xref: /freebsd/contrib/llvm-project/llvm/lib/Target/Hexagon/HexagonMachineScheduler.h (revision 25ecdc7d52770caf1c9b44b5ec11f468f6b636f3)
1 //===- HexagonMachineScheduler.h - Custom Hexagon MI scheduler --*- 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 // Custom Hexagon MI scheduler.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #ifndef LLVM_LIB_TARGET_HEXAGON_HEXAGONMACHINESCHEDULER_H
14 #define LLVM_LIB_TARGET_HEXAGON_HEXAGONMACHINESCHEDULER_H
15 
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/Twine.h"
18 #include "llvm/CodeGen/DFAPacketizer.h"
19 #include "llvm/CodeGen/MachineScheduler.h"
20 #include "llvm/CodeGen/RegisterPressure.h"
21 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
22 #include "llvm/CodeGen/TargetInstrInfo.h"
23 #include "llvm/CodeGen/TargetSchedule.h"
24 #include "llvm/CodeGen/TargetSubtargetInfo.h"
25 #include <algorithm>
26 #include <cassert>
27 #include <limits>
28 #include <memory>
29 #include <vector>
30 
31 namespace llvm {
32 
33 class SUnit;
34 
35 class VLIWResourceModel {
36   /// ResourcesModel - Represents VLIW state.
37   /// Not limited to VLIW targets per se, but assumes
38   /// definition of DFA by a target.
39   DFAPacketizer *ResourcesModel;
40 
41   const TargetSchedModel *SchedModel;
42 
43   /// Local packet/bundle model. Purely
44   /// internal to the MI schedulre at the time.
45   std::vector<SUnit *> Packet;
46 
47   /// Total packets created.
48   unsigned TotalPackets = 0;
49 
50 public:
51   VLIWResourceModel(const TargetSubtargetInfo &STI, const TargetSchedModel *SM)
52       : SchedModel(SM) {
53     ResourcesModel = STI.getInstrInfo()->CreateTargetScheduleState(STI);
54 
55     // This hard requirement could be relaxed,
56     // but for now do not let it proceed.
57     assert(ResourcesModel && "Unimplemented CreateTargetScheduleState.");
58 
59     Packet.resize(SchedModel->getIssueWidth());
60     Packet.clear();
61     ResourcesModel->clearResources();
62   }
63 
64   ~VLIWResourceModel() {
65     delete ResourcesModel;
66   }
67 
68   void resetPacketState() {
69     Packet.clear();
70   }
71 
72   void resetDFA() {
73     ResourcesModel->clearResources();
74   }
75 
76   void reset() {
77     Packet.clear();
78     ResourcesModel->clearResources();
79   }
80 
81   bool isResourceAvailable(SUnit *SU, bool IsTop);
82   bool reserveResources(SUnit *SU, bool IsTop);
83   unsigned getTotalPackets() const { return TotalPackets; }
84   bool isInPacket(SUnit *SU) const { return is_contained(Packet, SU); }
85 };
86 
87 /// Extend the standard ScheduleDAGMI to provide more context and override the
88 /// top-level schedule() driver.
89 class VLIWMachineScheduler : public ScheduleDAGMILive {
90 public:
91   VLIWMachineScheduler(MachineSchedContext *C,
92                        std::unique_ptr<MachineSchedStrategy> S)
93       : ScheduleDAGMILive(C, std::move(S)) {}
94 
95   /// Schedule - This is called back from ScheduleDAGInstrs::Run() when it's
96   /// time to do some work.
97   void schedule() override;
98 
99   RegisterClassInfo *getRegClassInfo() { return RegClassInfo; }
100   int getBBSize() { return BB->size(); }
101 };
102 
103 //===----------------------------------------------------------------------===//
104 // ConvergingVLIWScheduler - Implementation of the standard
105 // MachineSchedStrategy.
106 //===----------------------------------------------------------------------===//
107 
108 /// ConvergingVLIWScheduler shrinks the unscheduled zone using heuristics
109 /// to balance the schedule.
110 class ConvergingVLIWScheduler : public MachineSchedStrategy {
111   /// Store the state used by ConvergingVLIWScheduler heuristics, required
112   ///  for the lifetime of one invocation of pickNode().
113   struct SchedCandidate {
114     // The best SUnit candidate.
115     SUnit *SU = nullptr;
116 
117     // Register pressure values for the best candidate.
118     RegPressureDelta RPDelta;
119 
120     // Best scheduling cost.
121     int SCost = 0;
122 
123     SchedCandidate() = default;
124   };
125   /// Represent the type of SchedCandidate found within a single queue.
126   enum CandResult {
127     NoCand, NodeOrder, SingleExcess, SingleCritical, SingleMax, MultiPressure,
128     BestCost, Weak};
129 
130   /// Each Scheduling boundary is associated with ready queues. It tracks the
131   /// current cycle in whichever direction at has moved, and maintains the state
132   /// of "hazards" and other interlocks at the current cycle.
