| 1 | //===- AMDGPURegBankLegalizeHelper ------------------------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUREGBANKLEGALIZEHELPER_H |
| 10 | #define LLVM_LIB_TARGET_AMDGPU_AMDGPUREGBANKLEGALIZEHELPER_H |
| 11 | |
| 12 | #include "AMDGPURegBankLegalizeRules.h" |
| 13 | #include "llvm/ADT/SmallSet.h" |
| 14 | #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h" |
| 15 | #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" |
| 16 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 17 | |
| 18 | namespace llvm { |
| 19 | |
| 20 | class MachineIRBuilder; |
| 21 | |
| 22 | namespace AMDGPU { |
| 23 | |
| 24 | // Receives list of RegBankLLTMappingApplyID and applies register banks on all |
| 25 | // operands. It is user's responsibility to provide RegBankLLTMappingApplyIDs |
| 26 | // for all register operands, there is no need to specify NonReg for trailing |
| 27 | // imm operands. This finishes selection of register banks if there is no need |
| 28 | // to replace instruction. In other case InstApplyMethod will create new |
| 29 | // instruction(s). |
| 30 | class RegBankLegalizeHelper { |
| 31 | MachineFunction &MF; |
| 32 | const GCNSubtarget &ST; |
| 33 | MachineIRBuilder &B; |
| 34 | MachineRegisterInfo &MRI; |
| 35 | const MachineUniformityInfo &MUI; |
| 36 | const RegisterBankInfo &RBI; |
| 37 | MachineOptimizationRemarkEmitter MORE; |
| 38 | const RegBankLegalizeRules &RBLRules; |
| 39 | const bool IsWave32; |
| 40 | const RegisterBank *SgprRB; |
| 41 | const RegisterBank *VgprRB; |
| 42 | const RegisterBank *VccRB; |
| 43 | |
| 44 | static constexpr LLT S1 = LLT::scalar(SizeInBits: 1); |
| 45 | static constexpr LLT S16 = LLT::scalar(SizeInBits: 16); |
| 46 | static constexpr LLT S32 = LLT::scalar(SizeInBits: 32); |
| 47 | static constexpr LLT S64 = LLT::scalar(SizeInBits: 64); |
| 48 | static constexpr LLT S96 = LLT::scalar(SizeInBits: 96); |
| 49 | static constexpr LLT S128 = LLT::scalar(SizeInBits: 128); |
| 50 | static constexpr LLT S256 = LLT::scalar(SizeInBits: 256); |
| 51 | |
| 52 | static constexpr LLT V2S16 = LLT::fixed_vector(NumElements: 2, ScalarSizeInBits: 16); |
| 53 | static constexpr LLT V4S16 = LLT::fixed_vector(NumElements: 4, ScalarSizeInBits: 16); |
| 54 | static constexpr LLT V6S16 = LLT::fixed_vector(NumElements: 6, ScalarSizeInBits: 16); |
| 55 | static constexpr LLT V8S16 = LLT::fixed_vector(NumElements: 8, ScalarSizeInBits: 16); |
| 56 | static constexpr LLT V16S16 = LLT::fixed_vector(NumElements: 16, ScalarSizeInBits: 16); |
| 57 | static constexpr LLT V32S16 = LLT::fixed_vector(NumElements: 32, ScalarSizeInBits: 16); |
| 58 | |
| 59 | static constexpr LLT V2S32 = LLT::fixed_vector(NumElements: 2, ScalarSizeInBits: 32); |
| 60 | static constexpr LLT V3S32 = LLT::fixed_vector(NumElements: 3, ScalarSizeInBits: 32); |
| 61 | static constexpr LLT V4S32 = LLT::fixed_vector(NumElements: 4, ScalarSizeInBits: 32); |
| 62 | static constexpr LLT V6S32 = LLT::fixed_vector(NumElements: 6, ScalarSizeInBits: 32); |
| 63 | static constexpr LLT V7S32 = LLT::fixed_vector(NumElements: 7, ScalarSizeInBits: 32); |
| 64 | static constexpr LLT V8S32 = LLT::fixed_vector(NumElements: 8, ScalarSizeInBits: 32); |
| 65 | static constexpr LLT V16S32 = LLT::fixed_vector(NumElements: 16, ScalarSizeInBits: 32); |
| 66 | |
| 67 | static constexpr LLT V2S64 = LLT::fixed_vector(NumElements: 2, ScalarSizeInBits: 64); |
| 68 | static constexpr LLT V3S64 = LLT::fixed_vector(NumElements: 3, ScalarSizeInBits: 64); |
