File: | lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp |
Location: | line 1424, column 11 |
Description: | Access to field 'isAvailable' results in a dereference of a null pointer (loaded from variable 'BtSU') |
1 | //===----- ScheduleDAGRRList.cpp - Reg pressure reduction list scheduler --===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This implements bottom-up and top-down register pressure reduction list | |||
11 | // schedulers, using standard algorithms. The basic approach uses a priority | |||
12 | // queue of available nodes to schedule. One at a time, nodes are taken from | |||
13 | // the priority queue (thus in priority order), checked for legality to | |||
14 | // schedule, and emitted if legal. | |||
15 | // | |||
16 | //===----------------------------------------------------------------------===// | |||
17 | ||||
18 | #include "llvm/CodeGen/SchedulerRegistry.h" | |||
19 | #include "ScheduleDAGSDNodes.h" | |||
20 | #include "llvm/ADT/STLExtras.h" | |||
21 | #include "llvm/ADT/SmallSet.h" | |||
22 | #include "llvm/ADT/Statistic.h" | |||
23 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
24 | #include "llvm/CodeGen/ScheduleHazardRecognizer.h" | |||
25 | #include "llvm/CodeGen/SelectionDAGISel.h" | |||
26 | #include "llvm/IR/DataLayout.h" | |||
27 | #include "llvm/IR/InlineAsm.h" | |||
28 | #include "llvm/Support/Debug.h" | |||
29 | #include "llvm/Support/ErrorHandling.h" | |||
30 | #include "llvm/Support/raw_ostream.h" | |||
31 | #include "llvm/Target/TargetInstrInfo.h" | |||
32 | #include "llvm/Target/TargetLowering.h" | |||
33 | #include "llvm/Target/TargetRegisterInfo.h" | |||
34 | #include "llvm/Target/TargetSubtargetInfo.h" | |||
35 | #include <climits> | |||
36 | using namespace llvm; | |||
37 | ||||
38 | #define DEBUG_TYPE"pre-RA-sched" "pre-RA-sched" | |||
39 | ||||
40 | STATISTIC(NumBacktracks, "Number of times scheduler backtracked")static llvm::Statistic NumBacktracks = { "pre-RA-sched", "Number of times scheduler backtracked" , 0, 0 }; | |||
41 | STATISTIC(NumUnfolds, "Number of nodes unfolded")static llvm::Statistic NumUnfolds = { "pre-RA-sched", "Number of nodes unfolded" , 0, 0 }; | |||
42 | STATISTIC(NumDups, "Number of duplicated nodes")static llvm::Statistic NumDups = { "pre-RA-sched", "Number of duplicated nodes" , 0, 0 }; | |||
43 | STATISTIC(NumPRCopies, "Number of physical register copies")static llvm::Statistic NumPRCopies = { "pre-RA-sched", "Number of physical register copies" , 0, 0 }; | |||
44 | ||||
45 | static RegisterScheduler | |||
46 | burrListDAGScheduler("list-burr", | |||
47 | "Bottom-up register reduction list scheduling", | |||
48 | createBURRListDAGScheduler); | |||
49 | static RegisterScheduler | |||
50 | sourceListDAGScheduler("source", | |||
51 | "Similar to list-burr but schedules in source " | |||
52 | "order when possible", | |||
53 | createSourceListDAGScheduler); | |||
54 | ||||
55 | static RegisterScheduler | |||
56 | hybridListDAGScheduler("list-hybrid", | |||
57 | "Bottom-up register pressure aware list scheduling " | |||
58 | "which tries to balance latency and register pressure", | |||
59 | createHybridListDAGScheduler); | |||
60 | ||||
61 | static RegisterScheduler | |||
62 | ILPListDAGScheduler("list-ilp", | |||
63 | "Bottom-up register pressure aware list scheduling " | |||
64 | "which tries to balance ILP and register pressure", | |||
65 | createILPListDAGScheduler); | |||
66 | ||||
67 | static cl::opt<bool> DisableSchedCycles( | |||
68 | "disable-sched-cycles", cl::Hidden, cl::init(false), | |||
69 | cl::desc("Disable cycle-level precision during preRA scheduling")); | |||
70 | ||||
71 | // Temporary sched=list-ilp flags until the heuristics are robust. | |||
72 | // Some options are also available under sched=list-hybrid. | |||
73 | static cl::opt<bool> DisableSchedRegPressure( | |||
74 | "disable-sched-reg-pressure", cl::Hidden, cl::init(false), | |||
75 | cl::desc("Disable regpressure priority in sched=list-ilp")); | |||
76 | static cl::opt<bool> DisableSchedLiveUses( | |||
77 | "disable-sched-live-uses", cl::Hidden, cl::init(true), | |||
78 | cl::desc("Disable live use priority in sched=list-ilp")); | |||
79 | static cl::opt<bool> DisableSchedVRegCycle( | |||
80 | "disable-sched-vrcycle", cl::Hidden, cl::init(false), | |||
81 | cl::desc("Disable virtual register cycle interference checks")); | |||
82 | static cl::opt<bool> DisableSchedPhysRegJoin( | |||
83 | "disable-sched-physreg-join", cl::Hidden, cl::init(false), | |||
84 | cl::desc("Disable physreg def-use affinity")); | |||
85 | static cl::opt<bool> DisableSchedStalls( | |||
86 | "disable-sched-stalls", cl::Hidden, cl::init(true), | |||
87 | cl::desc("Disable no-stall priority in sched=list-ilp")); | |||
88 | static cl::opt<bool> DisableSchedCriticalPath( | |||
89 | "disable-sched-critical-path", cl::Hidden, cl::init(false), | |||
90 | cl::desc("Disable critical path priority in sched=list-ilp")); | |||
91 | static cl::opt<bool> DisableSchedHeight( | |||
92 | "disable-sched-height", cl::Hidden, cl::init(false), | |||
93 | cl::desc("Disable scheduled-height priority in sched=list-ilp")); | |||
94 | static cl::opt<bool> Disable2AddrHack( | |||
95 | "disable-2addr-hack", cl::Hidden, cl::init(true), | |||
96 | cl::desc("Disable scheduler's two-address hack")); | |||
97 | ||||
98 | static cl::opt<int> MaxReorderWindow( | |||
99 | "max-sched-reorder", cl::Hidden, cl::init(6), | |||
100 | cl::desc("Number of instructions to allow ahead of the critical path " | |||
101 | "in sched=list-ilp")); | |||
102 | ||||
103 | static cl::opt<unsigned> AvgIPC( | |||
104 | "sched-avg-ipc", cl::Hidden, cl::init(1), | |||
105 | cl::desc("Average inst/cycle whan no target itinerary exists.")); | |||
106 | ||||
107 | namespace { | |||
108 | //===----------------------------------------------------------------------===// | |||
109 | /// ScheduleDAGRRList - The actual register reduction list scheduler | |||
110 | /// implementation. This supports both top-down and bottom-up scheduling. | |||
111 | /// | |||
112 | class ScheduleDAGRRList : public ScheduleDAGSDNodes { | |||
113 | private: | |||
114 | /// NeedLatency - True if the scheduler will make use of latency information. | |||
115 | /// | |||
116 | bool NeedLatency; | |||
117 | ||||
118 | /// AvailableQueue - The priority queue to use for the available SUnits. | |||
119 | SchedulingPriorityQueue *AvailableQueue; | |||
120 | ||||
121 | /// PendingQueue - This contains all of the instructions whose operands have | |||
122 | /// been issued, but their results are not ready yet (due to the latency of | |||
123 | /// the operation). Once the operands becomes available, the instruction is | |||
124 | /// added to the AvailableQueue. | |||
125 | std::vector<SUnit*> PendingQueue; | |||
126 | ||||
127 | /// HazardRec - The hazard recognizer to use. | |||
128 | ScheduleHazardRecognizer *HazardRec; | |||
129 | ||||
130 | /// CurCycle - The current scheduler state corresponds to this cycle. | |||
131 | unsigned CurCycle; | |||
132 | ||||
133 | /// MinAvailableCycle - Cycle of the soonest available instruction. | |||
134 | unsigned MinAvailableCycle; | |||
135 | ||||
136 | /// IssueCount - Count instructions issued in this cycle | |||
137 | /// Currently valid only for bottom-up scheduling. | |||
138 | unsigned IssueCount; | |||
139 | ||||
140 | /// LiveRegDefs - A set of physical registers and their definition | |||
141 | /// that are "live". These nodes must be scheduled before any other nodes that | |||
142 | /// modifies the registers can be scheduled. | |||
143 | unsigned NumLiveRegs; | |||
144 | std::unique_ptr<SUnit*[]> LiveRegDefs; | |||
145 | std::unique_ptr<SUnit*[]> LiveRegGens; | |||
146 | ||||
147 | // Collect interferences between physical register use/defs. | |||
148 | // Each interference is an SUnit and set of physical registers. | |||
149 | SmallVector<SUnit*, 4> Interferences; | |||
150 | typedef DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMapT; | |||
151 | LRegsMapT LRegsMap; | |||
152 | ||||
153 | /// Topo - A topological ordering for SUnits which permits fast IsReachable | |||
154 | /// and similar queries. | |||
155 | ScheduleDAGTopologicalSort Topo; | |||
156 | ||||
157 | // Hack to keep track of the inverse of FindCallSeqStart without more crazy | |||
158 | // DAG crawling. | |||
159 | DenseMap<SUnit*, SUnit*> CallSeqEndForStart; | |||
160 | ||||
161 | public: | |||
162 | ScheduleDAGRRList(MachineFunction &mf, bool needlatency, | |||
163 | SchedulingPriorityQueue *availqueue, | |||
164 | CodeGenOpt::Level OptLevel) | |||
165 | : ScheduleDAGSDNodes(mf), | |||
166 | NeedLatency(needlatency), AvailableQueue(availqueue), CurCycle(0), | |||
167 | Topo(SUnits, nullptr) { | |||
168 | ||||
169 | const TargetSubtargetInfo &STI = mf.getSubtarget(); | |||
170 | if (DisableSchedCycles || !NeedLatency) | |||
171 | HazardRec = new ScheduleHazardRecognizer(); | |||
172 | else | |||
173 | HazardRec = STI.getInstrInfo()->CreateTargetHazardRecognizer(&STI, this); | |||
174 | } | |||
175 | ||||
176 | ~ScheduleDAGRRList() override { | |||
177 | delete HazardRec; | |||
178 | delete AvailableQueue; | |||
179 | } | |||
180 | ||||
181 | void Schedule() override; | |||
182 | ||||
183 | ScheduleHazardRecognizer *getHazardRec() { return HazardRec; } | |||
184 | ||||
185 | /// IsReachable - Checks if SU is reachable from TargetSU. | |||
186 | bool IsReachable(const SUnit *SU, const SUnit *TargetSU) { | |||
187 | return Topo.IsReachable(SU, TargetSU); | |||
188 | } | |||
189 | ||||
190 | /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will | |||
191 | /// create a cycle. | |||
192 | bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) { | |||
193 | return Topo.WillCreateCycle(SU, TargetSU); | |||
194 | } | |||
195 | ||||
196 | /// AddPred - adds a predecessor edge to SUnit SU. | |||
197 | /// This returns true if this is a new predecessor. | |||
198 | /// Updates the topological ordering if required. | |||
199 | void AddPred(SUnit *SU, const SDep &D) { | |||
200 | Topo.AddPred(SU, D.getSUnit()); | |||
201 | SU->addPred(D); | |||
202 | } | |||
203 | ||||
204 | /// RemovePred - removes a predecessor edge from SUnit SU. | |||
205 | /// This returns true if an edge was removed. | |||
206 | /// Updates the topological ordering if required. | |||
207 | void RemovePred(SUnit *SU, const SDep &D) { | |||
208 | Topo.RemovePred(SU, D.getSUnit()); | |||
209 | SU->removePred(D); | |||
210 | } | |||
211 | ||||
212 | private: | |||
213 | bool isReady(SUnit *SU) { | |||
214 | return DisableSchedCycles || !AvailableQueue->hasReadyFilter() || | |||
215 | AvailableQueue->isReady(SU); | |||
216 | } | |||
217 | ||||
218 | void ReleasePred(SUnit *SU, const SDep *PredEdge); | |||
219 | void ReleasePredecessors(SUnit *SU); | |||
220 | void ReleasePending(); | |||
221 | void AdvanceToCycle(unsigned NextCycle); | |||
222 | void AdvancePastStalls(SUnit *SU); | |||
223 | void EmitNode(SUnit *SU); | |||
224 | void ScheduleNodeBottomUp(SUnit*); | |||
225 | void CapturePred(SDep *PredEdge); | |||
226 | void UnscheduleNodeBottomUp(SUnit*); | |||
227 | void RestoreHazardCheckerBottomUp(); | |||
228 | void BacktrackBottomUp(SUnit*, SUnit*); | |||
229 | SUnit *CopyAndMoveSuccessors(SUnit*); | |||
230 | void InsertCopiesAndMoveSuccs(SUnit*, unsigned, | |||
231 | const TargetRegisterClass*, | |||
232 | const TargetRegisterClass*, | |||
233 | SmallVectorImpl<SUnit*>&); | |||
234 | bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&); | |||
235 | ||||
236 | void releaseInterferences(unsigned Reg = 0); | |||
237 | ||||
238 | SUnit *PickNodeToScheduleBottomUp(); | |||
239 | void ListScheduleBottomUp(); | |||
240 | ||||
241 | /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it. | |||
242 | /// Updates the topological ordering if required. | |||
243 | SUnit *CreateNewSUnit(SDNode *N) { | |||
244 | unsigned NumSUnits = SUnits.size(); | |||
245 | SUnit *NewNode = newSUnit(N); | |||
246 | // Update the topological ordering. | |||
247 | if (NewNode->NodeNum >= NumSUnits) | |||
248 | Topo.InitDAGTopologicalSorting(); | |||
249 | return NewNode; | |||
250 | } | |||
251 | ||||
252 | /// CreateClone - Creates a new SUnit from an existing one. | |||
253 | /// Updates the topological ordering if required. | |||
254 | SUnit *CreateClone(SUnit *N) { | |||
255 | unsigned NumSUnits = SUnits.size(); | |||
256 | SUnit *NewNode = Clone(N); | |||
257 | // Update the topological ordering. | |||
258 | if (NewNode->NodeNum >= NumSUnits) | |||
259 | Topo.InitDAGTopologicalSorting(); | |||
260 | return NewNode; | |||
261 | } | |||
262 | ||||
263 | /// forceUnitLatencies - Register-pressure-reducing scheduling doesn't | |||
264 | /// need actual latency information but the hybrid scheduler does. | |||
265 | bool forceUnitLatencies() const override { | |||
266 | return !NeedLatency; | |||
267 | } | |||
268 | }; | |||
269 | } // end anonymous namespace | |||
270 | ||||
271 | /// GetCostForDef - Looks up the register class and cost for a given definition. | |||
272 | /// Typically this just means looking up the representative register class, | |||
273 | /// but for untyped values (MVT::Untyped) it means inspecting the node's | |||
274 | /// opcode to determine what register class is being generated. | |||
275 | static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos, | |||
276 | const TargetLowering *TLI, | |||
277 | const TargetInstrInfo *TII, | |||
278 | const TargetRegisterInfo *TRI, | |||
279 | unsigned &RegClass, unsigned &Cost, | |||
280 | const MachineFunction &MF) { | |||
281 | MVT VT = RegDefPos.GetValue(); | |||
282 | ||||
283 | // Special handling for untyped values. These values can only come from | |||
284 | // the expansion of custom DAG-to-DAG patterns. | |||
285 | if (VT == MVT::Untyped) { | |||
286 | const SDNode *Node = RegDefPos.GetNode(); | |||
287 | ||||
288 | // Special handling for CopyFromReg of untyped values. | |||
289 | if (!Node->isMachineOpcode() && Node->getOpcode() == ISD::CopyFromReg) { | |||
290 | unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg(); | |||
291 | const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(Reg); | |||
292 | RegClass = RC->getID(); | |||
293 | Cost = 1; | |||
294 | return; | |||
295 | } | |||
296 | ||||
297 | unsigned Opcode = Node->getMachineOpcode(); | |||
298 | if (Opcode == TargetOpcode::REG_SEQUENCE) { | |||
299 | unsigned DstRCIdx = cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue(); | |||
300 | const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx); | |||
301 | RegClass = RC->getID(); | |||
302 | Cost = 1; | |||
303 | return; | |||
304 | } | |||
305 | ||||
306 | unsigned Idx = RegDefPos.GetIdx(); | |||
307 | const MCInstrDesc Desc = TII->get(Opcode); | |||
308 | const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI, MF); | |||
309 | RegClass = RC->getID(); | |||
310 | // FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a | |||
311 | // better way to determine it. | |||
312 | Cost = 1; | |||
313 | } else { | |||
314 | RegClass = TLI->getRepRegClassFor(VT)->getID(); | |||
315 | Cost = TLI->getRepRegClassCostFor(VT); | |||
316 | } | |||
317 | } | |||
318 | ||||
319 | /// Schedule - Schedule the DAG using list scheduling. | |||
320 | void ScheduleDAGRRList::Schedule() { | |||
321 | DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "********** List Scheduling BB#" << BB->getNumber() << " '" << BB->getName () << "' **********\n"; } } while (0) | |||
322 | << "********** List Scheduling BB#" << BB->getNumber()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "********** List Scheduling BB#" << BB->getNumber() << " '" << BB->getName () << "' **********\n"; } } while (0) | |||
323 | << " '" << BB->getName() << "' **********\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "********** List Scheduling BB#" << BB->getNumber() << " '" << BB->getName () << "' **********\n"; } } while (0); | |||
324 | ||||
325 | CurCycle = 0; | |||
326 | IssueCount = 0; | |||
327 | MinAvailableCycle = DisableSchedCycles ? 0 : UINT_MAX(2147483647 *2U +1U); | |||
328 | NumLiveRegs = 0; | |||
329 | // Allocate slots for each physical register, plus one for a special register | |||
330 | // to track the virtual resource of a calling sequence. | |||
331 | LiveRegDefs.reset(new SUnit*[TRI->getNumRegs() + 1]()); | |||
332 | LiveRegGens.reset(new SUnit*[TRI->getNumRegs() + 1]()); | |||
333 | CallSeqEndForStart.clear(); | |||
334 | assert(Interferences.empty() && LRegsMap.empty() && "stale Interferences")((Interferences.empty() && LRegsMap.empty() && "stale Interferences") ? static_cast<void> (0) : __assert_fail ("Interferences.empty() && LRegsMap.empty() && \"stale Interferences\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 334, __PRETTY_FUNCTION__)); | |||
335 | ||||
336 | // Build the scheduling graph. | |||
337 | BuildSchedGraph(nullptr); | |||
338 | ||||
339 | DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { for (unsigned su = 0, e = SUnits.size(); su != e; ++su) SUnits[su].dumpAll(this); } } while (0) | |||
340 | SUnits[su].dumpAll(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { for (unsigned su = 0, e = SUnits.size(); su != e; ++su) SUnits[su].dumpAll(this); } } while (0); | |||
341 | Topo.InitDAGTopologicalSorting(); | |||
342 | ||||
343 | AvailableQueue->initNodes(SUnits); | |||
344 | ||||
345 | HazardRec->Reset(); | |||
346 | ||||
347 | // Execute the actual scheduling loop. | |||
348 | ListScheduleBottomUp(); | |||
349 | ||||
350 | AvailableQueue->releaseState(); | |||
351 | ||||
352 | DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (0) | |||
353 | dbgs() << "*** Final schedule ***\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (0) | |||
354 | dumpSchedule();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (0) | |||
355 | dbgs() << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (0) | |||
356 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (0); | |||
357 | } | |||
358 | ||||
359 | //===----------------------------------------------------------------------===// | |||
360 | // Bottom-Up Scheduling | |||
361 | //===----------------------------------------------------------------------===// | |||
362 | ||||
363 | /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to | |||
364 | /// the AvailableQueue if the count reaches zero. Also update its cycle bound. | |||
365 | void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) { | |||
366 | SUnit *PredSU = PredEdge->getSUnit(); | |||
367 | ||||
368 | #ifndef NDEBUG | |||
369 | if (PredSU->NumSuccsLeft == 0) { | |||
370 | dbgs() << "*** Scheduling failed! ***\n"; | |||
371 | PredSU->dump(this); | |||
372 | dbgs() << " has been released too many times!\n"; | |||
373 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 373); | |||
374 | } | |||
375 | #endif | |||
