File: | lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp |
Warning: | line 1156, column 5 Value stored to 'N' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
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 "ScheduleDAGSDNodes.h" |
19 | #include "llvm/ADT/ArrayRef.h" |
20 | #include "llvm/ADT/DenseMap.h" |
21 | #include "llvm/ADT/STLExtras.h" |
22 | #include "llvm/ADT/SmallSet.h" |
23 | #include "llvm/ADT/SmallVector.h" |
24 | #include "llvm/ADT/Statistic.h" |
25 | #include "llvm/CodeGen/ISDOpcodes.h" |
26 | #include "llvm/CodeGen/MachineFunction.h" |
27 | #include "llvm/CodeGen/MachineOperand.h" |
28 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
29 | #include "llvm/CodeGen/ScheduleDAG.h" |
30 | #include "llvm/CodeGen/ScheduleHazardRecognizer.h" |
31 | #include "llvm/CodeGen/SchedulerRegistry.h" |
32 | #include "llvm/CodeGen/SelectionDAGISel.h" |
33 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
34 | #include "llvm/CodeGen/TargetInstrInfo.h" |
35 | #include "llvm/CodeGen/TargetLowering.h" |
36 | #include "llvm/CodeGen/TargetOpcodes.h" |
37 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
38 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
39 | #include "llvm/IR/InlineAsm.h" |
40 | #include "llvm/MC/MCInstrDesc.h" |
41 | #include "llvm/MC/MCRegisterInfo.h" |
42 | #include "llvm/Support/Casting.h" |
43 | #include "llvm/Support/CodeGen.h" |
44 | #include "llvm/Support/CommandLine.h" |
45 | #include "llvm/Support/Compiler.h" |
46 | #include "llvm/Support/Debug.h" |
47 | #include "llvm/Support/ErrorHandling.h" |
48 | #include "llvm/Support/MachineValueType.h" |
49 | #include "llvm/Support/raw_ostream.h" |
50 | #include <algorithm> |
51 | #include <cassert> |
52 | #include <cstdint> |
53 | #include <cstdlib> |
54 | #include <iterator> |
55 | #include <limits> |
56 | #include <memory> |
57 | #include <utility> |
58 | #include <vector> |
59 | |
60 | using namespace llvm; |
61 | |
62 | #define DEBUG_TYPE"pre-RA-sched" "pre-RA-sched" |
63 | |
64 | STATISTIC(NumBacktracks, "Number of times scheduler backtracked")static llvm::Statistic NumBacktracks = {"pre-RA-sched", "NumBacktracks" , "Number of times scheduler backtracked", {0}, {false}}; |
65 | STATISTIC(NumUnfolds, "Number of nodes unfolded")static llvm::Statistic NumUnfolds = {"pre-RA-sched", "NumUnfolds" , "Number of nodes unfolded", {0}, {false}}; |
66 | STATISTIC(NumDups, "Number of duplicated nodes")static llvm::Statistic NumDups = {"pre-RA-sched", "NumDups", "Number of duplicated nodes" , {0}, {false}}; |
67 | STATISTIC(NumPRCopies, "Number of physical register copies")static llvm::Statistic NumPRCopies = {"pre-RA-sched", "NumPRCopies" , "Number of physical register copies", {0}, {false}}; |
68 | |
69 | static RegisterScheduler |
70 | burrListDAGScheduler("list-burr", |
71 | "Bottom-up register reduction list scheduling", |
72 | createBURRListDAGScheduler); |
73 | |
74 | static RegisterScheduler |
75 | sourceListDAGScheduler("source", |
76 | "Similar to list-burr but schedules in source " |
77 | "order when possible", |
78 | createSourceListDAGScheduler); |
79 | |
80 | static RegisterScheduler |
81 | hybridListDAGScheduler("list-hybrid", |
82 | "Bottom-up register pressure aware list scheduling " |
83 | "which tries to balance latency and register pressure", |
84 | createHybridListDAGScheduler); |
85 | |
86 | static RegisterScheduler |
87 | ILPListDAGScheduler("list-ilp", |
88 | "Bottom-up register pressure aware list scheduling " |
89 | "which tries to balance ILP and register pressure", |
90 | createILPListDAGScheduler); |
91 | |
92 | static cl::opt<bool> DisableSchedCycles( |
93 | "disable-sched-cycles", cl::Hidden, cl::init(false), |
94 | cl::desc("Disable cycle-level precision during preRA scheduling")); |
95 | |
96 | // Temporary sched=list-ilp flags until the heuristics are robust. |
97 | // Some options are also available under sched=list-hybrid. |
98 | static cl::opt<bool> DisableSchedRegPressure( |
99 | "disable-sched-reg-pressure", cl::Hidden, cl::init(false), |
100 | cl::desc("Disable regpressure priority in sched=list-ilp")); |
101 | static cl::opt<bool> DisableSchedLiveUses( |
102 | "disable-sched-live-uses", cl::Hidden, cl::init(true), |
103 | cl::desc("Disable live use priority in sched=list-ilp")); |
104 | static cl::opt<bool> DisableSchedVRegCycle( |
105 | "disable-sched-vrcycle", cl::Hidden, cl::init(false), |
106 | cl::desc("Disable virtual register cycle interference checks")); |
107 | static cl::opt<bool> DisableSchedPhysRegJoin( |
108 | "disable-sched-physreg-join", cl::Hidden, cl::init(false), |
109 | cl::desc("Disable physreg def-use affinity")); |
110 | static cl::opt<bool> DisableSchedStalls( |
111 | "disable-sched-stalls", cl::Hidden, cl::init(true), |
112 | cl::desc("Disable no-stall priority in sched=list-ilp")); |
113 | static cl::opt<bool> DisableSchedCriticalPath( |
114 | "disable-sched-critical-path", cl::Hidden, cl::init(false), |
115 | cl::desc("Disable critical path priority in sched=list-ilp")); |
116 | static cl::opt<bool> DisableSchedHeight( |
117 | "disable-sched-height", cl::Hidden, cl::init(false), |
118 | cl::desc("Disable scheduled-height priority in sched=list-ilp")); |
119 | static cl::opt<bool> Disable2AddrHack( |
120 | "disable-2addr-hack", cl::Hidden, cl::init(true), |
121 | cl::desc("Disable scheduler's two-address hack")); |
122 | |
123 | static cl::opt<int> MaxReorderWindow( |
124 | "max-sched-reorder", cl::Hidden, cl::init(6), |
125 | cl::desc("Number of instructions to allow ahead of the critical path " |
126 | "in sched=list-ilp")); |
127 | |
128 | static cl::opt<unsigned> AvgIPC( |
129 | "sched-avg-ipc", cl::Hidden, cl::init(1), |
130 | cl::desc("Average inst/cycle whan no target itinerary exists.")); |
131 | |
132 | namespace { |
133 | |
134 | //===----------------------------------------------------------------------===// |
135 | /// ScheduleDAGRRList - The actual register reduction list scheduler |
136 | /// implementation. This supports both top-down and bottom-up scheduling. |
137 | /// |
138 | class ScheduleDAGRRList : public ScheduleDAGSDNodes { |
139 | private: |
140 | /// NeedLatency - True if the scheduler will make use of latency information. |
141 | bool NeedLatency; |
142 | |
143 | /// AvailableQueue - The priority queue to use for the available SUnits. |
144 | SchedulingPriorityQueue *AvailableQueue; |
145 | |
146 | /// PendingQueue - This contains all of the instructions whose operands have |
147 | /// been issued, but their results are not ready yet (due to the latency of |
148 | /// the operation). Once the operands becomes available, the instruction is |
149 | /// added to the AvailableQueue. |
150 | std::vector<SUnit *> PendingQueue; |
151 | |
152 | /// HazardRec - The hazard recognizer to use. |
153 | ScheduleHazardRecognizer *HazardRec; |
154 | |
155 | /// CurCycle - The current scheduler state corresponds to this cycle. |
156 | unsigned CurCycle = 0; |
157 | |
158 | /// MinAvailableCycle - Cycle of the soonest available instruction. |
159 | unsigned MinAvailableCycle; |
160 | |
161 | /// IssueCount - Count instructions issued in this cycle |
162 | /// Currently valid only for bottom-up scheduling. |
163 | unsigned IssueCount; |
164 | |
165 | /// LiveRegDefs - A set of physical registers and their definition |
166 | /// that are "live". These nodes must be scheduled before any other nodes that |
167 | /// modifies the registers can be scheduled. |
168 | unsigned NumLiveRegs; |
169 | std::unique_ptr<SUnit*[]> LiveRegDefs; |
170 | std::unique_ptr<SUnit*[]> LiveRegGens; |
171 | |
172 | // Collect interferences between physical register use/defs. |
173 | // Each interference is an SUnit and set of physical registers. |
174 | SmallVector<SUnit*, 4> Interferences; |
175 | |
176 | using LRegsMapT = DenseMap<SUnit *, SmallVector<unsigned, 4>>; |
177 | |
178 | LRegsMapT LRegsMap; |
179 | |
180 | /// Topo - A topological ordering for SUnits which permits fast IsReachable |
181 | /// and similar queries. |
182 | ScheduleDAGTopologicalSort Topo; |
183 | |
184 | // Hack to keep track of the inverse of FindCallSeqStart without more crazy |
185 | // DAG crawling. |
186 | DenseMap<SUnit*, SUnit*> CallSeqEndForStart; |
187 | |
188 | public: |
189 | ScheduleDAGRRList(MachineFunction &mf, bool needlatency, |
190 | SchedulingPriorityQueue *availqueue, |
191 | CodeGenOpt::Level OptLevel) |
192 | : ScheduleDAGSDNodes(mf), |
193 | NeedLatency(needlatency), AvailableQueue(availqueue), |
194 | Topo(SUnits, nullptr) { |
195 | const TargetSubtargetInfo &STI = mf.getSubtarget(); |
196 | if (DisableSchedCycles || !NeedLatency) |
197 | HazardRec = new ScheduleHazardRecognizer(); |
198 | else |
199 | HazardRec = STI.getInstrInfo()->CreateTargetHazardRecognizer(&STI, this); |
200 | } |
201 | |
202 | ~ScheduleDAGRRList() override { |
203 | delete HazardRec; |
204 | delete AvailableQueue; |
205 | } |
206 | |
207 | void Schedule() override; |
208 | |
209 | ScheduleHazardRecognizer *getHazardRec() { return HazardRec; } |
210 | |
211 | /// IsReachable - Checks if SU is reachable from TargetSU. |
212 | bool IsReachable(const SUnit *SU, const SUnit *TargetSU) { |
213 | return Topo.IsReachable(SU, TargetSU); |
214 | } |
215 | |
216 | /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will |
217 | /// create a cycle. |
218 | bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) { |
219 | return Topo.WillCreateCycle(SU, TargetSU); |
220 | } |
221 | |
222 | /// AddPred - adds a predecessor edge to SUnit SU. |
223 | /// This returns true if this is a new predecessor. |
224 | /// Updates the topological ordering if required. |
225 | void AddPred(SUnit *SU, const SDep &D) { |
226 | Topo.AddPred(SU, D.getSUnit()); |
227 | SU->addPred(D); |
228 | } |
229 | |
230 | /// RemovePred - removes a predecessor edge from SUnit SU. |
231 | /// This returns true if an edge was removed. |
232 | /// Updates the topological ordering if required. |
233 | void RemovePred(SUnit *SU, const SDep &D) { |
234 | Topo.RemovePred(SU, D.getSUnit()); |
235 | SU->removePred(D); |
236 | } |
237 | |
238 | private: |
239 | bool isReady(SUnit *SU) { |
240 | return DisableSchedCycles || !AvailableQueue->hasReadyFilter() || |
241 | AvailableQueue->isReady(SU); |
242 | } |
243 | |
244 | void ReleasePred(SUnit *SU, const SDep *PredEdge); |
245 | void ReleasePredecessors(SUnit *SU); |
246 | void ReleasePending(); |
247 | void AdvanceToCycle(unsigned NextCycle); |
248 | void AdvancePastStalls(SUnit *SU); |
249 | void EmitNode(SUnit *SU); |
250 | void ScheduleNodeBottomUp(SUnit*); |
251 | void CapturePred(SDep *PredEdge); |
252 | void UnscheduleNodeBottomUp(SUnit*); |
253 | void RestoreHazardCheckerBottomUp(); |
254 | void BacktrackBottomUp(SUnit*, SUnit*); |
255 | SUnit *TryUnfoldSU(SUnit *); |
256 | SUnit *CopyAndMoveSuccessors(SUnit*); |
257 | void InsertCopiesAndMoveSuccs(SUnit*, unsigned, |
258 | const TargetRegisterClass*, |
259 | const TargetRegisterClass*, |
260 | SmallVectorImpl<SUnit*>&); |
261 | bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&); |
262 | |
263 | void releaseInterferences(unsigned Reg = 0); |
264 | |
265 | SUnit *PickNodeToScheduleBottomUp(); |
266 | void ListScheduleBottomUp(); |
267 | |
268 | /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it. |
269 | /// Updates the topological ordering if required. |
