| 1 | //===-- GCNHazardRecognizers.h - GCN Hazard Recognizers ---------*- 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 | // This file defines hazard recognizers for scheduling on GCN processors. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_LIB_TARGET_AMDGPUHAZARDRECOGNIZERS_H |
| 14 | #define LLVM_LIB_TARGET_AMDGPUHAZARDRECOGNIZERS_H |
| 15 | |
| 16 | #include "llvm/ADT/BitVector.h" |
| 17 | #include "llvm/ADT/STLExtras.h" |
| 18 | #include "llvm/CodeGen/ScheduleHazardRecognizer.h" |
| 19 | #include "llvm/CodeGen/TargetSchedule.h" |
| 20 | #include <list> |
| 21 | |
| 22 | namespace llvm { |
| 23 | |
| 24 | class MachineFunction; |
| 25 | class MachineInstr; |
| 26 | class MachineOperand; |
| 27 | class MachineRegisterInfo; |
| 28 | class SIInstrInfo; |
| 29 | class SIRegisterInfo; |
| 30 | class GCNSubtarget; |
| 31 | |
| 32 | class GCNHazardRecognizer final : public ScheduleHazardRecognizer { |
| 33 | public: |
| 34 | typedef function_ref<bool(const MachineInstr &)> IsHazardFn; |
| 35 | |
| 36 | private: |
| 37 | // Distinguish if we are called from scheduler or hazard recognizer |
| 38 | bool IsHazardRecognizerMode; |
| 39 | |
| 40 | // This variable stores the instruction that has been emitted this cycle. It |
| 41 | // will be added to EmittedInstrs, when AdvanceCycle() or RecedeCycle() is |
| 42 | // called. |
| 43 | MachineInstr *CurrCycleInstr; |
| 44 | std::list<MachineInstr*> EmittedInstrs; |
| 45 | const MachineFunction &MF; |
| 46 | const GCNSubtarget &ST; |
| 47 | const SIInstrInfo &TII; |
| 48 | const SIRegisterInfo &TRI; |
| 49 | const TargetSchedModel &TSchedModel; |
| 50 | bool RunLdsBranchVmemWARHazardFixup; |
| 51 | |
| 52 | /// RegUnits of uses in the current soft memory clause. |
| 53 | BitVector ClauseUses; |
| 54 | |
| 55 | /// RegUnits of defs in the current soft memory clause. |
| 56 | BitVector ClauseDefs; |
| 57 | |
| 58 | void resetClause() { |
| 59 | ClauseUses.reset(); |
| 60 | ClauseDefs.reset(); |
| 61 | } |
| 62 | |
| 63 | void addClauseInst(const MachineInstr &MI); |
| 64 | |
| 65 | /// \returns the number of wait states before another MFMA instruction can be |
| 66 | /// issued after \p MI. |
| 67 | unsigned getMFMAPipelineWaitStates(const MachineInstr &MI) const; |
| 68 | |
| 69 | // Advance over a MachineInstr bundle. Look for hazards in the bundled |
| 70 | // instructions. |
| 71 | void processBundle(); |
| 72 | |
| 73 | // Run on an individual instruction in hazard recognizer mode. This can be |
| 74 | // used on a newly inserted instruction before returning from PreEmitNoops. |
| 75 | void runOnInstruction(MachineInstr *MI); |
| 76 | |
| 77 | int getWaitStatesSince(IsHazardFn IsHazard, int Limit); |
| 78 | int getWaitStatesSinceDef(unsigned Reg, IsHazardFn IsHazardDef, int Limit); |
| 79 | int getWaitStatesSinceSetReg(IsHazardFn IsHazard, int Limit); |
| 80 | |
| 81 | int checkSoftClauseHazards(MachineInstr *SMEM); |
| 82 | int checkSMRDHazards(MachineInstr *SMRD); |
| 83 | int checkVMEMHazards(MachineInstr* VMEM); |
| 84 | int checkDPPHazards(MachineInstr *DPP); |
| 85 | int checkDivFMasHazards(MachineInstr *DivFMas); |
| 86 | int checkGetRegHazards(MachineInstr *GetRegInstr); |
| 87 | int checkSetRegHazards(MachineInstr *SetRegInstr); |