376 | --PredSU->NumSuccsLeft; | |||
377 | ||||
378 | if (!forceUnitLatencies()) { | |||
379 | // Updating predecessor's height. This is now the cycle when the | |||
380 | // predecessor can be scheduled without causing a pipeline stall. | |||
381 | PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency()); | |||
382 | } | |||
383 | ||||
384 | // If all the node's successors are scheduled, this node is ready | |||
385 | // to be scheduled. Ignore the special EntrySU node. | |||
386 | if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) { | |||
387 | PredSU->isAvailable = true; | |||
388 | ||||
389 | unsigned Height = PredSU->getHeight(); | |||
390 | if (Height < MinAvailableCycle) | |||
391 | MinAvailableCycle = Height; | |||
392 | ||||
393 | if (isReady(PredSU)) { | |||
394 | AvailableQueue->push(PredSU); | |||
395 | } | |||
396 | // CapturePred and others may have left the node in the pending queue, avoid | |||
397 | // adding it twice. | |||
398 | else if (!PredSU->isPending) { | |||
399 | PredSU->isPending = true; | |||
400 | PendingQueue.push_back(PredSU); | |||
401 | } | |||
402 | } | |||
403 | } | |||
404 | ||||
405 | /// IsChainDependent - Test if Outer is reachable from Inner through | |||
406 | /// chain dependencies. | |||
407 | static bool IsChainDependent(SDNode *Outer, SDNode *Inner, | |||
408 | unsigned NestLevel, | |||
409 | const TargetInstrInfo *TII) { | |||
410 | SDNode *N = Outer; | |||
411 | for (;;) { | |||
412 | if (N == Inner) | |||
413 | return true; | |||
414 | // For a TokenFactor, examine each operand. There may be multiple ways | |||
415 | // to get to the CALLSEQ_BEGIN, but we need to find the path with the | |||
416 | // most nesting in order to ensure that we find the corresponding match. | |||
417 | if (N->getOpcode() == ISD::TokenFactor) { | |||
418 | for (const SDValue &Op : N->op_values()) | |||
419 | if (IsChainDependent(Op.getNode(), Inner, NestLevel, TII)) | |||
420 | return true; | |||
421 | return false; | |||
422 | } | |||
423 | // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END. | |||
424 | if (N->isMachineOpcode()) { | |||
425 | if (N->getMachineOpcode() == | |||
426 | (unsigned)TII->getCallFrameDestroyOpcode()) { | |||
427 | ++NestLevel; | |||
428 | } else if (N->getMachineOpcode() == | |||
429 | (unsigned)TII->getCallFrameSetupOpcode()) { | |||
430 | if (NestLevel == 0) | |||
431 | return false; | |||
432 | --NestLevel; | |||
433 | } | |||
434 | } | |||
435 | // Otherwise, find the chain and continue climbing. | |||
436 | for (const SDValue &Op : N->op_values()) | |||
437 | if (Op.getValueType() == MVT::Other) { | |||
438 | N = Op.getNode(); | |||
439 | goto found_chain_operand; | |||
440 | } | |||
441 | return false; | |||
442 | found_chain_operand:; | |||
443 | if (N->getOpcode() == ISD::EntryToken) | |||
444 | return false; | |||
445 | } | |||
446 | } | |||
447 | ||||
448 | /// FindCallSeqStart - Starting from the (lowered) CALLSEQ_END node, locate | |||
449 | /// the corresponding (lowered) CALLSEQ_BEGIN node. | |||
450 | /// | |||
451 | /// NestLevel and MaxNested are used in recursion to indcate the current level | |||
452 | /// of nesting of CALLSEQ_BEGIN and CALLSEQ_END pairs, as well as the maximum | |||
453 | /// level seen so far. | |||
454 | /// | |||
455 | /// TODO: It would be better to give CALLSEQ_END an explicit operand to point | |||
456 | /// to the corresponding CALLSEQ_BEGIN to avoid needing to search for it. | |||
457 | static SDNode * | |||
458 | FindCallSeqStart(SDNode *N, unsigned &NestLevel, unsigned &MaxNest, | |||
459 | const TargetInstrInfo *TII) { | |||
460 | for (;;) { | |||
461 | // For a TokenFactor, examine each operand. There may be multiple ways | |||
462 | // to get to the CALLSEQ_BEGIN, but we need to find the path with the | |||
463 | // most nesting in order to ensure that we find the corresponding match. | |||
464 | if (N->getOpcode() == ISD::TokenFactor) { | |||
465 | SDNode *Best = nullptr; | |||
466 | unsigned BestMaxNest = MaxNest; | |||
467 | for (const SDValue &Op : N->op_values()) { | |||
468 | unsigned MyNestLevel = NestLevel; | |||
469 | unsigned MyMaxNest = MaxNest; | |||
470 | if (SDNode *New = FindCallSeqStart(Op.getNode(), | |||
471 | MyNestLevel, MyMaxNest, TII)) | |||
472 | if (!Best || (MyMaxNest > BestMaxNest)) { | |||
473 | Best = New; | |||
474 | BestMaxNest = MyMaxNest; | |||
475 | } | |||
476 | } | |||
477 | assert(Best)((Best) ? static_cast<void> (0) : __assert_fail ("Best" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 477, __PRETTY_FUNCTION__)); | |||
478 | MaxNest = BestMaxNest; | |||
479 | return Best; | |||
480 | } | |||
481 | // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END. | |||
482 | if (N->isMachineOpcode()) { | |||
483 | if (N->getMachineOpcode() == | |||
484 | (unsigned)TII->getCallFrameDestroyOpcode()) { | |||
485 | ++NestLevel; | |||
486 | MaxNest = std::max(MaxNest, NestLevel); | |||
487 | } else if (N->getMachineOpcode() == | |||
488 | (unsigned)TII->getCallFrameSetupOpcode()) { | |||
489 | assert(NestLevel != 0)((NestLevel != 0) ? static_cast<void> (0) : __assert_fail ("NestLevel != 0", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 489, __PRETTY_FUNCTION__)); | |||
490 | --NestLevel; | |||
491 | if (NestLevel == 0) | |||
492 | return N; | |||
493 | } | |||
494 | } | |||
495 | // Otherwise, find the chain and continue climbing. | |||
496 | for (const SDValue &Op : N->op_values()) | |||
497 | if (Op.getValueType() == MVT::Other) { | |||
498 | N = Op.getNode(); | |||
499 | goto found_chain_operand; | |||
500 | } | |||
501 | return nullptr; | |||
502 | found_chain_operand:; | |||
503 | if (N->getOpcode() == ISD::EntryToken) | |||
504 | return nullptr; | |||
505 | } | |||
506 | } | |||
507 | ||||
508 | /// Call ReleasePred for each predecessor, then update register live def/gen. | |||
509 | /// Always update LiveRegDefs for a register dependence even if the current SU | |||
510 | /// also defines the register. This effectively create one large live range | |||
511 | /// across a sequence of two-address node. This is important because the | |||
512 | /// entire chain must be scheduled together. Example: | |||
513 | /// | |||
514 | /// flags = (3) add | |||
515 | /// flags = (2) addc flags | |||
516 | /// flags = (1) addc flags | |||
517 | /// | |||
518 | /// results in | |||
519 | /// | |||
520 | /// LiveRegDefs[flags] = 3 | |||
521 | /// LiveRegGens[flags] = 1 | |||
522 | /// | |||
523 | /// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid | |||
524 | /// interference on flags. | |||
525 | void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) { | |||
526 | // Bottom up: release predecessors | |||
527 | for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
528 | I != E; ++I) { | |||
529 | ReleasePred(SU, &*I); | |||
530 | if (I->isAssignedRegDep()) { | |||
531 | // This is a physical register dependency and it's impossible or | |||
532 | // expensive to copy the register. Make sure nothing that can | |||
533 | // clobber the register is scheduled between the predecessor and | |||
534 | // this node. | |||
535 | SUnit *RegDef = LiveRegDefs[I->getReg()]; (void)RegDef; | |||
536 | assert((!RegDef || RegDef == SU || RegDef == I->getSUnit()) &&(((!RegDef || RegDef == SU || RegDef == I->getSUnit()) && "interference on register dependence") ? static_cast<void > (0) : __assert_fail ("(!RegDef || RegDef == SU || RegDef == I->getSUnit()) && \"interference on register dependence\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 537, __PRETTY_FUNCTION__)) | |||
537 | "interference on register dependence")(((!RegDef || RegDef == SU || RegDef == I->getSUnit()) && "interference on register dependence") ? static_cast<void > (0) : __assert_fail ("(!RegDef || RegDef == SU || RegDef == I->getSUnit()) && \"interference on register dependence\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 537, __PRETTY_FUNCTION__)); | |||
538 | LiveRegDefs[I->getReg()] = I->getSUnit(); | |||
539 | if (!LiveRegGens[I->getReg()]) { | |||
540 | ++NumLiveRegs; | |||
541 | LiveRegGens[I->getReg()] = SU; | |||
542 | } | |||
543 | } | |||
544 | } | |||
545 | ||||
546 | // If we're scheduling a lowered CALLSEQ_END, find the corresponding | |||
547 | // CALLSEQ_BEGIN. Inject an artificial physical register dependence between | |||
548 | // these nodes, to prevent other calls from being interscheduled with them. | |||
549 | unsigned CallResource = TRI->getNumRegs(); | |||
550 | if (!LiveRegDefs[CallResource]) | |||
551 | for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) | |||
552 | if (Node->isMachineOpcode() && | |||
553 | Node->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) { | |||
554 | unsigned NestLevel = 0; | |||
555 | unsigned MaxNest = 0; | |||
556 | SDNode *N = FindCallSeqStart(Node, NestLevel, MaxNest, TII); | |||
557 | ||||
558 | SUnit *Def = &SUnits[N->getNodeId()]; | |||
559 | CallSeqEndForStart[Def] = SU; | |||
560 | ||||
561 | ++NumLiveRegs; | |||
562 | LiveRegDefs[CallResource] = Def; | |||
563 | LiveRegGens[CallResource] = SU; | |||
564 | break; | |||
565 | } | |||
566 | } | |||
567 | ||||
568 | /// Check to see if any of the pending instructions are ready to issue. If | |||
569 | /// so, add them to the available queue. | |||
570 | void ScheduleDAGRRList::ReleasePending() { | |||
571 | if (DisableSchedCycles) { | |||
572 | assert(PendingQueue.empty() && "pending instrs not allowed in this mode")((PendingQueue.empty() && "pending instrs not allowed in this mode" ) ? static_cast<void> (0) : __assert_fail ("PendingQueue.empty() && \"pending instrs not allowed in this mode\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 572, __PRETTY_FUNCTION__)); | |||
573 | return; | |||
574 | } | |||
575 | ||||
576 | // If the available queue is empty, it is safe to reset MinAvailableCycle. | |||
577 | if (AvailableQueue->empty()) | |||
578 | MinAvailableCycle = UINT_MAX(2147483647 *2U +1U); | |||
579 | ||||
580 | // Check to see if any of the pending instructions are ready to issue. If | |||
581 | // so, add them to the available queue. | |||
582 | for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) { | |||
583 | unsigned ReadyCycle = PendingQueue[i]->getHeight(); | |||
584 | if (ReadyCycle < MinAvailableCycle) | |||
585 | MinAvailableCycle = ReadyCycle; | |||
586 | ||||
587 | if (PendingQueue[i]->isAvailable) { | |||
588 | if (!isReady(PendingQueue[i])) | |||
589 | continue; | |||
590 | AvailableQueue->push(PendingQueue[i]); | |||
591 | } | |||
592 | PendingQueue[i]->isPending = false; | |||
593 | PendingQueue[i] = PendingQueue.back(); | |||
594 | PendingQueue.pop_back(); | |||
595 | --i; --e; | |||
596 | } | |||
597 | } | |||
598 | ||||
599 | /// Move the scheduler state forward by the specified number of Cycles. | |||
600 | void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) { | |||
601 | if (NextCycle <= CurCycle) | |||
602 | return; | |||
603 | ||||
604 | IssueCount = 0; | |||
605 | AvailableQueue->setCurCycle(NextCycle); | |||
606 | if (!HazardRec->isEnabled()) { | |||
607 | // Bypass lots of virtual calls in case of long latency. | |||
608 | CurCycle = NextCycle; | |||
609 | } | |||
610 | else { | |||
611 | for (; CurCycle != NextCycle; ++CurCycle) { | |||
612 | HazardRec->RecedeCycle(); | |||
613 | } | |||
614 | } | |||
615 | // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the | |||
616 | // available Q to release pending nodes at least once before popping. | |||
617 | ReleasePending(); | |||
618 | } | |||
619 | ||||
620 | /// Move the scheduler state forward until the specified node's dependents are | |||
621 | /// ready and can be scheduled with no resource conflicts. | |||
622 | void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) { | |||
623 | if (DisableSchedCycles) | |||
624 | return; | |||
625 | ||||
626 | // FIXME: Nodes such as CopyFromReg probably should not advance the current | |||
627 | // cycle. Otherwise, we can wrongly mask real stalls. If the non-machine node | |||
628 | // has predecessors the cycle will be advanced when they are scheduled. | |||
629 | // But given the crude nature of modeling latency though such nodes, we | |||
630 | // currently need to treat these nodes like real instructions. | |||
631 | // if (!SU->getNode() || !SU->getNode()->isMachineOpcode()) return; | |||
632 | ||||
633 | unsigned ReadyCycle = SU->getHeight(); | |||
634 | ||||
635 | // Bump CurCycle to account for latency. We assume the latency of other | |||
636 | // available instructions may be hidden by the stall (not a full pipe stall). | |||
637 | // This updates the hazard recognizer's cycle before reserving resources for | |||
638 | // this instruction. | |||
639 | AdvanceToCycle(ReadyCycle); | |||
640 | ||||
641 | // Calls are scheduled in their preceding cycle, so don't conflict with | |||
642 | // hazards from instructions after the call. EmitNode will reset the | |||
643 | // scoreboard state before emitting the call. | |||
644 | if (SU->isCall) | |||
645 | return; | |||
646 | ||||
647 | // FIXME: For resource conflicts in very long non-pipelined stages, we | |||
648 | // should probably skip ahead here to avoid useless scoreboard checks. | |||
649 | int Stalls = 0; | |||
650 | while (true) { | |||
651 | ScheduleHazardRecognizer::HazardType HT = | |||
652 | HazardRec->getHazardType(SU, -Stalls); | |||
653 | ||||
654 | if (HT == ScheduleHazardRecognizer::NoHazard) | |||
655 | break; | |||
656 | ||||
657 | ++Stalls; | |||
658 | } | |||
659 | AdvanceToCycle(CurCycle + Stalls); | |||
660 | } | |||
661 | ||||
662 | /// Record this SUnit in the HazardRecognizer. | |||
663 | /// Does not update CurCycle. | |||
664 | void ScheduleDAGRRList::EmitNode(SUnit *SU) { | |||
665 | if (!HazardRec->isEnabled()) | |||
666 | return; | |||
667 | ||||
668 | // Check for phys reg copy. | |||
669 | if (!SU->getNode()) | |||
670 | return; | |||
671 | ||||
672 | switch (SU->getNode()->getOpcode()) { | |||
673 | default: | |||
674 | assert(SU->getNode()->isMachineOpcode() &&((SU->getNode()->isMachineOpcode() && "This target-independent node should not be scheduled." ) ? static_cast<void> (0) : __assert_fail ("SU->getNode()->isMachineOpcode() && \"This target-independent node should not be scheduled.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 675, __PRETTY_FUNCTION__)) | |||
675 | "This target-independent node should not be scheduled.")((SU->getNode()->isMachineOpcode() && "This target-independent node should not be scheduled." ) ? static_cast<void> (0) : __assert_fail ("SU->getNode()->isMachineOpcode() && \"This target-independent node should not be scheduled.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 675, __PRETTY_FUNCTION__)); | |||
676 | break; | |||
677 | case ISD::MERGE_VALUES: | |||
678 | case ISD::TokenFactor: | |||
679 | case ISD::LIFETIME_START: | |||
680 | case ISD::LIFETIME_END: | |||
681 | case ISD::CopyToReg: | |||
682 | case ISD::CopyFromReg: | |||
683 | case ISD::EH_LABEL: | |||
684 | // Noops don't affect the scoreboard state. Copies are likely to be | |||
685 | // removed. | |||
686 | return; | |||
687 | case ISD::INLINEASM: | |||
688 | // For inline asm, clear the pipeline state. | |||
689 | HazardRec->Reset(); | |||
690 | return; | |||
691 | } | |||
692 | if (SU->isCall) { | |||
693 | // Calls are scheduled with their preceding instructions. For bottom-up | |||
694 | // scheduling, clear the pipeline state before emitting. | |||
695 | HazardRec->Reset(); | |||
696 | } | |||
697 | ||||
698 | HazardRec->EmitInstruction(SU); | |||
699 | } | |||
700 | ||||
701 | static void resetVRegCycle(SUnit *SU); | |||
702 | ||||
703 | /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending | |||
704 | /// count of its predecessors. If a predecessor pending count is zero, add it to | |||
705 | /// the Available queue. | |||
706 | void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) { | |||
707 | DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "\n*** Scheduling [" << CurCycle << "]: "; } } while (0); | |||
708 | DEBUG(SU->dump(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { SU->dump(this); } } while (0); | |||
709 | ||||
710 | #ifndef NDEBUG | |||
711 | if (CurCycle < SU->getHeight()) | |||
712 | DEBUG(dbgs() << " Height [" << SU->getHeight()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Height [" << SU ->getHeight() << "] pipeline stall!\n"; } } while (0 ) | |||
713 | << "] pipeline stall!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Height [" << SU ->getHeight() << "] pipeline stall!\n"; } } while (0 ); | |||
714 | #endif | |||
715 | ||||
716 | // FIXME: Do not modify node height. It may interfere with | |||
717 | // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the | |||
718 | // node its ready cycle can aid heuristics, and after scheduling it can | |||
719 | // indicate the scheduled cycle. | |||
720 | SU->setHeightToAtLeast(CurCycle); | |||
721 | ||||
722 | // Reserve resources for the scheduled instruction. | |||
723 | EmitNode(SU); | |||
724 | ||||
725 | Sequence.push_back(SU); | |||
726 | ||||
727 | AvailableQueue->scheduledNode(SU); | |||
728 | ||||
729 | // If HazardRec is disabled, and each inst counts as one cycle, then | |||
730 | // advance CurCycle before ReleasePredecessors to avoid useless pushes to | |||
731 | // PendingQueue for schedulers that implement HasReadyFilter. | |||
732 | if (!HazardRec->isEnabled() && AvgIPC < 2) | |||
733 | AdvanceToCycle(CurCycle + 1); | |||
734 | ||||
735 | // Update liveness of predecessors before successors to avoid treating a | |||
736 | // two-address node as a live range def. | |||
737 | ReleasePredecessors(SU); | |||
738 | ||||
739 | // Release all the implicit physical register defs that are live. | |||
740 | for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); | |||
741 | I != E; ++I) { | |||
742 | // LiveRegDegs[I->getReg()] != SU when SU is a two-address node. | |||
743 | if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] == SU) { | |||
744 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")((NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? static_cast<void> (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 744, __PRETTY_FUNCTION__)); | |||
745 | --NumLiveRegs; | |||
746 | LiveRegDefs[I->getReg()] = nullptr; | |||
747 | LiveRegGens[I->getReg()] = nullptr; | |||
748 | releaseInterferences(I->getReg()); | |||
749 | } | |||
750 | } | |||
751 | // Release the special call resource dependence, if this is the beginning | |||
752 | // of a call. | |||
753 | unsigned CallResource = TRI->getNumRegs(); | |||
754 | if (LiveRegDefs[CallResource] == SU) | |||
755 | for (const SDNode *SUNode = SU->getNode(); SUNode; | |||
756 | SUNode = SUNode->getGluedNode()) { | |||
757 | if (SUNode->isMachineOpcode() && | |||
758 | SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode()) { | |||