270 | SUnit *CreateNewSUnit(SDNode *N) { |
271 | unsigned NumSUnits = SUnits.size(); |
272 | SUnit *NewNode = newSUnit(N); |
273 | // Update the topological ordering. |
274 | if (NewNode->NodeNum >= NumSUnits) |
275 | Topo.InitDAGTopologicalSorting(); |
276 | return NewNode; |
277 | } |
278 | |
279 | /// CreateClone - Creates a new SUnit from an existing one. |
280 | /// Updates the topological ordering if required. |
281 | SUnit *CreateClone(SUnit *N) { |
282 | unsigned NumSUnits = SUnits.size(); |
283 | SUnit *NewNode = Clone(N); |
284 | // Update the topological ordering. |
285 | if (NewNode->NodeNum >= NumSUnits) |
286 | Topo.InitDAGTopologicalSorting(); |
287 | return NewNode; |
288 | } |
289 | |
290 | /// forceUnitLatencies - Register-pressure-reducing scheduling doesn't |
291 | /// need actual latency information but the hybrid scheduler does. |
292 | bool forceUnitLatencies() const override { |
293 | return !NeedLatency; |
294 | } |
295 | }; |
296 | |
297 | } // end anonymous namespace |
298 | |
299 | /// GetCostForDef - Looks up the register class and cost for a given definition. |
300 | /// Typically this just means looking up the representative register class, |
301 | /// but for untyped values (MVT::Untyped) it means inspecting the node's |
302 | /// opcode to determine what register class is being generated. |
303 | static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos, |
304 | const TargetLowering *TLI, |
305 | const TargetInstrInfo *TII, |
306 | const TargetRegisterInfo *TRI, |
307 | unsigned &RegClass, unsigned &Cost, |
308 | const MachineFunction &MF) { |
309 | MVT VT = RegDefPos.GetValue(); |
310 | |
311 | // Special handling for untyped values. These values can only come from |
312 | // the expansion of custom DAG-to-DAG patterns. |
313 | if (VT == MVT::Untyped) { |
314 | const SDNode *Node = RegDefPos.GetNode(); |
315 | |
316 | // Special handling for CopyFromReg of untyped values. |
317 | if (!Node->isMachineOpcode() && Node->getOpcode() == ISD::CopyFromReg) { |
318 | unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg(); |
319 | const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(Reg); |
320 | RegClass = RC->getID(); |
321 | Cost = 1; |
322 | return; |
323 | } |
324 | |
325 | unsigned Opcode = Node->getMachineOpcode(); |
326 | if (Opcode == TargetOpcode::REG_SEQUENCE) { |
327 | unsigned DstRCIdx = cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue(); |
328 | const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx); |
329 | RegClass = RC->getID(); |
330 | Cost = 1; |
331 | return; |
332 | } |
333 | |
334 | unsigned Idx = RegDefPos.GetIdx(); |
335 | const MCInstrDesc Desc = TII->get(Opcode); |
336 | const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI, MF); |
337 | RegClass = RC->getID(); |
338 | // FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a |
339 | // better way to determine it. |
340 | Cost = 1; |
341 | } else { |
342 | RegClass = TLI->getRepRegClassFor(VT)->getID(); |
343 | Cost = TLI->getRepRegClassCostFor(VT); |
344 | } |
345 | } |
346 | |
347 | /// Schedule - Schedule the DAG using list scheduling. |
348 | void ScheduleDAGRRList::Schedule() { |
349 | DEBUG(dbgs() << "********** List Scheduling " << printMBBReference(*BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "********** List Scheduling " << printMBBReference(*BB) << " '" << BB-> getName() << "' **********\n"; } } while (false) |
350 | << " '" << BB->getName() << "' **********\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "********** List Scheduling " << printMBBReference(*BB) << " '" << BB-> getName() << "' **********\n"; } } while (false); |
351 | |
352 | CurCycle = 0; |
353 | IssueCount = 0; |
354 | MinAvailableCycle = |
355 | DisableSchedCycles ? 0 : std::numeric_limits<unsigned>::max(); |
356 | NumLiveRegs = 0; |
357 | // Allocate slots for each physical register, plus one for a special register |
358 | // to track the virtual resource of a calling sequence. |
359 | LiveRegDefs.reset(new SUnit*[TRI->getNumRegs() + 1]()); |
360 | LiveRegGens.reset(new SUnit*[TRI->getNumRegs() + 1]()); |
361 | CallSeqEndForStart.clear(); |
362 | assert(Interferences.empty() && LRegsMap.empty() && "stale Interferences")(static_cast <bool> (Interferences.empty() && LRegsMap .empty() && "stale Interferences") ? void (0) : __assert_fail ("Interferences.empty() && LRegsMap.empty() && \"stale Interferences\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 362, __extension__ __PRETTY_FUNCTION__)); |
363 | |
364 | // Build the scheduling graph. |
365 | BuildSchedGraph(nullptr); |
366 | |
367 | DEBUG(for (SUnit &SU : SUnits)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { for (SUnit &SU : SUnits) SU.dumpAll(this ); } } while (false) |
368 | SU.dumpAll(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { for (SUnit &SU : SUnits) SU.dumpAll(this ); } } while (false); |
369 | Topo.InitDAGTopologicalSorting(); |
370 | |
371 | AvailableQueue->initNodes(SUnits); |
372 | |
373 | HazardRec->Reset(); |
374 | |
375 | // Execute the actual scheduling loop. |
376 | ListScheduleBottomUp(); |
377 | |
378 | AvailableQueue->releaseState(); |
379 | |
380 | DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (false) |
381 | dbgs() << "*** Final schedule ***\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (false) |
382 | dumpSchedule();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (false) |
383 | dbgs() << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (false) |
384 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { { dbgs() << "*** Final schedule ***\n" ; dumpSchedule(); dbgs() << '\n'; }; } } while (false); |
385 | } |
386 | |
387 | //===----------------------------------------------------------------------===// |
388 | // Bottom-Up Scheduling |
389 | //===----------------------------------------------------------------------===// |
390 | |
391 | /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to |
392 | /// the AvailableQueue if the count reaches zero. Also update its cycle bound. |
393 | void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) { |
394 | SUnit *PredSU = PredEdge->getSUnit(); |
395 | |
396 | #ifndef NDEBUG |
397 | if (PredSU->NumSuccsLeft == 0) { |
398 | dbgs() << "*** Scheduling failed! ***\n"; |
399 | PredSU->dump(this); |
400 | dbgs() << " has been released too many times!\n"; |
401 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 401); |
402 | } |
403 | #endif |
404 | --PredSU->NumSuccsLeft; |
405 | |
406 | if (!forceUnitLatencies()) { |
407 | // Updating predecessor's height. This is now the cycle when the |
408 | // predecessor can be scheduled without causing a pipeline stall. |
409 | PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency()); |
410 | } |
411 | |
412 | // If all the node's successors are scheduled, this node is ready |
413 | // to be scheduled. Ignore the special EntrySU node. |
414 | if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) { |
415 | PredSU->isAvailable = true; |
416 | |
417 | unsigned Height = PredSU->getHeight(); |
418 | if (Height < MinAvailableCycle) |
419 | MinAvailableCycle = Height; |
420 | |
421 | if (isReady(PredSU)) { |
422 | AvailableQueue->push(PredSU); |
423 | } |
424 | // CapturePred and others may have left the node in the pending queue, avoid |
425 | // adding it twice. |
426 | else if (!PredSU->isPending) { |
427 | PredSU->isPending = true; |
428 | PendingQueue.push_back(PredSU); |
429 | } |
430 | } |
431 | } |
432 | |
433 | /// IsChainDependent - Test if Outer is reachable from Inner through |
434 | /// chain dependencies. |
435 | static bool IsChainDependent(SDNode *Outer, SDNode *Inner, |
436 | unsigned NestLevel, |
437 | const TargetInstrInfo *TII) { |
438 | SDNode *N = Outer; |
439 | while (true) { |
440 | if (N == Inner) |
441 | return true; |
442 | // For a TokenFactor, examine each operand. There may be multiple ways |
443 | // to get to the CALLSEQ_BEGIN, but we need to find the path with the |
444 | // most nesting in order to ensure that we find the corresponding match. |
445 | if (N->getOpcode() == ISD::TokenFactor) { |
446 | for (const SDValue &Op : N->op_values()) |
447 | if (IsChainDependent(Op.getNode(), Inner, NestLevel, TII)) |
448 | return true; |
449 | return false; |
450 | } |
451 | // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END. |
452 | if (N->isMachineOpcode()) { |
453 | if (N->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) { |
454 | ++NestLevel; |
455 | } else if (N->getMachineOpcode() == TII->getCallFrameSetupOpcode()) { |
456 | if (NestLevel == 0) |
457 | return false; |
458 | --NestLevel; |
459 | } |
460 | } |
461 | // Otherwise, find the chain and continue climbing. |
462 | for (const SDValue &Op : N->op_values()) |
463 | if (Op.getValueType() == MVT::Other) { |
464 | N = Op.getNode(); |
465 | goto found_chain_operand; |
466 | } |
467 | return false; |
468 | found_chain_operand:; |
469 | if (N->getOpcode() == ISD::EntryToken) |
470 | return false; |
471 | } |
472 | } |
473 | |
474 | /// FindCallSeqStart - Starting from the (lowered) CALLSEQ_END node, locate |
475 | /// the corresponding (lowered) CALLSEQ_BEGIN node. |
476 | /// |
477 | /// NestLevel and MaxNested are used in recursion to indcate the current level |
478 | /// of nesting of CALLSEQ_BEGIN and CALLSEQ_END pairs, as well as the maximum |
479 | /// level seen so far. |
480 | /// |
481 | /// TODO: It would be better to give CALLSEQ_END an explicit operand to point |
482 | /// to the corresponding CALLSEQ_BEGIN to avoid needing to search for it. |
483 | static SDNode * |
484 | FindCallSeqStart(SDNode *N, unsigned &NestLevel, unsigned &MaxNest, |
485 | const TargetInstrInfo *TII) { |
486 | while (true) { |
487 | // For a TokenFactor, examine each operand. There may be multiple ways |
488 | // to get to the CALLSEQ_BEGIN, but we need to find the path with the |
489 | // most nesting in order to ensure that we find the corresponding match. |
490 | if (N->getOpcode() == ISD::TokenFactor) { |
491 | SDNode *Best = nullptr; |
492 | unsigned BestMaxNest = MaxNest; |
493 | for (const SDValue &Op : N->op_values()) { |
494 | unsigned MyNestLevel = NestLevel; |
495 | unsigned MyMaxNest = MaxNest; |
496 | if (SDNode *New = FindCallSeqStart(Op.getNode(), |
497 | MyNestLevel, MyMaxNest, TII)) |
498 | if (!Best || (MyMaxNest > BestMaxNest)) { |
499 | Best = New; |
500 | BestMaxNest = MyMaxNest; |
501 | } |
502 | } |
503 | assert(Best)(static_cast <bool> (Best) ? void (0) : __assert_fail ( "Best", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 503, __extension__ __PRETTY_FUNCTION__)); |
504 | MaxNest = BestMaxNest; |
505 | return Best; |
506 | } |
507 | // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END. |
508 | if (N->isMachineOpcode()) { |
509 | if (N->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) { |
510 | ++NestLevel; |
511 | MaxNest = std::max(MaxNest, NestLevel); |
512 | } else if (N->getMachineOpcode() == TII->getCallFrameSetupOpcode()) { |
513 | assert(NestLevel != 0)(static_cast <bool> (NestLevel != 0) ? void (0) : __assert_fail ("NestLevel != 0", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 513, __extension__ __PRETTY_FUNCTION__)); |
514 | --NestLevel; |
515 | if (NestLevel == 0) |
516 | return N; |
517 | } |
518 | } |
