DL, TII->get(MovExecOpc), Exec)
    .addReg(SaveExec);

  return InsPt;
}

// Returns subreg index, offset
static std::pair<unsigned, int>
computeIndirectRegAndOffset(const SIRegisterInfo &TRI,
                            const TargetRegisterClass *SuperRC,
                            unsigned VecReg,
                            int Offset) {
  int NumElts = TRI.getRegSizeInBits(*SuperRC) / 32;

  // Skip out of bounds offsets, or else we would end up using an undefined
  // register.
  if (Offset >= NumElts || Offset < 0)
    return std::make_pair(AMDGPU::sub0, Offset);

  return std::make_pair(AMDGPU::sub0 + Offset, 0);
}

// Return true if the index is an SGPR and was set.
static bool setM0ToIndexFromSGPR(const SIInstrInfo *TII,
                                 MachineRegisterInfo &MRI,
                                 MachineInstr &MI,
                                 int Offset,
                                 bool UseGPRIdxMode,
                                 bool IsIndirectSrc) {
  MachineBasicBlock *MBB = MI.getParent();
  const DebugLoc &DL = MI.getDebugLoc();
  MachineBasicBlock::iterator I(&MI);

  const MachineOperand *Idx = TII->getNamedOperand(MI, AMDGPU::OpName::idx);
  const TargetRegisterClass *IdxRC = MRI.getRegClass(Idx->getReg());

  assert(Idx->getReg() != AMDGPU::NoRegister);

  if (!TII->getRegisterInfo().isSGPRClass(IdxRC))
    return false;

  if (UseGPRIdxMode) {
    unsigned IdxMode = IsIndirectSrc ?
      AMDGPU::VGPRIndexMode::SRC0_ENABLE : AMDGPU::VGPRIndexMode::DST_ENABLE;
    if (Offset == 0) {
      MachineInstr *SetOn =
          BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_ON))
              .add(*Idx)
              .addImm(IdxMode);

      SetOn->getOperand(3).setIsUndef();
    } else {
      Register Tmp = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
      BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_ADD_I32), Tmp)
          .add(*Idx)
          .addImm(Offset);
      MachineInstr *SetOn =
        BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_ON))
        .addReg(Tmp, RegState::Kill)
        .addImm(IdxMode);

      SetOn->getOperand(3).setIsUndef();
    }

    return true;
  }

  if (Offset == 0) {
    BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
      .add(*Idx);
  } else {
    BuildMI(*MBB, I, DL, TII->get(AMDGPU::S_ADD_I32), AMDGPU::M0)
      .add(*Idx)
      .addImm(Offset);
  }

  return true;
}

// Control flow needs to be inserted if indexing with a VGPR.
static MachineBasicBlock *emitIndirectSrc(MachineInstr &MI,
                                          MachineBasicBlock &MBB,
                                          const GCNSubtarget &ST) {
  const SIInstrInfo *TII = ST.getInstrInfo();
  const SIRegisterInfo &TRI = TII->getRegisterInfo();
  MachineFunction *MF = MBB.getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();

  Register Dst = MI.getOperand(0).getReg();
  Register SrcReg = TII->getNamedOperand(MI, AMDGPU::OpName::src)->getReg();
  int Offset = TII->getNamedOperand(MI, AMDGPU::OpName::offset)->getImm();

  const TargetRegisterClass *VecRC = MRI.getRegClass(SrcReg);

  unsigned SubReg;
  std::tie(SubReg, Offset)
    = computeIndirectRegAndOffset(TRI, VecRC, SrcReg, Offset);

  bool UseGPRIdxMode = ST.useVGPRIndexMode(EnableVGPRIndexMode);

  if (setM0ToIndexFromSGPR(TII, MRI, MI, Offset, UseGPRIdxMode, true)) {
    MachineBasicBlock::iterator I(&MI);
    const DebugLoc &DL = MI.getDebugLoc();

    if (UseGPRIdxMode) {
      // TODO: Look at the uses to avoid the copy. This may require rescheduling
      // to avoid interfering with other uses, so probably requires a new
      // optimization pass.
      BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOV_B32_e32), Dst)
        .addReg(SrcReg, RegState::Undef, SubReg)
        .addReg(SrcReg, RegState::Implicit)
        .addReg(AMDGPU::M0, RegState::Implicit);
      BuildMI(MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
    } else {
      BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
        .addReg(SrcReg, RegState::Undef, SubReg)
        .addReg(SrcReg, RegState::Implicit);
    }

    MI.eraseFromParent();

    return &MBB;
  }

  const DebugLoc &DL = MI.getDebugLoc();
  MachineBasicBlock::iterator I(&MI);

  Register PhiReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
  Register InitReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);

  BuildMI(MBB, I, DL, TII->get(TargetOpcode::IMPLICIT_DEF), InitReg);

  auto InsPt = loadM0FromVGPR(TII, MBB, MI, InitReg, PhiReg,
                              Offset, UseGPRIdxMode, true);
  MachineBasicBlock *LoopBB = InsPt->getParent();

  if (UseGPRIdxMode) {
    BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::V_MOV_B32_e32), Dst)
      .addReg(SrcReg, RegState::Undef, SubReg)
      .addReg(SrcReg, RegState::Implicit)
      .addReg(AMDGPU::M0, RegState::Implicit);
    BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
  } else {
    BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
      .addReg(SrcReg, RegState::Undef, SubReg)
      .addReg(SrcReg, RegState::Implicit);
  }

