Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
14 | pmbaty | 1 | //===- ScheduleDAGInstrs.h - MachineInstr Scheduling ------------*- 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 | /// \file Implements the ScheduleDAGInstrs class, which implements scheduling |
||
10 | /// for a MachineInstr-based dependency graph. |
||
11 | // |
||
12 | //===----------------------------------------------------------------------===// |
||
13 | |||
14 | #ifndef LLVM_CODEGEN_SCHEDULEDAGINSTRS_H |
||
15 | #define LLVM_CODEGEN_SCHEDULEDAGINSTRS_H |
||
16 | |||
17 | #include "llvm/ADT/DenseMap.h" |
||
18 | #include "llvm/ADT/PointerIntPair.h" |
||
19 | #include "llvm/ADT/SmallVector.h" |
||
20 | #include "llvm/ADT/SparseMultiSet.h" |
||
21 | #include "llvm/ADT/SparseSet.h" |
||
22 | #include "llvm/ADT/identity.h" |
||
23 | #include "llvm/CodeGen/LivePhysRegs.h" |
||
24 | #include "llvm/CodeGen/MachineBasicBlock.h" |
||
25 | #include "llvm/CodeGen/ScheduleDAG.h" |
||
26 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
||
27 | #include "llvm/CodeGen/TargetSchedule.h" |
||
28 | #include "llvm/MC/LaneBitmask.h" |
||
29 | #include <cassert> |
||
30 | #include <cstdint> |
||
31 | #include <list> |
||
32 | #include <string> |
||
33 | #include <utility> |
||
34 | #include <vector> |
||
35 | |||
36 | namespace llvm { |
||
37 | |||
38 | class AAResults; |
||
39 | class LiveIntervals; |
||
40 | class MachineFrameInfo; |
||
41 | class MachineFunction; |
||
42 | class MachineInstr; |
||
43 | class MachineLoopInfo; |
||
44 | class MachineOperand; |
||
45 | struct MCSchedClassDesc; |
||
46 | class PressureDiffs; |
||
47 | class PseudoSourceValue; |
||
48 | class RegPressureTracker; |
||
49 | class UndefValue; |
||
50 | class Value; |
||
51 | |||
52 | /// An individual mapping from virtual register number to SUnit. |
||
53 | struct VReg2SUnit { |
||
54 | unsigned VirtReg; |
||
55 | LaneBitmask LaneMask; |
||
56 | SUnit *SU; |
||
57 | |||
58 | VReg2SUnit(unsigned VReg, LaneBitmask LaneMask, SUnit *SU) |
||
59 | : VirtReg(VReg), LaneMask(LaneMask), SU(SU) {} |
||
60 | |||
61 | unsigned getSparseSetIndex() const { |
||
62 | return Register::virtReg2Index(VirtReg); |
||
63 | } |
||
64 | }; |
||
65 | |||
66 | /// Mapping from virtual register to SUnit including an operand index. |
||
67 | struct VReg2SUnitOperIdx : public VReg2SUnit { |
||
68 | unsigned OperandIndex; |
||
69 | |||
70 | VReg2SUnitOperIdx(unsigned VReg, LaneBitmask LaneMask, |
||
71 | unsigned OperandIndex, SUnit *SU) |
||
72 | : VReg2SUnit(VReg, LaneMask, SU), OperandIndex(OperandIndex) {} |
||
73 | }; |
||
74 | |||
75 | /// Record a physical register access. |
||
76 | /// For non-data-dependent uses, OpIdx == -1. |
||
77 | struct PhysRegSUOper { |
||
78 | SUnit *SU; |
||
79 | int OpIdx; |
||
80 | unsigned Reg; |
||
81 | |||
82 | PhysRegSUOper(SUnit *su, int op, unsigned R): SU(su), OpIdx(op), Reg(R) {} |
||
83 | |||
84 | unsigned getSparseSetIndex() const { return Reg; } |
||
85 | }; |
||
86 | |||
87 | /// Use a SparseMultiSet to track physical registers. Storage is only |
||
88 | /// allocated once for the pass. It can be cleared in constant time and reused |
||
89 | /// without any frees. |
||
90 | using Reg2SUnitsMap = |
||
91 | SparseMultiSet<PhysRegSUOper, identity<unsigned>, uint16_t>; |
||
92 | |||
93 | /// Use SparseSet as a SparseMap by relying on the fact that it never |
||
94 | /// compares ValueT's, only unsigned keys. This allows the set to be cleared |
||
95 | /// between scheduling regions in constant time as long as ValueT does not |
||
96 | /// require a destructor. |
||