519 | // Otherwise, find the chain and continue climbing. |
520 | for (const SDValue &Op : N->op_values()) |
521 | if (Op.getValueType() == MVT::Other) { |
522 | N = Op.getNode(); |
523 | goto found_chain_operand; |
524 | } |
525 | return nullptr; |
526 | found_chain_operand:; |
527 | if (N->getOpcode() == ISD::EntryToken) |
528 | return nullptr; |
529 | } |
530 | } |
531 | |
532 | /// Call ReleasePred for each predecessor, then update register live def/gen. |
533 | /// Always update LiveRegDefs for a register dependence even if the current SU |
534 | /// also defines the register. This effectively create one large live range |
535 | /// across a sequence of two-address node. This is important because the |
536 | /// entire chain must be scheduled together. Example: |
537 | /// |
538 | /// flags = (3) add |
539 | /// flags = (2) addc flags |
540 | /// flags = (1) addc flags |
541 | /// |
542 | /// results in |
543 | /// |
544 | /// LiveRegDefs[flags] = 3 |
545 | /// LiveRegGens[flags] = 1 |
546 | /// |
547 | /// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid |
548 | /// interference on flags. |
549 | void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) { |
550 | // Bottom up: release predecessors |
551 | for (SDep &Pred : SU->Preds) { |
552 | ReleasePred(SU, &Pred); |
553 | if (Pred.isAssignedRegDep()) { |
554 | // This is a physical register dependency and it's impossible or |
555 | // expensive to copy the register. Make sure nothing that can |
556 | // clobber the register is scheduled between the predecessor and |
557 | // this node. |
558 | SUnit *RegDef = LiveRegDefs[Pred.getReg()]; (void)RegDef; |
559 | assert((!RegDef || RegDef == SU || RegDef == Pred.getSUnit()) &&(static_cast <bool> ((!RegDef || RegDef == SU || RegDef == Pred.getSUnit()) && "interference on register dependence" ) ? void (0) : __assert_fail ("(!RegDef || RegDef == SU || RegDef == Pred.getSUnit()) && \"interference on register dependence\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 560, __extension__ __PRETTY_FUNCTION__)) |
560 | "interference on register dependence")(static_cast <bool> ((!RegDef || RegDef == SU || RegDef == Pred.getSUnit()) && "interference on register dependence" ) ? void (0) : __assert_fail ("(!RegDef || RegDef == SU || RegDef == Pred.getSUnit()) && \"interference on register dependence\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 560, __extension__ __PRETTY_FUNCTION__)); |
561 | LiveRegDefs[Pred.getReg()] = Pred.getSUnit(); |
562 | if (!LiveRegGens[Pred.getReg()]) { |
563 | ++NumLiveRegs; |
564 | LiveRegGens[Pred.getReg()] = SU; |
565 | } |
566 | } |
567 | } |
568 | |
569 | // If we're scheduling a lowered CALLSEQ_END, find the corresponding |
570 | // CALLSEQ_BEGIN. Inject an artificial physical register dependence between |
571 | // these nodes, to prevent other calls from being interscheduled with them. |
572 | unsigned CallResource = TRI->getNumRegs(); |
573 | if (!LiveRegDefs[CallResource]) |
574 | for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) |
575 | if (Node->isMachineOpcode() && |
576 | Node->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) { |
577 | unsigned NestLevel = 0; |
578 | unsigned MaxNest = 0; |
579 | SDNode *N = FindCallSeqStart(Node, NestLevel, MaxNest, TII); |
580 | assert(N && "Must find call sequence start")(static_cast <bool> (N && "Must find call sequence start" ) ? void (0) : __assert_fail ("N && \"Must find call sequence start\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 580, __extension__ __PRETTY_FUNCTION__)); |
581 | |
582 | SUnit *Def = &SUnits[N->getNodeId()]; |
583 | CallSeqEndForStart[Def] = SU; |
584 | |
585 | ++NumLiveRegs; |
586 | LiveRegDefs[CallResource] = Def; |
587 | LiveRegGens[CallResource] = SU; |
588 | break; |
589 | } |
590 | } |
591 | |
592 | /// Check to see if any of the pending instructions are ready to issue. If |
593 | /// so, add them to the available queue. |
594 | void ScheduleDAGRRList::ReleasePending() { |
595 | if (DisableSchedCycles) { |
596 | assert(PendingQueue.empty() && "pending instrs not allowed in this mode")(static_cast <bool> (PendingQueue.empty() && "pending instrs not allowed in this mode" ) ? void (0) : __assert_fail ("PendingQueue.empty() && \"pending instrs not allowed in this mode\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 596, __extension__ __PRETTY_FUNCTION__)); |
597 | return; |
598 | } |
599 | |
600 | // If the available queue is empty, it is safe to reset MinAvailableCycle. |
601 | if (AvailableQueue->empty()) |
602 | MinAvailableCycle = std::numeric_limits<unsigned>::max(); |
603 | |
604 | // Check to see if any of the pending instructions are ready to issue. If |
605 | // so, add them to the available queue. |
606 | for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) { |
607 | unsigned ReadyCycle = PendingQueue[i]->getHeight(); |
608 | if (ReadyCycle < MinAvailableCycle) |
609 | MinAvailableCycle = ReadyCycle; |
610 | |
611 | if (PendingQueue[i]->isAvailable) { |
612 | if (!isReady(PendingQueue[i])) |
613 | continue; |
614 | AvailableQueue->push(PendingQueue[i]); |
615 | } |
616 | PendingQueue[i]->isPending = false; |
617 | PendingQueue[i] = PendingQueue.back(); |
618 | PendingQueue.pop_back(); |
619 | --i; --e; |
620 | } |
621 | } |
622 | |
623 | /// Move the scheduler state forward by the specified number of Cycles. |
624 | void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) { |
625 | if (NextCycle <= CurCycle) |
626 | return; |
627 | |
628 | IssueCount = 0; |
629 | AvailableQueue->setCurCycle(NextCycle); |
630 | if (!HazardRec->isEnabled()) { |
631 | // Bypass lots of virtual calls in case of long latency. |
632 | CurCycle = NextCycle; |
633 | } |
634 | else { |
635 | for (; CurCycle != NextCycle; ++CurCycle) { |
636 | HazardRec->RecedeCycle(); |
637 | } |
638 | } |
639 | // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the |
640 | // available Q to release pending nodes at least once before popping. |
641 | ReleasePending(); |
642 | } |
643 | |
644 | /// Move the scheduler state forward until the specified node's dependents are |
645 | /// ready and can be scheduled with no resource conflicts. |
646 | void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) { |
647 | if (DisableSchedCycles) |
648 | return; |
649 | |
650 | // FIXME: Nodes such as CopyFromReg probably should not advance the current |
651 | // cycle. Otherwise, we can wrongly mask real stalls. If the non-machine node |
652 | // has predecessors the cycle will be advanced when they are scheduled. |
653 | // But given the crude nature of modeling latency though such nodes, we |
654 | // currently need to treat these nodes like real instructions. |
655 | // if (!SU->getNode() || !SU->getNode()->isMachineOpcode()) return; |
656 | |
657 | unsigned ReadyCycle = SU->getHeight(); |
658 | |
659 | // Bump CurCycle to account for latency. We assume the latency of other |
660 | // available instructions may be hidden by the stall (not a full pipe stall). |
661 | // This updates the hazard recognizer's cycle before reserving resources for |
662 | // this instruction. |
663 | AdvanceToCycle(ReadyCycle); |
664 | |
665 | // Calls are scheduled in their preceding cycle, so don't conflict with |
666 | // hazards from instructions after the call. EmitNode will reset the |
667 | // scoreboard state before emitting the call. |
668 | if (SU->isCall) |
669 | return; |
670 | |
671 | // FIXME: For resource conflicts in very long non-pipelined stages, we |
672 | // should probably skip ahead here to avoid useless scoreboard checks. |
673 | int Stalls = 0; |
674 | while (true) { |
675 | ScheduleHazardRecognizer::HazardType HT = |
676 | HazardRec->getHazardType(SU, -Stalls); |
677 | |
678 | if (HT == ScheduleHazardRecognizer::NoHazard) |
679 | break; |
680 | |
681 | ++Stalls; |
682 | } |
683 | AdvanceToCycle(CurCycle + Stalls); |
684 | } |
685 | |
686 | /// Record this SUnit in the HazardRecognizer. |
687 | /// Does not update CurCycle. |
688 | void ScheduleDAGRRList::EmitNode(SUnit *SU) { |
689 | if (!HazardRec->isEnabled()) |
690 | return; |
691 | |
692 | // Check for phys reg copy. |
693 | if (!SU->getNode()) |
694 | return; |
695 | |
696 | switch (SU->getNode()->getOpcode()) { |
697 | default: |
698 | assert(SU->getNode()->isMachineOpcode() &&(static_cast <bool> (SU->getNode()->isMachineOpcode () && "This target-independent node should not be scheduled." ) ? void (0) : __assert_fail ("SU->getNode()->isMachineOpcode() && \"This target-independent node should not be scheduled.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 699, __extension__ __PRETTY_FUNCTION__)) |
699 | "This target-independent node should not be scheduled.")(static_cast <bool> (SU->getNode()->isMachineOpcode () && "This target-independent node should not be scheduled." ) ? void (0) : __assert_fail ("SU->getNode()->isMachineOpcode() && \"This target-independent node should not be scheduled.\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 699, __extension__ __PRETTY_FUNCTION__)); |
700 | break; |
701 | case ISD::MERGE_VALUES: |
702 | case ISD::TokenFactor: |
703 | case ISD::LIFETIME_START: |
704 | case ISD::LIFETIME_END: |
705 | case ISD::CopyToReg: |
706 | case ISD::CopyFromReg: |
707 | case ISD::EH_LABEL: |
708 | // Noops don't affect the scoreboard state. Copies are likely to be |
709 | // removed. |
710 | return; |
711 | case ISD::INLINEASM: |
712 | // For inline asm, clear the pipeline state. |
713 | HazardRec->Reset(); |
714 | return; |
715 | } |
716 | if (SU->isCall) { |
717 | // Calls are scheduled with their preceding instructions. For bottom-up |
718 | // scheduling, clear the pipeline state before emitting. |
719 | HazardRec->Reset(); |
720 | } |
721 | |
722 | HazardRec->EmitInstruction(SU); |
723 | } |
724 | |
725 | static void resetVRegCycle(SUnit *SU); |
726 | |
727 | /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending |
728 | /// count of its predecessors. If a predecessor pending count is zero, add it to |
729 | /// the Available queue. |
730 | void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) { |
731 | DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "\n*** Scheduling [" << CurCycle << "]: "; } } while (false); |
732 | DEBUG(SU->dump(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { SU->dump(this); } } while (false); |
733 | |
734 | #ifndef NDEBUG |
735 | if (CurCycle < SU->getHeight()) |
736 | DEBUG(dbgs() << " Height [" << SU->getHeight()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Height [" << SU ->getHeight() << "] pipeline stall!\n"; } } while (false ) |
737 | << "] pipeline stall!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Height [" << SU ->getHeight() << "] pipeline stall!\n"; } } while (false ); |
738 | #endif |
739 | |
740 | // FIXME: Do not modify node height. It may interfere with |
741 | // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the |
742 | // node its ready cycle can aid heuristics, and after scheduling it can |
743 | // indicate the scheduled cycle. |
744 | SU->setHeightToAtLeast(CurCycle); |
745 | |
746 | // Reserve resources for the scheduled instruction. |
747 | EmitNode(SU); |
748 | |
749 | Sequence.push_back(SU); |
750 | |
751 | AvailableQueue->scheduledNode(SU); |
752 | |
753 | // If HazardRec is disabled, and each inst counts as one cycle, then |