  MI.eraseFromParent();

  return LoopBB;
}

static unsigned getMOVRELDPseudo(const SIRegisterInfo &TRI,
                                 const TargetRegisterClass *VecRC) {
  switch (TRI.getRegSizeInBits(*VecRC)) {
  case 32: // 4 bytes
    return AMDGPU::V_MOVRELD_B32_V1;
  case 64: // 8 bytes
    return AMDGPU::V_MOVRELD_B32_V2;
  case 128: // 16 bytes
    return AMDGPU::V_MOVRELD_B32_V4;
  case 256: // 32 bytes
    return AMDGPU::V_MOVRELD_B32_V8;
  case 512: // 64 bytes
    return AMDGPU::V_MOVRELD_B32_V16;
  default:
    llvm_unreachable("unsupported size for MOVRELD pseudos");
  }
}

static MachineBasicBlock *emitIndirectDst(MachineInstr &MI,
                                          MachineBasicBlock &MBB,
                                          const GCNSubtarget &ST) {
  const SIInstrInfo *TII = ST.getInstrInfo();
  const SIRegisterInfo &TRI = TII->getRegisterInfo();
  MachineFunction *MF = MBB.getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();

  Register Dst = MI.getOperand(0).getReg();
  const MachineOperand *SrcVec = TII->getNamedOperand(MI, AMDGPU::OpName::src);
  const MachineOperand *Idx = TII->getNamedOperand(MI, AMDGPU::OpName::idx);
  const MachineOperand *Val = TII->getNamedOperand(MI, AMDGPU::OpName::val);
  int Offset = TII->getNamedOperand(MI, AMDGPU::OpName::offset)->getImm();
  const TargetRegisterClass *VecRC = MRI.getRegClass(SrcVec->getReg());

  // This can be an immediate, but will be folded later.
  assert(Val->getReg());

  unsigned SubReg;
  std::tie(SubReg, Offset) = computeIndirectRegAndOffset(TRI, VecRC,
                                                         SrcVec->getReg(),
                                                         Offset);
  bool UseGPRIdxMode = ST.useVGPRIndexMode(EnableVGPRIndexMode);

  if (Idx->getReg() == AMDGPU::NoRegister) {
    MachineBasicBlock::iterator I(&MI);
    const DebugLoc &DL = MI.getDebugLoc();

    assert(Offset == 0);

    BuildMI(MBB, I, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dst)
        .add(*SrcVec)
        .add(*Val)
        .addImm(SubReg);

    MI.eraseFromParent();
    return &MBB;
  }

  if (setM0ToIndexFromSGPR(TII, MRI, MI, Offset, UseGPRIdxMode, false)) {
    MachineBasicBlock::iterator I(&MI);
    const DebugLoc &DL = MI.getDebugLoc();

    if (UseGPRIdxMode) {
      BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOV_B32_indirect))
          .addReg(SrcVec->getReg(), RegState::Undef, SubReg) // vdst
          .add(*Val)
          .addReg(Dst, RegState::ImplicitDefine)
          .addReg(SrcVec->getReg(), RegState::Implicit)
          .addReg(AMDGPU::M0, RegState::Implicit);

      BuildMI(MBB, I, DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
    } else {
      const MCInstrDesc &MovRelDesc = TII->get(getMOVRELDPseudo(TRI, VecRC));

      BuildMI(MBB, I, DL, MovRelDesc)
          .addReg(Dst, RegState::Define)
          .addReg(SrcVec->getReg())
          .add(*Val)
          .addImm(SubReg - AMDGPU::sub0);
    }

    MI.eraseFromParent();
    return &MBB;
  }

  if (Val->isReg())
    MRI.clearKillFlags(Val->getReg());

  const DebugLoc &DL = MI.getDebugLoc();

  Register PhiReg = MRI.createVirtualRegister(VecRC);

  auto InsPt = loadM0FromVGPR(TII, MBB, MI, SrcVec->getReg(), PhiReg,
                              Offset, UseGPRIdxMode, false);
  MachineBasicBlock *LoopBB = InsPt->getParent();

  if (UseGPRIdxMode) {
    BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::V_MOV_B32_indirect))
        .addReg(PhiReg, RegState::Undef, SubReg) // vdst
        .add(*Val)                               // src0
        .addReg(Dst, RegState::ImplicitDefine)
        .addReg(PhiReg, RegState::Implicit)
        .addReg(AMDGPU::M0, RegState::Implicit);
    BuildMI(*LoopBB, InsPt, DL, TII->get(AMDGPU::S_SET_GPR_IDX_OFF));
  } else {
    const MCInstrDesc &MovRelDesc = TII->get(getMOVRELDPseudo(TRI, VecRC));

    BuildMI(*LoopBB, InsPt, DL, MovRelDesc)
        .addReg(Dst, RegState::Define)
        .addReg(PhiReg)
        .add(*Val)
        .addImm(SubReg - AMDGPU::sub0);
  }

  MI.eraseFromParent();

  return LoopBB;
}

MachineBasicBlock *SITargetLowering::EmitInstrWithCustomInserter(
  MachineInstr &MI, MachineBasicBlock *BB) const {

  const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
  MachineFunction *MF = BB->getParent();
  SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();

  if (TII->isMIMG(MI)) {
    if (MI.memoperands_empty() && MI.mayLoadOrStore()) {
      report_fatal_error("missing mem operand from MIMG instruction");
    }
    // Add a memoperand for mimg instructions so that they aren't assumed to
    // be ordered memory instuctions.