| 88 | int createsVALUHazard(const MachineInstr &MI); |
| 89 | int checkVALUHazards(MachineInstr *VALU); |
| 90 | int checkVALUHazardsHelper(const MachineOperand &Def, const MachineRegisterInfo &MRI); |
| 91 | int checkRWLaneHazards(MachineInstr *RWLane); |
| 92 | int checkRFEHazards(MachineInstr *RFE); |
| 93 | int checkInlineAsmHazards(MachineInstr *IA); |
| 94 | int checkReadM0Hazards(MachineInstr *SMovRel); |
| 95 | int checkNSAtoVMEMHazard(MachineInstr *MI); |
| 96 | int checkFPAtomicToDenormModeHazard(MachineInstr *MI); |
| 97 | void fixHazards(MachineInstr *MI); |
| 98 | bool fixVcmpxPermlaneHazards(MachineInstr *MI); |
| 99 | bool fixVMEMtoScalarWriteHazards(MachineInstr *MI); |
| 100 | bool fixSMEMtoVectorWriteHazards(MachineInstr *MI); |
| 101 | bool fixVcmpxExecWARHazard(MachineInstr *MI); |
| 102 | bool fixLdsBranchVmemWARHazard(MachineInstr *MI); |
| 103 | bool fixLdsDirectVALUHazard(MachineInstr *MI); |
| 104 | bool fixLdsDirectVMEMHazard(MachineInstr *MI); |
| 105 | bool fixVALUPartialForwardingHazard(MachineInstr *MI); |
| 106 | bool fixVALUTransUseHazard(MachineInstr *MI); |
| 107 | bool fixWMMAHazards(MachineInstr *MI); |
| 108 | bool fixShift64HighRegBug(MachineInstr *MI); |
| 109 | bool fixVALUMaskWriteHazard(MachineInstr *MI); |
| 110 | bool fixRequiredExportPriority(MachineInstr *MI); |
| 111 | |
| 112 | int checkMAIHazards(MachineInstr *MI); |
| 113 | int checkMAIHazards908(MachineInstr *MI); |
| 114 | int checkMAIHazards90A(MachineInstr *MI); |
| 115 | /// Pad the latency between neighboring MFMA instructions with s_nops. The |
| 116 | /// percentage of wait states to fill with s_nops is specified by the command |
| 117 | /// line option '-amdgpu-mfma-padding-ratio'. |
| 118 | /// |
| 119 | /// For example, with '-amdgpu-mfma-padding-ratio=100': |
| 120 | /// |
| 121 | /// 2 pass MFMA instructions have a latency of 2 wait states. Therefore, a |
| 122 | /// 'S_NOP 1' will be added between sequential MFMA instructions. |
| 123 | /// |
| 124 | /// V_MFMA_F32_4X4X1F32 |
| 125 | /// V_MFMA_F32_4X4X1F32 |
| 126 | ///--> |
| 127 | /// V_MFMA_F32_4X4X1F32 |
| 128 | /// S_NOP 1 |
| 129 | /// V_MFMA_F32_4X4X1F32 |
| 130 | int checkMFMAPadding(MachineInstr *MI); |
| 131 | int checkMAIVALUHazards(MachineInstr *MI); |
| 132 | int checkMAILdStHazards(MachineInstr *MI); |
| 133 | int checkPermlaneHazards(MachineInstr *MI); |
| 134 | |
| 135 | public: |
| 136 | GCNHazardRecognizer(const MachineFunction &MF); |
| 137 | // We can only issue one instruction per cycle. |
| 138 | bool atIssueLimit() const override { return true; } |
| 139 | void EmitInstruction(SUnit *SU) override; |
| 140 | void EmitInstruction(MachineInstr *MI) override; |
| 141 | HazardType getHazardType(SUnit *SU, int Stalls) override; |
| 142 | void EmitNoop() override; |
| 143 | unsigned PreEmitNoops(MachineInstr *) override; |
| 144 | unsigned PreEmitNoopsCommon(MachineInstr *); |
| 145 | void AdvanceCycle() override; |
| 146 | void RecedeCycle() override; |
| 147 | bool ShouldPreferAnother(SUnit *SU) override; |
| 148 | void Reset() override; |
| 149 | }; |
| 150 | |
| 151 | } // end namespace llvm |
| 152 | |
| 153 | #endif //LLVM_LIB_TARGET_AMDGPUHAZARDRECOGNIZERS_H |
| 154 | |