754 | // advance CurCycle before ReleasePredecessors to avoid useless pushes to |
755 | // PendingQueue for schedulers that implement HasReadyFilter. |
756 | if (!HazardRec->isEnabled() && AvgIPC < 2) |
757 | AdvanceToCycle(CurCycle + 1); |
758 | |
759 | // Update liveness of predecessors before successors to avoid treating a |
760 | // two-address node as a live range def. |
761 | ReleasePredecessors(SU); |
762 | |
763 | // Release all the implicit physical register defs that are live. |
764 | for (SDep &Succ : SU->Succs) { |
765 | // LiveRegDegs[Succ.getReg()] != SU when SU is a two-address node. |
766 | if (Succ.isAssignedRegDep() && LiveRegDefs[Succ.getReg()] == SU) { |
767 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")(static_cast <bool> (NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? void (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 767, __extension__ __PRETTY_FUNCTION__)); |
768 | --NumLiveRegs; |
769 | LiveRegDefs[Succ.getReg()] = nullptr; |
770 | LiveRegGens[Succ.getReg()] = nullptr; |
771 | releaseInterferences(Succ.getReg()); |
772 | } |
773 | } |
774 | // Release the special call resource dependence, if this is the beginning |
775 | // of a call. |
776 | unsigned CallResource = TRI->getNumRegs(); |
777 | if (LiveRegDefs[CallResource] == SU) |
778 | for (const SDNode *SUNode = SU->getNode(); SUNode; |
779 | SUNode = SUNode->getGluedNode()) { |
780 | if (SUNode->isMachineOpcode() && |
781 | SUNode->getMachineOpcode() == TII->getCallFrameSetupOpcode()) { |
782 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")(static_cast <bool> (NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? void (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 782, __extension__ __PRETTY_FUNCTION__)); |
783 | --NumLiveRegs; |
784 | LiveRegDefs[CallResource] = nullptr; |
785 | LiveRegGens[CallResource] = nullptr; |
786 | releaseInterferences(CallResource); |
787 | } |
788 | } |
789 | |
790 | resetVRegCycle(SU); |
791 | |
792 | SU->isScheduled = true; |
793 | |
794 | // Conditions under which the scheduler should eagerly advance the cycle: |
795 | // (1) No available instructions |
796 | // (2) All pipelines full, so available instructions must have hazards. |
797 | // |
798 | // If HazardRec is disabled, the cycle was pre-advanced before calling |
799 | // ReleasePredecessors. In that case, IssueCount should remain 0. |
800 | // |
801 | // Check AvailableQueue after ReleasePredecessors in case of zero latency. |
802 | if (HazardRec->isEnabled() || AvgIPC > 1) { |
803 | if (SU->getNode() && SU->getNode()->isMachineOpcode()) |
804 | ++IssueCount; |
805 | if ((HazardRec->isEnabled() && HazardRec->atIssueLimit()) |
806 | || (!HazardRec->isEnabled() && IssueCount == AvgIPC)) |
807 | AdvanceToCycle(CurCycle + 1); |
808 | } |
809 | } |
810 | |
811 | /// CapturePred - This does the opposite of ReleasePred. Since SU is being |
812 | /// unscheduled, increase the succ left count of its predecessors. Remove |
813 | /// them from AvailableQueue if necessary. |
814 | void ScheduleDAGRRList::CapturePred(SDep *PredEdge) { |
815 | SUnit *PredSU = PredEdge->getSUnit(); |
816 | if (PredSU->isAvailable) { |
817 | PredSU->isAvailable = false; |
818 | if (!PredSU->isPending) |
819 | AvailableQueue->remove(PredSU); |
820 | } |
821 | |
822 | assert(PredSU->NumSuccsLeft < std::numeric_limits<unsigned>::max() &&(static_cast <bool> (PredSU->NumSuccsLeft < std:: numeric_limits<unsigned>::max() && "NumSuccsLeft will overflow!" ) ? void (0) : __assert_fail ("PredSU->NumSuccsLeft < std::numeric_limits<unsigned>::max() && \"NumSuccsLeft will overflow!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 823, __extension__ __PRETTY_FUNCTION__)) |
823 | "NumSuccsLeft will overflow!")(static_cast <bool> (PredSU->NumSuccsLeft < std:: numeric_limits<unsigned>::max() && "NumSuccsLeft will overflow!" ) ? void (0) : __assert_fail ("PredSU->NumSuccsLeft < std::numeric_limits<unsigned>::max() && \"NumSuccsLeft will overflow!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 823, __extension__ __PRETTY_FUNCTION__)); |
824 | ++PredSU->NumSuccsLeft; |
825 | } |
826 | |
827 | /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and |
828 | /// its predecessor states to reflect the change. |
829 | void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) { |
830 | DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: "; } } while (false); |
831 | DEBUG(SU->dump(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { SU->dump(this); } } while (false); |
832 | |
833 | for (SDep &Pred : SU->Preds) { |
834 | CapturePred(&Pred); |
835 | if (Pred.isAssignedRegDep() && SU == LiveRegGens[Pred.getReg()]){ |
836 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")(static_cast <bool> (NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? void (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 836, __extension__ __PRETTY_FUNCTION__)); |
837 | assert(LiveRegDefs[Pred.getReg()] == Pred.getSUnit() &&(static_cast <bool> (LiveRegDefs[Pred.getReg()] == Pred .getSUnit() && "Physical register dependency violated?" ) ? void (0) : __assert_fail ("LiveRegDefs[Pred.getReg()] == Pred.getSUnit() && \"Physical register dependency violated?\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 838, __extension__ __PRETTY_FUNCTION__)) |
838 | "Physical register dependency violated?")(static_cast <bool> (LiveRegDefs[Pred.getReg()] == Pred .getSUnit() && "Physical register dependency violated?" ) ? void (0) : __assert_fail ("LiveRegDefs[Pred.getReg()] == Pred.getSUnit() && \"Physical register dependency violated?\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 838, __extension__ __PRETTY_FUNCTION__)); |
839 | --NumLiveRegs; |
840 | LiveRegDefs[Pred.getReg()] = nullptr; |
841 | LiveRegGens[Pred.getReg()] = nullptr; |
842 | releaseInterferences(Pred.getReg()); |
843 | } |
844 | } |
845 | |
846 | // Reclaim the special call resource dependence, if this is the beginning |
847 | // of a call. |
848 | unsigned CallResource = TRI->getNumRegs(); |
849 | for (const SDNode *SUNode = SU->getNode(); SUNode; |
850 | SUNode = SUNode->getGluedNode()) { |
851 | if (SUNode->isMachineOpcode() && |
852 | SUNode->getMachineOpcode() == TII->getCallFrameSetupOpcode()) { |
853 | SUnit *SeqEnd = CallSeqEndForStart[SU]; |
854 | assert(SeqEnd && "Call sequence start/end must be known")(static_cast <bool> (SeqEnd && "Call sequence start/end must be known" ) ? void (0) : __assert_fail ("SeqEnd && \"Call sequence start/end must be known\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 854, __extension__ __PRETTY_FUNCTION__)); |
855 | assert(!LiveRegDefs[CallResource])(static_cast <bool> (!LiveRegDefs[CallResource]) ? void (0) : __assert_fail ("!LiveRegDefs[CallResource]", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 855, __extension__ __PRETTY_FUNCTION__)); |
856 | assert(!LiveRegGens[CallResource])(static_cast <bool> (!LiveRegGens[CallResource]) ? void (0) : __assert_fail ("!LiveRegGens[CallResource]", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 856, __extension__ __PRETTY_FUNCTION__)); |
857 | ++NumLiveRegs; |
858 | LiveRegDefs[CallResource] = SU; |
859 | LiveRegGens[CallResource] = SeqEnd; |
860 | } |
861 | } |
862 | |
863 | // Release the special call resource dependence, if this is the end |
864 | // of a call. |
865 | if (LiveRegGens[CallResource] == SU) |
866 | for (const SDNode *SUNode = SU->getNode(); SUNode; |
867 | SUNode = SUNode->getGluedNode()) { |
868 | if (SUNode->isMachineOpcode() && |
869 | SUNode->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) { |
870 | assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!")(static_cast <bool> (NumLiveRegs > 0 && "NumLiveRegs is already zero!" ) ? void (0) : __assert_fail ("NumLiveRegs > 0 && \"NumLiveRegs is already zero!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 870, __extension__ __PRETTY_FUNCTION__)); |
871 | assert(LiveRegDefs[CallResource])(static_cast <bool> (LiveRegDefs[CallResource]) ? void ( 0) : __assert_fail ("LiveRegDefs[CallResource]", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 871, __extension__ __PRETTY_FUNCTION__)); |
872 | assert(LiveRegGens[CallResource])(static_cast <bool> (LiveRegGens[CallResource]) ? void ( 0) : __assert_fail ("LiveRegGens[CallResource]", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 872, __extension__ __PRETTY_FUNCTION__)); |
873 | --NumLiveRegs; |
874 | LiveRegDefs[CallResource] = nullptr; |
875 | LiveRegGens[CallResource] = nullptr; |
876 | releaseInterferences(CallResource); |
877 | } |
878 | } |
879 | |
880 | for (auto &Succ : SU->Succs) { |
881 | if (Succ.isAssignedRegDep()) { |
882 | auto Reg = Succ.getReg(); |
883 | if (!LiveRegDefs[Reg]) |
884 | ++NumLiveRegs; |
885 | // This becomes the nearest def. Note that an earlier def may still be |
886 | // pending if this is a two-address node. |
887 | LiveRegDefs[Reg] = SU; |
888 | |
889 | // Update LiveRegGen only if was empty before this unscheduling. |
890 | // This is to avoid incorrect updating LiveRegGen set in previous run. |
891 | if (!LiveRegGens[Reg]) { |
892 | // Find the successor with the lowest height. |
893 | LiveRegGens[Reg] = Succ.getSUnit(); |
894 | for (auto &Succ2 : SU->Succs) { |
895 | if (Succ2.isAssignedRegDep() && Succ2.getReg() == Reg && |
896 | Succ2.getSUnit()->getHeight() < LiveRegGens[Reg]->getHeight()) |
897 | LiveRegGens[Reg] = Succ2.getSUnit(); |
898 | } |
899 | } |
900 | } |
901 | } |
902 | if (SU->getHeight() < MinAvailableCycle) |
903 | MinAvailableCycle = SU->getHeight(); |
904 | |
905 | SU->setHeightDirty(); |
906 | SU->isScheduled = false; |
907 | SU->isAvailable = true; |
908 | if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) { |
909 | // Don't make available until backtracking is complete. |
910 | SU->isPending = true; |
911 | PendingQueue.push_back(SU); |
912 | } |
913 | else { |
914 | AvailableQueue->push(SU); |
915 | } |
916 | AvailableQueue->unscheduledNode(SU); |
917 | } |
918 | |
919 | /// After backtracking, the hazard checker needs to be restored to a state |
920 | /// corresponding the current cycle. |
921 | void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() { |
922 | HazardRec->Reset(); |
923 | |
924 | unsigned LookAhead = std::min((unsigned)Sequence.size(), |
925 | HazardRec->getMaxLookAhead()); |
926 | if (LookAhead == 0) |
927 | return; |
928 | |
929 | std::vector<SUnit *>::const_iterator I = (Sequence.end() - LookAhead); |
930 | unsigned HazardCycle = (*I)->getHeight(); |
931 | for (auto E = Sequence.end(); I != E; ++I) { |
932 | SUnit *SU = *I; |
933 | for (; SU->getHeight() > HazardCycle; ++HazardCycle) { |
934 | HazardRec->RecedeCycle(); |
935 | } |
936 | EmitNode(SU); |
937 | } |
938 | } |
939 | |
940 | /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in |
941 | /// BTCycle in order to schedule a specific node. |
942 | void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) { |
943 | SUnit *OldSU = Sequence.back(); |
944 | while (true) { |
945 | Sequence.pop_back(); |
946 | // FIXME: use ready cycle instead of height |
947 | CurCycle = OldSU->getHeight(); |
948 | UnscheduleNodeBottomUp(OldSU); |
949 | AvailableQueue->setCurCycle(CurCycle); |
950 | if (OldSU == BtSU) |
951 | break; |
952 | OldSU = Sequence.back(); |
953 | } |
954 | |