759 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")((NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? static_cast<void> (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 759, __PRETTY_FUNCTION__)); | |||
760 | --NumLiveRegs; | |||
761 | LiveRegDefs[CallResource] = nullptr; | |||
762 | LiveRegGens[CallResource] = nullptr; | |||
763 | releaseInterferences(CallResource); | |||
764 | } | |||
765 | } | |||
766 | ||||
767 | resetVRegCycle(SU); | |||
768 | ||||
769 | SU->isScheduled = true; | |||
770 | ||||
771 | // Conditions under which the scheduler should eagerly advance the cycle: | |||
772 | // (1) No available instructions | |||
773 | // (2) All pipelines full, so available instructions must have hazards. | |||
774 | // | |||
775 | // If HazardRec is disabled, the cycle was pre-advanced before calling | |||
776 | // ReleasePredecessors. In that case, IssueCount should remain 0. | |||
777 | // | |||
778 | // Check AvailableQueue after ReleasePredecessors in case of zero latency. | |||
779 | if (HazardRec->isEnabled() || AvgIPC > 1) { | |||
780 | if (SU->getNode() && SU->getNode()->isMachineOpcode()) | |||
781 | ++IssueCount; | |||
782 | if ((HazardRec->isEnabled() && HazardRec->atIssueLimit()) | |||
783 | || (!HazardRec->isEnabled() && IssueCount == AvgIPC)) | |||
784 | AdvanceToCycle(CurCycle + 1); | |||
785 | } | |||
786 | } | |||
787 | ||||
788 | /// CapturePred - This does the opposite of ReleasePred. Since SU is being | |||
789 | /// unscheduled, incrcease the succ left count of its predecessors. Remove | |||
790 | /// them from AvailableQueue if necessary. | |||
791 | void ScheduleDAGRRList::CapturePred(SDep *PredEdge) { | |||
792 | SUnit *PredSU = PredEdge->getSUnit(); | |||
793 | if (PredSU->isAvailable) { | |||
794 | PredSU->isAvailable = false; | |||
795 | if (!PredSU->isPending) | |||
796 | AvailableQueue->remove(PredSU); | |||
797 | } | |||
798 | ||||
799 | assert(PredSU->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!")((PredSU->NumSuccsLeft < (2147483647 *2U +1U) && "NumSuccsLeft will overflow!") ? static_cast<void> (0) : __assert_fail ("PredSU->NumSuccsLeft < (2147483647 *2U +1U) && \"NumSuccsLeft will overflow!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 799, __PRETTY_FUNCTION__)); | |||
800 | ++PredSU->NumSuccsLeft; | |||
801 | } | |||
802 | ||||
803 | /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and | |||
804 | /// its predecessor states to reflect the change. | |||
805 | void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) { | |||
806 | DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: "; } } while (0); | |||
807 | DEBUG(SU->dump(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { SU->dump(this); } } while (0); | |||
808 | ||||
809 | for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
810 | I != E; ++I) { | |||
811 | CapturePred(&*I); | |||
812 | if (I->isAssignedRegDep() && SU == LiveRegGens[I->getReg()]){ | |||
813 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")((NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? static_cast<void> (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 813, __PRETTY_FUNCTION__)); | |||
814 | assert(LiveRegDefs[I->getReg()] == I->getSUnit() &&((LiveRegDefs[I->getReg()] == I->getSUnit() && "Physical register dependency violated?" ) ? static_cast<void> (0) : __assert_fail ("LiveRegDefs[I->getReg()] == I->getSUnit() && \"Physical register dependency violated?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 815, __PRETTY_FUNCTION__)) | |||
815 | "Physical register dependency violated?")((LiveRegDefs[I->getReg()] == I->getSUnit() && "Physical register dependency violated?" ) ? static_cast<void> (0) : __assert_fail ("LiveRegDefs[I->getReg()] == I->getSUnit() && \"Physical register dependency violated?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 815, __PRETTY_FUNCTION__)); | |||
816 | --NumLiveRegs; | |||
817 | LiveRegDefs[I->getReg()] = nullptr; | |||
818 | LiveRegGens[I->getReg()] = nullptr; | |||
819 | releaseInterferences(I->getReg()); | |||
820 | } | |||
821 | } | |||
822 | ||||
823 | // Reclaim the special call resource dependence, if this is the beginning | |||
824 | // of a call. | |||
825 | unsigned CallResource = TRI->getNumRegs(); | |||
826 | for (const SDNode *SUNode = SU->getNode(); SUNode; | |||
827 | SUNode = SUNode->getGluedNode()) { | |||
828 | if (SUNode->isMachineOpcode() && | |||
829 | SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode()) { | |||
830 | ++NumLiveRegs; | |||
831 | LiveRegDefs[CallResource] = SU; | |||
832 | LiveRegGens[CallResource] = CallSeqEndForStart[SU]; | |||
833 | } | |||
834 | } | |||
835 | ||||
836 | // Release the special call resource dependence, if this is the end | |||
837 | // of a call. | |||
838 | if (LiveRegGens[CallResource] == SU) | |||
839 | for (const SDNode *SUNode = SU->getNode(); SUNode; | |||
840 | SUNode = SUNode->getGluedNode()) { | |||
841 | if (SUNode->isMachineOpcode() && | |||
842 | SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) { | |||
843 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")((NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? static_cast<void> (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 843, __PRETTY_FUNCTION__)); | |||
844 | --NumLiveRegs; | |||
845 | LiveRegDefs[CallResource] = nullptr; | |||
846 | LiveRegGens[CallResource] = nullptr; | |||
847 | releaseInterferences(CallResource); | |||
848 | } | |||
849 | } | |||
850 | ||||
851 | for (auto &Succ : SU->Succs) { | |||
852 | if (Succ.isAssignedRegDep()) { | |||
853 | auto Reg = Succ.getReg(); | |||
854 | if (!LiveRegDefs[Reg]) | |||
855 | ++NumLiveRegs; | |||
856 | // This becomes the nearest def. Note that an earlier def may still be | |||
857 | // pending if this is a two-address node. | |||
858 | LiveRegDefs[Reg] = SU; | |||
859 | ||||
860 | // Update LiveRegGen only if was empty before this unscheduling. | |||
861 | // This is to avoid incorrect updating LiveRegGen set in previous run. | |||
862 | if (!LiveRegGens[Reg]) { | |||
863 | // Find the successor with the lowest height. | |||
864 | LiveRegGens[Reg] = Succ.getSUnit(); | |||
865 | for (auto &Succ2 : SU->Succs) { | |||
866 | if (Succ2.isAssignedRegDep() && Succ2.getReg() == Reg && | |||
867 | Succ2.getSUnit()->getHeight() < LiveRegGens[Reg]->getHeight()) | |||
868 | LiveRegGens[Reg] = Succ2.getSUnit(); | |||
869 | } | |||
870 | } | |||
871 | } | |||
872 | } | |||
873 | if (SU->getHeight() < MinAvailableCycle) | |||
874 | MinAvailableCycle = SU->getHeight(); | |||
875 | ||||
876 | SU->setHeightDirty(); | |||
877 | SU->isScheduled = false; | |||
878 | SU->isAvailable = true; | |||
879 | if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) { | |||
880 | // Don't make available until backtracking is complete. | |||
881 | SU->isPending = true; | |||
882 | PendingQueue.push_back(SU); | |||
883 | } | |||
884 | else { | |||
885 | AvailableQueue->push(SU); | |||
886 | } | |||
887 | AvailableQueue->unscheduledNode(SU); | |||
888 | } | |||
889 | ||||
890 | /// After backtracking, the hazard checker needs to be restored to a state | |||
891 | /// corresponding the current cycle. | |||
892 | void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() { | |||
893 | HazardRec->Reset(); | |||
894 | ||||
895 | unsigned LookAhead = std::min((unsigned)Sequence.size(), | |||
896 | HazardRec->getMaxLookAhead()); | |||
897 | if (LookAhead == 0) | |||
898 | return; | |||
899 | ||||
900 | std::vector<SUnit*>::const_iterator I = (Sequence.end() - LookAhead); | |||
901 | unsigned HazardCycle = (*I)->getHeight(); | |||
902 | for (std::vector<SUnit*>::const_iterator E = Sequence.end(); I != E; ++I) { | |||
903 | SUnit *SU = *I; | |||
904 | for (; SU->getHeight() > HazardCycle; ++HazardCycle) { | |||
905 | HazardRec->RecedeCycle(); | |||
906 | } | |||
907 | EmitNode(SU); | |||
908 | } | |||
909 | } | |||
910 | ||||
911 | /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in | |||
912 | /// BTCycle in order to schedule a specific node. | |||
913 | void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) { | |||
914 | SUnit *OldSU = Sequence.back(); | |||
915 | while (true) { | |||
916 | Sequence.pop_back(); | |||
917 | // FIXME: use ready cycle instead of height | |||
918 | CurCycle = OldSU->getHeight(); | |||
919 | UnscheduleNodeBottomUp(OldSU); | |||
920 | AvailableQueue->setCurCycle(CurCycle); | |||
921 | if (OldSU == BtSU) | |||
922 | break; | |||
923 | OldSU = Sequence.back(); | |||
924 | } | |||
925 | ||||
926 | assert(!SU->isSucc(OldSU) && "Something is wrong!")((!SU->isSucc(OldSU) && "Something is wrong!") ? static_cast <void> (0) : __assert_fail ("!SU->isSucc(OldSU) && \"Something is wrong!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 926, __PRETTY_FUNCTION__)); | |||
927 | ||||
928 | RestoreHazardCheckerBottomUp(); | |||
929 | ||||
930 | ReleasePending(); | |||
931 | ||||
932 | ++NumBacktracks; | |||
933 | } | |||
934 | ||||
935 | static bool isOperandOf(const SUnit *SU, SDNode *N) { | |||
936 | for (const SDNode *SUNode = SU->getNode(); SUNode; | |||
937 | SUNode = SUNode->getGluedNode()) { | |||
938 | if (SUNode->isOperandOf(N)) | |||
939 | return true; | |||
940 | } | |||
941 | return false; | |||
942 | } | |||
943 | ||||
944 | /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled | |||
945 | /// successors to the newly created node. | |||
946 | SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) { | |||
947 | SDNode *N = SU->getNode(); | |||
948 | if (!N) | |||
949 | return nullptr; | |||
950 | ||||
951 | if (SU->getNode()->getGluedNode()) | |||
952 | return nullptr; | |||
953 | ||||
954 | SUnit *NewSU; | |||
955 | bool TryUnfold = false; | |||
956 | for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) { | |||
957 | MVT VT = N->getSimpleValueType(i); | |||
958 | if (VT == MVT::Glue) | |||
959 | return nullptr; | |||
960 | else if (VT == MVT::Other) | |||
961 | TryUnfold = true; | |||
962 | } | |||
963 | for (const SDValue &Op : N->op_values()) { | |||
964 | MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo()); | |||
965 | if (VT == MVT::Glue) | |||
966 | return nullptr; | |||
967 | } | |||
968 | ||||
969 | if (TryUnfold) { | |||
970 | SmallVector<SDNode*, 2> NewNodes; | |||
971 | if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes)) | |||
972 | return nullptr; | |||
973 | ||||
974 | // unfolding an x86 DEC64m operation results in store, dec, load which | |||
975 | // can't be handled here so quit | |||
976 | if (NewNodes.size() == 3) | |||
977 | return nullptr; | |||
978 | ||||
979 | DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Unfolding SU #" << SU->NodeNum << "\n"; } } while (0); | |||
980 | assert(NewNodes.size() == 2 && "Expected a load folding node!")((NewNodes.size() == 2 && "Expected a load folding node!" ) ? static_cast<void> (0) : __assert_fail ("NewNodes.size() == 2 && \"Expected a load folding node!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 980, __PRETTY_FUNCTION__)); | |||
981 | ||||
982 | N = NewNodes[1]; | |||
983 | SDNode *LoadNode = NewNodes[0]; | |||
984 | unsigned NumVals = N->getNumValues(); | |||
985 | unsigned OldNumVals = SU->getNode()->getNumValues(); | |||
986 | for (unsigned i = 0; i != NumVals; ++i) | |||
987 | DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i)); | |||
988 | DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1), | |||
989 | SDValue(LoadNode, 1)); | |||
990 | ||||
991 | // LoadNode may already exist. This can happen when there is another | |||
992 | // load from the same location and producing the same type of value | |||
993 | // but it has different alignment or volatileness. | |||
994 | bool isNewLoad = true; | |||
995 | SUnit *LoadSU; | |||
996 | if (LoadNode->getNodeId() != -1) { | |||
997 | LoadSU = &SUnits[LoadNode->getNodeId()]; | |||
998 | isNewLoad = false; | |||
999 | } else { | |||
1000 | LoadSU = CreateNewSUnit(LoadNode); | |||
1001 | LoadNode->setNodeId(LoadSU->NodeNum); | |||
1002 | ||||
1003 | InitNumRegDefsLeft(LoadSU); | |||
1004 | computeLatency(LoadSU); | |||
1005 | } | |||
1006 | ||||
1007 | SUnit *NewSU = CreateNewSUnit(N); | |||
1008 | assert(N->getNodeId() == -1 && "Node already inserted!")((N->getNodeId() == -1 && "Node already inserted!" ) ? static_cast<void> (0) : __assert_fail ("N->getNodeId() == -1 && \"Node already inserted!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1008, __PRETTY_FUNCTION__)); | |||
1009 | N->setNodeId(NewSU->NodeNum); | |||
1010 | ||||
1011 | const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); | |||
1012 | for (unsigned i = 0; i != MCID.getNumOperands(); ++i) { | |||
1013 | if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) { | |||
1014 | NewSU->isTwoAddress = true; | |||
1015 | break; | |||
1016 | } | |||
1017 | } | |||
1018 | if (MCID.isCommutable()) | |||
1019 | NewSU->isCommutable = true; | |||
1020 | ||||
1021 | InitNumRegDefsLeft(NewSU); | |||
1022 | computeLatency(NewSU); | |||
1023 | ||||
1024 | // Record all the edges to and from the old SU, by category. | |||
1025 | SmallVector<SDep, 4> ChainPreds; | |||
1026 | SmallVector<SDep, 4> ChainSuccs; | |||
1027 | SmallVector<SDep, 4> LoadPreds; | |||
1028 | SmallVector<SDep, 4> NodePreds; | |||
1029 | SmallVector<SDep, 4> NodeSuccs; | |||
1030 | for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
1031 | I != E; ++I) { | |||
1032 | if (I->isCtrl()) | |||
1033 | ChainPreds.push_back(*I); | |||
1034 | else if (isOperandOf(I->getSUnit(), LoadNode)) | |||
1035 | LoadPreds.push_back(*I); | |||
1036 | else | |||
1037 | NodePreds.push_back(*I); | |||
1038 | } | |||
1039 | for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); | |||
1040 | I != E; ++I) { | |||
1041 | if (I->isCtrl()) | |||
1042 | ChainSuccs.push_back(*I); | |||
1043 | else | |||
1044 | NodeSuccs.push_back(*I); | |||
1045 | } | |||
1046 | ||||
1047 | // Now assign edges to the newly-created nodes. | |||
1048 | for (unsigned i = 0, e = ChainPreds.size(); i != e; ++i) { | |||
1049 | const SDep &Pred = ChainPreds[i]; | |||
1050 | RemovePred(SU, Pred); | |||
1051 | if (isNewLoad) | |||
1052 | AddPred(LoadSU, Pred); | |||
1053 | } | |||
1054 | for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) { | |||
1055 | const SDep &Pred = LoadPreds[i]; | |||
1056 | RemovePred(SU, Pred); | |||
1057 | if (isNewLoad) | |||
1058 | AddPred(LoadSU, Pred); | |||
1059 | } | |||
1060 | for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) { | |||
1061 | const SDep &Pred = NodePreds[i]; | |||
1062 | RemovePred(SU, Pred); | |||
1063 | AddPred(NewSU, Pred); | |||
1064 | } | |||
1065 | for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) { | |||
1066 | SDep D = NodeSuccs[i]; | |||
1067 | SUnit *SuccDep = D.getSUnit(); | |||
1068 | D.setSUnit(SU); | |||
1069 | RemovePred(SuccDep, D); | |||
1070 | D.setSUnit(NewSU); | |||
1071 | AddPred(SuccDep, D); | |||
1072 | // Balance register pressure. | |||
1073 | if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled | |||
1074 | && !D.isCtrl() && NewSU->NumRegDefsLeft > 0) | |||
1075 | --NewSU->NumRegDefsLeft; | |||
1076 | } | |||
1077 | for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) { | |||
1078 | SDep D = ChainSuccs[i]; | |||
1079 | SUnit *SuccDep = D.getSUnit(); | |||
1080 | D.setSUnit(SU); | |||
1081 | RemovePred(SuccDep, D); | |||
1082 | if (isNewLoad) { | |||
1083 | D.setSUnit(LoadSU); | |||
1084 | AddPred(SuccDep, D); | |||
1085 | } | |||
1086 | } | |||
1087 | ||||
1088 | // Add a data dependency to reflect that NewSU reads the value defined | |||
1089 | // by LoadSU. | |||
1090 | SDep D(LoadSU, SDep::Data, 0); | |||
1091 | D.setLatency(LoadSU->Latency); | |||
1092 | AddPred(NewSU, D); | |||
1093 | ||||
1094 | if (isNewLoad) | |||
1095 | AvailableQueue->addNode(LoadSU); | |||
1096 | AvailableQueue->addNode(NewSU); | |||
1097 | ||||
1098 | ++NumUnfolds; | |||
1099 | ||||
1100 | if (NewSU->NumSuccsLeft == 0) { | |||
1101 | NewSU->isAvailable = true; | |||
1102 | return NewSU; | |||
1103 | } | |||
1104 | SU = NewSU; | |||
1105 | } | |||
1106 | ||||
1107 | DEBUG(dbgs() << " Duplicating SU #" << SU->NodeNum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Duplicating SU #" << SU->NodeNum << "\n"; } } while (0); | |||
1108 | NewSU = CreateClone(SU); | |||
1109 | ||||
1110 | // New SUnit has the exact same predecessors. | |||
1111 | for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
1112 | I != E; ++I) | |||
1113 | if (!I->isArtificial()) | |||
1114 | AddPred(NewSU, *I); | |||
1115 | ||||
1116 | // Only copy scheduled successors. Cut them from old node's successor | |||
1117 | // list and move them over. | |||
1118 | SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps; | |||
1119 | for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); | |||
1120 | I != E; ++I) { | |||
1121 | if (I->isArtificial()) | |||
1122 | continue; | |||
1123 | SUnit *SuccSU = I->getSUnit(); | |||
1124 | if (SuccSU->isScheduled) { | |||
1125 | SDep D = *I; | |||
1126 | D.setSUnit(NewSU); | |||
1127 | AddPred(SuccSU, D); | |||
1128 | D.setSUnit(SU); | |||
1129 | DelDeps.push_back(std::make_pair(SuccSU, D)); | |||
1130 | } | |||
1131 | } | |||
1132 | for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) | |||
1133 | RemovePred(DelDeps[i].first, DelDeps[i].second); | |||
1134 | ||||
1135 | AvailableQueue->updateNode(SU); | |||
1136 | AvailableQueue->addNode(NewSU); | |||
1137 | ||||
1138 | ++NumDups; | |||
1139 | return NewSU; | |||
1140 | } | |||
1141 | ||||
1142 | /// InsertCopiesAndMoveSuccs - Insert register copies and move all | |||
1143 | /// scheduled successors of the given SUnit to the last copy. | |||
1144 | void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg, | |||
1145 | const TargetRegisterClass *DestRC, | |||
1146 | const TargetRegisterClass *SrcRC, | |||
1147 | SmallVectorImpl<SUnit*> &Copies) { | |||
1148 | SUnit *CopyFromSU = CreateNewSUnit(nullptr); | |||
1149 | CopyFromSU->CopySrcRC = SrcRC; | |||
1150 | CopyFromSU->CopyDstRC = DestRC; | |||
1151 | ||||
1152 | SUnit *CopyToSU = CreateNewSUnit(nullptr); | |||
1153 | CopyToSU->CopySrcRC = DestRC; | |||
1154 | CopyToSU->CopyDstRC = SrcRC; | |||
1155 | ||||
1156 | // Only copy scheduled successors. Cut them from old node's successor | |||
1157 | // list and move them over. | |||
1158 | SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps; | |||
1159 | for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); | |||
1160 | I != E; ++I) { | |||
1161 | if (I->isArtificial()) | |||
1162 | continue; | |||
1163 | SUnit *SuccSU = I->getSUnit(); | |||
1164 | if (SuccSU->isScheduled) { | |||
1165 | SDep D = *I; | |||
1166 | D.setSUnit(CopyToSU); | |||
1167 | AddPred(SuccSU, D); | |||
1168 | DelDeps.push_back(std::make_pair(SuccSU, *I)); | |||
1169 | } | |||
1170 | else { | |||