    return BB;
  }

  switch (MI.getOpcode()) {
  case AMDGPU::S_ADD_U64_PSEUDO:
  case AMDGPU::S_SUB_U64_PSEUDO: {
    MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
    const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
    const SIRegisterInfo *TRI = ST.getRegisterInfo();
    const TargetRegisterClass *BoolRC = TRI->getBoolRC();
    const DebugLoc &DL = MI.getDebugLoc();

    MachineOperand &Dest = MI.getOperand(0);
    MachineOperand &Src0 = MI.getOperand(1);
    MachineOperand &Src1 = MI.getOperand(2);

    Register DestSub0 = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
    Register DestSub1 = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);

    MachineOperand Src0Sub0 = TII->buildExtractSubRegOrImm(MI, MRI,
     Src0, BoolRC, AMDGPU::sub0,
     &AMDGPU::SReg_32_XM0RegClass);
    MachineOperand Src0Sub1 = TII->buildExtractSubRegOrImm(MI, MRI,
      Src0, BoolRC, AMDGPU::sub1,
      &AMDGPU::SReg_32_XM0RegClass);

    MachineOperand Src1Sub0 = TII->buildExtractSubRegOrImm(MI, MRI,
      Src1, BoolRC, AMDGPU::sub0,
      &AMDGPU::SReg_32_XM0RegClass);
    MachineOperand Src1Sub1 = TII->buildExtractSubRegOrImm(MI, MRI,
      Src1, BoolRC, AMDGPU::sub1,
      &AMDGPU::SReg_32_XM0RegClass);

    bool IsAdd = (MI.getOpcode() == AMDGPU::S_ADD_U64_PSEUDO);

    unsigned LoOpc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
    unsigned HiOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
    BuildMI(*BB, MI, DL, TII->get(LoOpc), DestSub0)
      .add(Src0Sub0)
      .add(Src1Sub0);
    BuildMI(*BB, MI, DL, TII->get(HiOpc), DestSub1)
      .add(Src0Sub1)
      .add(Src1Sub1);
    BuildMI(*BB, MI, DL, TII->get(TargetOpcode::REG_SEQUENCE), Dest.getReg())
      .addReg(DestSub0)
      .addImm(AMDGPU::sub0)
      .addReg(DestSub1)
      .addImm(AMDGPU::sub1);
    MI.eraseFromParent();
    return BB;
  }
  case AMDGPU::SI_INIT_M0: {
    BuildMI(*BB, MI.getIterator(), MI.getDebugLoc(),
            TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
        .add(MI.getOperand(0));
    MI.eraseFromParent();
    return BB;
  }
  case AMDGPU::SI_INIT_EXEC:
    // This should be before all vector instructions.
    BuildMI(*BB, &*BB->begin(), MI.getDebugLoc(), TII->get(AMDGPU::S_MOV_B64),
            AMDGPU::EXEC)
        .addImm(MI.getOperand(0).getImm());
    MI.eraseFromParent();
    return BB;

  case AMDGPU::SI_INIT_EXEC_LO:
    // This should be before all vector instructions.
    BuildMI(*BB, &*BB->begin(), MI.getDebugLoc(), TII->get(AMDGPU::S_MOV_B32),
            AMDGPU::EXEC_LO)
        .addImm(MI.getOperand(0).getImm());
    MI.eraseFromParent();
    return BB;

  case AMDGPU::SI_INIT_EXEC_FROM_INPUT: {
    // Extract the thread count from an SGPR input and set EXEC accordingly.
    // Since BFM can't shift by 64, handle that case with CMP + CMOV.
    //
    // S_BFE_U32 count, input, {shift, 7}
    // S_BFM_B64 exec, count, 0
    // S_CMP_EQ_U32 count, 64
    // S_CMOV_B64 exec, -1
    MachineInstr *FirstMI = &*BB->begin();
    MachineRegisterInfo &MRI = MF->getRegInfo();
    Register InputReg = MI.getOperand(0).getReg();
    Register CountReg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
    bool Found = false;

    // Move the COPY of the input reg to the beginning, so that we can use it.
    for (auto I = BB->begin(); I != &MI; I++) {
      if (I->getOpcode() != TargetOpcode::COPY ||
          I->getOperand(0).getReg() != InputReg)
        continue;

      if (I == FirstMI) {
        FirstMI = &*++BB->begin();
      } else {
        I->removeFromParent();
        BB->insert(FirstMI, &*I);
      }
      Found = true;
      break;
    }
    assert(Found);
    (void)Found;

    // This should be before all vector instructions.
    unsigned Mask = (getSubtarget()->getWavefrontSize() << 1) - 1;
    bool isWave32 = getSubtarget()->isWave32();
    unsigned Exec = isWave32 ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
    BuildMI(*BB, FirstMI, DebugLoc(), TII->get(AMDGPU::S_BFE_U32), CountReg)
        .addReg(InputReg)
        .addImm((MI.getOperand(1).getImm() & Mask) | 0x70000);
    BuildMI(*BB, FirstMI, DebugLoc(),
            TII->get(isWave32 ? AMDGPU::S_BFM_B32 : AMDGPU::S_BFM_B64),
            Exec)
        .addReg(CountReg)
        .addImm(0);
    BuildMI(*BB, FirstMI, DebugLoc(), TII->get(AMDGPU::S_CMP_EQ_U32))
        .addReg(CountReg, RegState::Kill)
        .addImm(getSubtarget()->getWavefrontSize());
    BuildMI(*BB, FirstMI, DebugLoc(),
            TII->get(isWave32 ? AMDGPU::S_CMOV_B32 : AMDGPU::S_CMOV_B64),
            Exec)
        .addImm(-1);
    MI.eraseFromParent();
    return BB;
  }