955 | assert(!SU->isSucc(OldSU) && "Something is wrong!")(static_cast <bool> (!SU->isSucc(OldSU) && "Something is wrong!" ) ? void (0) : __assert_fail ("!SU->isSucc(OldSU) && \"Something is wrong!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 955, __extension__ __PRETTY_FUNCTION__)); |
956 | |
957 | RestoreHazardCheckerBottomUp(); |
958 | |
959 | ReleasePending(); |
960 | |
961 | ++NumBacktracks; |
962 | } |
963 | |
964 | static bool isOperandOf(const SUnit *SU, SDNode *N) { |
965 | for (const SDNode *SUNode = SU->getNode(); SUNode; |
966 | SUNode = SUNode->getGluedNode()) { |
967 | if (SUNode->isOperandOf(N)) |
968 | return true; |
969 | } |
970 | return false; |
971 | } |
972 | |
973 | /// TryUnfold - Attempt to unfold |
974 | SUnit *ScheduleDAGRRList::TryUnfoldSU(SUnit *SU) { |
975 | SDNode *N = SU->getNode(); |
976 | // Use while over if to ease fall through. |
977 | SmallVector<SDNode *, 2> NewNodes; |
978 | if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes)) |
979 | return nullptr; |
980 | |
981 | // unfolding an x86 DEC64m operation results in store, dec, load which |
982 | // can't be handled here so quit |
983 | if (NewNodes.size() == 3) |
984 | return nullptr; |
985 | |
986 | assert(NewNodes.size() == 2 && "Expected a load folding node!")(static_cast <bool> (NewNodes.size() == 2 && "Expected a load folding node!" ) ? void (0) : __assert_fail ("NewNodes.size() == 2 && \"Expected a load folding node!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 986, __extension__ __PRETTY_FUNCTION__)); |
987 | |
988 | N = NewNodes[1]; |
989 | SDNode *LoadNode = NewNodes[0]; |
990 | unsigned NumVals = N->getNumValues(); |
991 | unsigned OldNumVals = SU->getNode()->getNumValues(); |
992 | |
993 | // LoadNode may already exist. This can happen when there is another |
994 | // load from the same location and producing the same type of value |
995 | // but it has different alignment or volatileness. |
996 | bool isNewLoad = true; |
997 | SUnit *LoadSU; |
998 | if (LoadNode->getNodeId() != -1) { |
999 | LoadSU = &SUnits[LoadNode->getNodeId()]; |
1000 | // If LoadSU has already been scheduled, we should clone it but |
1001 | // this would negate the benefit to unfolding so just return SU. |
1002 | if (LoadSU->isScheduled) |
1003 | return SU; |
1004 | isNewLoad = false; |
1005 | } else { |
1006 | LoadSU = CreateNewSUnit(LoadNode); |
1007 | LoadNode->setNodeId(LoadSU->NodeNum); |
1008 | |
1009 | InitNumRegDefsLeft(LoadSU); |
1010 | computeLatency(LoadSU); |
1011 | } |
1012 | |
1013 | DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Unfolding SU #" << SU->NodeNum << "\n"; } } while (false); |
1014 | |
1015 | // Now that we are committed to unfolding replace DAG Uses. |
1016 | for (unsigned i = 0; i != NumVals; ++i) |
1017 | DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i)); |
1018 | DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals - 1), |
1019 | SDValue(LoadNode, 1)); |
1020 | |
1021 | SUnit *NewSU = CreateNewSUnit(N); |
1022 | assert(N->getNodeId() == -1 && "Node already inserted!")(static_cast <bool> (N->getNodeId() == -1 && "Node already inserted!") ? void (0) : __assert_fail ("N->getNodeId() == -1 && \"Node already inserted!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1022, __extension__ __PRETTY_FUNCTION__)); |
1023 | N->setNodeId(NewSU->NodeNum); |
1024 | |
1025 | const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); |
1026 | for (unsigned i = 0; i != MCID.getNumOperands(); ++i) { |
1027 | if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) { |
1028 | NewSU->isTwoAddress = true; |
1029 | break; |
1030 | } |
1031 | } |
1032 | if (MCID.isCommutable()) |
1033 | NewSU->isCommutable = true; |
1034 | |
1035 | InitNumRegDefsLeft(NewSU); |
1036 | computeLatency(NewSU); |
1037 | |
1038 | // Record all the edges to and from the old SU, by category. |
1039 | SmallVector<SDep, 4> ChainPreds; |
1040 | SmallVector<SDep, 4> ChainSuccs; |
1041 | SmallVector<SDep, 4> LoadPreds; |
1042 | SmallVector<SDep, 4> NodePreds; |
1043 | SmallVector<SDep, 4> NodeSuccs; |
1044 | for (SDep &Pred : SU->Preds) { |
1045 | if (Pred.isCtrl()) |
1046 | ChainPreds.push_back(Pred); |
1047 | else if (isOperandOf(Pred.getSUnit(), LoadNode)) |
1048 | LoadPreds.push_back(Pred); |
1049 | else |
1050 | NodePreds.push_back(Pred); |
1051 | } |
1052 | for (SDep &Succ : SU->Succs) { |
1053 | if (Succ.isCtrl()) |
1054 | ChainSuccs.push_back(Succ); |
1055 | else |
1056 | NodeSuccs.push_back(Succ); |
1057 | } |
1058 | |
1059 | // Now assign edges to the newly-created nodes. |
1060 | for (const SDep &Pred : ChainPreds) { |
1061 | RemovePred(SU, Pred); |
1062 | if (isNewLoad) |
1063 | AddPred(LoadSU, Pred); |
1064 | } |
1065 | for (const SDep &Pred : LoadPreds) { |
1066 | RemovePred(SU, Pred); |
1067 | if (isNewLoad) |
1068 | AddPred(LoadSU, Pred); |
1069 | } |
1070 | for (const SDep &Pred : NodePreds) { |
1071 | RemovePred(SU, Pred); |
1072 | AddPred(NewSU, Pred); |
1073 | } |
1074 | for (SDep D : NodeSuccs) { |
1075 | SUnit *SuccDep = D.getSUnit(); |
1076 | D.setSUnit(SU); |
1077 | RemovePred(SuccDep, D); |
1078 | D.setSUnit(NewSU); |
1079 | AddPred(SuccDep, D); |
1080 | // Balance register pressure. |
1081 | if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled && |
1082 | !D.isCtrl() && NewSU->NumRegDefsLeft > 0) |
1083 | --NewSU->NumRegDefsLeft; |
1084 | } |
1085 | for (SDep D : ChainSuccs) { |
1086 | SUnit *SuccDep = D.getSUnit(); |
1087 | D.setSUnit(SU); |
1088 | RemovePred(SuccDep, D); |
1089 | if (isNewLoad) { |
1090 | D.setSUnit(LoadSU); |
1091 | AddPred(SuccDep, D); |
1092 | } |
1093 | } |
1094 | |
1095 | // Add a data dependency to reflect that NewSU reads the value defined |
1096 | // by LoadSU. |
1097 | SDep D(LoadSU, SDep::Data, 0); |
1098 | D.setLatency(LoadSU->Latency); |
1099 | AddPred(NewSU, D); |
1100 | |
1101 | if (isNewLoad) |
1102 | AvailableQueue->addNode(LoadSU); |
1103 | AvailableQueue->addNode(NewSU); |
1104 | |
1105 | ++NumUnfolds; |
1106 | |
1107 | if (NewSU->NumSuccsLeft == 0) |
1108 | NewSU->isAvailable = true; |
1109 | |
1110 | return NewSU; |
1111 | } |
1112 | |
1113 | /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled |
1114 | /// successors to the newly created node. |
1115 | SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) { |
1116 | SDNode *N = SU->getNode(); |
1117 | if (!N) |
1118 | return nullptr; |
1119 | |
1120 | DEBUG(dbgs() << "Considering duplicating the SU\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Considering duplicating the SU\n" ; } } while (false); |
1121 | DEBUG(SU->dump(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { SU->dump(this); } } while (false); |
1122 | |
1123 | if (N->getGluedNode() && |
1124 | !TII->canCopyGluedNodeDuringSchedule(N)) { |
1125 | DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Giving up because it has incoming glue and the target does not " "want to copy it\n"; } } while (false) |
1126 | << "Giving up because it has incoming glue and the target does not "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Giving up because it has incoming glue and the target does not " "want to copy it\n"; } } while (false) |
1127 | "want to copy it\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Giving up because it has incoming glue and the target does not " "want to copy it\n"; } } while (false); |
1128 | return nullptr; |
1129 | } |
1130 | |
1131 | SUnit *NewSU; |
1132 | bool TryUnfold = false; |
1133 | for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) { |
1134 | MVT VT = N->getSimpleValueType(i); |
1135 | if (VT == MVT::Glue) { |
1136 | DEBUG(dbgs() << "Giving up because it has outgoing glue\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Giving up because it has outgoing glue\n" ; } } while (false); |
1137 | return nullptr; |
1138 | } else if (VT == MVT::Other) |
1139 | TryUnfold = true; |
1140 | } |
1141 | for (const SDValue &Op : N->op_values()) { |
1142 | MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo()); |
1143 | if (VT == MVT::Glue && !TII->canCopyGluedNodeDuringSchedule(N)) { |
1144 | DEBUG(dbgs() << "Giving up because it one of the operands is glue and "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Giving up because it one of the operands is glue and " "the target does not want to copy it\n"; } } while (false) |
1145 | "the target does not want to copy it\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "Giving up because it one of the operands is glue and " "the target does not want to copy it\n"; } } while (false); |
1146 | return nullptr; |
1147 | } |
1148 | } |
1149 | |
1150 | // If possible unfold instruction. |
1151 | if (TryUnfold) { |
1152 | SUnit *UnfoldSU = TryUnfoldSU(SU); |
1153 | if (!UnfoldSU) |
1154 | return nullptr; |
1155 | SU = UnfoldSU; |
1156 | N = SU->getNode(); |
Value stored to 'N' is never read | |
1157 | // If this can be scheduled don't bother duplicating and just return |
1158 | if (SU->NumSuccsLeft == 0) |
1159 | return SU; |
1160 | } |
1161 | |
1162 | DEBUG(dbgs() << " Duplicating SU #" << SU->NodeNum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Duplicating SU #" << SU->NodeNum << "\n"; } } while (false); |
1163 | NewSU = CreateClone(SU); |
1164 | |
1165 | // New SUnit has the exact same predecessors. |
1166 | for (SDep &Pred : SU->Preds) |
1167 | if (!Pred.isArtificial()) |
1168 | AddPred(NewSU, Pred); |
1169 | |
1170 | // Only copy scheduled successors. Cut them from old node's successor |
1171 | // list and move them over. |
1172 | SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps; |
1173 | for (SDep &Succ : SU->Succs) { |
1174 | if (Succ.isArtificial()) |
1175 | continue; |
1176 | SUnit *SuccSU = Succ.getSUnit(); |
1177 | if (SuccSU->isScheduled) { |
1178 | SDep D = Succ; |
1179 | D.setSUnit(NewSU); |
1180 | AddPred(SuccSU, D); |
1181 | D.setSUnit(SU); |
1182 | DelDeps.push_back(std::make_pair(SuccSU, D)); |
1183 | } |
1184 | } |
1185 | for (auto &DelDep : DelDeps) |
1186 | RemovePred(DelDep.first, DelDep.second); |
1187 | |
1188 | AvailableQueue->updateNode(SU); |
1189 | AvailableQueue->addNode(NewSU); |
1190 | |
1191 | ++NumDups; |
1192 | return NewSU; |
1193 | } |
1194 | |
1195 | /// InsertCopiesAndMoveSuccs - Insert register copies and move all |
1196 | /// scheduled successors of the given SUnit to the last copy. |
1197 | void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg, |
1198 | const TargetRegisterClass *DestRC, |
1199 | const TargetRegisterClass *SrcRC, |
1200 | SmallVectorImpl<SUnit*> &Copies) { |
1201 | SUnit *CopyFromSU = CreateNewSUnit(nullptr); |
1202 | CopyFromSU->CopySrcRC = SrcRC; |
1203 | CopyFromSU->CopyDstRC = DestRC; |
1204 | |
1205 | SUnit *CopyToSU = CreateNewSUnit(nullptr); |
1206 | CopyToSU->CopySrcRC = DestRC; |
1207 | CopyToSU->CopyDstRC = SrcRC; |
1208 | |
1209 | // Only copy scheduled successors. Cut them from old node's successor |
1210 | // list and move them over. |
1211 | SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps; |
1212 | for (SDep &Succ : SU->Succs) { |
1213 | if (Succ.isArtificial()) |
1214 | continue; |
1215 | SUnit *SuccSU = Succ.getSUnit(); |
1216 | if (SuccSU->isScheduled) { |
1217 | SDep D = Succ; |
1218 | D.setSUnit(CopyToSU); |
1219 | AddPred(SuccSU, D); |
1220 | DelDeps.push_back(std::make_pair(SuccSU, Succ)); |
1221 | } |
1222 | else { |