1171 | // Avoid scheduling the def-side copy before other successors. Otherwise | |||
1172 | // we could introduce another physreg interference on the copy and | |||
1173 | // continue inserting copies indefinitely. | |||
1174 | AddPred(SuccSU, SDep(CopyFromSU, SDep::Artificial)); | |||
1175 | } | |||
1176 | } | |||
1177 | for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) | |||
1178 | RemovePred(DelDeps[i].first, DelDeps[i].second); | |||
1179 | ||||
1180 | SDep FromDep(SU, SDep::Data, Reg); | |||
1181 | FromDep.setLatency(SU->Latency); | |||
1182 | AddPred(CopyFromSU, FromDep); | |||
1183 | SDep ToDep(CopyFromSU, SDep::Data, 0); | |||
1184 | ToDep.setLatency(CopyFromSU->Latency); | |||
1185 | AddPred(CopyToSU, ToDep); | |||
1186 | ||||
1187 | AvailableQueue->updateNode(SU); | |||
1188 | AvailableQueue->addNode(CopyFromSU); | |||
1189 | AvailableQueue->addNode(CopyToSU); | |||
1190 | Copies.push_back(CopyFromSU); | |||
1191 | Copies.push_back(CopyToSU); | |||
1192 | ||||
1193 | ++NumPRCopies; | |||
1194 | } | |||
1195 | ||||
1196 | /// getPhysicalRegisterVT - Returns the ValueType of the physical register | |||
1197 | /// definition of the specified node. | |||
1198 | /// FIXME: Move to SelectionDAG? | |||
1199 | static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg, | |||
1200 | const TargetInstrInfo *TII) { | |||
1201 | unsigned NumRes; | |||
1202 | if (N->getOpcode() == ISD::CopyFromReg) { | |||
1203 | // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type. | |||
1204 | NumRes = 1; | |||
1205 | } else { | |||
1206 | const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); | |||
1207 | assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!")((MCID.ImplicitDefs && "Physical reg def must be in implicit def list!" ) ? static_cast<void> (0) : __assert_fail ("MCID.ImplicitDefs && \"Physical reg def must be in implicit def list!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1207, __PRETTY_FUNCTION__)); | |||
1208 | NumRes = MCID.getNumDefs(); | |||
1209 | for (const MCPhysReg *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) { | |||
1210 | if (Reg == *ImpDef) | |||
1211 | break; | |||
1212 | ++NumRes; | |||
1213 | } | |||
1214 | } | |||
1215 | return N->getSimpleValueType(NumRes); | |||
1216 | } | |||
1217 | ||||
1218 | /// CheckForLiveRegDef - Return true and update live register vector if the | |||
1219 | /// specified register def of the specified SUnit clobbers any "live" registers. | |||
1220 | static void CheckForLiveRegDef(SUnit *SU, unsigned Reg, | |||
1221 | SUnit **LiveRegDefs, | |||
1222 | SmallSet<unsigned, 4> &RegAdded, | |||
1223 | SmallVectorImpl<unsigned> &LRegs, | |||
1224 | const TargetRegisterInfo *TRI) { | |||
1225 | for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) { | |||
1226 | ||||
1227 | // Check if Ref is live. | |||
1228 | if (!LiveRegDefs[*AliasI]) continue; | |||
1229 | ||||
1230 | // Allow multiple uses of the same def. | |||
1231 | if (LiveRegDefs[*AliasI] == SU) continue; | |||
1232 | ||||
1233 | // Add Reg to the set of interfering live regs. | |||
1234 | if (RegAdded.insert(*AliasI).second) { | |||
1235 | LRegs.push_back(*AliasI); | |||
1236 | } | |||
1237 | } | |||
1238 | } | |||
1239 | ||||
1240 | /// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered | |||
1241 | /// by RegMask, and add them to LRegs. | |||
1242 | static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask, | |||
1243 | ArrayRef<SUnit*> LiveRegDefs, | |||
1244 | SmallSet<unsigned, 4> &RegAdded, | |||
1245 | SmallVectorImpl<unsigned> &LRegs) { | |||
1246 | // Look at all live registers. Skip Reg0 and the special CallResource. | |||
1247 | for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) { | |||
1248 | if (!LiveRegDefs[i]) continue; | |||
1249 | if (LiveRegDefs[i] == SU) continue; | |||
1250 | if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue; | |||
1251 | if (RegAdded.insert(i).second) | |||
1252 | LRegs.push_back(i); | |||
1253 | } | |||
1254 | } | |||
1255 | ||||
1256 | /// getNodeRegMask - Returns the register mask attached to an SDNode, if any. | |||
1257 | static const uint32_t *getNodeRegMask(const SDNode *N) { | |||
1258 | for (const SDValue &Op : N->op_values()) | |||
1259 | if (const auto *RegOp = dyn_cast<RegisterMaskSDNode>(Op.getNode())) | |||
1260 | return RegOp->getRegMask(); | |||
1261 | return nullptr; | |||
1262 | } | |||
1263 | ||||
1264 | /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay | |||
1265 | /// scheduling of the given node to satisfy live physical register dependencies. | |||
1266 | /// If the specific node is the last one that's available to schedule, do | |||
1267 | /// whatever is necessary (i.e. backtracking or cloning) to make it possible. | |||
1268 | bool ScheduleDAGRRList:: | |||
1269 | DelayForLiveRegsBottomUp(SUnit *SU, SmallVectorImpl<unsigned> &LRegs) { | |||
1270 | if (NumLiveRegs == 0) | |||
1271 | return false; | |||
1272 | ||||
1273 | SmallSet<unsigned, 4> RegAdded; | |||
1274 | // If this node would clobber any "live" register, then it's not ready. | |||
1275 | // | |||
1276 | // If SU is the currently live definition of the same register that it uses, | |||
1277 | // then we are free to schedule it. | |||
1278 | for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
1279 | I != E; ++I) { | |||
1280 | if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] != SU) | |||
1281 | CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs.get(), | |||
1282 | RegAdded, LRegs, TRI); | |||
1283 | } | |||
1284 | ||||
1285 | for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) { | |||
1286 | if (Node->getOpcode() == ISD::INLINEASM) { | |||
1287 | // Inline asm can clobber physical defs. | |||
1288 | unsigned NumOps = Node->getNumOperands(); | |||
1289 | if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue) | |||
1290 | --NumOps; // Ignore the glue operand. | |||
1291 | ||||
1292 | for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) { | |||
1293 | unsigned Flags = | |||
1294 | cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue(); | |||
1295 | unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags); | |||
1296 | ||||
1297 | ++i; // Skip the ID value. | |||
1298 | if (InlineAsm::isRegDefKind(Flags) || | |||
1299 | InlineAsm::isRegDefEarlyClobberKind(Flags) || | |||
1300 | InlineAsm::isClobberKind(Flags)) { | |||
1301 | // Check for def of register or earlyclobber register. | |||
1302 | for (; NumVals; --NumVals, ++i) { | |||
1303 | unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg(); | |||
1304 | if (TargetRegisterInfo::isPhysicalRegister(Reg)) | |||
1305 | CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI); | |||
1306 | } | |||
1307 | } else | |||
1308 | i += NumVals; | |||
1309 | } | |||
1310 | continue; | |||
1311 | } | |||
1312 | ||||
1313 | if (!Node->isMachineOpcode()) | |||
1314 | continue; | |||
1315 | // If we're in the middle of scheduling a call, don't begin scheduling | |||
1316 | // another call. Also, don't allow any physical registers to be live across | |||
1317 | // the call. | |||
1318 | if (Node->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) { | |||
1319 | // Check the special calling-sequence resource. | |||
1320 | unsigned CallResource = TRI->getNumRegs(); | |||
1321 | if (LiveRegDefs[CallResource]) { | |||
1322 | SDNode *Gen = LiveRegGens[CallResource]->getNode(); | |||
1323 | while (SDNode *Glued = Gen->getGluedNode()) | |||
1324 | Gen = Glued; | |||
1325 | if (!IsChainDependent(Gen, Node, 0, TII) && | |||
1326 | RegAdded.insert(CallResource).second) | |||
1327 | LRegs.push_back(CallResource); | |||
1328 | } | |||
1329 | } | |||
1330 | if (const uint32_t *RegMask = getNodeRegMask(Node)) | |||
1331 | CheckForLiveRegDefMasked(SU, RegMask, | |||
1332 | makeArrayRef(LiveRegDefs.get(), TRI->getNumRegs()), | |||
1333 | RegAdded, LRegs); | |||
1334 | ||||
1335 | const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode()); | |||
1336 | if (!MCID.ImplicitDefs) | |||
1337 | continue; | |||
1338 | for (const MCPhysReg *Reg = MCID.getImplicitDefs(); *Reg; ++Reg) | |||
1339 | CheckForLiveRegDef(SU, *Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI); | |||
1340 | } | |||
1341 | ||||
1342 | return !LRegs.empty(); | |||
1343 | } | |||
1344 | ||||
1345 | void ScheduleDAGRRList::releaseInterferences(unsigned Reg) { | |||
1346 | // Add the nodes that aren't ready back onto the available list. | |||
1347 | for (unsigned i = Interferences.size(); i > 0; --i) { | |||
1348 | SUnit *SU = Interferences[i-1]; | |||
1349 | LRegsMapT::iterator LRegsPos = LRegsMap.find(SU); | |||
1350 | if (Reg) { | |||
1351 | SmallVectorImpl<unsigned> &LRegs = LRegsPos->second; | |||
1352 | if (std::find(LRegs.begin(), LRegs.end(), Reg) == LRegs.end()) | |||
1353 | continue; | |||
1354 | } | |||
1355 | SU->isPending = false; | |||
1356 | // The interfering node may no longer be available due to backtracking. | |||
1357 | // Furthermore, it may have been made available again, in which case it is | |||
1358 | // now already in the AvailableQueue. | |||
1359 | if (SU->isAvailable && !SU->NodeQueueId) { | |||
1360 | DEBUG(dbgs() << " Repushing SU #" << SU->NodeNum << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Repushing SU #" << SU->NodeNum << '\n'; } } while (0); | |||
1361 | AvailableQueue->push(SU); | |||
1362 | } | |||
1363 | if (i < Interferences.size()) | |||
1364 | Interferences[i-1] = Interferences.back(); | |||
1365 | Interferences.pop_back(); | |||
1366 | LRegsMap.erase(LRegsPos); | |||
1367 | } | |||
1368 | } | |||
1369 | ||||
1370 | /// Return a node that can be scheduled in this cycle. Requirements: | |||
1371 | /// (1) Ready: latency has been satisfied | |||
1372 | /// (2) No Hazards: resources are available | |||
1373 | /// (3) No Interferences: may unschedule to break register interferences. | |||
1374 | SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() { | |||
1375 | SUnit *CurSU = AvailableQueue->empty() ? nullptr : AvailableQueue->pop(); | |||
| ||||
1376 | while (CurSU) { | |||
1377 | SmallVector<unsigned, 4> LRegs; | |||
1378 | if (!DelayForLiveRegsBottomUp(CurSU, LRegs)) | |||
1379 | break; | |||
1380 | DEBUG(dbgs() << " Interfering reg " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg " << (LRegs[0] == TRI->getNumRegs() ? "CallResource" : TRI-> getName(LRegs[0])) << " SU #" << CurSU->NodeNum << '\n'; } } while (0) | |||
1381 | (LRegs[0] == TRI->getNumRegs() ? "CallResource"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg " << (LRegs[0] == TRI->getNumRegs() ? "CallResource" : TRI-> getName(LRegs[0])) << " SU #" << CurSU->NodeNum << '\n'; } } while (0) | |||
1382 | : TRI->getName(LRegs[0]))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg " << (LRegs[0] == TRI->getNumRegs() ? "CallResource" : TRI-> getName(LRegs[0])) << " SU #" << CurSU->NodeNum << '\n'; } } while (0) | |||
1383 | << " SU #" << CurSU->NodeNum << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg " << (LRegs[0] == TRI->getNumRegs() ? "CallResource" : TRI-> getName(LRegs[0])) << " SU #" << CurSU->NodeNum << '\n'; } } while (0); | |||
1384 | std::pair<LRegsMapT::iterator, bool> LRegsPair = | |||
1385 | LRegsMap.insert(std::make_pair(CurSU, LRegs)); | |||
1386 | if (LRegsPair.second) { | |||
1387 | CurSU->isPending = true; // This SU is not in AvailableQueue right now. | |||
1388 | Interferences.push_back(CurSU); | |||
1389 | } | |||
1390 | else { | |||
1391 | assert(CurSU->isPending && "Interferences are pending")((CurSU->isPending && "Interferences are pending") ? static_cast<void> (0) : __assert_fail ("CurSU->isPending && \"Interferences are pending\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1391, __PRETTY_FUNCTION__)); | |||
1392 | // Update the interference with current live regs. | |||
1393 | LRegsPair.first->second = LRegs; | |||
1394 | } | |||
1395 | CurSU = AvailableQueue->pop(); | |||
1396 | } | |||
1397 | if (CurSU) | |||
1398 | return CurSU; | |||
1399 | ||||
1400 | // All candidates are delayed due to live physical reg dependencies. | |||
1401 | // Try backtracking, code duplication, or inserting cross class copies | |||
1402 | // to resolve it. | |||
1403 | for (unsigned i = 0, e = Interferences.size(); i != e; ++i) { | |||
1404 | SUnit *TrySU = Interferences[i]; | |||
1405 | SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU]; | |||
1406 | ||||
1407 | // Try unscheduling up to the point where it's safe to schedule | |||
1408 | // this node. | |||
1409 | SUnit *BtSU = nullptr; | |||
1410 | unsigned LiveCycle = UINT_MAX(2147483647 *2U +1U); | |||
1411 | for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) { | |||
1412 | unsigned Reg = LRegs[j]; | |||
1413 | if (LiveRegGens[Reg]->getHeight() < LiveCycle) { | |||
1414 | BtSU = LiveRegGens[Reg]; | |||
1415 | LiveCycle = BtSU->getHeight(); | |||
1416 | } | |||
1417 | } | |||
1418 | if (!WillCreateCycle(TrySU, BtSU)) { | |||
1419 | // BacktrackBottomUp mutates Interferences! | |||
1420 | BacktrackBottomUp(TrySU, BtSU); | |||
1421 | ||||
1422 | // Force the current node to be scheduled before the node that | |||
1423 | // requires the physical reg dep. | |||
1424 | if (BtSU->isAvailable) { | |||
| ||||
1425 | BtSU->isAvailable = false; | |||
1426 | if (!BtSU->isPending) | |||
1427 | AvailableQueue->remove(BtSU); | |||
1428 | } | |||
1429 | DEBUG(dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum << ") to SU("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum << ") to SU(" << TrySU ->NodeNum << ")\n"; } } while (0) | |||
1430 | << TrySU->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum << ") to SU(" << TrySU ->NodeNum << ")\n"; } } while (0); | |||
1431 | AddPred(TrySU, SDep(BtSU, SDep::Artificial)); | |||
1432 | ||||
1433 | // If one or more successors has been unscheduled, then the current | |||
1434 | // node is no longer available. | |||
1435 | if (!TrySU->isAvailable || !TrySU->NodeQueueId) | |||
1436 | CurSU = AvailableQueue->pop(); | |||
1437 | else { | |||
1438 | // Available and in AvailableQueue | |||
1439 | AvailableQueue->remove(TrySU); | |||
1440 | CurSU = TrySU; | |||
1441 | } | |||
1442 | // Interferences has been mutated. We must break. | |||
1443 | break; | |||
1444 | } | |||
1445 | } | |||
1446 | ||||
1447 | if (!CurSU) { | |||
1448 | // Can't backtrack. If it's too expensive to copy the value, then try | |||
1449 | // duplicate the nodes that produces these "too expensive to copy" | |||
1450 | // values to break the dependency. In case even that doesn't work, | |||
1451 | // insert cross class copies. | |||
1452 | // If it's not too expensive, i.e. cost != -1, issue copies. | |||
1453 | SUnit *TrySU = Interferences[0]; | |||
1454 | SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU]; | |||
1455 | assert(LRegs.size() == 1 && "Can't handle this yet!")((LRegs.size() == 1 && "Can't handle this yet!") ? static_cast <void> (0) : __assert_fail ("LRegs.size() == 1 && \"Can't handle this yet!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1455, __PRETTY_FUNCTION__)); | |||
1456 | unsigned Reg = LRegs[0]; | |||
1457 | SUnit *LRDef = LiveRegDefs[Reg]; | |||
1458 | MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII); | |||
1459 | const TargetRegisterClass *RC = | |||
1460 | TRI->getMinimalPhysRegClass(Reg, VT); | |||
1461 | const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC); | |||
1462 | ||||
1463 | // If cross copy register class is the same as RC, then it must be possible | |||
1464 | // copy the value directly. Do not try duplicate the def. | |||
1465 | // If cross copy register class is not the same as RC, then it's possible to | |||
1466 | // copy the value but it require cross register class copies and it is | |||
1467 | // expensive. | |||
1468 | // If cross copy register class is null, then it's not possible to copy | |||
1469 | // the value at all. | |||
1470 | SUnit *NewDef = nullptr; | |||
1471 | if (DestRC != RC) { | |||
1472 | NewDef = CopyAndMoveSuccessors(LRDef); | |||
1473 | if (!DestRC && !NewDef) | |||
1474 | report_fatal_error("Can't handle live physical register dependency!"); | |||
1475 | } | |||
1476 | if (!NewDef) { | |||
1477 | // Issue copies, these can be expensive cross register class copies. | |||
1478 | SmallVector<SUnit*, 2> Copies; | |||
1479 | InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies); | |||
1480 | DEBUG(dbgs() << " Adding an edge from SU #" << TrySU->NodeNumdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Adding an edge from SU #" << TrySU->NodeNum << " to SU #" << Copies .front()->NodeNum << "\n"; } } while (0) | |||
1481 | << " to SU #" << Copies.front()->NodeNum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Adding an edge from SU #" << TrySU->NodeNum << " to SU #" << Copies .front()->NodeNum << "\n"; } } while (0); | |||
1482 | AddPred(TrySU, SDep(Copies.front(), SDep::Artificial)); | |||
1483 | NewDef = Copies.back(); | |||
1484 | } | |||
1485 | ||||
1486 | DEBUG(dbgs() << " Adding an edge from SU #" << NewDef->NodeNumdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Adding an edge from SU #" << NewDef->NodeNum << " to SU #" << TrySU ->NodeNum << "\n"; } } while (0) | |||
1487 | << " to SU #" << TrySU->NodeNum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Adding an edge from SU #" << NewDef->NodeNum << " to SU #" << TrySU ->NodeNum << "\n"; } } while (0); | |||
1488 | LiveRegDefs[Reg] = NewDef; | |||
1489 | AddPred(NewDef, SDep(TrySU, SDep::Artificial)); | |||
1490 | TrySU->isAvailable = false; | |||
1491 | CurSU = NewDef; | |||
1492 | } | |||
1493 | assert(CurSU && "Unable to resolve live physical register dependencies!")((CurSU && "Unable to resolve live physical register dependencies!" ) ? static_cast<void> (0) : __assert_fail ("CurSU && \"Unable to resolve live physical register dependencies!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1493, __PRETTY_FUNCTION__)); | |||
1494 | return CurSU; | |||
1495 | } | |||
1496 | ||||
1497 | /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up | |||
1498 | /// schedulers. | |||
1499 | void ScheduleDAGRRList::ListScheduleBottomUp() { | |||
1500 | // Release any predecessors of the special Exit node. | |||
1501 | ReleasePredecessors(&ExitSU); | |||
1502 | ||||
1503 | // Add root to Available queue. | |||
1504 | if (!SUnits.empty()) { | |||
1505 | SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()]; | |||
1506 | assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!")((RootSU->Succs.empty() && "Graph root shouldn't have successors!" ) ? static_cast<void> (0) : __assert_fail ("RootSU->Succs.empty() && \"Graph root shouldn't have successors!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1506, __PRETTY_FUNCTION__)); | |||
1507 | RootSU->isAvailable = true; | |||
1508 | AvailableQueue->push(RootSU); | |||
1509 | } | |||
1510 | ||||