  case AMDGPU::GET_GROUPSTATICSIZE: {
    assert(getTargetMachine().getTargetTriple().getOS() == Triple::AMDHSA ||
           getTargetMachine().getTargetTriple().getOS() == Triple::AMDPAL);
    DebugLoc DL = MI.getDebugLoc();
    BuildMI(*BB, MI, DL, TII->get(AMDGPU::S_MOV_B32))
        .add(MI.getOperand(0))
        .addImm(MFI->getLDSSize());
    MI.eraseFromParent();
    return BB;
  }
  case AMDGPU::SI_INDIRECT_SRC_V1:
  case AMDGPU::SI_INDIRECT_SRC_V2:
  case AMDGPU::SI_INDIRECT_SRC_V4:
  case AMDGPU::SI_INDIRECT_SRC_V8:
  case AMDGPU::SI_INDIRECT_SRC_V16:
    return emitIndirectSrc(MI, *BB, *getSubtarget());
  case AMDGPU::SI_INDIRECT_DST_V1:
  case AMDGPU::SI_INDIRECT_DST_V2:
  case AMDGPU::SI_INDIRECT_DST_V4:
  case AMDGPU::SI_INDIRECT_DST_V8:
  case AMDGPU::SI_INDIRECT_DST_V16:
    return emitIndirectDst(MI, *BB, *getSubtarget());
  case AMDGPU::SI_KILL_F32_COND_IMM_PSEUDO:
  case AMDGPU::SI_KILL_I1_PSEUDO:
    return splitKillBlock(MI, BB);
  case AMDGPU::V_CNDMASK_B64_PSEUDO: {
    MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
    const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
    const SIRegisterInfo *TRI = ST.getRegisterInfo();

    Register Dst = MI.getOperand(0).getReg();
    Register Src0 = MI.getOperand(1).getReg();
    Register Src1 = MI.getOperand(2).getReg();
    const DebugLoc &DL = MI.getDebugLoc();
    Register SrcCond = MI.getOperand(3).getReg();

    Register DstLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
    Register DstHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
    const auto *CondRC = TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
    Register SrcCondCopy = MRI.createVirtualRegister(CondRC);

    BuildMI(*BB, MI, DL, TII->get(AMDGPU::COPY), SrcCondCopy)
      .addReg(SrcCond);
    BuildMI(*BB, MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64), DstLo)
      .addImm(0)
      .addReg(Src0, 0, AMDGPU::sub0)
      .addImm(0)
      .addReg(Src1, 0, AMDGPU::sub0)
      .addReg(SrcCondCopy);
    BuildMI(*BB, MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64), DstHi)
      .addImm(0)
      .addReg(Src0, 0, AMDGPU::sub1)
      .addImm(0)
      .addReg(Src1, 0, AMDGPU::sub1)
      .addReg(SrcCondCopy);

    BuildMI(*BB, MI, DL, TII->get(AMDGPU::REG_SEQUENCE), Dst)
      .addReg(DstLo)
      .addImm(AMDGPU::sub0)
      .addReg(DstHi)
      .addImm(AMDGPU::sub1);
    MI.eraseFromParent();
    return BB;
  }
  case AMDGPU::SI_BR_UNDEF: {
    const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
    const DebugLoc &DL = MI.getDebugLoc();
    MachineInstr *Br = BuildMI(*BB, MI, DL, TII->get(AMDGPU::S_CBRANCH_SCC1))
                           .add(MI.getOperand(0));
    Br->getOperand(1).setIsUndef(true); // read undef SCC
    MI.eraseFromParent();
    return BB;
  }
  case AMDGPU::ADJCALLSTACKUP:
  case AMDGPU::ADJCALLSTACKDOWN: {
    const SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
    MachineInstrBuilder MIB(*MF, &MI);

    // Add an implicit use of the frame offset reg to prevent the restore copy
    // inserted after the call from being reorderd after stack operations in the
    // the caller's frame.
    MIB.addReg(Info->getStackPtrOffsetReg(), RegState::ImplicitDefine)
        .addReg(Info->getStackPtrOffsetReg(), RegState::Implicit)
        .addReg(Info->getFrameOffsetReg(), RegState::Implicit);
    return BB;
  }
  case AMDGPU::SI_CALL_ISEL: {
    const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
    const DebugLoc &DL = MI.getDebugLoc();

    unsigned ReturnAddrReg = TII->getRegisterInfo().getReturnAddressReg(*MF);

    MachineInstrBuilder MIB;
    MIB = BuildMI(*BB, MI, DL, TII->get(AMDGPU::SI_CALL), ReturnAddrReg);

    for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I)
      MIB.add(MI.getOperand(I));

    MIB.cloneMemRefs(MI);
    MI.eraseFromParent();
    return BB;
  }
  case AMDGPU::V_ADD_I32_e32:
  case AMDGPU::V_SUB_I32_e32:
  case AMDGPU::V_SUBREV_I32_e32: {
    // TODO: Define distinct V_*_I32_Pseudo instructions instead.
    const DebugLoc &DL = MI.getDebugLoc();
    unsigned Opc = MI.getOpcode();

    bool NeedClampOperand = false;
    if (TII->pseudoToMCOpcode(Opc) == -1) {
      Opc = AMDGPU::getVOPe64(Opc);
      NeedClampOperand = true;
    }

    auto I = BuildMI(*BB, MI, DL, TII->get(Opc), MI.getOperand(0).getReg());
    if (TII->isVOP3(*I)) {
      const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
      const SIRegisterInfo *TRI = ST.getRegisterInfo();
      I.addReg(TRI->getVCC(), RegState::Define);
    }
    I.add(MI.getOperand(1))
     .add(MI.getOperand(2));
    if (NeedClampOperand)
      I.addImm(0); // clamp bit for e64 encoding