1223 | // Avoid scheduling the def-side copy before other successors. Otherwise |
1224 | // we could introduce another physreg interference on the copy and |
1225 | // continue inserting copies indefinitely. |
1226 | AddPred(SuccSU, SDep(CopyFromSU, SDep::Artificial)); |
1227 | } |
1228 | } |
1229 | for (auto &DelDep : DelDeps) |
1230 | RemovePred(DelDep.first, DelDep.second); |
1231 | |
1232 | SDep FromDep(SU, SDep::Data, Reg); |
1233 | FromDep.setLatency(SU->Latency); |
1234 | AddPred(CopyFromSU, FromDep); |
1235 | SDep ToDep(CopyFromSU, SDep::Data, 0); |
1236 | ToDep.setLatency(CopyFromSU->Latency); |
1237 | AddPred(CopyToSU, ToDep); |
1238 | |
1239 | AvailableQueue->updateNode(SU); |
1240 | AvailableQueue->addNode(CopyFromSU); |
1241 | AvailableQueue->addNode(CopyToSU); |
1242 | Copies.push_back(CopyFromSU); |
1243 | Copies.push_back(CopyToSU); |
1244 | |
1245 | ++NumPRCopies; |
1246 | } |
1247 | |
1248 | /// getPhysicalRegisterVT - Returns the ValueType of the physical register |
1249 | /// definition of the specified node. |
1250 | /// FIXME: Move to SelectionDAG? |
1251 | static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg, |
1252 | const TargetInstrInfo *TII) { |
1253 | unsigned NumRes; |
1254 | if (N->getOpcode() == ISD::CopyFromReg) { |
1255 | // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type. |
1256 | NumRes = 1; |
1257 | } else { |
1258 | const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); |
1259 | assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!")(static_cast <bool> (MCID.ImplicitDefs && "Physical reg def must be in implicit def list!" ) ? void (0) : __assert_fail ("MCID.ImplicitDefs && \"Physical reg def must be in implicit def list!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1259, __extension__ __PRETTY_FUNCTION__)); |
1260 | NumRes = MCID.getNumDefs(); |
1261 | for (const MCPhysReg *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) { |
1262 | if (Reg == *ImpDef) |
1263 | break; |
1264 | ++NumRes; |
1265 | } |
1266 | } |
1267 | return N->getSimpleValueType(NumRes); |
1268 | } |
1269 | |
1270 | /// CheckForLiveRegDef - Return true and update live register vector if the |
1271 | /// specified register def of the specified SUnit clobbers any "live" registers. |
1272 | static void CheckForLiveRegDef(SUnit *SU, unsigned Reg, |
1273 | SUnit **LiveRegDefs, |
1274 | SmallSet<unsigned, 4> &RegAdded, |
1275 | SmallVectorImpl<unsigned> &LRegs, |
1276 | const TargetRegisterInfo *TRI) { |
1277 | for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) { |
1278 | |
1279 | // Check if Ref is live. |
1280 | if (!LiveRegDefs[*AliasI]) continue; |
1281 | |
1282 | // Allow multiple uses of the same def. |
1283 | if (LiveRegDefs[*AliasI] == SU) continue; |
1284 | |
1285 | // Add Reg to the set of interfering live regs. |
1286 | if (RegAdded.insert(*AliasI).second) { |
1287 | LRegs.push_back(*AliasI); |
1288 | } |
1289 | } |
1290 | } |
1291 | |
1292 | /// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered |
1293 | /// by RegMask, and add them to LRegs. |
1294 | static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask, |
1295 | ArrayRef<SUnit*> LiveRegDefs, |
1296 | SmallSet<unsigned, 4> &RegAdded, |
1297 | SmallVectorImpl<unsigned> &LRegs) { |
1298 | // Look at all live registers. Skip Reg0 and the special CallResource. |
1299 | for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) { |
1300 | if (!LiveRegDefs[i]) continue; |
1301 | if (LiveRegDefs[i] == SU) continue; |
1302 | if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue; |
1303 | if (RegAdded.insert(i).second) |
1304 | LRegs.push_back(i); |
1305 | } |
1306 | } |
1307 | |
1308 | /// getNodeRegMask - Returns the register mask attached to an SDNode, if any. |
1309 | static const uint32_t *getNodeRegMask(const SDNode *N) { |
1310 | for (const SDValue &Op : N->op_values()) |
1311 | if (const auto *RegOp = dyn_cast<RegisterMaskSDNode>(Op.getNode())) |
1312 | return RegOp->getRegMask(); |
1313 | return nullptr; |
1314 | } |
1315 | |
1316 | /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay |
1317 | /// scheduling of the given node to satisfy live physical register dependencies. |
1318 | /// If the specific node is the last one that's available to schedule, do |
1319 | /// whatever is necessary (i.e. backtracking or cloning) to make it possible. |
1320 | bool ScheduleDAGRRList:: |
1321 | DelayForLiveRegsBottomUp(SUnit *SU, SmallVectorImpl<unsigned> &LRegs) { |
1322 | if (NumLiveRegs == 0) |
1323 | return false; |
1324 | |
1325 | SmallSet<unsigned, 4> RegAdded; |
1326 | // If this node would clobber any "live" register, then it's not ready. |
1327 | // |
1328 | // If SU is the currently live definition of the same register that it uses, |
1329 | // then we are free to schedule it. |
1330 | for (SDep &Pred : SU->Preds) { |
1331 | if (Pred.isAssignedRegDep() && LiveRegDefs[Pred.getReg()] != SU) |
1332 | CheckForLiveRegDef(Pred.getSUnit(), Pred.getReg(), LiveRegDefs.get(), |
1333 | RegAdded, LRegs, TRI); |
1334 | } |
1335 | |
1336 | for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) { |
1337 | if (Node->getOpcode() == ISD::INLINEASM) { |
1338 | // Inline asm can clobber physical defs. |
1339 | unsigned NumOps = Node->getNumOperands(); |
1340 | if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue) |
1341 | --NumOps; // Ignore the glue operand. |
1342 | |
1343 | for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) { |
1344 | unsigned Flags = |
1345 | cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue(); |
1346 | unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags); |
1347 | |
1348 | ++i; // Skip the ID value. |
1349 | if (InlineAsm::isRegDefKind(Flags) || |
1350 | InlineAsm::isRegDefEarlyClobberKind(Flags) || |
1351 | InlineAsm::isClobberKind(Flags)) { |
1352 | // Check for def of register or earlyclobber register. |
1353 | for (; NumVals; --NumVals, ++i) { |
1354 | unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg(); |
1355 | if (TargetRegisterInfo::isPhysicalRegister(Reg)) |
1356 | CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI); |
1357 | } |
1358 | } else |
1359 | i += NumVals; |
1360 | } |
1361 | continue; |
1362 | } |
1363 | |
1364 | if (!Node->isMachineOpcode()) |
1365 | continue; |
1366 | // If we're in the middle of scheduling a call, don't begin scheduling |
1367 | // another call. Also, don't allow any physical registers to be live across |
1368 | // the call. |
1369 | if (Node->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) { |
1370 | // Check the special calling-sequence resource. |
1371 | unsigned CallResource = TRI->getNumRegs(); |
1372 | if (LiveRegDefs[CallResource]) { |
1373 | SDNode *Gen = LiveRegGens[CallResource]->getNode(); |
1374 | while (SDNode *Glued = Gen->getGluedNode()) |
1375 | Gen = Glued; |
1376 | if (!IsChainDependent(Gen, Node, 0, TII) && |
1377 | RegAdded.insert(CallResource).second) |
1378 | LRegs.push_back(CallResource); |
1379 | } |
1380 | } |
1381 | if (const uint32_t *RegMask = getNodeRegMask(Node)) |
1382 | CheckForLiveRegDefMasked(SU, RegMask, |
1383 | makeArrayRef(LiveRegDefs.get(), TRI->getNumRegs()), |
1384 | RegAdded, LRegs); |
1385 | |
1386 | const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode()); |
1387 | if (MCID.hasOptionalDef()) { |
1388 | // Most ARM instructions have an OptionalDef for CPSR, to model the S-bit. |
1389 | // This operand can be either a def of CPSR, if the S bit is set; or a use |
1390 | // of %noreg. When the OptionalDef is set to a valid register, we need to |
1391 | // handle it in the same way as an ImplicitDef. |
1392 | for (unsigned i = 0; i < MCID.getNumDefs(); ++i) |
1393 | if (MCID.OpInfo[i].isOptionalDef()) { |
1394 | const SDValue &OptionalDef = Node->getOperand(i - Node->getNumValues()); |
1395 | unsigned Reg = cast<RegisterSDNode>(OptionalDef)->getReg(); |
1396 | CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI); |
1397 | } |
1398 | } |
1399 | if (!MCID.ImplicitDefs) |
1400 | continue; |
1401 | for (const MCPhysReg *Reg = MCID.getImplicitDefs(); *Reg; ++Reg) |
1402 | CheckForLiveRegDef(SU, *Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI); |
1403 | } |
1404 | |
1405 | return !LRegs.empty(); |
1406 | } |
1407 | |
1408 | void ScheduleDAGRRList::releaseInterferences(unsigned Reg) { |
1409 | // Add the nodes that aren't ready back onto the available list. |
1410 | for (unsigned i = Interferences.size(); i > 0; --i) { |
1411 | SUnit *SU = Interferences[i-1]; |
1412 | LRegsMapT::iterator LRegsPos = LRegsMap.find(SU); |
1413 | if (Reg) { |
1414 | SmallVectorImpl<unsigned> &LRegs = LRegsPos->second; |
1415 | if (!is_contained(LRegs, Reg)) |
1416 | continue; |
1417 | } |
1418 | SU->isPending = false; |
1419 | // The interfering node may no longer be available due to backtracking. |
1420 | // Furthermore, it may have been made available again, in which case it is |
1421 | // now already in the AvailableQueue. |
1422 | if (SU->isAvailable && !SU->NodeQueueId) { |
1423 | DEBUG(dbgs() << " Repushing SU #" << SU->NodeNum << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Repushing SU #" << SU->NodeNum << '\n'; } } while (false); |
1424 | AvailableQueue->push(SU); |
1425 | } |
1426 | if (i < Interferences.size()) |
1427 | Interferences[i-1] = Interferences.back(); |
1428 | Interferences.pop_back(); |
1429 | LRegsMap.erase(LRegsPos); |
1430 | } |
1431 | } |
1432 | |
1433 | /// Return a node that can be scheduled in this cycle. Requirements: |
1434 | /// (1) Ready: latency has been satisfied |
1435 | /// (2) No Hazards: resources are available |
1436 | /// (3) No Interferences: may unschedule to break register interferences. |
1437 | SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() { |
1438 | SUnit *CurSU = AvailableQueue->empty() ? nullptr : AvailableQueue->pop(); |
1439 | auto FindAvailableNode = [&]() { |
1440 | while (CurSU) { |
1441 | SmallVector<unsigned, 4> LRegs; |
1442 | if (!DelayForLiveRegsBottomUp(CurSU, LRegs)) |
1443 | break; |
1444 | DEBUG(dbgs() << " Interfering reg ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg "; if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource" ; else dbgs() << printReg(LRegs[0], TRI); dbgs() << " SU #" << CurSU->NodeNum << '\n'; } } while ( false) |
1445 | if (LRegs[0] == TRI->getNumRegs())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg "; if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource" ; else dbgs() << printReg(LRegs[0], TRI); dbgs() << " SU #" << CurSU->NodeNum << '\n'; } } while ( false) |
1446 | dbgs() << "CallResource";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg "; if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource" ; else dbgs() << printReg(LRegs[0], TRI); dbgs() << " SU #" << CurSU->NodeNum << '\n'; } } while ( false) |
1447 | elsedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg "; if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource" ; else dbgs() << printReg(LRegs[0], TRI); dbgs() << " SU #" << CurSU->NodeNum << '\n'; } } while ( false) |
1448 | dbgs() << printReg(LRegs[0], TRI);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg "; if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource" ; else dbgs() << printReg(LRegs[0], TRI); dbgs() << " SU #" << CurSU->NodeNum << '\n'; } } while ( false) |