1511 | // While Available queue is not empty, grab the node with the highest | |||
1512 | // priority. If it is not ready put it back. Schedule the node. | |||
1513 | Sequence.reserve(SUnits.size()); | |||
1514 | while (!AvailableQueue->empty() || !Interferences.empty()) { | |||
1515 | DEBUG(dbgs() << "\nExamining Available:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "\nExamining Available:\n" ; AvailableQueue->dump(this); } } while (0) | |||
1516 | AvailableQueue->dump(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "\nExamining Available:\n" ; AvailableQueue->dump(this); } } while (0); | |||
1517 | ||||
1518 | // Pick the best node to schedule taking all constraints into | |||
1519 | // consideration. | |||
1520 | SUnit *SU = PickNodeToScheduleBottomUp(); | |||
1521 | ||||
1522 | AdvancePastStalls(SU); | |||
1523 | ||||
1524 | ScheduleNodeBottomUp(SU); | |||
1525 | ||||
1526 | while (AvailableQueue->empty() && !PendingQueue.empty()) { | |||
1527 | // Advance the cycle to free resources. Skip ahead to the next ready SU. | |||
1528 | assert(MinAvailableCycle < UINT_MAX && "MinAvailableCycle uninitialized")((MinAvailableCycle < (2147483647 *2U +1U) && "MinAvailableCycle uninitialized" ) ? static_cast<void> (0) : __assert_fail ("MinAvailableCycle < (2147483647 *2U +1U) && \"MinAvailableCycle uninitialized\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1528, __PRETTY_FUNCTION__)); | |||
1529 | AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle)); | |||
1530 | } | |||
1531 | } | |||
1532 | ||||
1533 | // Reverse the order if it is bottom up. | |||
1534 | std::reverse(Sequence.begin(), Sequence.end()); | |||
1535 | ||||
1536 | #ifndef NDEBUG | |||
1537 | VerifyScheduledSequence(/*isBottomUp=*/true); | |||
1538 | #endif | |||
1539 | } | |||
1540 | ||||
1541 | //===----------------------------------------------------------------------===// | |||
1542 | // RegReductionPriorityQueue Definition | |||
1543 | //===----------------------------------------------------------------------===// | |||
1544 | // | |||
1545 | // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers | |||
1546 | // to reduce register pressure. | |||
1547 | // | |||
1548 | namespace { | |||
1549 | class RegReductionPQBase; | |||
1550 | ||||
1551 | struct queue_sort : public std::binary_function<SUnit*, SUnit*, bool> { | |||
1552 | bool isReady(SUnit* SU, unsigned CurCycle) const { return true; } | |||
1553 | }; | |||
1554 | ||||
1555 | #ifndef NDEBUG | |||
1556 | template<class SF> | |||
1557 | struct reverse_sort : public queue_sort { | |||
1558 | SF &SortFunc; | |||
1559 | reverse_sort(SF &sf) : SortFunc(sf) {} | |||
1560 | ||||
1561 | bool operator()(SUnit* left, SUnit* right) const { | |||
1562 | // reverse left/right rather than simply !SortFunc(left, right) | |||
1563 | // to expose different paths in the comparison logic. | |||
1564 | return SortFunc(right, left); | |||
1565 | } | |||
1566 | }; | |||
1567 | #endif // NDEBUG | |||
1568 | ||||
1569 | /// bu_ls_rr_sort - Priority function for bottom up register pressure | |||
1570 | // reduction scheduler. | |||
1571 | struct bu_ls_rr_sort : public queue_sort { | |||
1572 | enum { | |||
1573 | IsBottomUp = true, | |||
1574 | HasReadyFilter = false | |||
1575 | }; | |||
1576 | ||||
1577 | RegReductionPQBase *SPQ; | |||
1578 | bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {} | |||
1579 | ||||
1580 | bool operator()(SUnit* left, SUnit* right) const; | |||
1581 | }; | |||
1582 | ||||
1583 | // src_ls_rr_sort - Priority function for source order scheduler. | |||
1584 | struct src_ls_rr_sort : public queue_sort { | |||
1585 | enum { | |||
1586 | IsBottomUp = true, | |||
1587 | HasReadyFilter = false | |||
1588 | }; | |||
1589 | ||||
1590 | RegReductionPQBase *SPQ; | |||
1591 | src_ls_rr_sort(RegReductionPQBase *spq) | |||
1592 | : SPQ(spq) {} | |||
1593 | ||||
1594 | bool operator()(SUnit* left, SUnit* right) const; | |||
1595 | }; | |||
1596 | ||||
1597 | // hybrid_ls_rr_sort - Priority function for hybrid scheduler. | |||
1598 | struct hybrid_ls_rr_sort : public queue_sort { | |||
1599 | enum { | |||
1600 | IsBottomUp = true, | |||
1601 | HasReadyFilter = false | |||
1602 | }; | |||
1603 | ||||
1604 | RegReductionPQBase *SPQ; | |||
1605 | hybrid_ls_rr_sort(RegReductionPQBase *spq) | |||
1606 | : SPQ(spq) {} | |||
1607 | ||||
1608 | bool isReady(SUnit *SU, unsigned CurCycle) const; | |||
1609 | ||||
1610 | bool operator()(SUnit* left, SUnit* right) const; | |||
1611 | }; | |||
1612 | ||||
1613 | // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism) | |||
1614 | // scheduler. | |||
1615 | struct ilp_ls_rr_sort : public queue_sort { | |||
1616 | enum { | |||
1617 | IsBottomUp = true, | |||
1618 | HasReadyFilter = false | |||
1619 | }; | |||
1620 | ||||
1621 | RegReductionPQBase *SPQ; | |||
1622 | ilp_ls_rr_sort(RegReductionPQBase *spq) | |||
1623 | : SPQ(spq) {} | |||
1624 | ||||
1625 | bool isReady(SUnit *SU, unsigned CurCycle) const; | |||
1626 | ||||
1627 | bool operator()(SUnit* left, SUnit* right) const; | |||
1628 | }; | |||
1629 | ||||
1630 | class RegReductionPQBase : public SchedulingPriorityQueue { | |||
1631 | protected: | |||
1632 | std::vector<SUnit*> Queue; | |||
1633 | unsigned CurQueueId; | |||
1634 | bool TracksRegPressure; | |||
1635 | bool SrcOrder; | |||
1636 | ||||
1637 | // SUnits - The SUnits for the current graph. | |||
1638 | std::vector<SUnit> *SUnits; | |||
1639 | ||||
1640 | MachineFunction &MF; | |||
1641 | const TargetInstrInfo *TII; | |||
1642 | const TargetRegisterInfo *TRI; | |||
1643 | const TargetLowering *TLI; | |||
1644 | ScheduleDAGRRList *scheduleDAG; | |||
1645 | ||||
1646 | // SethiUllmanNumbers - The SethiUllman number for each node. | |||
1647 | std::vector<unsigned> SethiUllmanNumbers; | |||
1648 | ||||
1649 | /// RegPressure - Tracking current reg pressure per register class. | |||
1650 | /// | |||
1651 | std::vector<unsigned> RegPressure; | |||
1652 | ||||
1653 | /// RegLimit - Tracking the number of allocatable registers per register | |||
1654 | /// class. | |||
1655 | std::vector<unsigned> RegLimit; | |||
1656 | ||||
1657 | public: | |||
1658 | RegReductionPQBase(MachineFunction &mf, | |||
1659 | bool hasReadyFilter, | |||
1660 | bool tracksrp, | |||
1661 | bool srcorder, | |||
1662 | const TargetInstrInfo *tii, | |||
1663 | const TargetRegisterInfo *tri, | |||
1664 | const TargetLowering *tli) | |||
1665 | : SchedulingPriorityQueue(hasReadyFilter), | |||
1666 | CurQueueId(0), TracksRegPressure(tracksrp), SrcOrder(srcorder), | |||
1667 | MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(nullptr) { | |||
1668 | if (TracksRegPressure) { | |||
1669 | unsigned NumRC = TRI->getNumRegClasses(); | |||
1670 | RegLimit.resize(NumRC); | |||
1671 | RegPressure.resize(NumRC); | |||
1672 | std::fill(RegLimit.begin(), RegLimit.end(), 0); | |||
1673 | std::fill(RegPressure.begin(), RegPressure.end(), 0); | |||
1674 | for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(), | |||
1675 | E = TRI->regclass_end(); I != E; ++I) | |||
1676 | RegLimit[(*I)->getID()] = tri->getRegPressureLimit(*I, MF); | |||
1677 | } | |||
1678 | } | |||
1679 | ||||
1680 | void setScheduleDAG(ScheduleDAGRRList *scheduleDag) { | |||
1681 | scheduleDAG = scheduleDag; | |||
1682 | } | |||
1683 | ||||
1684 | ScheduleHazardRecognizer* getHazardRec() { | |||
1685 | return scheduleDAG->getHazardRec(); | |||
1686 | } | |||
1687 | ||||
1688 | void initNodes(std::vector<SUnit> &sunits) override; | |||
1689 | ||||
1690 | void addNode(const SUnit *SU) override; | |||
1691 | ||||
1692 | void updateNode(const SUnit *SU) override; | |||
1693 | ||||
1694 | void releaseState() override { | |||
1695 | SUnits = nullptr; | |||
1696 | SethiUllmanNumbers.clear(); | |||
1697 | std::fill(RegPressure.begin(), RegPressure.end(), 0); | |||
1698 | } | |||
1699 | ||||
1700 | unsigned getNodePriority(const SUnit *SU) const; | |||
1701 | ||||
1702 | unsigned getNodeOrdering(const SUnit *SU) const { | |||
1703 | if (!SU->getNode()) return 0; | |||
1704 | ||||
1705 | return SU->getNode()->getIROrder(); | |||
1706 | } | |||
1707 | ||||
1708 | bool empty() const override { return Queue.empty(); } | |||
1709 | ||||
1710 | void push(SUnit *U) override { | |||
1711 | assert(!U->NodeQueueId && "Node in the queue already")((!U->NodeQueueId && "Node in the queue already") ? static_cast<void> (0) : __assert_fail ("!U->NodeQueueId && \"Node in the queue already\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1711, __PRETTY_FUNCTION__)); | |||
1712 | U->NodeQueueId = ++CurQueueId; | |||
1713 | Queue.push_back(U); | |||
1714 | } | |||
1715 | ||||
1716 | void remove(SUnit *SU) override { | |||
1717 | assert(!Queue.empty() && "Queue is empty!")((!Queue.empty() && "Queue is empty!") ? static_cast< void> (0) : __assert_fail ("!Queue.empty() && \"Queue is empty!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1717, __PRETTY_FUNCTION__)); | |||
1718 | assert(SU->NodeQueueId != 0 && "Not in queue!")((SU->NodeQueueId != 0 && "Not in queue!") ? static_cast <void> (0) : __assert_fail ("SU->NodeQueueId != 0 && \"Not in queue!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1718, __PRETTY_FUNCTION__)); | |||
1719 | std::vector<SUnit *>::iterator I = std::find(Queue.begin(), Queue.end(), | |||
1720 | SU); | |||
1721 | if (I != std::prev(Queue.end())) | |||
1722 | std::swap(*I, Queue.back()); | |||
1723 | Queue.pop_back(); | |||
1724 | SU->NodeQueueId = 0; | |||
1725 | } | |||
1726 | ||||
1727 | bool tracksRegPressure() const override { return TracksRegPressure; } | |||
1728 | ||||
1729 | void dumpRegPressure() const; | |||
1730 | ||||
1731 | bool HighRegPressure(const SUnit *SU) const; | |||
1732 | ||||
1733 | bool MayReduceRegPressure(SUnit *SU) const; | |||
1734 | ||||
1735 | int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const; | |||
1736 | ||||
1737 | void scheduledNode(SUnit *SU) override; | |||
1738 | ||||
1739 | void unscheduledNode(SUnit *SU) override; | |||
1740 | ||||
1741 | protected: | |||
1742 | bool canClobber(const SUnit *SU, const SUnit *Op); | |||
1743 | void AddPseudoTwoAddrDeps(); | |||
1744 | void PrescheduleNodesWithMultipleUses(); | |||
1745 | void CalculateSethiUllmanNumbers(); | |||
1746 | }; | |||
1747 | ||||
1748 | template<class SF> | |||
1749 | static SUnit *popFromQueueImpl(std::vector<SUnit*> &Q, SF &Picker) { | |||
1750 | std::vector<SUnit *>::iterator Best = Q.begin(); | |||
1751 | for (std::vector<SUnit *>::iterator I = std::next(Q.begin()), | |||
1752 | E = Q.end(); I != E; ++I) | |||
1753 | if (Picker(*Best, *I)) | |||
1754 | Best = I; | |||
1755 | SUnit *V = *Best; | |||
1756 | if (Best != std::prev(Q.end())) | |||
1757 | std::swap(*Best, Q.back()); | |||
1758 | Q.pop_back(); | |||
1759 | return V; | |||
1760 | } | |||
1761 | ||||
1762 | template<class SF> | |||
1763 | SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker, ScheduleDAG *DAG) { | |||
1764 | #ifndef NDEBUG | |||
1765 | if (DAG->StressSched) { | |||
1766 | reverse_sort<SF> RPicker(Picker); | |||
1767 | return popFromQueueImpl(Q, RPicker); | |||
1768 | } | |||
1769 | #endif | |||
1770 | (void)DAG; | |||
1771 | return popFromQueueImpl(Q, Picker); | |||
1772 | } | |||
1773 | ||||
1774 | template<class SF> | |||
1775 | class RegReductionPriorityQueue : public RegReductionPQBase { | |||
1776 | SF Picker; | |||
1777 | ||||
1778 | public: | |||
1779 | RegReductionPriorityQueue(MachineFunction &mf, | |||
1780 | bool tracksrp, | |||
1781 | bool srcorder, | |||
1782 | const TargetInstrInfo *tii, | |||
1783 | const TargetRegisterInfo *tri, | |||
1784 | const TargetLowering *tli) | |||
1785 | : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder, | |||
1786 | tii, tri, tli), | |||
1787 | Picker(this) {} | |||
1788 | ||||
1789 | bool isBottomUp() const override { return SF::IsBottomUp; } | |||
1790 | ||||
1791 | bool isReady(SUnit *U) const override { | |||
1792 | return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle()); | |||
1793 | } | |||
1794 | ||||
1795 | SUnit *pop() override { | |||
1796 | if (Queue.empty()) return nullptr; | |||
1797 | ||||
1798 | SUnit *V = popFromQueue(Queue, Picker, scheduleDAG); | |||
1799 | V->NodeQueueId = 0; | |||
1800 | return V; | |||
1801 | } | |||
1802 | ||||
1803 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
1804 | void dump(ScheduleDAG *DAG) const override { | |||
1805 | // Emulate pop() without clobbering NodeQueueIds. | |||
1806 | std::vector<SUnit*> DumpQueue = Queue; | |||
1807 | SF DumpPicker = Picker; | |||
1808 | while (!DumpQueue.empty()) { | |||
1809 | SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG); | |||
1810 | dbgs() << "Height " << SU->getHeight() << ": "; | |||
1811 | SU->dump(DAG); | |||
1812 | } | |||
1813 | } | |||
1814 | #endif | |||
1815 | }; | |||
1816 | ||||
1817 | typedef RegReductionPriorityQueue<bu_ls_rr_sort> | |||
1818 | BURegReductionPriorityQueue; | |||
1819 | ||||
1820 | typedef RegReductionPriorityQueue<src_ls_rr_sort> | |||
1821 | SrcRegReductionPriorityQueue; | |||
1822 | ||||
1823 | typedef RegReductionPriorityQueue<hybrid_ls_rr_sort> | |||
1824 | HybridBURRPriorityQueue; | |||
1825 | ||||
1826 | typedef RegReductionPriorityQueue<ilp_ls_rr_sort> | |||
1827 | ILPBURRPriorityQueue; | |||
1828 | } // end anonymous namespace | |||
1829 | ||||
1830 | //===----------------------------------------------------------------------===// | |||
1831 | // Static Node Priority for Register Pressure Reduction | |||
1832 | //===----------------------------------------------------------------------===// | |||
1833 | ||||
1834 | // Check for special nodes that bypass scheduling heuristics. | |||
1835 | // Currently this pushes TokenFactor nodes down, but may be used for other | |||
1836 | // pseudo-ops as well. | |||
1837 | // | |||
1838 | // Return -1 to schedule right above left, 1 for left above right. | |||
1839 | // Return 0 if no bias exists. | |||
1840 | static int checkSpecialNodes(const SUnit *left, const SUnit *right) { | |||
1841 | bool LSchedLow = left->isScheduleLow; | |||
1842 | bool RSchedLow = right->isScheduleLow; | |||
1843 | if (LSchedLow != RSchedLow) | |||
1844 | return LSchedLow < RSchedLow ? 1 : -1; | |||
1845 | return 0; | |||
1846 | } | |||
1847 | ||||
1848 | /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number. | |||
1849 | /// Smaller number is the higher priority. | |||
1850 | static unsigned | |||
1851 | CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) { | |||
1852 | unsigned &SethiUllmanNumber = SUNumbers[SU->NodeNum]; | |||
1853 | if (SethiUllmanNumber != 0) | |||
1854 | return SethiUllmanNumber; | |||
1855 | ||||
1856 | unsigned Extra = 0; | |||
1857 | for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
1858 | I != E; ++I) { | |||
1859 | if (I->isCtrl()) continue; // ignore chain preds | |||
1860 | SUnit *PredSU = I->getSUnit(); | |||
1861 | unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU, SUNumbers); | |||
1862 | if (PredSethiUllman > SethiUllmanNumber) { | |||
1863 | SethiUllmanNumber = PredSethiUllman; | |||
1864 | Extra = 0; | |||
1865 | } else if (PredSethiUllman == SethiUllmanNumber) | |||
1866 | ++Extra; | |||
1867 | } | |||
1868 | ||||
1869 | SethiUllmanNumber += Extra; | |||
1870 | ||||
1871 | if (SethiUllmanNumber == 0) | |||
1872 | SethiUllmanNumber = 1; | |||
1873 | ||||
1874 | return SethiUllmanNumber; | |||
1875 | } | |||
1876 | ||||
1877 | /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all | |||
1878 | /// scheduling units. | |||
1879 | void RegReductionPQBase::CalculateSethiUllmanNumbers() { | |||
1880 | SethiUllmanNumbers.assign(SUnits->size(), 0); | |||
1881 | ||||
1882 | for (unsigned i = 0, e = SUnits->size(); i != e; ++i) | |||
1883 | CalcNodeSethiUllmanNumber(&(*SUnits)[i], SethiUllmanNumbers); | |||
1884 | } | |||
1885 | ||||
1886 | void RegReductionPQBase::addNode(const SUnit *SU) { | |||
1887 | unsigned SUSize = SethiUllmanNumbers.size(); | |||
1888 | if (SUnits->size() > SUSize) | |||
1889 | SethiUllmanNumbers.resize(SUSize*2, 0); | |||
1890 | CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers); | |||
1891 | } | |||
1892 | ||||
1893 | void RegReductionPQBase::updateNode(const SUnit *SU) { | |||
1894 | SethiUllmanNumbers[SU->NodeNum] = 0; | |||
1895 | CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers); | |||
1896 | } | |||
1897 | ||||
1898 | // Lower priority means schedule further down. For bottom-up scheduling, lower | |||
1899 | // priority SUs are scheduled before higher priority SUs. | |||
1900 | unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const { | |||
1901 | assert(SU->NodeNum < SethiUllmanNumbers.size())((SU->NodeNum < SethiUllmanNumbers.size()) ? static_cast <void> (0) : __assert_fail ("SU->NodeNum < SethiUllmanNumbers.size()" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1901, __PRETTY_FUNCTION__)); | |||
1902 | unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0; | |||
1903 | if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg) | |||
1904 | // CopyToReg should be close to its uses to facilitate coalescing and | |||
1905 | // avoid spilling. | |||
1906 | return 0; | |||
1907 | if (Opc == TargetOpcode::EXTRACT_SUBREG || | |||
1908 | Opc == TargetOpcode::SUBREG_TO_REG || | |||
1909 | Opc == TargetOpcode::INSERT_SUBREG) | |||
1910 | // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be | |||
1911 | // close to their uses to facilitate coalescing. | |||
1912 | return 0; | |||
1913 | if (SU->NumSuccs == 0 && SU->NumPreds != 0) | |||
1914 | // If SU does not have a register use, i.e. it doesn't produce a value | |||
1915 | // that would be consumed (e.g. store), then it terminates a chain of | |||
1916 | // computation. Give it a large SethiUllman number so it will be | |||
1917 | // scheduled right before its predecessors that it doesn't lengthen | |||
1918 | // their live ranges. | |||
1919 | return 0xffff; | |||
1920 | if (SU->NumPreds == 0 && SU->NumSuccs != 0) | |||
1921 | // If SU does not have a register def, schedule it close to its uses | |||
1922 | // because it does not lengthen any live ranges. | |||
1923 | return 0; | |||
1924 | #if 1 | |||
1925 | return SethiUllmanNumbers[SU->NodeNum]; | |||
1926 | #else | |||
1927 | unsigned Priority = SethiUllmanNumbers[SU->NodeNum]; | |||
1928 | if (SU->isCallOp) { | |||
1929 | // FIXME: This assumes all of the defs are used as call operands. | |||
1930 | int NP = (int)Priority - SU->getNode()->getNumValues(); | |||
1931 | return (NP > 0) ? NP : 0; | |||
1932 | } | |||
1933 | return Priority; | |||
1934 | #endif | |||
1935 | } | |||
1936 | ||||
1937 | //===----------------------------------------------------------------------===// | |||
1938 | // Register Pressure Tracking | |||
1939 | //===----------------------------------------------------------------------===// | |||