    TII->legalizeOperands(*I);

    MI.eraseFromParent();
    return BB;
  }
  case AMDGPU::DS_GWS_INIT:
  case AMDGPU::DS_GWS_SEMA_V:
  case AMDGPU::DS_GWS_SEMA_BR:
  case AMDGPU::DS_GWS_SEMA_P:
  case AMDGPU::DS_GWS_SEMA_RELEASE_ALL:
  case AMDGPU::DS_GWS_BARRIER:
    // A s_waitcnt 0 is required to be the instruction immediately following.
    if (getSubtarget()->hasGWSAutoReplay()) {
      bundleInstWithWaitcnt(MI);
      return BB;
    }

    return emitGWSMemViolTestLoop(MI, BB);
  default:
    return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
  }
}

bool SITargetLowering::hasBitPreservingFPLogic(EVT VT) const {
  return isTypeLegal(VT.getScalarType());
}

bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
  // This currently forces unfolding various combinations of fsub into fma with
  // free fneg'd operands. As long as we have fast FMA (controlled by
  // isFMAFasterThanFMulAndFAdd), we should perform these.

  // When fma is quarter rate, for f64 where add / sub are at best half rate,
  // most of these combines appear to be cycle neutral but save on instruction
  // count / code size.
  return true;
}

EVT SITargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &Ctx,
                                         EVT VT) const {
  if (!VT.isVector()) {
    return MVT::i1;
  }
  return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements());
}

MVT SITargetLowering::getScalarShiftAmountTy(const DataLayout &, EVT VT) const {
  // TODO: Should i16 be used always if legal? For now it would force VALU
  // shifts.
  return (VT == MVT::i16) ? MVT::i16 : MVT::i32;
}

// Answering this is somewhat tricky and depends on the specific device which
// have different rates for fma or all f64 operations.
//
// v_fma_f64 and v_mul_f64 always take the same number of cycles as each other
// regardless of which device (although the number of cycles differs between
// devices), so it is always profitable for f64.
//
// v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable
// only on full rate devices. Normally, we should prefer selecting v_mad_f32
// which we can always do even without fused FP ops since it returns the same
// result as the separate operations and since it is always full
// rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32
// however does not support denormals, so we do report fma as faster if we have
// a fast fma device and require denormals.
//
bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
  VT = VT.getScalarType();

  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::f32: {
    // This is as fast on some subtargets. However, we always have full rate f32
    // mad available which returns the same result as the separate operations
    // which we should prefer over fma. We can't use this if we want to support
    // denormals, so only report this in these cases.
    if (Subtarget->hasFP32Denormals())
      return Subtarget->hasFastFMAF32() || Subtarget->hasDLInsts();

    // If the subtarget has v_fmac_f32, that's just as good as v_mac_f32.
    return Subtarget->hasFastFMAF32() && Subtarget->hasDLInsts();
  }
  case MVT::f64:
    return true;
  case MVT::f16:
    return Subtarget->has16BitInsts() && Subtarget->hasFP16Denormals();
  default:
    break;
  }

  return false;
}

//===----------------------------------------------------------------------===//
// Custom DAG Lowering Operations
//===----------------------------------------------------------------------===//

// Work around LegalizeDAG doing the wrong thing and fully scalarizing if the
// wider vector type is legal.
SDValue SITargetLowering::splitUnaryVectorOp(SDValue Op,
                                             SelectionDAG &DAG) const {
  unsigned Opc = Op.getOpcode();
  EVT VT = Op.getValueType();
  assert(VT == MVT::v4f16);

  SDValue Lo, Hi;
  std::tie(Lo, Hi) = DAG.SplitVectorOperand(Op.getNode(), 0);

  SDLoc SL(Op);
  SDValue OpLo = DAG.getNode(Opc, SL, Lo.getValueType(), Lo,
                             Op->getFlags());
  SDValue OpHi = DAG.getNode(Opc, SL, Hi.getValueType(), Hi,
                             Op->getFlags());

  return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(Op), VT, OpLo, OpHi);
}

// Work around LegalizeDAG doing the wrong thing and fully scalarizing if the
// wider vector type is legal.
SDValue SITargetLowering::splitBinaryVectorOp(SDValue Op,
                                              SelectionDAG &DAG) const {
  unsigned Opc = Op.getOpcode();
  EVT VT = Op.getValueType();
  assert(VT == MVT::v4i16 || VT == MVT::v4f16);

  SDValue Lo0, Hi0;
  std::tie(Lo0, Hi0) = DAG.SplitVectorOperand(Op.getNode(), 0);
  SDValue Lo1, Hi1;
  std::tie(Lo1, Hi1) = DAG.SplitVectorOperand(Op.getNode(), 1);

  SDLoc SL(Op);

  SDValue OpLo = DAG.getNode(Opc, SL, Lo0.getValueType(), Lo0, Lo1,
                             Op->getFlags());
  SDValue OpHi = DAG.getNode(Opc, SL, Hi0.getValueType(), Hi0, Hi1,
                             Op->getFlags());

  return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(Op), VT, OpLo, OpHi);
}

SDValue SITargetLowering::splitTernaryVectorOp(SDValue Op,
                                              SelectionDAG &DAG) const {
  unsigned Opc = Op.getOpcode();
  EVT VT = Op.getValueType();
  assert(VT == MVT::v4i16 || VT == MVT::v4f16);