133   struct VLIWSchedBoundary {
134     VLIWMachineScheduler *DAG = nullptr;
135     const TargetSchedModel *SchedModel = nullptr;
136 
137     ReadyQueue Available;
138     ReadyQueue Pending;
139     bool CheckPending = false;
140 
141     ScheduleHazardRecognizer *HazardRec = nullptr;
142     VLIWResourceModel *ResourceModel = nullptr;
143 
144     unsigned CurrCycle = 0;
145     unsigned IssueCount = 0;
146     unsigned CriticalPathLength = 0;
147 
148     /// MinReadyCycle - Cycle of the soonest available instruction.
149     unsigned MinReadyCycle = std::numeric_limits<unsigned>::max();
150 
151     // Remember the greatest min operand latency.
152     unsigned MaxMinLatency = 0;
153 
154     /// Pending queues extend the ready queues with the same ID and the
155     /// PendingFlag set.
156     VLIWSchedBoundary(unsigned ID, const Twine &Name)
157         : Available(ID, Name+".A"),
158           Pending(ID << ConvergingVLIWScheduler::LogMaxQID, Name+".P") {}
159 
160     ~VLIWSchedBoundary() {
161       delete ResourceModel;
162       delete HazardRec;
163     }
164 
165     void init(VLIWMachineScheduler *dag, const TargetSchedModel *smodel) {
166       DAG = dag;
167       SchedModel = smodel;
168       CurrCycle = 0;
169       IssueCount = 0;
170       // Initialize the critical path length limit, which used by the scheduling
171       // cost model to determine the value for scheduling an instruction. We use
172       // a slightly different heuristic for small and large functions. For small
173       // functions, it's important to use the height/depth of the instruction.
174       // For large functions, prioritizing by height or depth increases spills.
175       CriticalPathLength = DAG->getBBSize() / SchedModel->getIssueWidth();
176       if (DAG->getBBSize() < 50)
177         // We divide by two as a cheap and simple heuristic to reduce the
178         // critcal path length, which increases the priority of using the graph
179         // height/depth in the scheduler's cost computation.
180         CriticalPathLength >>= 1;
181       else {
182         // For large basic blocks, we prefer a larger critical path length to
183         // decrease the priority of using the graph height/depth.
184         unsigned MaxPath = 0;
185         for (auto &SU : DAG->SUnits)
186           MaxPath = std::max(MaxPath, isTop() ? SU.getHeight() : SU.getDepth());
187         CriticalPathLength = std::max(CriticalPathLength, MaxPath) + 1;
188       }
189     }
190 
191     bool isTop() const {
192       return Available.getID() == ConvergingVLIWScheduler::TopQID;
193     }
194 
195     bool checkHazard(SUnit *SU);
196 
197     void releaseNode(SUnit *SU, unsigned ReadyCycle);
198 
199     void bumpCycle();
200 
201     void bumpNode(SUnit *SU);
202 
203     void releasePending();
204 
205     void removeReady(SUnit *SU);
206 
207     SUnit *pickOnlyChoice();
208 
209     bool isLatencyBound(SUnit *SU) {
210       if (CurrCycle >= CriticalPathLength)
211         return true;
212       unsigned PathLength = isTop() ? SU->getHeight() : SU->getDepth();
213       return CriticalPathLength - CurrCycle <= PathLength;
214     }
215   };
216 
217   VLIWMachineScheduler *DAG = nullptr;
218   const TargetSchedModel *SchedModel = nullptr;
219 
220   // State of the top and bottom scheduled instruction boundaries.
221   VLIWSchedBoundary Top;
222   VLIWSchedBoundary Bot;
223 
224   /// List of pressure sets that have a high pressure level in the region.
225   std::vector<bool> HighPressureSets;
226 
227 public:
228   /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
229   enum {
230     TopQID = 1,
231     BotQID = 2,
232     LogMaxQID = 2
233   };
234 
235   ConvergingVLIWScheduler() : Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
236 
237   void initialize(ScheduleDAGMI *dag) override;
238 
239   SUnit *pickNode(bool &IsTopNode) override;
240 
241   void schedNode(SUnit *SU, bool IsTopNode) override;
242 
243   void releaseTopNode(SUnit *SU) override;
244 
245   void releaseBottomNode(SUnit *SU) override;
246 
247   unsigned reportPackets() {
248     return Top.ResourceModel->getTotalPackets() +
249            Bot.ResourceModel->getTotalPackets();
250   }
251 
252 protected:
253   SUnit *pickNodeBidrectional(bool &IsTopNode);
254 
255   int pressureChange(const SUnit *SU, bool isBotUp);
256 
257   int SchedulingCost(ReadyQueue &Q,
258                      SUnit *SU, SchedCandidate &Candidate,
259                      RegPressureDelta &Delta, bool verbose);
260 
261   CandResult pickNodeFromQueue(VLIWSchedBoundary &Zone,
262                                const RegPressureTracker &RPTracker,
263                                SchedCandidate &Candidate);
264 #ifndef NDEBUG
265   void traceCandidate(const char *Label, const ReadyQueue &Q, SUnit *SU,
266                       int Cost, PressureChange P = PressureChange());
267 
268   void readyQueueVerboseDump(const RegPressureTracker &RPTracker,
269                              SchedCandidate &Candidate, ReadyQueue &Q);
270 #endif
271 };
272 
273 } // end namespace llvm
274 
275 #endif // LLVM_LIB_TARGET_HEXAGON_HEXAGONMACHINESCHEDULER_H
276