| 1 | //===- RISCVMachineScheduler.cpp - MI Scheduler for RISC-V ----------------===// |
| 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 | #include "RISCVMachineScheduler.h" |
| 10 | #include "llvm/CodeGen/ScheduleDAG.h" |
| 11 | |
| 12 | using namespace llvm; |
| 13 | |
| 14 | #define DEBUG_TYPE "riscv-prera-sched-strategy" |
| 15 | |
| 16 | RISCV::VSETVLIInfo |
| 17 | RISCVPreRAMachineSchedStrategy::getVSETVLIInfo(const MachineInstr *MI) const { |
| 18 | unsigned TSFlags = MI->getDesc().TSFlags; |
| 19 | if (!RISCVII::hasSEWOp(TSFlags)) |
| 20 | return RISCV::VSETVLIInfo(); |
| 21 | return VIA.computeInfoForInstr(MI: *MI); |
| 22 | } |
| 23 | |
| 24 | bool RISCVPreRAMachineSchedStrategy::tryVSETVLIInfo( |
| 25 | const RISCV::VSETVLIInfo &TryInfo, const RISCV::VSETVLIInfo &CandInfo, |
| 26 | SchedCandidate &TryCand, SchedCandidate &Cand, CandReason Reason) const { |
| 27 | // Do not compare the vsetvli info changes between top and bottom |
| 28 | // boundary. |
| 29 | if (Cand.AtTop != TryCand.AtTop) |
| 30 | return false; |
| 31 | |
| 32 | auto IsCompatible = [&](const RISCV::VSETVLIInfo &FirstInfo, |
| 33 | const RISCV::VSETVLIInfo &SecondInfo) { |
| 34 | return FirstInfo.isValid() && SecondInfo.isValid() && |
| 35 | FirstInfo.isCompatible(Used: RISCV::DemandedFields::all(), Require: SecondInfo, |
| 36 | LIS: Context->LIS); |
| 37 | }; |
| 38 | |
| 39 | // Try Cand first. |
| 40 | // We prefer the top node as it is straightforward from the perspective of |
| 41 | // vsetvli dataflow. |
| 42 | if (Cand.AtTop && IsCompatible(CandInfo, TopInfo)) |
| 43 | return true; |
| 44 | |
| 45 | if (!Cand.AtTop && IsCompatible(CandInfo, BottomInfo)) |
| 46 | return true; |
| 47 | |
| 48 | // Then try TryCand. |
| 49 | if (TryCand.AtTop && IsCompatible(TryInfo, TopInfo)) { |
| 50 | TryCand.Reason = Reason; |
| 51 | return true; |
| 52 | } |
| 53 | |
| 54 | if (!TryCand.AtTop && IsCompatible(TryInfo, BottomInfo)) { |
| 55 | TryCand.Reason = Reason; |
| 56 | return true; |
| 57 | } |
| 58 | |
| 59 | return false; |
| 60 | } |
| 61 | |
| 62 | bool RISCVPreRAMachineSchedStrategy::tryCandidate(SchedCandidate &Cand, |
| 63 | SchedCandidate &TryCand, |
| 64 | SchedBoundary *Zone) const { |
| 65 | //-------------------------------------------------------------------------// |
| 66 | // Below is copied from `GenericScheduler::tryCandidate`. |
| 67 | // FIXME: Is there a way to not replicate this? |
| 68 | //-------------------------------------------------------------------------// |
| 69 | // Initialize the candidate if needed. |
| 70 | if (!Cand.isValid()) { |
| 71 | TryCand.Reason = FirstValid; |
| 72 | return true; |
| 73 | } |
| 74 | |
| 75 | // Bias PhysReg Defs and copies to their uses and defined respectively. |
| 76 | if (tryGreater(TryVal: biasPhysReg(SU: TryCand.SU, isTop: TryCand.AtTop), |
| 77 | CandVal: biasPhysReg(SU: Cand.SU, isTop: Cand.AtTop), TryCand, Cand, Reason: PhysReg)) |
| 78 | return TryCand.Reason != NoCand; |
| 79 | |
| 80 | // Avoid exceeding the target's limit. |
| 81 | if (DAG->isTrackingPressure() && |
| 82 | tryPressure(TryP: TryCand.RPDelta.Excess, CandP: Cand.RPDelta.Excess, TryCand, Cand, |
| 83 | Reason: RegExcess, TRI, MF: DAG->MF)) |