  SDValue Lo0, Hi0;
  std::tie(Lo0, Hi0) = DAG.SplitVectorOperand(Op.getNode(), 0);
  SDValue Lo1, Hi1;
  std::tie(Lo1, Hi1) = DAG.SplitVectorOperand(Op.getNode(), 1);
  SDValue Lo2, Hi2;
  std::tie(Lo2, Hi2) = DAG.SplitVectorOperand(Op.getNode(), 2);

  SDLoc SL(Op);

  SDValue OpLo = DAG.getNode(Opc, SL, Lo0.getValueType(), Lo0, Lo1, Lo2,
                             Op->getFlags());
  SDValue OpHi = DAG.getNode(Opc, SL, Hi0.getValueType(), Hi0, Hi1, Hi2,
                             Op->getFlags());

  return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(Op), VT, OpLo, OpHi);
}


SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
  switch (Op.getOpcode()) {
  default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
  case ISD::BRCOND: return LowerBRCOND(Op, DAG);
  case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
  case ISD::LOAD: {
    SDValue Result = LowerLOAD(Op, DAG);
    assert((!Result.getNode() ||
            Result.getNode()->getNumValues() == 2) &&
           "Load should return a value and a chain");
    return Result;
  }

  case ISD::FSIN:
  case ISD::FCOS:
    return LowerTrig(Op, DAG);
  case ISD::SELECT: return LowerSELECT(Op, DAG);
  case ISD::FDIV: return LowerFDIV(Op, DAG);
  case ISD::ATOMIC_CMP_SWAP: return LowerATOMIC_CMP_SWAP(Op, DAG);
  case ISD::STORE: return LowerSTORE(Op, DAG);
  case ISD::GlobalAddress: {
    MachineFunction &MF = DAG.getMachineFunction();
    SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
    return LowerGlobalAddress(MFI, Op, DAG);
  }
  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
  case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
  case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG);
  case ISD::ADDRSPACECAST: return lowerADDRSPACECAST(Op, DAG);
  case ISD::INSERT_SUBVECTOR:
    return lowerINSERT_SUBVECTOR(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::VECTOR_SHUFFLE:
    return lowerVECTOR_SHUFFLE(Op, DAG);
  case ISD::BUILD_VECTOR:
    return lowerBUILD_VECTOR(Op, DAG);
  case ISD::FP_ROUND:
    return lowerFP_ROUND(Op, DAG);
  case ISD::TRAP:
    return lowerTRAP(Op, DAG);
  case ISD::DEBUGTRAP:
    return lowerDEBUGTRAP(Op, DAG);
  case ISD::FABS:
  case ISD::FNEG:
  case ISD::FCANONICALIZE:
    return splitUnaryVectorOp(Op, DAG);
  case ISD::FMINNUM:
  case ISD::FMAXNUM:
    return lowerFMINNUM_FMAXNUM(Op, DAG);
  case ISD::FMA:
    return splitTernaryVectorOp(Op, DAG);
  case ISD::SHL:
  case ISD::SRA:
  case ISD::SRL:
  case ISD::ADD:
  case ISD::SUB:
  case ISD::MUL:
  case ISD::SMIN:
  case ISD::SMAX:
  case ISD::UMIN:
  case ISD::UMAX:
  case ISD::FADD:
  case ISD::FMUL:
  case ISD::FMINNUM_IEEE:
  case ISD::FMAXNUM_IEEE:
    return splitBinaryVectorOp(Op, DAG);
  }
  return SDValue();
}

static SDValue adjustLoadValueTypeImpl(SDValue Result, EVT LoadVT,
                                       const SDLoc &DL,
                                       SelectionDAG &DAG, bool Unpacked) {
  if (!LoadVT.isVector())
    return Result;

  if (Unpacked) { // From v2i32/v4i32 back to v2f16/v4f16.
    // Truncate to v2i16/v4i16.
    EVT IntLoadVT = LoadVT.changeTypeToInteger();

    // Workaround legalizer not scalarizing truncate after vector op
    // legalization byt not creating intermediate vector trunc.
    SmallVector<SDValue, 4> Elts;
    DAG.ExtractVectorElements(Result, Elts);
    for (SDValue &Elt : Elts)
      Elt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Elt);

    Result = DAG.getBuildVector(IntLoadVT, DL, Elts);

    // Bitcast to original type (v2f16/v4f16).
    return DAG.getNode(ISD::BITCAST, DL, LoadVT, Result);
  }

  // Cast back to the original packed type.
  return DAG.getNode(ISD::BITCAST, DL, LoadVT, Result);
}

SDValue SITargetLowering::adjustLoadValueType(unsigned Opcode,
                                              MemSDNode *M,
                                              SelectionDAG &DAG,
                                              ArrayRef<SDValue> Ops,
                                              bool IsIntrinsic) const {
  SDLoc DL(M);

  bool Unpacked = Subtarget->hasUnpackedD16VMem();
  EVT LoadVT = M->getValueType(0);

  EVT EquivLoadVT = LoadVT;
  if (Unpacked && LoadVT.isVector()) {
    EquivLoadVT = LoadVT.isVector() ?
      EVT::getVectorVT(*DAG.getContext(), MVT::i32,
                       LoadVT.getVectorNumElements()) : LoadVT;
  }

  // Change from v4f16/v2f16 to EquivLoadVT.
  SDVTList VTList = DAG.getVTList(EquivLoadVT, MVT::Other);