97 | using VReg2SUnitMap = SparseSet<VReg2SUnit, VirtReg2IndexFunctor>; |
||
98 | |||
99 | /// Track local uses of virtual registers. These uses are gathered by the DAG |
||
100 | /// builder and may be consulted by the scheduler to avoid iterating an entire |
||
101 | /// vreg use list. |
||
102 | using VReg2SUnitMultiMap = SparseMultiSet<VReg2SUnit, VirtReg2IndexFunctor>; |
||
103 | |||
104 | using VReg2SUnitOperIdxMultiMap = |
||
105 | SparseMultiSet<VReg2SUnitOperIdx, VirtReg2IndexFunctor>; |
||
106 | |||
107 | using ValueType = PointerUnion<const Value *, const PseudoSourceValue *>; |
||
108 | |||
109 | struct UnderlyingObject : PointerIntPair<ValueType, 1, bool> { |
||
110 | UnderlyingObject(ValueType V, bool MayAlias) |
||
111 | : PointerIntPair<ValueType, 1, bool>(V, MayAlias) {} |
||
112 | |||
113 | ValueType getValue() const { return getPointer(); } |
||
114 | bool mayAlias() const { return getInt(); } |
||
115 | }; |
||
116 | |||
117 | using UnderlyingObjectsVector = SmallVector<UnderlyingObject, 4>; |
||
118 | |||
119 | /// A ScheduleDAG for scheduling lists of MachineInstr. |
||
120 | class ScheduleDAGInstrs : public ScheduleDAG { |
||
121 | protected: |
||
122 | const MachineLoopInfo *MLI; |
||
123 | const MachineFrameInfo &MFI; |
||
124 | |||
125 | /// TargetSchedModel provides an interface to the machine model. |
||
126 | TargetSchedModel SchedModel; |
||
127 | |||
128 | /// True if the DAG builder should remove kill flags (in preparation for |
||
129 | /// rescheduling). |
||
130 | bool RemoveKillFlags; |
||
131 | |||
132 | /// The standard DAG builder does not normally include terminators as DAG |
||
133 | /// nodes because it does not create the necessary dependencies to prevent |
||
134 | /// reordering. A specialized scheduler can override |
||
135 | /// TargetInstrInfo::isSchedulingBoundary then enable this flag to indicate |
||
136 | /// it has taken responsibility for scheduling the terminator correctly. |
||
137 | bool CanHandleTerminators = false; |
||
138 | |||
139 | /// Whether lane masks should get tracked. |
||
140 | bool TrackLaneMasks = false; |
||
141 | |||
142 | // State specific to the current scheduling region. |
||
143 | // ------------------------------------------------ |
||
144 | |||
145 | /// The block in which to insert instructions |
||
146 | MachineBasicBlock *BB; |
||
147 | |||
148 | /// The beginning of the range to be scheduled. |
||
149 | MachineBasicBlock::iterator RegionBegin; |
||
150 | |||
151 | /// The end of the range to be scheduled. |
||
152 | MachineBasicBlock::iterator RegionEnd; |
||
153 | |||
154 | /// Instructions in this region (distance(RegionBegin, RegionEnd)). |
||
155 | unsigned NumRegionInstrs; |
||
156 | |||
157 | /// After calling BuildSchedGraph, each machine instruction in the current |
||
158 | /// scheduling region is mapped to an SUnit. |
||
159 | DenseMap<MachineInstr*, SUnit*> MISUnitMap; |
||
160 | |||
161 | // State internal to DAG building. |
||
162 | // ------------------------------- |
||
163 | |||
164 | /// Defs, Uses - Remember where defs and uses of each register are as we |
||
165 | /// iterate upward through the instructions. This is allocated here instead |
||
166 | /// of inside BuildSchedGraph to avoid the need for it to be initialized and |
||
167 | /// destructed for each block. |
||
168 | Reg2SUnitsMap Defs; |
||
169 | Reg2SUnitsMap Uses; |
||
170 | |||
171 | /// Tracks the last instruction(s) in this region defining each virtual |
||
172 | /// register. There may be multiple current definitions for a register with |
||
173 | /// disjunct lanemasks. |