| 69 | static constexpr LLT V4S64 = LLT::fixed_vector(NumElements: 4, ScalarSizeInBits: 64); |
| 70 | static constexpr LLT V8S64 = LLT::fixed_vector(NumElements: 8, ScalarSizeInBits: 64); |
| 71 | static constexpr LLT V16S64 = LLT::fixed_vector(NumElements: 16, ScalarSizeInBits: 64); |
| 72 | |
| 73 | static constexpr LLT P1 = LLT::pointer(AddressSpace: 1, SizeInBits: 64); |
| 74 | static constexpr LLT P4 = LLT::pointer(AddressSpace: 4, SizeInBits: 64); |
| 75 | static constexpr LLT P6 = LLT::pointer(AddressSpace: 6, SizeInBits: 32); |
| 76 | |
| 77 | MachineRegisterInfo::VRegAttrs SgprRB_S32 = {.RCOrRB: SgprRB, .Ty: S32}; |
| 78 | MachineRegisterInfo::VRegAttrs SgprRB_S16 = {.RCOrRB: SgprRB, .Ty: S16}; |
| 79 | MachineRegisterInfo::VRegAttrs VgprRB_S32 = {.RCOrRB: VgprRB, .Ty: S32}; |
| 80 | MachineRegisterInfo::VRegAttrs VccRB_S1 = {.RCOrRB: VccRB, .Ty: S1}; |
| 81 | |
| 82 | public: |
| 83 | RegBankLegalizeHelper(MachineIRBuilder &B, const MachineUniformityInfo &MUI, |
| 84 | const RegisterBankInfo &RBI, |
| 85 | const RegBankLegalizeRules &RBLRules); |
| 86 | |
| 87 | bool findRuleAndApplyMapping(MachineInstr &MI); |
| 88 | |
| 89 | // Manual apply helpers. |
| 90 | bool applyMappingPHI(MachineInstr &MI); |
| 91 | void applyMappingTrivial(MachineInstr &MI); |
| 92 | |
| 93 | private: |
| 94 | bool executeInWaterfallLoop(MachineIRBuilder &B, |
| 95 | iterator_range<MachineBasicBlock::iterator> Range, |
| 96 | SmallSet<Register, 4> &SgprOperandRegs); |
| 97 | |
| 98 | LLT getTyFromID(RegBankLLTMappingApplyID ID); |
| 99 | LLT getBTyFromID(RegBankLLTMappingApplyID ID, LLT Ty); |
| 100 | |
| 101 | const RegisterBank *getRegBankFromID(RegBankLLTMappingApplyID ID); |
| 102 | |
| 103 | bool |
| 104 | applyMappingDst(MachineInstr &MI, unsigned &OpIdx, |
| 105 | const SmallVectorImpl<RegBankLLTMappingApplyID> &MethodIDs); |
| 106 | |
| 107 | bool |
| 108 | applyMappingSrc(MachineInstr &MI, unsigned &OpIdx, |
| 109 | const SmallVectorImpl<RegBankLLTMappingApplyID> &MethodIDs, |
| 110 | SmallSet<Register, 4> &SgprWaterfallOperandRegs); |
| 111 | |
| 112 | bool splitLoad(MachineInstr &MI, ArrayRef<LLT> LLTBreakdown, |
| 113 | LLT MergeTy = LLT()); |
| 114 | bool widenLoad(MachineInstr &MI, LLT WideTy, LLT MergeTy = LLT()); |
| 115 | bool widenMMOToS32(GAnyLoad &MI) const; |
| 116 | |
| 117 | bool lower(MachineInstr &MI, const RegBankLLTMapping &Mapping, |
| 118 | SmallSet<Register, 4> &SgprWaterfallOperandRegs); |
| 119 | |
| 120 | bool lowerVccExtToSel(MachineInstr &MI); |
| 121 | std::pair<Register, Register> unpackZExt(Register Reg); |
| 122 | std::pair<Register, Register> unpackSExt(Register Reg); |
| 123 | std::pair<Register, Register> unpackAExt(Register Reg); |
| 124 | std::pair<Register, Register> unpackAExtTruncS16(Register Reg); |
| 125 | bool lowerUnpackBitShift(MachineInstr &MI); |
| 126 | bool lowerV_BFE(MachineInstr &MI); |
| 127 | bool lowerS_BFE(MachineInstr &MI); |
| 128 | bool lowerUniMAD64(MachineInstr &MI); |
| 129 | bool lowerSplitTo32(MachineInstr &MI); |
| 130 | bool lowerSplitTo32Mul(MachineInstr &MI); |
| 131 | bool lowerSplitTo16(MachineInstr &MI); |
| 132 | bool lowerSplitTo32Select(MachineInstr &MI); |
| 133 | bool lowerSplitTo32SExtInReg(MachineInstr &MI); |
| 134 | bool lowerUnpackMinMax(MachineInstr &MI); |
| 135 | bool lowerUnpackAExt(MachineInstr &MI); |
| 136 | }; |
| 137 | |
| 138 | } // end namespace AMDGPU |
| 139 | } // end namespace llvm |
| 140 | |
| 141 | #endif |
| 142 | |