  SDValue Load
    = DAG.getMemIntrinsicNode(
      IsIntrinsic ? (unsigned)ISD::INTRINSIC_W_CHAIN : Opcode, DL,
      VTList, Ops, M->getMemoryVT(),
      M->getMemOperand());
  if (!Unpacked) // Just adjusted the opcode.
    return Load;

  SDValue Adjusted = adjustLoadValueTypeImpl(Load, LoadVT, DL, DAG, Unpacked);

  return DAG.getMergeValues({ Adjusted, Load.getValue(1) }, DL);
}

SDValue SITargetLowering::lowerIntrinsicLoad(MemSDNode *M, bool IsFormat,
                                             SelectionDAG &DAG,
                                             ArrayRef<SDValue> Ops) const {
  SDLoc DL(M);
  EVT LoadVT = M->getValueType(0);
  EVT EltType = LoadVT.getScalarType();
  EVT IntVT = LoadVT.changeTypeToInteger();

  bool IsD16 = IsFormat && (EltType.getSizeInBits() == 16);

  unsigned Opc =
      IsFormat ? AMDGPUISD::BUFFER_LOAD_FORMAT : AMDGPUISD::BUFFER_LOAD;

  if (IsD16) {
    return adjustLoadValueType(AMDGPUISD::BUFFER_LOAD_FORMAT_D16, M, DAG, Ops);
  }

  // Handle BUFFER_LOAD_BYTE/UBYTE/SHORT/USHORT overloaded intrinsics
  if (!IsD16 && !LoadVT.isVector() && EltType.getSizeInBits() < 32)
    return handleByteShortBufferLoads(DAG, LoadVT, DL, Ops, M);

  if (isTypeLegal(LoadVT)) {
    return getMemIntrinsicNode(Opc, DL, M->getVTList(), Ops, IntVT,
                               M->getMemOperand(), DAG);
  }

  EVT CastVT = getEquivalentMemType(*DAG.getContext(), LoadVT);
  SDVTList VTList = DAG.getVTList(CastVT, MVT::Other);
  SDValue MemNode = getMemIntrinsicNode(Opc, DL, VTList, Ops, CastVT,
                                        M->getMemOperand(), DAG);
  return DAG.getMergeValues(
      {DAG.getNode(ISD::BITCAST, DL, LoadVT, MemNode), MemNode.getValue(1)},
      DL);
}

static SDValue lowerICMPIntrinsic(const SITargetLowering &TLI,
                                  SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  const auto *CD = cast<ConstantSDNode>(N->getOperand(3));
  int CondCode = CD->getSExtValue();
  if (CondCode < ICmpInst::Predicate::FIRST_ICMP_PREDICATE ||
      CondCode > ICmpInst::Predicate::LAST_ICMP_PREDICATE)
    return DAG.getUNDEF(VT);

  ICmpInst::Predicate IcInput = static_cast<ICmpInst::Predicate>(CondCode);

  SDValue LHS = N->getOperand(1);
  SDValue RHS = N->getOperand(2);

  SDLoc DL(N);

  EVT CmpVT = LHS.getValueType();
  if (CmpVT == MVT::i16 && !TLI.isTypeLegal(MVT::i16)) {
    unsigned PromoteOp = ICmpInst::isSigned(IcInput) ?
      ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
    LHS = DAG.getNode(PromoteOp, DL, MVT::i32, LHS);
    RHS = DAG.getNode(PromoteOp, DL, MVT::i32, RHS);
  }

  ISD::CondCode CCOpcode = getICmpCondCode(IcInput);

  unsigned WavefrontSize = TLI.getSubtarget()->getWavefrontSize();
  EVT CCVT = EVT::getIntegerVT(*DAG.getContext(), WavefrontSize);

  SDValue SetCC = DAG.getNode(AMDGPUISD::SETCC, DL, CCVT, LHS, RHS,
                              DAG.getCondCode(CCOpcode));
  if (VT.bitsEq(CCVT))
    return SetCC;
  return DAG.getZExtOrTrunc(SetCC, DL, VT);
}

static SDValue lowerFCMPIntrinsic(const SITargetLowering &TLI,
                                  SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  const auto *CD = cast<ConstantSDNode>(N->getOperand(3));

  int CondCode = CD->getSExtValue();
  if (CondCode < FCmpInst::Predicate::FIRST_FCMP_PREDICATE ||
      CondCode > FCmpInst::Predicate::LAST_FCMP_PREDICATE) {
    return DAG.getUNDEF(VT);
  }

  SDValue Src0 = N->getOperand(1);
  SDValue Src1 = N->getOperand(2);
  EVT CmpVT = Src0.getValueType();
  SDLoc SL(N);

  if (CmpVT == MVT::f16 && !TLI.isTypeLegal(CmpVT)) {
    Src0 = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, Src0);
    Src1 = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, Src1);
  }