| 84 | return TryCand.Reason != NoCand; |
| 85 | |
| 86 | // Avoid increasing the max critical pressure in the scheduled region. |
| 87 | if (DAG->isTrackingPressure() && |
| 88 | tryPressure(TryP: TryCand.RPDelta.CriticalMax, CandP: Cand.RPDelta.CriticalMax, |
| 89 | TryCand, Cand, Reason: RegCritical, TRI, MF: DAG->MF)) |
| 90 | return TryCand.Reason != NoCand; |
| 91 | |
| 92 | // We only compare a subset of features when comparing nodes between |
| 93 | // Top and Bottom boundary. Some properties are simply incomparable, in many |
| 94 | // other instances we should only override the other boundary if something |
| 95 | // is a clear good pick on one boundary. Skip heuristics that are more |
| 96 | // "tie-breaking" in nature. |
| 97 | bool SameBoundary = Zone != nullptr; |
| 98 | if (SameBoundary) { |
| 99 | // For loops that are acyclic path limited, aggressively schedule for |
| 100 | // latency. Within an single cycle, whenever CurrMOps > 0, allow normal |
| 101 | // heuristics to take precedence. |
| 102 | if (Rem.IsAcyclicLatencyLimited && !Zone->getCurrMOps() && |
| 103 | tryLatency(TryCand, Cand, Zone&: *Zone)) |
| 104 | return TryCand.Reason != NoCand; |
| 105 | |
| 106 | // Prioritize instructions that read unbuffered resources by stall cycles. |
| 107 | if (tryLess(TryVal: Zone->getLatencyStallCycles(SU: TryCand.SU), |
| 108 | CandVal: Zone->getLatencyStallCycles(SU: Cand.SU), TryCand, Cand, Reason: Stall)) |
| 109 | return TryCand.Reason != NoCand; |
| 110 | } |
| 111 | |
| 112 | // Keep clustered nodes together to encourage downstream peephole |
| 113 | // optimizations which may reduce resource requirements. |
| 114 | // |
| 115 | // This is a best effort to set things up for a post-RA pass. Optimizations |
| 116 | // like generating loads of multiple registers should ideally be done within |
| 117 | // the scheduler pass by combining the loads during DAG postprocessing. |
| 118 | unsigned CandZoneCluster = Cand.AtTop ? TopClusterID : BotClusterID; |
| 119 | unsigned TryCandZoneCluster = TryCand.AtTop ? TopClusterID : BotClusterID; |
| 120 | bool CandIsClusterSucc = |
| 121 | isTheSameCluster(A: CandZoneCluster, B: Cand.SU->ParentClusterIdx); |
| 122 | bool TryCandIsClusterSucc = |
| 123 | isTheSameCluster(A: TryCandZoneCluster, B: TryCand.SU->ParentClusterIdx); |
| 124 | |
| 125 | if (tryGreater(TryVal: TryCandIsClusterSucc, CandVal: CandIsClusterSucc, TryCand, Cand, |
| 126 | Reason: Cluster)) |
| 127 | return TryCand.Reason != NoCand; |
| 128 | |
| 129 | if (SameBoundary) { |
| 130 | // Weak edges are for clustering and other constraints. |
| 131 | if (tryLess(TryVal: getWeakLeft(SU: TryCand.SU, isTop: TryCand.AtTop), |
| 132 | CandVal: getWeakLeft(SU: Cand.SU, isTop: Cand.AtTop), TryCand, Cand, Reason: Weak)) |
| 133 | return TryCand.Reason != NoCand; |
| 134 | } |
| 135 | |
| 136 | // Avoid increasing the max pressure of the entire region. |
| 137 | if (DAG->isTrackingPressure() && |
| 138 | tryPressure(TryP: TryCand.RPDelta.CurrentMax, CandP: Cand.RPDelta.CurrentMax, TryCand, |
| 139 | Cand, Reason: RegMax, TRI, MF: DAG->MF)) |
| 140 | return TryCand.Reason != NoCand; |
| 141 | |
| 142 | if (SameBoundary) { |