||
174 | VReg2SUnitMultiMap CurrentVRegDefs; |
||
175 | /// Tracks the last instructions in this region using each virtual register. |
||
176 | VReg2SUnitOperIdxMultiMap CurrentVRegUses; |
||
177 | |||
178 | AAResults *AAForDep = nullptr; |
||
179 | |||
180 | /// Remember a generic side-effecting instruction as we proceed. |
||
181 | /// No other SU ever gets scheduled around it (except in the special |
||
182 | /// case of a huge region that gets reduced). |
||
183 | SUnit *BarrierChain = nullptr; |
||
184 | |||
185 | public: |
||
186 | /// A list of SUnits, used in Value2SUsMap, during DAG construction. |
||
187 | /// Note: to gain speed it might be worth investigating an optimized |
||
188 | /// implementation of this data structure, such as a singly linked list |
||
189 | /// with a memory pool (SmallVector was tried but slow and SparseSet is not |
||
190 | /// applicable). |
||
191 | using SUList = std::list<SUnit *>; |
||
192 | |||
193 | protected: |
||
194 | /// A map from ValueType to SUList, used during DAG construction, as |
||
195 | /// a means of remembering which SUs depend on which memory locations. |
||
196 | class Value2SUsMap; |
||
197 | |||
198 | /// Reduces maps in FIFO order, by N SUs. This is better than turning |
||
199 | /// every Nth memory SU into BarrierChain in buildSchedGraph(), since |
||
200 | /// it avoids unnecessary edges between seen SUs above the new BarrierChain, |
||
201 | /// and those below it. |
||
202 | void reduceHugeMemNodeMaps(Value2SUsMap &stores, |
||
203 | Value2SUsMap &loads, unsigned N); |
||
204 | |||
205 | /// Adds a chain edge between SUa and SUb, but only if both |
||
206 | /// AAResults and Target fail to deny the dependency. |
||
207 | void addChainDependency(SUnit *SUa, SUnit *SUb, |
||
208 | unsigned Latency = 0); |
||
209 | |||
210 | /// Adds dependencies as needed from all SUs in list to SU. |
||
211 | void addChainDependencies(SUnit *SU, SUList &SUs, unsigned Latency) { |
||
212 | for (SUnit *Entry : SUs) |
||
213 | addChainDependency(SU, Entry, Latency); |
||
214 | } |
||
215 | |||
216 | /// Adds dependencies as needed from all SUs in map, to SU. |
||
217 | void addChainDependencies(SUnit *SU, Value2SUsMap &Val2SUsMap); |
||
218 | |||
219 | /// Adds dependencies as needed to SU, from all SUs mapped to V. |
||
220 | void addChainDependencies(SUnit *SU, Value2SUsMap &Val2SUsMap, |
||
221 | ValueType V); |
||
222 | |||
223 | /// Adds barrier chain edges from all SUs in map, and then clear the map. |
||
224 | /// This is equivalent to insertBarrierChain(), but optimized for the common |
||
225 | /// case where the new BarrierChain (a global memory object) has a higher |
||
226 | /// NodeNum than all SUs in map. It is assumed BarrierChain has been set |
||
227 | /// before calling this. |
||
228 | void addBarrierChain(Value2SUsMap &map); |
||
229 | |||
230 | /// Inserts a barrier chain in a huge region, far below current SU. |
||
231 | /// Adds barrier chain edges from all SUs in map with higher NodeNums than |
||
232 | /// this new BarrierChain, and remove them from map. It is assumed |
||
233 | /// BarrierChain has been set before calling this. |
||
234 | void insertBarrierChain(Value2SUsMap &map); |
||
235 | |||
236 | /// For an unanalyzable memory access, this Value is used in maps. |
||
237 | UndefValue *UnknownValue; |
||
238 | |||
239 | |||
240 | /// Topo - A topological ordering for SUnits which permits fast IsReachable |
||
241 | /// and similar queries. |
||
242 | ScheduleDAGTopologicalSort Topo; |
||
243 | |||
244 | using DbgValueVector = |
||