1940 | ||||
1941 | void RegReductionPQBase::dumpRegPressure() const { | |||
1942 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
1943 | for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(), | |||
1944 | E = TRI->regclass_end(); I != E; ++I) { | |||
1945 | const TargetRegisterClass *RC = *I; | |||
1946 | unsigned Id = RC->getID(); | |||
1947 | unsigned RP = RegPressure[Id]; | |||
1948 | if (!RP) continue; | |||
1949 | DEBUG(dbgs() << TRI->getRegClassName(RC) << ": " << RP << " / "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << TRI->getRegClassName(RC ) << ": " << RP << " / " << RegLimit[ Id] << '\n'; } } while (0) | |||
1950 | << RegLimit[Id] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << TRI->getRegClassName(RC ) << ": " << RP << " / " << RegLimit[ Id] << '\n'; } } while (0); | |||
1951 | } | |||
1952 | #endif | |||
1953 | } | |||
1954 | ||||
1955 | bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const { | |||
1956 | if (!TLI) | |||
1957 | return false; | |||
1958 | ||||
1959 | for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end(); | |||
1960 | I != E; ++I) { | |||
1961 | if (I->isCtrl()) | |||
1962 | continue; | |||
1963 | SUnit *PredSU = I->getSUnit(); | |||
1964 | // NumRegDefsLeft is zero when enough uses of this node have been scheduled | |||
1965 | // to cover the number of registers defined (they are all live). | |||
1966 | if (PredSU->NumRegDefsLeft == 0) { | |||
1967 | continue; | |||
1968 | } | |||
1969 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG); | |||
1970 | RegDefPos.IsValid(); RegDefPos.Advance()) { | |||
1971 | unsigned RCId, Cost; | |||
1972 | GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF); | |||
1973 | ||||
1974 | if ((RegPressure[RCId] + Cost) >= RegLimit[RCId]) | |||
1975 | return true; | |||
1976 | } | |||
1977 | } | |||
1978 | return false; | |||
1979 | } | |||
1980 | ||||
1981 | bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const { | |||
1982 | const SDNode *N = SU->getNode(); | |||
1983 | ||||
1984 | if (!N->isMachineOpcode() || !SU->NumSuccs) | |||
1985 | return false; | |||
1986 | ||||
1987 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); | |||
1988 | for (unsigned i = 0; i != NumDefs; ++i) { | |||
1989 | MVT VT = N->getSimpleValueType(i); | |||
1990 | if (!N->hasAnyUseOfValue(i)) | |||
1991 | continue; | |||
1992 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); | |||
1993 | if (RegPressure[RCId] >= RegLimit[RCId]) | |||
1994 | return true; | |||
1995 | } | |||
1996 | return false; | |||
1997 | } | |||
1998 | ||||
1999 | // Compute the register pressure contribution by this instruction by count up | |||
2000 | // for uses that are not live and down for defs. Only count register classes | |||
2001 | // that are already under high pressure. As a side effect, compute the number of | |||
2002 | // uses of registers that are already live. | |||
2003 | // | |||
2004 | // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure | |||
2005 | // so could probably be factored. | |||
2006 | int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const { | |||
2007 | LiveUses = 0; | |||
2008 | int PDiff = 0; | |||
2009 | for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end(); | |||
2010 | I != E; ++I) { | |||
2011 | if (I->isCtrl()) | |||
2012 | continue; | |||
2013 | SUnit *PredSU = I->getSUnit(); | |||
2014 | // NumRegDefsLeft is zero when enough uses of this node have been scheduled | |||
2015 | // to cover the number of registers defined (they are all live). | |||
2016 | if (PredSU->NumRegDefsLeft == 0) { | |||
2017 | if (PredSU->getNode()->isMachineOpcode()) | |||
2018 | ++LiveUses; | |||
2019 | continue; | |||
2020 | } | |||
2021 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG); | |||
2022 | RegDefPos.IsValid(); RegDefPos.Advance()) { | |||
2023 | MVT VT = RegDefPos.GetValue(); | |||
2024 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); | |||
2025 | if (RegPressure[RCId] >= RegLimit[RCId]) | |||
2026 | ++PDiff; | |||
2027 | } | |||
2028 | } | |||
2029 | const SDNode *N = SU->getNode(); | |||
2030 | ||||
2031 | if (!N || !N->isMachineOpcode() || !SU->NumSuccs) | |||
2032 | return PDiff; | |||
2033 | ||||
2034 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); | |||
2035 | for (unsigned i = 0; i != NumDefs; ++i) { | |||
2036 | MVT VT = N->getSimpleValueType(i); | |||
2037 | if (!N->hasAnyUseOfValue(i)) | |||
2038 | continue; | |||
2039 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); | |||
2040 | if (RegPressure[RCId] >= RegLimit[RCId]) | |||
2041 | --PDiff; | |||
2042 | } | |||
2043 | return PDiff; | |||
2044 | } | |||
2045 | ||||
2046 | void RegReductionPQBase::scheduledNode(SUnit *SU) { | |||
2047 | if (!TracksRegPressure) | |||
2048 | return; | |||
2049 | ||||
2050 | if (!SU->getNode()) | |||
2051 | return; | |||
2052 | ||||
2053 | for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
2054 | I != E; ++I) { | |||
2055 | if (I->isCtrl()) | |||
2056 | continue; | |||
2057 | SUnit *PredSU = I->getSUnit(); | |||
2058 | // NumRegDefsLeft is zero when enough uses of this node have been scheduled | |||
2059 | // to cover the number of registers defined (they are all live). | |||
2060 | if (PredSU->NumRegDefsLeft == 0) { | |||
2061 | continue; | |||
2062 | } | |||
2063 | // FIXME: The ScheduleDAG currently loses information about which of a | |||
2064 | // node's values is consumed by each dependence. Consequently, if the node | |||
2065 | // defines multiple register classes, we don't know which to pressurize | |||
2066 | // here. Instead the following loop consumes the register defs in an | |||
2067 | // arbitrary order. At least it handles the common case of clustered loads | |||
2068 | // to the same class. For precise liveness, each SDep needs to indicate the | |||
2069 | // result number. But that tightly couples the ScheduleDAG with the | |||
2070 | // SelectionDAG making updates tricky. A simpler hack would be to attach a | |||
2071 | // value type or register class to SDep. | |||
2072 | // | |||
2073 | // The most important aspect of register tracking is balancing the increase | |||
2074 | // here with the reduction further below. Note that this SU may use multiple | |||
2075 | // defs in PredSU. The can't be determined here, but we've already | |||
2076 | // compensated by reducing NumRegDefsLeft in PredSU during | |||
2077 | // ScheduleDAGSDNodes::AddSchedEdges. | |||
2078 | --PredSU->NumRegDefsLeft; | |||
2079 | unsigned SkipRegDefs = PredSU->NumRegDefsLeft; | |||
2080 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG); | |||
2081 | RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) { | |||
2082 | if (SkipRegDefs) | |||
2083 | continue; | |||
2084 | ||||
2085 | unsigned RCId, Cost; | |||
2086 | GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF); | |||
2087 | RegPressure[RCId] += Cost; | |||
2088 | break; | |||
2089 | } | |||
2090 | } | |||
2091 | ||||
2092 | // We should have this assert, but there may be dead SDNodes that never | |||
2093 | // materialize as SUnits, so they don't appear to generate liveness. | |||
2094 | //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses"); | |||
2095 | int SkipRegDefs = (int)SU->NumRegDefsLeft; | |||
2096 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG); | |||
2097 | RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) { | |||
2098 | if (SkipRegDefs > 0) | |||
2099 | continue; | |||
2100 | unsigned RCId, Cost; | |||
2101 | GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF); | |||
2102 | if (RegPressure[RCId] < Cost) { | |||
2103 | // Register pressure tracking is imprecise. This can happen. But we try | |||
2104 | // hard not to let it happen because it likely results in poor scheduling. | |||
2105 | DEBUG(dbgs() << " SU(" << SU->NodeNum << ") has too many regdefs\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " SU(" << SU->NodeNum << ") has too many regdefs\n"; } } while (0); | |||
2106 | RegPressure[RCId] = 0; | |||
2107 | } | |||
2108 | else { | |||
2109 | RegPressure[RCId] -= Cost; | |||
2110 | } | |||
2111 | } | |||
2112 | dumpRegPressure(); | |||
2113 | } | |||
2114 | ||||
2115 | void RegReductionPQBase::unscheduledNode(SUnit *SU) { | |||
2116 | if (!TracksRegPressure) | |||
2117 | return; | |||
2118 | ||||
2119 | const SDNode *N = SU->getNode(); | |||
2120 | if (!N) return; | |||
2121 | ||||
2122 | if (!N->isMachineOpcode()) { | |||
2123 | if (N->getOpcode() != ISD::CopyToReg) | |||
2124 | return; | |||
2125 | } else { | |||
2126 | unsigned Opc = N->getMachineOpcode(); | |||
2127 | if (Opc == TargetOpcode::EXTRACT_SUBREG || | |||
2128 | Opc == TargetOpcode::INSERT_SUBREG || | |||
2129 | Opc == TargetOpcode::SUBREG_TO_REG || | |||
2130 | Opc == TargetOpcode::REG_SEQUENCE || | |||
2131 | Opc == TargetOpcode::IMPLICIT_DEF) | |||
2132 | return; | |||
2133 | } | |||
2134 | ||||
2135 | for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
2136 | I != E; ++I) { | |||
2137 | if (I->isCtrl()) | |||
2138 | continue; | |||
2139 | SUnit *PredSU = I->getSUnit(); | |||
2140 | // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only | |||
2141 | // counts data deps. | |||
2142 | if (PredSU->NumSuccsLeft != PredSU->Succs.size()) | |||
2143 | continue; | |||
2144 | const SDNode *PN = PredSU->getNode(); | |||
2145 | if (!PN->isMachineOpcode()) { | |||
2146 | if (PN->getOpcode() == ISD::CopyFromReg) { | |||
2147 | MVT VT = PN->getSimpleValueType(0); | |||
2148 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); | |||
2149 | RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); | |||
2150 | } | |||
2151 | continue; | |||
2152 | } | |||
2153 | unsigned POpc = PN->getMachineOpcode(); | |||
2154 | if (POpc == TargetOpcode::IMPLICIT_DEF) | |||
2155 | continue; | |||
2156 | if (POpc == TargetOpcode::EXTRACT_SUBREG || | |||
2157 | POpc == TargetOpcode::INSERT_SUBREG || | |||
2158 | POpc == TargetOpcode::SUBREG_TO_REG) { | |||
2159 | MVT VT = PN->getSimpleValueType(0); | |||
2160 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); | |||
2161 | RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); | |||
2162 | continue; | |||
2163 | } | |||
2164 | unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs(); | |||
2165 | for (unsigned i = 0; i != NumDefs; ++i) { | |||
2166 | MVT VT = PN->getSimpleValueType(i); | |||
2167 | if (!PN->hasAnyUseOfValue(i)) | |||
2168 | continue; | |||
2169 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); | |||
2170 | if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT)) | |||
2171 | // Register pressure tracking is imprecise. This can happen. | |||
2172 | RegPressure[RCId] = 0; | |||
2173 | else | |||
2174 | RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT); | |||
2175 | } | |||
2176 | } | |||
2177 | ||||
2178 | // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses() | |||
2179 | // may transfer data dependencies to CopyToReg. | |||
2180 | if (SU->NumSuccs && N->isMachineOpcode()) { | |||
2181 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); | |||
2182 | for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) { | |||
2183 | MVT VT = N->getSimpleValueType(i); | |||
2184 | if (VT == MVT::Glue || VT == MVT::Other) | |||
2185 | continue; | |||
2186 | if (!N->hasAnyUseOfValue(i)) | |||
2187 | continue; | |||
2188 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); | |||
2189 | RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); | |||
2190 | } | |||
2191 | } | |||
2192 | ||||
2193 | dumpRegPressure(); | |||
2194 | } | |||
2195 | ||||
2196 | //===----------------------------------------------------------------------===// | |||
2197 | // Dynamic Node Priority for Register Pressure Reduction | |||
2198 | //===----------------------------------------------------------------------===// | |||
2199 | ||||
2200 | /// closestSucc - Returns the scheduled cycle of the successor which is | |||
2201 | /// closest to the current cycle. | |||
2202 | static unsigned closestSucc(const SUnit *SU) { | |||
2203 | unsigned MaxHeight = 0; | |||
2204 | for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); | |||
2205 | I != E; ++I) { | |||
2206 | if (I->isCtrl()) continue; // ignore chain succs | |||
2207 | unsigned Height = I->getSUnit()->getHeight(); | |||
2208 | // If there are bunch of CopyToRegs stacked up, they should be considered | |||
2209 | // to be at the same position. | |||
2210 | if (I->getSUnit()->getNode() && | |||
2211 | I->getSUnit()->getNode()->getOpcode() == ISD::CopyToReg) | |||
2212 | Height = closestSucc(I->getSUnit())+1; | |||
2213 | if (Height > MaxHeight) | |||
2214 | MaxHeight = Height; | |||
2215 | } | |||
2216 | return MaxHeight; | |||
2217 | } | |||
2218 | ||||
2219 | /// calcMaxScratches - Returns an cost estimate of the worse case requirement | |||
2220 | /// for scratch registers, i.e. number of data dependencies. | |||
2221 | static unsigned calcMaxScratches(const SUnit *SU) { | |||
2222 | unsigned Scratches = 0; | |||
2223 | for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
2224 | I != E; ++I) { | |||
2225 | if (I->isCtrl()) continue; // ignore chain preds | |||
2226 | Scratches++; | |||
2227 | } | |||
2228 | return Scratches; | |||
2229 | } | |||
2230 | ||||
2231 | /// hasOnlyLiveInOpers - Return true if SU has only value predecessors that are | |||
2232 | /// CopyFromReg from a virtual register. | |||
2233 | static bool hasOnlyLiveInOpers(const SUnit *SU) { | |||
2234 | bool RetVal = false; | |||
2235 | for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
2236 | I != E; ++I) { | |||
2237 | if (I->isCtrl()) continue; | |||
2238 | const SUnit *PredSU = I->getSUnit(); | |||
2239 | if (PredSU->getNode() && | |||
2240 | PredSU->getNode()->getOpcode() == ISD::CopyFromReg) { | |||
2241 | unsigned Reg = | |||
2242 | cast<RegisterSDNode>(PredSU->getNode()->getOperand(1))->getReg(); | |||
2243 | if (TargetRegisterInfo::isVirtualRegister(Reg)) { | |||
2244 | RetVal = true; | |||
2245 | continue; | |||
2246 | } | |||
2247 | } | |||
2248 | return false; | |||
2249 | } | |||
2250 | return RetVal; | |||
2251 | } | |||
2252 | ||||
2253 | /// hasOnlyLiveOutUses - Return true if SU has only value successors that are | |||
2254 | /// CopyToReg to a virtual register. This SU def is probably a liveout and | |||
2255 | /// it has no other use. It should be scheduled closer to the terminator. | |||
2256 | static bool hasOnlyLiveOutUses(const SUnit *SU) { | |||
2257 | bool RetVal = false; | |||
2258 | for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); | |||
2259 | I != E; ++I) { | |||
2260 | if (I->isCtrl()) continue; | |||
2261 | const SUnit *SuccSU = I->getSUnit(); | |||
2262 | if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) { | |||
2263 | unsigned Reg = | |||
2264 | cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg(); | |||
2265 | if (TargetRegisterInfo::isVirtualRegister(Reg)) { | |||
2266 | RetVal = true; | |||
2267 | continue; | |||
2268 | } | |||
2269 | } | |||
2270 | return false; | |||
2271 | } | |||
2272 | return RetVal; | |||
2273 | } | |||
2274 | ||||
2275 | // Set isVRegCycle for a node with only live in opers and live out uses. Also | |||
2276 | // set isVRegCycle for its CopyFromReg operands. | |||
2277 | // | |||
2278 | // This is only relevant for single-block loops, in which case the VRegCycle | |||
2279 | // node is likely an induction variable in which the operand and target virtual | |||
2280 | // registers should be coalesced (e.g. pre/post increment values). Setting the | |||
2281 | // isVRegCycle flag helps the scheduler prioritize other uses of the same | |||
2282 | // CopyFromReg so that this node becomes the virtual register "kill". This | |||
2283 | // avoids interference between the values live in and out of the block and | |||
2284 | // eliminates a copy inside the loop. | |||
2285 | static void initVRegCycle(SUnit *SU) { | |||
2286 | if (DisableSchedVRegCycle) | |||
2287 | return; | |||
2288 | ||||
2289 | if (!hasOnlyLiveInOpers(SU) || !hasOnlyLiveOutUses(SU)) | |||
2290 | return; | |||
2291 | ||||
2292 | DEBUG(dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n"; } } while (0); | |||
2293 | ||||
2294 | SU->isVRegCycle = true; | |||
2295 | ||||
2296 | for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); | |||
2297 | I != E; ++I) { | |||
2298 | if (I->isCtrl()) continue; | |||
2299 | I->getSUnit()->isVRegCycle = true; | |||
2300 | } | |||
2301 | } | |||
2302 | ||||
2303 | // After scheduling the definition of a VRegCycle, clear the isVRegCycle flag of | |||
2304 | // CopyFromReg operands. We should no longer penalize other uses of this VReg. | |||
2305 | static void resetVRegCycle(SUnit *SU) { | |||
2306 | if (!SU->isVRegCycle) | |||
2307 | return; | |||
2308 | ||||
2309 | for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end(); | |||
2310 | I != E; ++I) { | |||
2311 | if (I->isCtrl()) continue; // ignore chain preds | |||
2312 | SUnit *PredSU = I->getSUnit(); | |||
2313 | if (PredSU->isVRegCycle) { | |||
2314 | assert(PredSU->getNode()->getOpcode() == ISD::CopyFromReg &&((PredSU->getNode()->getOpcode() == ISD::CopyFromReg && "VRegCycle def must be CopyFromReg") ? static_cast<void> (0) : __assert_fail ("PredSU->getNode()->getOpcode() == ISD::CopyFromReg && \"VRegCycle def must be CopyFromReg\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2315, __PRETTY_FUNCTION__)) | |||
2315 | "VRegCycle def must be CopyFromReg")((PredSU->getNode()->getOpcode() == ISD::CopyFromReg && "VRegCycle def must be CopyFromReg") ? static_cast<void> (0) : __assert_fail ("PredSU->getNode()->getOpcode() == ISD::CopyFromReg && \"VRegCycle def must be CopyFromReg\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2315, __PRETTY_FUNCTION__)); | |||
2316 | I->getSUnit()->isVRegCycle = 0; | |||
2317 | } | |||
2318 | } | |||
2319 | } | |||
2320 | ||||
2321 | // Return true if this SUnit uses a CopyFromReg node marked as a VRegCycle. This | |||
2322 | // means a node that defines the VRegCycle has not been scheduled yet. | |||
2323 | static bool hasVRegCycleUse(const SUnit *SU) { | |||
2324 | // If this SU also defines the VReg, don't hoist it as a "use". | |||
2325 | if (SU->isVRegCycle) | |||
2326 | return false; | |||
2327 | ||||
2328 | for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end(); | |||
2329 | I != E; ++I) { | |||
2330 | if (I->isCtrl()) continue; // ignore chain preds | |||
2331 | if (I->getSUnit()->isVRegCycle && | |||
2332 | I->getSUnit()->getNode()->getOpcode() == ISD::CopyFromReg) { | |||
2333 | DEBUG(dbgs() << " VReg cycle use: SU (" << SU->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " VReg cycle use: SU (" << SU->NodeNum << ")\n"; } } while (0); | |||
2334 | return true; | |||
2335 | } | |||
2336 | } | |||
2337 | return false; | |||
2338 | } | |||
2339 | ||||
