llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeHelper.h

//===- AMDGPURegBankLegalizeHelper ------------------------------*- C++ -*-===//
//
// 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
//
//===----------------------------------------------------------------------===//

#ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUREGBANKLEGALIZEHELPER_H
#define LLVM_LIB_TARGET_AMDGPU_AMDGPUREGBANKLEGALIZEHELPER_H

#include "AMDGPURegBankLegalizeRules.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
#include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"

namespace llvm {

class MachineIRBuilder;
class SIInstrInfo;
class SIMachineFunctionInfo;
class GISelValueTracking;

namespace AMDGPU {

/// Holds waterfall loop information: the set of SGPR operand registers that
/// need waterfalling, and an instruction range [Start, End) to wrap in the
/// loop.
struct WaterfallInfo {
  SmallSet<Register, 4> SgprWaterfallOperandRegs;
  MachineBasicBlock::iterator Start;
  MachineBasicBlock::iterator End;
};

// Receives list of RegBankLLTMappingApplyID and applies register banks on all
// operands. It is user's responsibility to provide RegBankLLTMappingApplyIDs
// for all register operands, there is no need to specify NonReg for trailing
// imm operands. This finishes selection of register banks if there is no need
// to replace instruction. In other case InstApplyMethod will create new
// instruction(s).
class RegBankLegalizeHelper {
  MachineFunction &MF;
  const SIMachineFunctionInfo *MFI;
  const GCNSubtarget &ST;
  const SIInstrInfo &TII;
  MachineIRBuilder &B;
  MachineRegisterInfo &MRI;
  const MachineUniformityInfo &MUI;
  GISelValueTracking *VT;
  const RegisterBankInfo &RBI;
  MachineOptimizationRemarkEmitter MORE;
  const RegBankLegalizeRules &RBLRules;
  const bool IsWave32;
  const RegisterBank *SgprRB;
  const RegisterBank *VgprRB;
  const RegisterBank *AgprRB;
  const RegisterBank *VccRB;

  static constexpr LLT S1 = LLT::scalar(1);
  static constexpr LLT S16 = LLT::scalar(16);
  static constexpr LLT S32 = LLT::scalar(32);
  static constexpr LLT S64 = LLT::scalar(64);
  static constexpr LLT S96 = LLT::scalar(96);
  static constexpr LLT S128 = LLT::scalar(128);
  static constexpr LLT S256 = LLT::scalar(256);

  static constexpr LLT V2S16 = LLT::fixed_vector(2, 16);
  static constexpr LLT V4S16 = LLT::fixed_vector(4, 16);
  static constexpr LLT V6S16 = LLT::fixed_vector(6, 16);
  static constexpr LLT V8S16 = LLT::fixed_vector(8, 16);
  static constexpr LLT V16S16 = LLT::fixed_vector(16, 16);
  static constexpr LLT V32S16 = LLT::fixed_vector(32, 16);

  static constexpr LLT V2S32 = LLT::fixed_vector(2, 32);
  static constexpr LLT V3S32 = LLT::fixed_vector(3, 32);
  static constexpr LLT V4S32 = LLT::fixed_vector(4, 32);
  static constexpr LLT V6S32 = LLT::fixed_vector(6, 32);
  static constexpr LLT V7S32 = LLT::fixed_vector(7, 32);
  static constexpr LLT V8S32 = LLT::fixed_vector(8, 32);
  static constexpr LLT V16S32 = LLT::fixed_vector(16, 32);

  static constexpr LLT V2S64 = LLT::fixed_vector(2, 64);
  static constexpr LLT V3S64 = LLT::fixed_vector(3, 64);
  static constexpr LLT V4S64 = LLT::fixed_vector(4, 64);
  static constexpr LLT V8S64 = LLT::fixed_vector(8, 64);
  static constexpr LLT V16S64 = LLT::fixed_vector(16, 64);

  static constexpr LLT P1 = LLT::pointer(1, 64);
  static constexpr LLT P4 = LLT::pointer(4, 64);
  static constexpr LLT P6 = LLT::pointer(6, 32);

  const LLT I16 = LLT::integer(16);
  const LLT I32 = LLT::integer(32);
  const LLT I64 = LLT::integer(64);