245 | std::vector<std::pair<MachineInstr *, MachineInstr *>>; |
||
246 | /// Remember instruction that precedes DBG_VALUE. |
||
247 | /// These are generated by buildSchedGraph but persist so they can be |
||
248 | /// referenced when emitting the final schedule. |
||
249 | DbgValueVector DbgValues; |
||
250 | MachineInstr *FirstDbgValue = nullptr; |
||
251 | |||
252 | /// Set of live physical registers for updating kill flags. |
||
253 | LivePhysRegs LiveRegs; |
||
254 | |||
255 | public: |
||
256 | explicit ScheduleDAGInstrs(MachineFunction &mf, |
||
257 | const MachineLoopInfo *mli, |
||
258 | bool RemoveKillFlags = false); |
||
259 | |||
260 | ~ScheduleDAGInstrs() override = default; |
||
261 | |||
262 | /// Gets the machine model for instruction scheduling. |
||
263 | const TargetSchedModel *getSchedModel() const { return &SchedModel; } |
||
264 | |||
265 | /// Resolves and cache a resolved scheduling class for an SUnit. |
||
266 | const MCSchedClassDesc *getSchedClass(SUnit *SU) const { |
||
267 | if (!SU->SchedClass && SchedModel.hasInstrSchedModel()) |
||
268 | SU->SchedClass = SchedModel.resolveSchedClass(SU->getInstr()); |
||
269 | return SU->SchedClass; |
||
270 | } |
||
271 | |||
272 | /// IsReachable - Checks if SU is reachable from TargetSU. |
||
273 | bool IsReachable(SUnit *SU, SUnit *TargetSU) { |
||
274 | return Topo.IsReachable(SU, TargetSU); |
||
275 | } |
||
276 | |||
277 | /// Returns an iterator to the top of the current scheduling region. |
||
278 | MachineBasicBlock::iterator begin() const { return RegionBegin; } |
||
279 | |||
280 | /// Returns an iterator to the bottom of the current scheduling region. |
||
281 | MachineBasicBlock::iterator end() const { return RegionEnd; } |
||
282 | |||
283 | /// Creates a new SUnit and return a ptr to it. |
||
284 | SUnit *newSUnit(MachineInstr *MI); |
||
285 | |||
286 | /// Returns an existing SUnit for this MI, or nullptr. |
||
287 | SUnit *getSUnit(MachineInstr *MI) const; |
||
288 | |||
289 | /// If this method returns true, handling of the scheduling regions |
||
290 | /// themselves (in case of a scheduling boundary in MBB) will be done |
||
291 | /// beginning with the topmost region of MBB. |
||
292 | virtual bool doMBBSchedRegionsTopDown() const { return false; } |
||
293 | |||
294 | /// Prepares to perform scheduling in the given block. |
||
295 | virtual void startBlock(MachineBasicBlock *BB); |
||
296 | |||
297 | /// Cleans up after scheduling in the given block. |
||
298 | virtual void finishBlock(); |
||
299 | |||
300 | /// Initialize the DAG and common scheduler state for a new |
||
301 | /// scheduling region. This does not actually create the DAG, only clears |
||
302 | /// it. The scheduling driver may call BuildSchedGraph multiple times per |
||
303 | /// scheduling region. |
||
304 | virtual void enterRegion(MachineBasicBlock *bb, |
||
305 | MachineBasicBlock::iterator begin, |
||
306 | MachineBasicBlock::iterator end, |
||
307 | unsigned regioninstrs); |
||
308 | |||
309 | /// Called when the scheduler has finished scheduling the current region. |
||
310 | virtual void exitRegion(); |
||
311 | |||
312 | /// Builds SUnits for the current region. |
||
313 | /// If \p RPTracker is non-null, compute register pressure as a side effect. |
||
314 | /// The DAG builder is an efficient place to do it because it already visits |
||
315 | /// operands. |
||
316 | void buildSchedGraph(AAResults *AA, |
||
317 | RegPressureTracker *RPTracker = nullptr, |
||
318 | PressureDiffs *PDiffs = nullptr, |
||
319 | LiveIntervals *LIS = nullptr, |
||
320 | bool TrackLaneMasks = false); |