  FCmpInst::Predicate IcInput = static_cast<FCmpInst::Predicate>(CondCode);
  ISD::CondCode CCOpcode = getFCmpCondCode(IcInput);
  unsigned WavefrontSize = TLI.getSubtarget()->getWavefrontSize();
  EVT CCVT = EVT::getIntegerVT(*DAG.getContext(), WavefrontSize);
  SDValue SetCC = DAG.getNode(AMDGPUISD::SETCC, SL, CCVT, Src0,
                              Src1, DAG.getCondCode(CCOpcode));
  if (VT.bitsEq(CCVT))
    return SetCC;
  return DAG.getZExtOrTrunc(SetCC, SL, VT);
}

void SITargetLowering::ReplaceNodeResults(SDNode *N,
                                          SmallVectorImpl<SDValue> &Results,
                                          SelectionDAG &DAG) const {
  switch (N->getOpcode()) {
  case ISD::INSERT_VECTOR_ELT: {
    if (SDValue Res = lowerINSERT_VECTOR_ELT(SDValue(N, 0), DAG))
      Results.push_back(Res);
    return;
  }
  case ISD::EXTRACT_VECTOR_ELT: {
    if (SDValue Res = lowerEXTRACT_VECTOR_ELT(SDValue(N, 0), DAG))
      Results.push_back(Res);
    return;
  }
  case ISD::INTRINSIC_WO_CHAIN: {
    unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
    switch (IID) {
    case Intrinsic::amdgcn_cvt_pkrtz: {
      SDValue Src0 = N->getOperand(1);
      SDValue Src1 = N->getOperand(2);
      SDLoc SL(N);
      SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_PKRTZ_F16_F32, SL, MVT::i32,
                                Src0, Src1);
      Results.push_back(DAG.getNode(ISD::BITCAST, SL, MVT::v2f16, Cvt));
      return;
    }
    case Intrinsic::amdgcn_cvt_pknorm_i16:
    case Intrinsic::amdgcn_cvt_pknorm_u16:
    case Intrinsic::amdgcn_cvt_pk_i16:
    case Intrinsic::amdgcn_cvt_pk_u16: {
      SDValue Src0 = N->getOperand(1);
      SDValue Src1 = N->getOperand(2);
      SDLoc SL(N);
      unsigned Opcode;

      if (IID == Intrinsic::amdgcn_cvt_pknorm_i16)
        Opcode = AMDGPUISD::CVT_PKNORM_I16_F32;
      else if (IID == Intrinsic::amdgcn_cvt_pknorm_u16)
        Opcode = AMDGPUISD::CVT_PKNORM_U16_F32;
      else if (IID == Intrinsic::amdgcn_cvt_pk_i16)
        Opcode = AMDGPUISD::CVT_PK_I16_I32;
      else
        Opcode = AMDGPUISD::CVT_PK_U16_U32;

      EVT VT = N->getValueType(0);
      if (isTypeLegal(VT))
        Results.push_back(DAG.getNode(Opcode, SL, VT, Src0, Src1));
      else {
        SDValue Cvt = DAG.getNode(Opcode, SL, MVT::i32, Src0, Src1);
        Results.push_back(DAG.getNode(ISD::BITCAST, SL, MVT::v2i16, Cvt));
      }
      return;
    }
    }
    break;
  }
  case ISD::INTRINSIC_W_CHAIN: {
    if (SDValue Res = LowerINTRINSIC_W_CHAIN(SDValue(N, 0), DAG)) {
      if (Res.getOpcode() == ISD::MERGE_VALUES) {
        // FIXME: Hacky
        Results.push_back(Res.getOperand(0));
        Results.push_back(Res.getOperand(1));
      } else {
        Results.push_back(Res);
        Results.push_back(Res.getValue(1));
      }
      return;
    }

    break;
  }
  case ISD::SELECT: {
    SDLoc SL(N);
    EVT VT = N->getValueType(0);
    EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
    SDValue LHS = DAG.getNode(ISD::BITCAST, SL, NewVT, N->getOperand(1));
    SDValue RHS = DAG.getNode(ISD::BITCAST, SL, NewVT, N->getOperand(2));

    EVT SelectVT = NewVT;
    if (NewVT.bitsLT(MVT::i32)) {
      LHS = DAG.getNode(ISD::ANY_EXTEND, SL, MVT::i32, LHS);
      RHS = DAG.getNode(ISD::ANY_EXTEND, SL, MVT::i32, RHS);
      SelectVT = MVT::i32;
    }

    SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, SelectVT,
                                    N->getOperand(0), LHS, RHS);

    if (NewVT != SelectVT)
      NewSelect = DAG.getNode(ISD::TRUNCATE, SL, NewVT, NewSelect);
    Results.push_back(DAG.getNode(ISD::BITCAST, SL, VT, NewSelect));
    return;
  }
  case ISD::FNEG: {
    if (N->getValueType(0) != MVT::v2f16)
      break;

    SDLoc SL(N);
    SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::i32, N->getOperand(0));

    SDValue Op = DAG.getNode(ISD::XOR, SL, MVT::i32,
                             BC,
                             DAG.getConstant(0x80008000, SL, MVT::i32));
    Results.push_back(DAG.getNode(ISD::BITCAST, SL, MVT::v2f16, Op));
    return;
  }
  case ISD::FABS: {
    if (N->getValueType(0) != MVT::v2f16)
      break;

    SDLoc SL(N);
    SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::i32, N->getOperand(0));

    SDValue Op = DAG.getNode(ISD::AND, SL, MVT::i32,
                             BC,
                             DAG.getConstant(0x7fff7fff, SL, MVT::i32));
    Results.push_back(DAG.getNode(ISD::BITCAST, SL, MVT::v2f16, Op));
    return;
  }
  default:
    break;
  }
}

/// Helper function for LowerBRCOND
static SDNode *findUser(SDValue Value, unsigned Opcode) {

  SDNode *Parent = Value.getNode();
  for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
       I != E; ++I) {

    if (I.getUse().get() != Value)
      continue;

    if (I->getOpcode() == Opcode)
      return *I;
  }
  return nullptr;
}

unsigned SITargetLowering::isCFIntrinsic(const SDNode *Intr) const {
  if (Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
    switch (cast<ConstantSDNode>(Intr->getOperand(1))->getZExtValue()) {