| 143 | // Avoid critical resource consumption and balance the schedule. |
| 144 | TryCand.initResourceDelta(DAG, SchedModel); |
| 145 | if (tryLess(TryVal: TryCand.ResDelta.CritResources, CandVal: Cand.ResDelta.CritResources, |
| 146 | TryCand, Cand, Reason: ResourceReduce)) |
| 147 | return TryCand.Reason != NoCand; |
| 148 | if (tryGreater(TryVal: TryCand.ResDelta.DemandedResources, |
| 149 | CandVal: Cand.ResDelta.DemandedResources, TryCand, Cand, |
| 150 | Reason: ResourceDemand)) |
| 151 | return TryCand.Reason != NoCand; |
| 152 | |
| 153 | // Avoid serializing long latency dependence chains. |
| 154 | // For acyclic path limited loops, latency was already checked above. |
| 155 | if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency && |
| 156 | !Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, Zone&: *Zone)) |
| 157 | return TryCand.Reason != NoCand; |
| 158 | |
| 159 | // Fall through to original instruction order. |
| 160 | if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) || |
| 161 | (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) |
| 162 | TryCand.Reason = NodeOrder; |
| 163 | } |
| 164 | |
| 165 | //-------------------------------------------------------------------------// |
| 166 | // Below is RISC-V specific scheduling heuristics. |
| 167 | //-------------------------------------------------------------------------// |
| 168 | |
| 169 | // Add RISC-V specific heuristic only when TryCand isn't selected or |
| 170 | // selected as node order. |
| 171 | if (TryCand.Reason != NodeOrder && TryCand.Reason != NoCand) |
| 172 | return true; |
| 173 | |
| 174 | // TODO: We should not use `CandReason::Cluster` here, but is there a |
| 175 | // mechanism to extend this enum? |
| 176 | if (ST->enableVsetvliSchedHeuristic() && |
| 177 | tryVSETVLIInfo(TryInfo: getVSETVLIInfo(MI: TryCand.SU->getInstr()), |
| 178 | CandInfo: getVSETVLIInfo(MI: Cand.SU->getInstr()), TryCand, Cand, |
| 179 | Reason: Cluster)) |
| 180 | return TryCand.Reason != NoCand; |
| 181 | |
| 182 | return TryCand.Reason != NoCand; |
| 183 | } |
| 184 | |
| 185 | void RISCVPreRAMachineSchedStrategy::enterMBB(MachineBasicBlock *MBB) { |
| 186 | TopInfo = RISCV::VSETVLIInfo(); |
| 187 | BottomInfo = RISCV::VSETVLIInfo(); |
| 188 | } |
| 189 | |
| 190 | void RISCVPreRAMachineSchedStrategy::leaveMBB() { |
| 191 | TopInfo = RISCV::VSETVLIInfo(); |
| 192 | BottomInfo = RISCV::VSETVLIInfo(); |
| 193 | } |
| 194 | |
| 195 | void RISCVPreRAMachineSchedStrategy::schedNode(SUnit *SU, bool IsTopNode) { |
| 196 | GenericScheduler::schedNode(SU, IsTopNode); |
| 197 | if (ST->enableVsetvliSchedHeuristic()) { |
| 198 | MachineInstr *MI = SU->getInstr(); |
| 199 | const RISCV::VSETVLIInfo &Info = getVSETVLIInfo(MI); |
| 200 | if (Info.isValid()) { |
| 201 | if (IsTopNode) |
| 202 | TopInfo = Info; |
| 203 | else |
| 204 | BottomInfo = Info; |
| 205 | LLVM_DEBUG({ |
| 206 | dbgs() << "Previous scheduled Unit: \n" ; |
| 207 | dbgs() << " IsTop: " << IsTopNode << "\n" ; |
| 208 | dbgs() << " SU(" << SU->NodeNum << ") - " ; |
| 209 | MI->dump(); |
| 210 | dbgs() << " \n" ; |
| 211 | Info.dump(); |
| 212 | dbgs() << " \n" ; |
| 213 | }); |
| 214 | } |
| 215 | } |
| 216 | } |
| 217 | |