||
321 | |||
322 | /// Adds dependencies from instructions in the current list of |
||
323 | /// instructions being scheduled to scheduling barrier. We want to make sure |
||
324 | /// instructions which define registers that are either used by the |
||
325 | /// terminator or are live-out are properly scheduled. This is especially |
||
326 | /// important when the definition latency of the return value(s) are too |
||
327 | /// high to be hidden by the branch or when the liveout registers used by |
||
328 | /// instructions in the fallthrough block. |
||
329 | void addSchedBarrierDeps(); |
||
330 | |||
331 | /// Orders nodes according to selected style. |
||
332 | /// |
||
333 | /// Typically, a scheduling algorithm will implement schedule() without |
||
334 | /// overriding enterRegion() or exitRegion(). |
||
335 | virtual void schedule() = 0; |
||
336 | |||
337 | /// Allow targets to perform final scheduling actions at the level of the |
||
338 | /// whole MachineFunction. By default does nothing. |
||
339 | virtual void finalizeSchedule() {} |
||
340 | |||
341 | void dumpNode(const SUnit &SU) const override; |
||
342 | void dump() const override; |
||
343 | |||
344 | /// Returns a label for a DAG node that points to an instruction. |
||
345 | std::string getGraphNodeLabel(const SUnit *SU) const override; |
||
346 | |||
347 | /// Returns a label for the region of code covered by the DAG. |
||
348 | std::string getDAGName() const override; |
||
349 | |||
350 | /// Fixes register kill flags that scheduling has made invalid. |
||
351 | void fixupKills(MachineBasicBlock &MBB); |
||
352 | |||
353 | /// True if an edge can be added from PredSU to SuccSU without creating |
||
354 | /// a cycle. |
||
355 | bool canAddEdge(SUnit *SuccSU, SUnit *PredSU); |
||
356 | |||
357 | /// Add a DAG edge to the given SU with the given predecessor |
||
358 | /// dependence data. |
||
359 | /// |
||
360 | /// \returns true if the edge may be added without creating a cycle OR if an |
||
361 | /// equivalent edge already existed (false indicates failure). |
||
362 | bool addEdge(SUnit *SuccSU, const SDep &PredDep); |
||
363 | |||
364 | protected: |
||
365 | void initSUnits(); |
||
366 | void addPhysRegDataDeps(SUnit *SU, unsigned OperIdx); |
||
367 | void addPhysRegDeps(SUnit *SU, unsigned OperIdx); |
||
368 | void addVRegDefDeps(SUnit *SU, unsigned OperIdx); |
||
369 | void addVRegUseDeps(SUnit *SU, unsigned OperIdx); |
||
370 | |||
371 | /// Returns a mask for which lanes get read/written by the given (register) |
||
372 | /// machine operand. |
||
373 | LaneBitmask getLaneMaskForMO(const MachineOperand &MO) const; |
||
374 | |||
375 | /// Returns true if the def register in \p MO has no uses. |
||
376 | bool deadDefHasNoUse(const MachineOperand &MO); |
||
377 | }; |
||
378 | |||
379 | /// Creates a new SUnit and return a ptr to it. |
||
380 | inline SUnit *ScheduleDAGInstrs::newSUnit(MachineInstr *MI) { |
||
381 | #ifndef NDEBUG |
||
382 | const SUnit *Addr = SUnits.empty() ? nullptr : &SUnits[0]; |
||
383 | #endif |
||
384 | SUnits.emplace_back(MI, (unsigned)SUnits.size()); |
||
385 | assert((Addr == nullptr || Addr == &SUnits[0]) && |
||
386 | "SUnits std::vector reallocated on the fly!"); |
||
387 | return &SUnits.back(); |
||
388 | } |
||
389 | |||
390 | /// Returns an existing SUnit for this MI, or nullptr. |
||
391 | inline SUnit *ScheduleDAGInstrs::getSUnit(MachineInstr *MI) const { |
||
392 | return MISUnitMap.lookup(MI); |
||
393 | } |
||
394 | |||
395 | } // end namespace llvm |
||
396 | |||
397 | #endif // LLVM_CODEGEN_SCHEDULEDAGINSTRS_H |