2340 | // Check for either a dependence (latency) or resource (hazard) stall. | |||
2341 | // | |||
2342 | // Note: The ScheduleHazardRecognizer interface requires a non-const SU. | |||
2343 | static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) { | |||
2344 | if ((int)SPQ->getCurCycle() < Height) return true; | |||
2345 | if (SPQ->getHazardRec()->getHazardType(SU, 0) | |||
2346 | != ScheduleHazardRecognizer::NoHazard) | |||
2347 | return true; | |||
2348 | return false; | |||
2349 | } | |||
2350 | ||||
2351 | // Return -1 if left has higher priority, 1 if right has higher priority. | |||
2352 | // Return 0 if latency-based priority is equivalent. | |||
2353 | static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref, | |||
2354 | RegReductionPQBase *SPQ) { | |||
2355 | // Scheduling an instruction that uses a VReg whose postincrement has not yet | |||
2356 | // been scheduled will induce a copy. Model this as an extra cycle of latency. | |||
2357 | int LPenalty = hasVRegCycleUse(left) ? 1 : 0; | |||
2358 | int RPenalty = hasVRegCycleUse(right) ? 1 : 0; | |||
2359 | int LHeight = (int)left->getHeight() + LPenalty; | |||
2360 | int RHeight = (int)right->getHeight() + RPenalty; | |||
2361 | ||||
2362 | bool LStall = (!checkPref || left->SchedulingPref == Sched::ILP) && | |||
2363 | BUHasStall(left, LHeight, SPQ); | |||
2364 | bool RStall = (!checkPref || right->SchedulingPref == Sched::ILP) && | |||
2365 | BUHasStall(right, RHeight, SPQ); | |||
2366 | ||||
2367 | // If scheduling one of the node will cause a pipeline stall, delay it. | |||
2368 | // If scheduling either one of the node will cause a pipeline stall, sort | |||
2369 | // them according to their height. | |||
2370 | if (LStall) { | |||
2371 | if (!RStall) | |||
2372 | return 1; | |||
2373 | if (LHeight != RHeight) | |||
2374 | return LHeight > RHeight ? 1 : -1; | |||
2375 | } else if (RStall) | |||
2376 | return -1; | |||
2377 | ||||
2378 | // If either node is scheduling for latency, sort them by height/depth | |||
2379 | // and latency. | |||
2380 | if (!checkPref || (left->SchedulingPref == Sched::ILP || | |||
2381 | right->SchedulingPref == Sched::ILP)) { | |||
2382 | // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer | |||
2383 | // is enabled, grouping instructions by cycle, then its height is already | |||
2384 | // covered so only its depth matters. We also reach this point if both stall | |||
2385 | // but have the same height. | |||
2386 | if (!SPQ->getHazardRec()->isEnabled()) { | |||
2387 | if (LHeight != RHeight) | |||
2388 | return LHeight > RHeight ? 1 : -1; | |||
2389 | } | |||
2390 | int LDepth = left->getDepth() - LPenalty; | |||
2391 | int RDepth = right->getDepth() - RPenalty; | |||
2392 | if (LDepth != RDepth) { | |||
2393 | DEBUG(dbgs() << " Comparing latency of SU (" << left->NodeNumdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Comparing latency of SU (" << left->NodeNum << ") depth " << LDepth << " vs SU (" << right->NodeNum << ") depth " << RDepth << "\n"; } } while (0) | |||
2394 | << ") depth " << LDepth << " vs SU (" << right->NodeNumdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Comparing latency of SU (" << left->NodeNum << ") depth " << LDepth << " vs SU (" << right->NodeNum << ") depth " << RDepth << "\n"; } } while (0) | |||
2395 | << ") depth " << RDepth << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Comparing latency of SU (" << left->NodeNum << ") depth " << LDepth << " vs SU (" << right->NodeNum << ") depth " << RDepth << "\n"; } } while (0); | |||
2396 | return LDepth < RDepth ? 1 : -1; | |||
2397 | } | |||
2398 | if (left->Latency != right->Latency) | |||
2399 | return left->Latency > right->Latency ? 1 : -1; | |||
2400 | } | |||
2401 | return 0; | |||
2402 | } | |||
2403 | ||||
2404 | static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) { | |||
2405 | // Schedule physical register definitions close to their use. This is | |||
2406 | // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as | |||
2407 | // long as shortening physreg live ranges is generally good, we can defer | |||
2408 | // creating a subtarget hook. | |||
2409 | if (!DisableSchedPhysRegJoin) { | |||
2410 | bool LHasPhysReg = left->hasPhysRegDefs; | |||
2411 | bool RHasPhysReg = right->hasPhysRegDefs; | |||
2412 | if (LHasPhysReg != RHasPhysReg) { | |||
2413 | #ifndef NDEBUG | |||
2414 | static const char *const PhysRegMsg[] = { " has no physreg", | |||
2415 | " defines a physreg" }; | |||
2416 | #endif | |||
2417 | DEBUG(dbgs() << " SU (" << left->NodeNum << ") "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " SU (" << left-> NodeNum << ") " << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") " << PhysRegMsg [RHasPhysReg] << "\n"; } } while (0) | |||
2418 | << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " SU (" << left-> NodeNum << ") " << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") " << PhysRegMsg [RHasPhysReg] << "\n"; } } while (0) | |||
2419 | << PhysRegMsg[RHasPhysReg] << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " SU (" << left-> NodeNum << ") " << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") " << PhysRegMsg [RHasPhysReg] << "\n"; } } while (0); | |||
2420 | return LHasPhysReg < RHasPhysReg; | |||
2421 | } | |||
2422 | } | |||
2423 | ||||
2424 | // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down. | |||
2425 | unsigned LPriority = SPQ->getNodePriority(left); | |||
2426 | unsigned RPriority = SPQ->getNodePriority(right); | |||
2427 | ||||
2428 | // Be really careful about hoisting call operands above previous calls. | |||
2429 | // Only allows it if it would reduce register pressure. | |||
2430 | if (left->isCall && right->isCallOp) { | |||
2431 | unsigned RNumVals = right->getNode()->getNumValues(); | |||
2432 | RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0; | |||
2433 | } | |||
2434 | if (right->isCall && left->isCallOp) { | |||
2435 | unsigned LNumVals = left->getNode()->getNumValues(); | |||
2436 | LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0; | |||
2437 | } | |||
2438 | ||||
2439 | if (LPriority != RPriority) | |||
2440 | return LPriority > RPriority; | |||
2441 | ||||
2442 | // One or both of the nodes are calls and their sethi-ullman numbers are the | |||
2443 | // same, then keep source order. | |||
2444 | if (left->isCall || right->isCall) { | |||
2445 | unsigned LOrder = SPQ->getNodeOrdering(left); | |||
2446 | unsigned ROrder = SPQ->getNodeOrdering(right); | |||
2447 | ||||
2448 | // Prefer an ordering where the lower the non-zero order number, the higher | |||
2449 | // the preference. | |||
2450 | if ((LOrder || ROrder) && LOrder != ROrder) | |||
2451 | return LOrder != 0 && (LOrder < ROrder || ROrder == 0); | |||
2452 | } | |||
2453 | ||||
2454 | // Try schedule def + use closer when Sethi-Ullman numbers are the same. | |||
2455 | // e.g. | |||
2456 | // t1 = op t2, c1 | |||
2457 | // t3 = op t4, c2 | |||
2458 | // | |||
2459 | // and the following instructions are both ready. | |||
2460 | // t2 = op c3 | |||
2461 | // t4 = op c4 | |||
2462 | // | |||
2463 | // Then schedule t2 = op first. | |||
2464 | // i.e. | |||
2465 | // t4 = op c4 | |||
2466 | // t2 = op c3 | |||
2467 | // t1 = op t2, c1 | |||
2468 | // t3 = op t4, c2 | |||
2469 | // | |||
2470 | // This creates more short live intervals. | |||
2471 | unsigned LDist = closestSucc(left); | |||
2472 | unsigned RDist = closestSucc(right); | |||
2473 | if (LDist != RDist) | |||
2474 | return LDist < RDist; | |||
2475 | ||||
2476 | // How many registers becomes live when the node is scheduled. | |||
2477 | unsigned LScratch = calcMaxScratches(left); | |||
2478 | unsigned RScratch = calcMaxScratches(right); | |||
2479 | if (LScratch != RScratch) | |||
2480 | return LScratch > RScratch; | |||
2481 | ||||
2482 | // Comparing latency against a call makes little sense unless the node | |||
2483 | // is register pressure-neutral. | |||
2484 | if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0)) | |||
2485 | return (left->NodeQueueId > right->NodeQueueId); | |||
2486 | ||||
2487 | // Do not compare latencies when one or both of the nodes are calls. | |||
2488 | if (!DisableSchedCycles && | |||
2489 | !(left->isCall || right->isCall)) { | |||
2490 | int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ); | |||
2491 | if (result != 0) | |||
2492 | return result > 0; | |||
2493 | } | |||
2494 | else { | |||
2495 | if (left->getHeight() != right->getHeight()) | |||
2496 | return left->getHeight() > right->getHeight(); | |||
2497 | ||||
2498 | if (left->getDepth() != right->getDepth()) | |||
2499 | return left->getDepth() < right->getDepth(); | |||
2500 | } | |||
2501 | ||||
2502 | assert(left->NodeQueueId && right->NodeQueueId &&((left->NodeQueueId && right->NodeQueueId && "NodeQueueId cannot be zero") ? static_cast<void> (0) : __assert_fail ("left->NodeQueueId && right->NodeQueueId && \"NodeQueueId cannot be zero\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2503, __PRETTY_FUNCTION__)) | |||
2503 | "NodeQueueId cannot be zero")((left->NodeQueueId && right->NodeQueueId && "NodeQueueId cannot be zero") ? static_cast<void> (0) : __assert_fail ("left->NodeQueueId && right->NodeQueueId && \"NodeQueueId cannot be zero\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2503, __PRETTY_FUNCTION__)); | |||
2504 | return (left->NodeQueueId > right->NodeQueueId); | |||
2505 | } | |||
2506 | ||||
2507 | // Bottom up | |||
2508 | bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { | |||
2509 | if (int res = checkSpecialNodes(left, right)) | |||
2510 | return res > 0; | |||
2511 | ||||
2512 | return BURRSort(left, right, SPQ); | |||
2513 | } | |||
2514 | ||||
2515 | // Source order, otherwise bottom up. | |||
2516 | bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { | |||
2517 | if (int res = checkSpecialNodes(left, right)) | |||
2518 | return res > 0; | |||
2519 | ||||
2520 | unsigned LOrder = SPQ->getNodeOrdering(left); | |||
2521 | unsigned ROrder = SPQ->getNodeOrdering(right); | |||
2522 | ||||
2523 | // Prefer an ordering where the lower the non-zero order number, the higher | |||
2524 | // the preference. | |||
2525 | if ((LOrder || ROrder) && LOrder != ROrder) | |||
2526 | return LOrder != 0 && (LOrder < ROrder || ROrder == 0); | |||
2527 | ||||
2528 | return BURRSort(left, right, SPQ); | |||
2529 | } | |||
2530 | ||||
2531 | // If the time between now and when the instruction will be ready can cover | |||
2532 | // the spill code, then avoid adding it to the ready queue. This gives long | |||
2533 | // stalls highest priority and allows hoisting across calls. It should also | |||
2534 | // speed up processing the available queue. | |||
2535 | bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const { | |||
2536 | static const unsigned ReadyDelay = 3; | |||
2537 | ||||
2538 | if (SPQ->MayReduceRegPressure(SU)) return true; | |||
2539 | ||||
2540 | if (SU->getHeight() > (CurCycle + ReadyDelay)) return false; | |||
2541 | ||||
2542 | if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay) | |||
2543 | != ScheduleHazardRecognizer::NoHazard) | |||
2544 | return false; | |||
2545 | ||||
2546 | return true; | |||
2547 | } | |||
2548 | ||||
2549 | // Return true if right should be scheduled with higher priority than left. | |||
2550 | bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { | |||
2551 | if (int res = checkSpecialNodes(left, right)) | |||
2552 | return res > 0; | |||
2553 | ||||
2554 | if (left->isCall || right->isCall) | |||
2555 | // No way to compute latency of calls. | |||
2556 | return BURRSort(left, right, SPQ); | |||
2557 | ||||
2558 | bool LHigh = SPQ->HighRegPressure(left); | |||
2559 | bool RHigh = SPQ->HighRegPressure(right); | |||
2560 | // Avoid causing spills. If register pressure is high, schedule for | |||
2561 | // register pressure reduction. | |||
2562 | if (LHigh && !RHigh) { | |||
2563 | DEBUG(dbgs() << " pressure SU(" << left->NodeNum << ") > SU("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " pressure SU(" << left->NodeNum << ") > SU(" << right->NodeNum << ")\n"; } } while (0) | |||
2564 | << right->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " pressure SU(" << left->NodeNum << ") > SU(" << right->NodeNum << ")\n"; } } while (0); | |||
2565 | return true; | |||
2566 | } | |||
2567 | else if (!LHigh && RHigh) { | |||
2568 | DEBUG(dbgs() << " pressure SU(" << right->NodeNum << ") > SU("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " pressure SU(" << right->NodeNum << ") > SU(" << left->NodeNum << ")\n"; } } while (0) | |||
2569 | << left->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " pressure SU(" << right->NodeNum << ") > SU(" << left->NodeNum << ")\n"; } } while (0); | |||
2570 | return false; | |||
2571 | } | |||
2572 | if (!LHigh && !RHigh) { | |||
2573 | int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ); | |||
2574 | if (result != 0) | |||
2575 | return result > 0; | |||
2576 | } | |||
2577 | return BURRSort(left, right, SPQ); | |||
2578 | } | |||
2579 | ||||
2580 | // Schedule as many instructions in each cycle as possible. So don't make an | |||
2581 | // instruction available unless it is ready in the current cycle. | |||
2582 | bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const { | |||
2583 | if (SU->getHeight() > CurCycle) return false; | |||
2584 | ||||
2585 | if (SPQ->getHazardRec()->getHazardType(SU, 0) | |||
2586 | != ScheduleHazardRecognizer::NoHazard) | |||
2587 | return false; | |||
2588 | ||||
2589 | return true; | |||
2590 | } | |||
2591 | ||||
2592 | static bool canEnableCoalescing(SUnit *SU) { | |||
2593 | unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0; | |||
2594 | if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg) | |||
2595 | // CopyToReg should be close to its uses to facilitate coalescing and | |||
2596 | // avoid spilling. | |||
2597 | return true; | |||
2598 | ||||
2599 | if (Opc == TargetOpcode::EXTRACT_SUBREG || | |||
2600 | Opc == TargetOpcode::SUBREG_TO_REG || | |||
2601 | Opc == TargetOpcode::INSERT_SUBREG) | |||
2602 | // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be | |||
2603 | // close to their uses to facilitate coalescing. | |||
2604 | return true; | |||
2605 | ||||
2606 | if (SU->NumPreds == 0 && SU->NumSuccs != 0) | |||
2607 | // If SU does not have a register def, schedule it close to its uses | |||
2608 | // because it does not lengthen any live ranges. | |||
2609 | return true; | |||
2610 | ||||
2611 | return false; | |||
2612 | } | |||
2613 | ||||
2614 | // list-ilp is currently an experimental scheduler that allows various | |||
2615 | // heuristics to be enabled prior to the normal register reduction logic. | |||
2616 | bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { | |||
2617 | if (int res = checkSpecialNodes(left, right)) | |||
2618 | return res > 0; | |||
2619 | ||||
2620 | if (left->isCall || right->isCall) | |||
2621 | // No way to compute latency of calls. | |||
2622 | return BURRSort(left, right, SPQ); | |||
2623 | ||||
2624 | unsigned LLiveUses = 0, RLiveUses = 0; | |||
2625 | int LPDiff = 0, RPDiff = 0; | |||
2626 | if (!DisableSchedRegPressure || !DisableSchedLiveUses) { | |||
2627 | LPDiff = SPQ->RegPressureDiff(left, LLiveUses); | |||
2628 | RPDiff = SPQ->RegPressureDiff(right, RLiveUses); | |||
2629 | } | |||
2630 | if (!DisableSchedRegPressure && LPDiff != RPDiff) { | |||
2631 | DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiffdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiff << " != SU(" << right->NodeNum << "): " << RPDiff << "\n"; } } while (0) | |||
2632 | << " != SU(" << right->NodeNum << "): " << RPDiff << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiff << " != SU(" << right->NodeNum << "): " << RPDiff << "\n"; } } while (0); | |||
2633 | return LPDiff > RPDiff; | |||
2634 | } | |||
2635 | ||||
2636 | if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) { | |||
2637 | bool LReduce = canEnableCoalescing(left); | |||
2638 | bool RReduce = canEnableCoalescing(right); | |||
2639 | if (LReduce && !RReduce) return false; | |||
2640 | if (RReduce && !LReduce) return true; | |||
2641 | } | |||
2642 | ||||
2643 | if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) { | |||
2644 | DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUsesdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Live uses SU(" << left ->NodeNum << "): " << LLiveUses << " != SU(" << right->NodeNum << "): " << RLiveUses << "\n"; } } while (0) | |||
2645 | << " != SU(" << right->NodeNum << "): " << RLiveUses << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Live uses SU(" << left ->NodeNum << "): " << LLiveUses << " != SU(" << right->NodeNum << "): " << RLiveUses << "\n"; } } while (0); | |||
2646 | return LLiveUses < RLiveUses; | |||
2647 | } | |||
2648 | ||||
2649 | if (!DisableSchedStalls) { | |||
2650 | bool LStall = BUHasStall(left, left->getHeight(), SPQ); | |||
2651 | bool RStall = BUHasStall(right, right->getHeight(), SPQ); | |||
2652 | if (LStall != RStall) | |||
2653 | return left->getHeight() > right->getHeight(); | |||
2654 | } | |||
2655 | ||||
2656 | if (!DisableSchedCriticalPath) { | |||
2657 | int spread = (int)left->getDepth() - (int)right->getDepth(); | |||
2658 | if (std::abs(spread) > MaxReorderWindow) { | |||
2659 | DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Depth of SU(" << left ->NodeNum << "): " << left->getDepth() << " != SU(" << right->NodeNum << "): " << right->getDepth() << "\n"; } } while (0) | |||
2660 | << left->getDepth() << " != SU(" << right->NodeNum << "): "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Depth of SU(" << left ->NodeNum << "): " << left->getDepth() << " != SU(" << right->NodeNum << "): " << right->getDepth() << "\n"; } } while (0) | |||
2661 | << right->getDepth() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Depth of SU(" << left ->NodeNum << "): " << left->getDepth() << " != SU(" << right->NodeNum << "): " << right->getDepth() << "\n"; } } while (0); | |||
2662 | return left->getDepth() < right->getDepth(); | |||
2663 | } | |||
2664 | } | |||
2665 | ||||
2666 | if (!DisableSchedHeight && left->getHeight() != right->getHeight()) { | |||
2667 | int spread = (int)left->getHeight() - (int)right->getHeight(); | |||
2668 | if (std::abs(spread) > MaxReorderWindow) | |||
2669 | return left->getHeight() > right->getHeight(); | |||
2670 | } | |||
2671 | ||||
2672 | return BURRSort(left, right, SPQ); | |||
2673 | } | |||
2674 | ||||
2675 | void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) { | |||
2676 | SUnits = &sunits; | |||
2677 | // Add pseudo dependency edges for two-address nodes. | |||
2678 | if (!Disable2AddrHack) | |||
2679 | AddPseudoTwoAddrDeps(); | |||
2680 | // Reroute edges to nodes with multiple uses. | |||
2681 | if (!TracksRegPressure && !SrcOrder) | |||