  MachineRegisterInfo::VRegAttrs SgprRB_I32 = {SgprRB, I32};
  MachineRegisterInfo::VRegAttrs SgprRB_I64 = {SgprRB, I64};
  MachineRegisterInfo::VRegAttrs VgprRB_I32 = {VgprRB, I32};
  MachineRegisterInfo::VRegAttrs VgprRB_I64 = {VgprRB, I64};
  MachineRegisterInfo::VRegAttrs VccRB_S1 = {VccRB, S1};

public:
  RegBankLegalizeHelper(MachineIRBuilder &B, const MachineUniformityInfo &MUI,
                        GISelValueTracking *VT, const RegisterBankInfo &RBI,
                        const RegBankLegalizeRules &RBLRules);

  bool findRuleAndApplyMapping(MachineInstr &MI);

private:
  bool executeInWaterfallLoop(MachineIRBuilder &B, const WaterfallInfo &WFI);

  LLT getTyFromID(RegBankLLTMappingApplyID ID);
  LLT getBTyFromID(RegBankLLTMappingApplyID ID, LLT Ty);

  const RegisterBank *getRegBankFromID(RegBankLLTMappingApplyID ID);

  bool
  applyMappingDst(MachineInstr &MI, unsigned &OpIdx,
                  const SmallVectorImpl<RegBankLLTMappingApplyID> &MethodIDs);

  bool
  applyMappingSrc(MachineInstr &MI, unsigned &OpIdx,
                  const SmallVectorImpl<RegBankLLTMappingApplyID> &MethodIDs,
                  WaterfallInfo &WFI);

  unsigned setBufferOffsets(MachineIRBuilder &B, Register CombinedOffset,
                            Register &VOffsetReg, Register &SOffsetReg,
                            int64_t &InstOffsetVal, Align Alignment);

  bool splitLoad(MachineInstr &MI, ArrayRef<LLT> LLTBreakdown,
                 LLT MergeTy = LLT());
  bool widenLoad(MachineInstr &MI, LLT WideTy, LLT MergeTy = LLT());
  bool widenMMOToS32(GAnyLoad &MI) const;

  bool lower(MachineInstr &MI, const RegBankLLTMapping &Mapping,
             WaterfallInfo &WFI);

  bool lowerVccExtToSel(MachineInstr &MI);
  std::pair<Register, Register> unpackZExt(Register Reg);
  std::pair<Register, Register> unpackSExt(Register Reg);
  std::pair<Register, Register> unpackAExt(Register Reg);
  std::pair<Register, Register> unpackAExtTruncS16(Register Reg);
  bool lowerUnpackBitShift(MachineInstr &MI);
  bool lowerV_BFE(MachineInstr &MI);
  bool lowerS_BFE(MachineInstr &MI);
  bool lowerUniMAD64(MachineInstr &MI);
  bool lowerSplitTo32(MachineInstr &MI);
  bool lowerSplitTo32Mul(MachineInstr &MI);
  bool lowerSplitTo16(MachineInstr &MI);
  bool lowerSplitTo32Select(MachineInstr &MI);
  bool lowerSplitTo32SExtInReg(MachineInstr &MI);
  bool lowerSplitBitCount64To32(MachineInstr &MI);
  bool lowerUnpackMinMax(MachineInstr &MI);
  bool lowerUnpackAExt(MachineInstr &MI);
  bool lowerSBufToBuf(MachineInstr &MI, WaterfallInfo &WFI);
  bool lowerExtrVecEltToSel(MachineInstr &MI);
  bool lowerExtrVecEltTo32(MachineInstr &MI);
  bool lowerInsVecEltToSel(MachineInstr &MI);
  bool lowerInsVecEltTo32(MachineInstr &MI);
  bool lowerAbsToNegMax(MachineInstr &MI);
  bool lowerAbsToS32(MachineInstr &MI);
  bool lowerSetRounding(MachineInstr &MI);
  bool lowerGetRounding(MachineInstr &MI);
  bool applyRegisterBanksVgprWithSgprRsrc(MachineInstr &MI, unsigned RsrcIdx);
};

} // end namespace AMDGPU
} // end namespace llvm

#endif