1449 | dbgs() << " SU #" << CurSU->NodeNum << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " Interfering reg "; if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource" ; else dbgs() << printReg(LRegs[0], TRI); dbgs() << " SU #" << CurSU->NodeNum << '\n'; } } while ( false); |
1450 | std::pair<LRegsMapT::iterator, bool> LRegsPair = |
1451 | LRegsMap.insert(std::make_pair(CurSU, LRegs)); |
1452 | if (LRegsPair.second) { |
1453 | CurSU->isPending = true; // This SU is not in AvailableQueue right now. |
1454 | Interferences.push_back(CurSU); |
1455 | } |
1456 | else { |
1457 | assert(CurSU->isPending && "Interferences are pending")(static_cast <bool> (CurSU->isPending && "Interferences are pending" ) ? void (0) : __assert_fail ("CurSU->isPending && \"Interferences are pending\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1457, __extension__ __PRETTY_FUNCTION__)); |
1458 | // Update the interference with current live regs. |
1459 | LRegsPair.first->second = LRegs; |
1460 | } |
1461 | CurSU = AvailableQueue->pop(); |
1462 | } |
1463 | }; |
1464 | FindAvailableNode(); |
1465 | if (CurSU) |
1466 | return CurSU; |
1467 | |
1468 | // All candidates are delayed due to live physical reg dependencies. |
1469 | // Try backtracking, code duplication, or inserting cross class copies |
1470 | // to resolve it. |
1471 | for (SUnit *TrySU : Interferences) { |
1472 | SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU]; |
1473 | |
1474 | // Try unscheduling up to the point where it's safe to schedule |
1475 | // this node. |
1476 | SUnit *BtSU = nullptr; |
1477 | unsigned LiveCycle = std::numeric_limits<unsigned>::max(); |
1478 | for (unsigned Reg : LRegs) { |
1479 | if (LiveRegGens[Reg]->getHeight() < LiveCycle) { |
1480 | BtSU = LiveRegGens[Reg]; |
1481 | LiveCycle = BtSU->getHeight(); |
1482 | } |
1483 | } |
1484 | if (!WillCreateCycle(TrySU, BtSU)) { |
1485 | // BacktrackBottomUp mutates Interferences! |
1486 | BacktrackBottomUp(TrySU, BtSU); |
1487 | |
1488 | // Force the current node to be scheduled before the node that |
1489 | // requires the physical reg dep. |
1490 | if (BtSU->isAvailable) { |
1491 | BtSU->isAvailable = false; |
1492 | if (!BtSU->isPending) |
1493 | AvailableQueue->remove(BtSU); |
1494 | } |
1495 | 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 (false) |
1496 | << 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 (false); |
1497 | AddPred(TrySU, SDep(BtSU, SDep::Artificial)); |
1498 | |
1499 | // If one or more successors has been unscheduled, then the current |
1500 | // node is no longer available. |
1501 | if (!TrySU->isAvailable || !TrySU->NodeQueueId) { |
1502 | DEBUG(dbgs() << "TrySU not available; choosing node from queue\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "TrySU not available; choosing node from queue\n" ; } } while (false); |
1503 | CurSU = AvailableQueue->pop(); |
1504 | } else { |
1505 | DEBUG(dbgs() << "TrySU available\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "TrySU available\n"; } } while (false); |
1506 | // Available and in AvailableQueue |
1507 | AvailableQueue->remove(TrySU); |
1508 | CurSU = TrySU; |
1509 | } |
1510 | FindAvailableNode(); |
1511 | // Interferences has been mutated. We must break. |
1512 | break; |
1513 | } |
1514 | } |
1515 | |
1516 | if (!CurSU) { |
1517 | // Can't backtrack. If it's too expensive to copy the value, then try |
1518 | // duplicate the nodes that produces these "too expensive to copy" |
1519 | // values to break the dependency. In case even that doesn't work, |
1520 | // insert cross class copies. |
1521 | // If it's not too expensive, i.e. cost != -1, issue copies. |
1522 | SUnit *TrySU = Interferences[0]; |
1523 | SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU]; |
1524 | assert(LRegs.size() == 1 && "Can't handle this yet!")(static_cast <bool> (LRegs.size() == 1 && "Can't handle this yet!" ) ? void (0) : __assert_fail ("LRegs.size() == 1 && \"Can't handle this yet!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1524, __extension__ __PRETTY_FUNCTION__)); |
1525 | unsigned Reg = LRegs[0]; |
1526 | SUnit *LRDef = LiveRegDefs[Reg]; |
1527 | MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII); |
1528 | const TargetRegisterClass *RC = |
1529 | TRI->getMinimalPhysRegClass(Reg, VT); |
1530 | const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC); |
1531 | |
1532 | // If cross copy register class is the same as RC, then it must be possible |
1533 | // copy the value directly. Do not try duplicate the def. |
1534 | // If cross copy register class is not the same as RC, then it's possible to |
1535 | // copy the value but it require cross register class copies and it is |
1536 | // expensive. |
1537 | // If cross copy register class is null, then it's not possible to copy |
1538 | // the value at all. |
1539 | SUnit *NewDef = nullptr; |
1540 | if (DestRC != RC) { |
1541 | NewDef = CopyAndMoveSuccessors(LRDef); |
1542 | if (!DestRC && !NewDef) |
1543 | report_fatal_error("Can't handle live physical register dependency!"); |
1544 | } |
1545 | if (!NewDef) { |
1546 | // Issue copies, these can be expensive cross register class copies. |
1547 | SmallVector<SUnit*, 2> Copies; |
1548 | InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies); |
1549 | 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 (false) |
1550 | << " 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 (false); |
1551 | AddPred(TrySU, SDep(Copies.front(), SDep::Artificial)); |
1552 | NewDef = Copies.back(); |
1553 | } |
1554 | |
1555 | 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 (false) |
1556 | << " 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 (false); |
1557 | LiveRegDefs[Reg] = NewDef; |
1558 | AddPred(NewDef, SDep(TrySU, SDep::Artificial)); |
1559 | TrySU->isAvailable = false; |
1560 | CurSU = NewDef; |
1561 | } |
1562 | assert(CurSU && "Unable to resolve live physical register dependencies!")(static_cast <bool> (CurSU && "Unable to resolve live physical register dependencies!" ) ? void (0) : __assert_fail ("CurSU && \"Unable to resolve live physical register dependencies!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1562, __extension__ __PRETTY_FUNCTION__)); |
1563 | return CurSU; |
1564 | } |
1565 | |
1566 | /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up |
1567 | /// schedulers. |
1568 | void ScheduleDAGRRList::ListScheduleBottomUp() { |
1569 | // Release any predecessors of the special Exit node. |
1570 | ReleasePredecessors(&ExitSU); |
1571 | |
1572 | // Add root to Available queue. |
1573 | if (!SUnits.empty()) { |
1574 | SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()]; |
1575 | assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!")(static_cast <bool> (RootSU->Succs.empty() && "Graph root shouldn't have successors!") ? void (0) : __assert_fail ("RootSU->Succs.empty() && \"Graph root shouldn't have successors!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1575, __extension__ __PRETTY_FUNCTION__)); |
1576 | RootSU->isAvailable = true; |
1577 | AvailableQueue->push(RootSU); |
1578 | } |
1579 | |
1580 | // While Available queue is not empty, grab the node with the highest |
1581 | // priority. If it is not ready put it back. Schedule the node. |
1582 | Sequence.reserve(SUnits.size()); |
1583 | while (!AvailableQueue->empty() || !Interferences.empty()) { |
1584 | DEBUG(dbgs() << "\nExamining Available:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "\nExamining Available:\n" ; AvailableQueue->dump(this); } } while (false) |
1585 | AvailableQueue->dump(this))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "\nExamining Available:\n" ; AvailableQueue->dump(this); } } while (false); |
1586 | |
1587 | // Pick the best node to schedule taking all constraints into |
1588 | // consideration. |
1589 | SUnit *SU = PickNodeToScheduleBottomUp(); |
1590 | |
1591 | AdvancePastStalls(SU); |
1592 | |
1593 | ScheduleNodeBottomUp(SU); |
1594 | |
1595 | while (AvailableQueue->empty() && !PendingQueue.empty()) { |
1596 | // Advance the cycle to free resources. Skip ahead to the next ready SU. |
1597 | assert(MinAvailableCycle < std::numeric_limits<unsigned>::max() &&(static_cast <bool> (MinAvailableCycle < std::numeric_limits <unsigned>::max() && "MinAvailableCycle uninitialized" ) ? void (0) : __assert_fail ("MinAvailableCycle < std::numeric_limits<unsigned>::max() && \"MinAvailableCycle uninitialized\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1598, __extension__ __PRETTY_FUNCTION__)) |
1598 | "MinAvailableCycle uninitialized")(static_cast <bool> (MinAvailableCycle < std::numeric_limits <unsigned>::max() && "MinAvailableCycle uninitialized" ) ? void (0) : __assert_fail ("MinAvailableCycle < std::numeric_limits<unsigned>::max() && \"MinAvailableCycle uninitialized\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1598, __extension__ __PRETTY_FUNCTION__)); |
1599 | AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle)); |
1600 | } |
1601 | } |
1602 | |
1603 | // Reverse the order if it is bottom up. |
1604 | std::reverse(Sequence.begin(), Sequence.end()); |
1605 | |
1606 | #ifndef NDEBUG |
1607 | VerifyScheduledSequence(/*isBottomUp=*/true); |
1608 | #endif |
1609 | } |
1610 | |
1611 | namespace { |
1612 | |
1613 | class RegReductionPQBase; |
1614 | |
1615 | struct queue_sort { |
1616 | bool isReady(SUnit* SU, unsigned CurCycle) const { return true; } |
1617 | }; |
1618 | |
1619 | #ifndef NDEBUG |
1620 | template<class SF> |
1621 | struct reverse_sort : public queue_sort { |
1622 | SF &SortFunc; |
1623 | |
1624 | reverse_sort(SF &sf) : SortFunc(sf) {} |
1625 | |
1626 | bool operator()(SUnit* left, SUnit* right) const { |
1627 | // reverse left/right rather than simply !SortFunc(left, right) |
1628 | // to expose different paths in the comparison logic. |
1629 | return SortFunc(right, left); |
1630 | } |
1631 | }; |
1632 | #endif // NDEBUG |
1633 | |
1634 | /// bu_ls_rr_sort - Priority function for bottom up register pressure |
1635 | // reduction scheduler. |
1636 | struct bu_ls_rr_sort : public queue_sort { |
1637 | enum { |
1638 | IsBottomUp = true, |
1639 | HasReadyFilter = false |
1640 | }; |
1641 | |
1642 | RegReductionPQBase *SPQ; |
1643 | |
1644 | bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {} |
1645 | |
1646 | bool operator()(SUnit* left, SUnit* right) const; |
1647 | }; |
1648 | |
1649 | // src_ls_rr_sort - Priority function for source order scheduler. |
1650 | struct src_ls_rr_sort : public queue_sort { |
1651 | enum { |
1652 | IsBottomUp = true, |
1653 | HasReadyFilter = false |
1654 | }; |
1655 | |
1656 | RegReductionPQBase *SPQ; |
1657 | |
1658 | src_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {} |
1659 | |
1660 | bool operator()(SUnit* left, SUnit* right) const; |
1661 | }; |
1662 | |
1663 | // hybrid_ls_rr_sort - Priority function for hybrid scheduler. |
1664 | struct hybrid_ls_rr_sort : public queue_sort { |
1665 | enum { |
1666 | IsBottomUp = true, |
1667 | HasReadyFilter = false |
1668 | }; |
1669 | |
1670 | RegReductionPQBase *SPQ; |
1671 | |
1672 | hybrid_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {} |
1673 | |
1674 | bool isReady(SUnit *SU, unsigned CurCycle) const; |
1675 | |
1676 | bool operator()(SUnit* left, SUnit* right) const; |
1677 | }; |
1678 | |
1679 | // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism) |