2682 | PrescheduleNodesWithMultipleUses(); | |||
2683 | // Calculate node priorities. | |||
2684 | CalculateSethiUllmanNumbers(); | |||
2685 | ||||
2686 | // For single block loops, mark nodes that look like canonical IV increments. | |||
2687 | if (scheduleDAG->BB->isSuccessor(scheduleDAG->BB)) { | |||
2688 | for (unsigned i = 0, e = sunits.size(); i != e; ++i) { | |||
2689 | initVRegCycle(&sunits[i]); | |||
2690 | } | |||
2691 | } | |||
2692 | } | |||
2693 | ||||
2694 | //===----------------------------------------------------------------------===// | |||
2695 | // Preschedule for Register Pressure | |||
2696 | //===----------------------------------------------------------------------===// | |||
2697 | ||||
2698 | bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) { | |||
2699 | if (SU->isTwoAddress) { | |||
2700 | unsigned Opc = SU->getNode()->getMachineOpcode(); | |||
2701 | const MCInstrDesc &MCID = TII->get(Opc); | |||
2702 | unsigned NumRes = MCID.getNumDefs(); | |||
2703 | unsigned NumOps = MCID.getNumOperands() - NumRes; | |||
2704 | for (unsigned i = 0; i != NumOps; ++i) { | |||
2705 | if (MCID.getOperandConstraint(i+NumRes, MCOI::TIED_TO) != -1) { | |||
2706 | SDNode *DU = SU->getNode()->getOperand(i).getNode(); | |||
2707 | if (DU->getNodeId() != -1 && | |||
2708 | Op->OrigNode == &(*SUnits)[DU->getNodeId()]) | |||
2709 | return true; | |||
2710 | } | |||
2711 | } | |||
2712 | } | |||
2713 | return false; | |||
2714 | } | |||
2715 | ||||
2716 | /// canClobberReachingPhysRegUse - True if SU would clobber one of it's | |||
2717 | /// successor's explicit physregs whose definition can reach DepSU. | |||
2718 | /// i.e. DepSU should not be scheduled above SU. | |||
2719 | static bool canClobberReachingPhysRegUse(const SUnit *DepSU, const SUnit *SU, | |||
2720 | ScheduleDAGRRList *scheduleDAG, | |||
2721 | const TargetInstrInfo *TII, | |||
2722 | const TargetRegisterInfo *TRI) { | |||
2723 | const MCPhysReg *ImpDefs | |||
2724 | = TII->get(SU->getNode()->getMachineOpcode()).getImplicitDefs(); | |||
2725 | const uint32_t *RegMask = getNodeRegMask(SU->getNode()); | |||
2726 | if(!ImpDefs && !RegMask) | |||
2727 | return false; | |||
2728 | ||||
2729 | for (SUnit::const_succ_iterator SI = SU->Succs.begin(), SE = SU->Succs.end(); | |||
2730 | SI != SE; ++SI) { | |||
2731 | SUnit *SuccSU = SI->getSUnit(); | |||
2732 | for (SUnit::const_pred_iterator PI = SuccSU->Preds.begin(), | |||
2733 | PE = SuccSU->Preds.end(); PI != PE; ++PI) { | |||
2734 | if (!PI->isAssignedRegDep()) | |||
2735 | continue; | |||
2736 | ||||
2737 | if (RegMask && MachineOperand::clobbersPhysReg(RegMask, PI->getReg()) && | |||
2738 | scheduleDAG->IsReachable(DepSU, PI->getSUnit())) | |||
2739 | return true; | |||
2740 | ||||
2741 | if (ImpDefs) | |||
2742 | for (const MCPhysReg *ImpDef = ImpDefs; *ImpDef; ++ImpDef) | |||
2743 | // Return true if SU clobbers this physical register use and the | |||
2744 | // definition of the register reaches from DepSU. IsReachable queries | |||
2745 | // a topological forward sort of the DAG (following the successors). | |||
2746 | if (TRI->regsOverlap(*ImpDef, PI->getReg()) && | |||
2747 | scheduleDAG->IsReachable(DepSU, PI->getSUnit())) | |||
2748 | return true; | |||
2749 | } | |||
2750 | } | |||
2751 | return false; | |||
2752 | } | |||
2753 | ||||
2754 | /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's | |||
2755 | /// physical register defs. | |||
2756 | static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU, | |||
2757 | const TargetInstrInfo *TII, | |||
2758 | const TargetRegisterInfo *TRI) { | |||
2759 | SDNode *N = SuccSU->getNode(); | |||
2760 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); | |||
2761 | const MCPhysReg *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs(); | |||
2762 | assert(ImpDefs && "Caller should check hasPhysRegDefs")((ImpDefs && "Caller should check hasPhysRegDefs") ? static_cast <void> (0) : __assert_fail ("ImpDefs && \"Caller should check hasPhysRegDefs\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2762, __PRETTY_FUNCTION__)); | |||
2763 | for (const SDNode *SUNode = SU->getNode(); SUNode; | |||
2764 | SUNode = SUNode->getGluedNode()) { | |||
2765 | if (!SUNode->isMachineOpcode()) | |||
2766 | continue; | |||
2767 | const MCPhysReg *SUImpDefs = | |||
2768 | TII->get(SUNode->getMachineOpcode()).getImplicitDefs(); | |||
2769 | const uint32_t *SURegMask = getNodeRegMask(SUNode); | |||
2770 | if (!SUImpDefs && !SURegMask) | |||
2771 | continue; | |||
2772 | for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) { | |||
2773 | MVT VT = N->getSimpleValueType(i); | |||
2774 | if (VT == MVT::Glue || VT == MVT::Other) | |||
2775 | continue; | |||
2776 | if (!N->hasAnyUseOfValue(i)) | |||
2777 | continue; | |||
2778 | unsigned Reg = ImpDefs[i - NumDefs]; | |||
2779 | if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg)) | |||
2780 | return true; | |||
2781 | if (!SUImpDefs) | |||
2782 | continue; | |||
2783 | for (;*SUImpDefs; ++SUImpDefs) { | |||
2784 | unsigned SUReg = *SUImpDefs; | |||
2785 | if (TRI->regsOverlap(Reg, SUReg)) | |||
2786 | return true; | |||
2787 | } | |||
2788 | } | |||
2789 | } | |||
2790 | return false; | |||
2791 | } | |||
2792 | ||||
2793 | /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses | |||
2794 | /// are not handled well by the general register pressure reduction | |||
2795 | /// heuristics. When presented with code like this: | |||
2796 | /// | |||
2797 | /// N | |||
2798 | /// / | | |||
2799 | /// / | | |||
2800 | /// U store | |||
2801 | /// | | |||
2802 | /// ... | |||
2803 | /// | |||
2804 | /// the heuristics tend to push the store up, but since the | |||
2805 | /// operand of the store has another use (U), this would increase | |||
2806 | /// the length of that other use (the U->N edge). | |||
2807 | /// | |||
2808 | /// This function transforms code like the above to route U's | |||
2809 | /// dependence through the store when possible, like this: | |||
2810 | /// | |||
2811 | /// N | |||
2812 | /// || | |||
2813 | /// || | |||
2814 | /// store | |||
2815 | /// | | |||
2816 | /// U | |||
2817 | /// | | |||
2818 | /// ... | |||
2819 | /// | |||
2820 | /// This results in the store being scheduled immediately | |||
2821 | /// after N, which shortens the U->N live range, reducing | |||
2822 | /// register pressure. | |||
2823 | /// | |||
2824 | void RegReductionPQBase::PrescheduleNodesWithMultipleUses() { | |||
2825 | // Visit all the nodes in topological order, working top-down. | |||
2826 | for (unsigned i = 0, e = SUnits->size(); i != e; ++i) { | |||
2827 | SUnit *SU = &(*SUnits)[i]; | |||
2828 | // For now, only look at nodes with no data successors, such as stores. | |||
2829 | // These are especially important, due to the heuristics in | |||
2830 | // getNodePriority for nodes with no data successors. | |||
2831 | if (SU->NumSuccs != 0) | |||
2832 | continue; | |||
2833 | // For now, only look at nodes with exactly one data predecessor. | |||
2834 | if (SU->NumPreds != 1) | |||
2835 | continue; | |||
2836 | // Avoid prescheduling copies to virtual registers, which don't behave | |||
2837 | // like other nodes from the perspective of scheduling heuristics. | |||
2838 | if (SDNode *N = SU->getNode()) | |||
2839 | if (N->getOpcode() == ISD::CopyToReg && | |||
2840 | TargetRegisterInfo::isVirtualRegister | |||
2841 | (cast<RegisterSDNode>(N->getOperand(1))->getReg())) | |||
2842 | continue; | |||
2843 | ||||
2844 | // Locate the single data predecessor. | |||
2845 | SUnit *PredSU = nullptr; | |||
2846 | for (SUnit::const_pred_iterator II = SU->Preds.begin(), | |||
2847 | EE = SU->Preds.end(); II != EE; ++II) | |||
2848 | if (!II->isCtrl()) { | |||
2849 | PredSU = II->getSUnit(); | |||
2850 | break; | |||
2851 | } | |||
2852 | assert(PredSU)((PredSU) ? static_cast<void> (0) : __assert_fail ("PredSU" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2852, __PRETTY_FUNCTION__)); | |||
2853 | ||||
2854 | // Don't rewrite edges that carry physregs, because that requires additional | |||
2855 | // support infrastructure. | |||
2856 | if (PredSU->hasPhysRegDefs) | |||
2857 | continue; | |||
2858 | // Short-circuit the case where SU is PredSU's only data successor. | |||
2859 | if (PredSU->NumSuccs == 1) | |||
2860 | continue; | |||
2861 | // Avoid prescheduling to copies from virtual registers, which don't behave | |||
2862 | // like other nodes from the perspective of scheduling heuristics. | |||
2863 | if (SDNode *N = SU->getNode()) | |||
2864 | if (N->getOpcode() == ISD::CopyFromReg && | |||
2865 | TargetRegisterInfo::isVirtualRegister | |||
2866 | (cast<RegisterSDNode>(N->getOperand(1))->getReg())) | |||
2867 | continue; | |||
2868 | ||||
2869 | // Perform checks on the successors of PredSU. | |||
2870 | for (SUnit::const_succ_iterator II = PredSU->Succs.begin(), | |||
2871 | EE = PredSU->Succs.end(); II != EE; ++II) { | |||
2872 | SUnit *PredSuccSU = II->getSUnit(); | |||
2873 | if (PredSuccSU == SU) continue; | |||
2874 | // If PredSU has another successor with no data successors, for | |||
2875 | // now don't attempt to choose either over the other. | |||
2876 | if (PredSuccSU->NumSuccs == 0) | |||
2877 | goto outer_loop_continue; | |||
2878 | // Don't break physical register dependencies. | |||
2879 | if (SU->hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs) | |||
2880 | if (canClobberPhysRegDefs(PredSuccSU, SU, TII, TRI)) | |||
2881 | goto outer_loop_continue; | |||
2882 | // Don't introduce graph cycles. | |||
2883 | if (scheduleDAG->IsReachable(SU, PredSuccSU)) | |||
2884 | goto outer_loop_continue; | |||
2885 | } | |||
2886 | ||||
2887 | // Ok, the transformation is safe and the heuristics suggest it is | |||
2888 | // profitable. Update the graph. | |||
2889 | DEBUG(dbgs() << " Prescheduling SU #" << SU->NodeNumdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Prescheduling SU #" << SU->NodeNum << " next to PredSU #" << PredSU-> NodeNum << " to guide scheduling in the presence of multiple uses\n" ; } } while (0) | |||
2890 | << " next to PredSU #" << PredSU->NodeNumdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Prescheduling SU #" << SU->NodeNum << " next to PredSU #" << PredSU-> NodeNum << " to guide scheduling in the presence of multiple uses\n" ; } } while (0) | |||
2891 | << " to guide scheduling in the presence of multiple uses\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Prescheduling SU #" << SU->NodeNum << " next to PredSU #" << PredSU-> NodeNum << " to guide scheduling in the presence of multiple uses\n" ; } } while (0); | |||
2892 | for (unsigned i = 0; i != PredSU->Succs.size(); ++i) { | |||
2893 | SDep Edge = PredSU->Succs[i]; | |||
2894 | assert(!Edge.isAssignedRegDep())((!Edge.isAssignedRegDep()) ? static_cast<void> (0) : __assert_fail ("!Edge.isAssignedRegDep()", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn257205/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2894, __PRETTY_FUNCTION__)); | |||
2895 | SUnit *SuccSU = Edge.getSUnit(); | |||
2896 | if (SuccSU != SU) { | |||
2897 | Edge.setSUnit(PredSU); | |||
2898 | scheduleDAG->RemovePred(SuccSU, Edge); | |||
2899 | scheduleDAG->AddPred(SU, Edge); | |||
2900 | Edge.setSUnit(SU); | |||
2901 | scheduleDAG->AddPred(SuccSU, Edge); | |||
2902 | --i; | |||
2903 | } | |||
2904 | } | |||
2905 | outer_loop_continue:; | |||
2906 | } | |||
2907 | } | |||
2908 | ||||
2909 | /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses | |||
2910 | /// it as a def&use operand. Add a pseudo control edge from it to the other | |||
2911 | /// node (if it won't create a cycle) so the two-address one will be scheduled | |||
2912 | /// first (lower in the schedule). If both nodes are two-address, favor the | |||
2913 | /// one that has a CopyToReg use (more likely to be a loop induction update). | |||
2914 | /// If both are two-address, but one is commutable while the other is not | |||
2915 | /// commutable, favor the one that's not commutable. | |||
2916 | void RegReductionPQBase::AddPseudoTwoAddrDeps() { | |||
2917 | for (unsigned i = 0, e = SUnits->size(); i != e; ++i) { | |||
2918 | SUnit *SU = &(*SUnits)[i]; | |||
2919 | if (!SU->isTwoAddress) | |||
2920 | continue; | |||
2921 | ||||
2922 | SDNode *Node = SU->getNode(); | |||
2923 | if (!Node || !Node->isMachineOpcode() || SU->getNode()->getGluedNode()) | |||
2924 | continue; | |||
2925 | ||||
2926 | bool isLiveOut = hasOnlyLiveOutUses(SU); | |||
2927 | unsigned Opc = Node->getMachineOpcode(); | |||
2928 | const MCInstrDesc &MCID = TII->get(Opc); | |||
2929 | unsigned NumRes = MCID.getNumDefs(); | |||
2930 | unsigned NumOps = MCID.getNumOperands() - NumRes; | |||
2931 | for (unsigned j = 0; j != NumOps; ++j) { | |||
2932 | if (MCID.getOperandConstraint(j+NumRes, MCOI::TIED_TO) == -1) | |||
2933 | continue; | |||
2934 | SDNode *DU = SU->getNode()->getOperand(j).getNode(); | |||
2935 | if (DU->getNodeId() == -1) | |||
2936 | continue; | |||
2937 | const SUnit *DUSU = &(*SUnits)[DU->getNodeId()]; | |||
2938 | if (!DUSU) continue; | |||
2939 | for (SUnit::const_succ_iterator I = DUSU->Succs.begin(), | |||
2940 | E = DUSU->Succs.end(); I != E; ++I) { | |||
2941 | if (I->isCtrl()) continue; | |||
2942 | SUnit *SuccSU = I->getSUnit(); | |||
2943 | if (SuccSU == SU) | |||
2944 | continue; | |||
2945 | // Be conservative. Ignore if nodes aren't at roughly the same | |||
2946 | // depth and height. | |||
2947 | if (SuccSU->getHeight() < SU->getHeight() && | |||
2948 | (SU->getHeight() - SuccSU->getHeight()) > 1) | |||
2949 | continue; | |||
2950 | // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge | |||
2951 | // constrains whatever is using the copy, instead of the copy | |||
2952 | // itself. In the case that the copy is coalesced, this | |||
2953 | // preserves the intent of the pseudo two-address heurietics. | |||
2954 | while (SuccSU->Succs.size() == 1 && | |||
2955 | SuccSU->getNode()->isMachineOpcode() && | |||
2956 | SuccSU->getNode()->getMachineOpcode() == | |||
2957 | TargetOpcode::COPY_TO_REGCLASS) | |||
2958 | SuccSU = SuccSU->Succs.front().getSUnit(); | |||
2959 | // Don't constrain non-instruction nodes. | |||
2960 | if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode()) | |||
2961 | continue; | |||
2962 | // Don't constrain nodes with physical register defs if the | |||
2963 | // predecessor can clobber them. | |||
2964 | if (SuccSU->hasPhysRegDefs && SU->hasPhysRegClobbers) { | |||
2965 | if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI)) | |||
2966 | continue; | |||
2967 | } | |||
2968 | // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG; | |||
2969 | // these may be coalesced away. We want them close to their uses. | |||
2970 | unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode(); | |||
2971 | if (SuccOpc == TargetOpcode::EXTRACT_SUBREG || | |||
2972 | SuccOpc == TargetOpcode::INSERT_SUBREG || | |||
2973 | SuccOpc == TargetOpcode::SUBREG_TO_REG) | |||
2974 | continue; | |||
2975 | if (!canClobberReachingPhysRegUse(SuccSU, SU, scheduleDAG, TII, TRI) && | |||
2976 | (!canClobber(SuccSU, DUSU) || | |||
2977 | (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) || | |||
2978 | (!SU->isCommutable && SuccSU->isCommutable)) && | |||
2979 | !scheduleDAG->IsReachable(SuccSU, SU)) { | |||
2980 | DEBUG(dbgs() << " Adding a pseudo-two-addr edge from SU #"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Adding a pseudo-two-addr edge from SU #" << SU->NodeNum << " to SU #" << SuccSU-> NodeNum << "\n"; } } while (0) | |||
2981 | << SU->NodeNum << " to SU #" << SuccSU->NodeNum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Adding a pseudo-two-addr edge from SU #" << SU->NodeNum << " to SU #" << SuccSU-> NodeNum << "\n"; } } while (0); | |||
2982 | scheduleDAG->AddPred(SU, SDep(SuccSU, SDep::Artificial)); | |||
2983 | } | |||
2984 | } | |||
2985 | } | |||
2986 | } | |||
2987 | } | |||
2988 | ||||
2989 | //===----------------------------------------------------------------------===// | |||
2990 | // Public Constructor Functions | |||
2991 | //===----------------------------------------------------------------------===// | |||
2992 | ||||
2993 | llvm::ScheduleDAGSDNodes * | |||
2994 | llvm::createBURRListDAGScheduler(SelectionDAGISel *IS, | |||
2995 | CodeGenOpt::Level OptLevel) { | |||
2996 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); | |||
2997 | const TargetInstrInfo *TII = STI.getInstrInfo(); | |||
2998 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); | |||
2999 | ||||
3000 | BURegReductionPriorityQueue *PQ = | |||
3001 | new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, nullptr); | |||
3002 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel); | |||
3003 | PQ->setScheduleDAG(SD); | |||
3004 | return SD; | |||
3005 | } | |||
3006 | ||||
3007 | llvm::ScheduleDAGSDNodes * | |||
3008 | llvm::createSourceListDAGScheduler(SelectionDAGISel *IS, | |||
3009 | CodeGenOpt::Level OptLevel) { | |||
3010 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); | |||
3011 | const TargetInstrInfo *TII = STI.getInstrInfo(); | |||
3012 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); | |||
3013 | ||||
3014 | SrcRegReductionPriorityQueue *PQ = | |||
3015 | new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, nullptr); | |||
3016 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel); | |||
3017 | PQ->setScheduleDAG(SD); | |||
3018 | return SD; | |||
3019 | } | |||
3020 | ||||
3021 | llvm::ScheduleDAGSDNodes * | |||
3022 | llvm::createHybridListDAGScheduler(SelectionDAGISel *IS, | |||
3023 | CodeGenOpt::Level OptLevel) { | |||
3024 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); | |||
3025 | const TargetInstrInfo *TII = STI.getInstrInfo(); | |||
3026 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); | |||
3027 | const TargetLowering *TLI = IS->TLI; | |||
3028 | ||||
3029 | HybridBURRPriorityQueue *PQ = | |||
3030 | new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI); | |||
3031 | ||||
3032 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel); | |||
3033 | PQ->setScheduleDAG(SD); | |||
3034 | return SD; | |||
3035 | } | |||
3036 | ||||
3037 | llvm::ScheduleDAGSDNodes * | |||
3038 | llvm::createILPListDAGScheduler(SelectionDAGISel *IS, | |||
3039 | CodeGenOpt::Level OptLevel) { | |||
3040 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); | |||
3041 | const TargetInstrInfo *TII = STI.getInstrInfo(); | |||
3042 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); | |||
3043 | const TargetLowering *TLI = IS->TLI; | |||
3044 | ||||
3045 | ILPBURRPriorityQueue *PQ = | |||
3046 | new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI); | |||
3047 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel); | |||
3048 | PQ->setScheduleDAG(SD); | |||
3049 | return SD; | |||
3050 | } |