1680 | // scheduler. |
1681 | struct ilp_ls_rr_sort : public queue_sort { |
1682 | enum { |
1683 | IsBottomUp = true, |
1684 | HasReadyFilter = false |
1685 | }; |
1686 | |
1687 | RegReductionPQBase *SPQ; |
1688 | |
1689 | ilp_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {} |
1690 | |
1691 | bool isReady(SUnit *SU, unsigned CurCycle) const; |
1692 | |
1693 | bool operator()(SUnit* left, SUnit* right) const; |
1694 | }; |
1695 | |
1696 | class RegReductionPQBase : public SchedulingPriorityQueue { |
1697 | protected: |
1698 | std::vector<SUnit *> Queue; |
1699 | unsigned CurQueueId = 0; |
1700 | bool TracksRegPressure; |
1701 | bool SrcOrder; |
1702 | |
1703 | // SUnits - The SUnits for the current graph. |
1704 | std::vector<SUnit> *SUnits; |
1705 | |
1706 | MachineFunction &MF; |
1707 | const TargetInstrInfo *TII; |
1708 | const TargetRegisterInfo *TRI; |
1709 | const TargetLowering *TLI; |
1710 | ScheduleDAGRRList *scheduleDAG = nullptr; |
1711 | |
1712 | // SethiUllmanNumbers - The SethiUllman number for each node. |
1713 | std::vector<unsigned> SethiUllmanNumbers; |
1714 | |
1715 | /// RegPressure - Tracking current reg pressure per register class. |
1716 | std::vector<unsigned> RegPressure; |
1717 | |
1718 | /// RegLimit - Tracking the number of allocatable registers per register |
1719 | /// class. |
1720 | std::vector<unsigned> RegLimit; |
1721 | |
1722 | public: |
1723 | RegReductionPQBase(MachineFunction &mf, |
1724 | bool hasReadyFilter, |
1725 | bool tracksrp, |
1726 | bool srcorder, |
1727 | const TargetInstrInfo *tii, |
1728 | const TargetRegisterInfo *tri, |
1729 | const TargetLowering *tli) |
1730 | : SchedulingPriorityQueue(hasReadyFilter), TracksRegPressure(tracksrp), |
1731 | SrcOrder(srcorder), MF(mf), TII(tii), TRI(tri), TLI(tli) { |
1732 | if (TracksRegPressure) { |
1733 | unsigned NumRC = TRI->getNumRegClasses(); |
1734 | RegLimit.resize(NumRC); |
1735 | RegPressure.resize(NumRC); |
1736 | std::fill(RegLimit.begin(), RegLimit.end(), 0); |
1737 | std::fill(RegPressure.begin(), RegPressure.end(), 0); |
1738 | for (const TargetRegisterClass *RC : TRI->regclasses()) |
1739 | RegLimit[RC->getID()] = tri->getRegPressureLimit(RC, MF); |
1740 | } |
1741 | } |
1742 | |
1743 | void setScheduleDAG(ScheduleDAGRRList *scheduleDag) { |
1744 | scheduleDAG = scheduleDag; |
1745 | } |
1746 | |
1747 | ScheduleHazardRecognizer* getHazardRec() { |
1748 | return scheduleDAG->getHazardRec(); |
1749 | } |
1750 | |
1751 | void initNodes(std::vector<SUnit> &sunits) override; |
1752 | |
1753 | void addNode(const SUnit *SU) override; |
1754 | |
1755 | void updateNode(const SUnit *SU) override; |
1756 | |
1757 | void releaseState() override { |
1758 | SUnits = nullptr; |
1759 | SethiUllmanNumbers.clear(); |
1760 | std::fill(RegPressure.begin(), RegPressure.end(), 0); |
1761 | } |
1762 | |
1763 | unsigned getNodePriority(const SUnit *SU) const; |
1764 | |
1765 | unsigned getNodeOrdering(const SUnit *SU) const { |
1766 | if (!SU->getNode()) return 0; |
1767 | |
1768 | return SU->getNode()->getIROrder(); |
1769 | } |
1770 | |
1771 | bool empty() const override { return Queue.empty(); } |
1772 | |
1773 | void push(SUnit *U) override { |
1774 | assert(!U->NodeQueueId && "Node in the queue already")(static_cast <bool> (!U->NodeQueueId && "Node in the queue already" ) ? void (0) : __assert_fail ("!U->NodeQueueId && \"Node in the queue already\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1774, __extension__ __PRETTY_FUNCTION__)); |
1775 | U->NodeQueueId = ++CurQueueId; |
1776 | Queue.push_back(U); |
1777 | } |
1778 | |
1779 | void remove(SUnit *SU) override { |
1780 | assert(!Queue.empty() && "Queue is empty!")(static_cast <bool> (!Queue.empty() && "Queue is empty!" ) ? void (0) : __assert_fail ("!Queue.empty() && \"Queue is empty!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1780, __extension__ __PRETTY_FUNCTION__)); |
1781 | assert(SU->NodeQueueId != 0 && "Not in queue!")(static_cast <bool> (SU->NodeQueueId != 0 && "Not in queue!") ? void (0) : __assert_fail ("SU->NodeQueueId != 0 && \"Not in queue!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1781, __extension__ __PRETTY_FUNCTION__)); |
1782 | std::vector<SUnit *>::iterator I = llvm::find(Queue, SU); |
1783 | if (I != std::prev(Queue.end())) |
1784 | std::swap(*I, Queue.back()); |
1785 | Queue.pop_back(); |
1786 | SU->NodeQueueId = 0; |
1787 | } |
1788 | |
1789 | bool tracksRegPressure() const override { return TracksRegPressure; } |
1790 | |
1791 | void dumpRegPressure() const; |
1792 | |
1793 | bool HighRegPressure(const SUnit *SU) const; |
1794 | |
1795 | bool MayReduceRegPressure(SUnit *SU) const; |
1796 | |
1797 | int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const; |
1798 | |
1799 | void scheduledNode(SUnit *SU) override; |
1800 | |
1801 | void unscheduledNode(SUnit *SU) override; |
1802 | |
1803 | protected: |
1804 | bool canClobber(const SUnit *SU, const SUnit *Op); |
1805 | void AddPseudoTwoAddrDeps(); |
1806 | void PrescheduleNodesWithMultipleUses(); |
1807 | void CalculateSethiUllmanNumbers(); |
1808 | }; |
1809 | |
1810 | template<class SF> |
1811 | static SUnit *popFromQueueImpl(std::vector<SUnit *> &Q, SF &Picker) { |
1812 | std::vector<SUnit *>::iterator Best = Q.begin(); |
1813 | for (auto I = std::next(Q.begin()), E = Q.end(); I != E; ++I) |
1814 | if (Picker(*Best, *I)) |
1815 | Best = I; |
1816 | SUnit *V = *Best; |
1817 | if (Best != std::prev(Q.end())) |
1818 | std::swap(*Best, Q.back()); |
1819 | Q.pop_back(); |
1820 | return V; |
1821 | } |
1822 | |
1823 | template<class SF> |
1824 | SUnit *popFromQueue(std::vector<SUnit *> &Q, SF &Picker, ScheduleDAG *DAG) { |
1825 | #ifndef NDEBUG |
1826 | if (DAG->StressSched) { |
1827 | reverse_sort<SF> RPicker(Picker); |
1828 | return popFromQueueImpl(Q, RPicker); |
1829 | } |
1830 | #endif |
1831 | (void)DAG; |
1832 | return popFromQueueImpl(Q, Picker); |
1833 | } |
1834 | |
1835 | //===----------------------------------------------------------------------===// |
1836 | // RegReductionPriorityQueue Definition |
1837 | //===----------------------------------------------------------------------===// |
1838 | // |
1839 | // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers |
1840 | // to reduce register pressure. |
1841 | // |
1842 | template<class SF> |
1843 | class RegReductionPriorityQueue : public RegReductionPQBase { |
1844 | SF Picker; |
1845 | |
1846 | public: |
1847 | RegReductionPriorityQueue(MachineFunction &mf, |
1848 | bool tracksrp, |
1849 | bool srcorder, |
1850 | const TargetInstrInfo *tii, |
1851 | const TargetRegisterInfo *tri, |
1852 | const TargetLowering *tli) |
1853 | : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder, |
1854 | tii, tri, tli), |
1855 | Picker(this) {} |
1856 | |
1857 | bool isBottomUp() const override { return SF::IsBottomUp; } |
1858 | |
1859 | bool isReady(SUnit *U) const override { |
1860 | return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle()); |
1861 | } |
1862 | |
1863 | SUnit *pop() override { |
1864 | if (Queue.empty()) return nullptr; |
1865 | |
1866 | SUnit *V = popFromQueue(Queue, Picker, scheduleDAG); |
1867 | V->NodeQueueId = 0; |
1868 | return V; |
1869 | } |
1870 | |
1871 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1872 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump(ScheduleDAG *DAG) const override { |
1873 | // Emulate pop() without clobbering NodeQueueIds. |
1874 | std::vector<SUnit *> DumpQueue = Queue; |
1875 | SF DumpPicker = Picker; |
1876 | while (!DumpQueue.empty()) { |
1877 | SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG); |
1878 | dbgs() << "Height " << SU->getHeight() << ": "; |
1879 | SU->dump(DAG); |
1880 | } |
1881 | } |
1882 | #endif |
1883 | }; |
1884 | |
1885 | using BURegReductionPriorityQueue = RegReductionPriorityQueue<bu_ls_rr_sort>; |
1886 | using SrcRegReductionPriorityQueue = RegReductionPriorityQueue<src_ls_rr_sort>; |
1887 | using HybridBURRPriorityQueue = RegReductionPriorityQueue<hybrid_ls_rr_sort>; |
1888 | using ILPBURRPriorityQueue = RegReductionPriorityQueue<ilp_ls_rr_sort>; |
1889 | |
1890 | } // end anonymous namespace |
1891 | |
1892 | //===----------------------------------------------------------------------===// |
1893 | // Static Node Priority for Register Pressure Reduction |
1894 | //===----------------------------------------------------------------------===// |
1895 | |
1896 | // Check for special nodes that bypass scheduling heuristics. |
1897 | // Currently this pushes TokenFactor nodes down, but may be used for other |
1898 | // pseudo-ops as well. |
1899 | // |
1900 | // Return -1 to schedule right above left, 1 for left above right. |
1901 | // Return 0 if no bias exists. |
1902 | static int checkSpecialNodes(const SUnit *left, const SUnit *right) { |
1903 | bool LSchedLow = left->isScheduleLow; |
1904 | bool RSchedLow = right->isScheduleLow; |
1905 | if (LSchedLow != RSchedLow) |
1906 | return LSchedLow < RSchedLow ? 1 : -1; |
1907 | return 0; |
1908 | } |
1909 | |
1910 | /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number. |
1911 | /// Smaller number is the higher priority. |
1912 | static unsigned |
1913 | CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) { |
1914 | if (SUNumbers[SU->NodeNum] != 0) |
1915 | return SUNumbers[SU->NodeNum]; |
1916 | |
1917 | // Use WorkList to avoid stack overflow on excessively large IRs. |
1918 | struct WorkState { |
1919 | WorkState(const SUnit *SU) : SU(SU) {} |
1920 | const SUnit *SU; |
1921 | unsigned PredsProcessed = 0; |
1922 | }; |
1923 | |
1924 | SmallVector<WorkState, 16> WorkList; |
1925 | WorkList.push_back(SU); |
1926 | while (!WorkList.empty()) { |
1927 | auto &Temp = WorkList.back(); |
1928 | auto *TempSU = Temp.SU; |
1929 | bool AllPredsKnown = true; |
1930 | // Try to find a non-evaluated pred and push it into the processing stack. |
1931 | for (unsigned P = Temp.PredsProcessed; P < TempSU->Preds.size(); ++P) { |
1932 | auto &Pred = TempSU->Preds[P]; |
1933 | if (Pred.isCtrl()) continue; // ignore chain preds |
1934 | SUnit *PredSU = Pred.getSUnit(); |
1935 | if (SUNumbers[PredSU->NodeNum] == 0) { |
1936 | #ifndef NDEBUG |
1937 | // In debug mode, check that we don't have such element in the stack. |
1938 | for (auto It : WorkList) |
1939 | assert(It.SU != PredSU && "Trying to push an element twice?")(static_cast <bool> (It.SU != PredSU && "Trying to push an element twice?" ) ? void (0) : __assert_fail ("It.SU != PredSU && \"Trying to push an element twice?\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1939, __extension__ __PRETTY_FUNCTION__)); |
1940 | #endif |
1941 | // Next time start processing this one starting from the next pred. |
1942 | Temp.PredsProcessed = P + 1; |
1943 | WorkList.push_back(PredSU); |
1944 | AllPredsKnown = false; |
1945 | break; |
1946 | } |
1947 | } |
1948 | |
1949 | if (!AllPredsKnown) |
1950 | continue; |
1951 | |
1952 | // Once all preds are known, we can calculate the answer for this one. |
1953 | unsigned SethiUllmanNumber = 0; |
1954 | unsigned Extra = 0; |
1955 | for (const SDep &Pred : TempSU->Preds) { |
1956 | if (Pred.isCtrl()) continue; // ignore chain preds |
1957 | SUnit *PredSU = Pred.getSUnit(); |
1958 | unsigned PredSethiUllman = SUNumbers[PredSU->NodeNum]; |
1959 | assert(PredSethiUllman > 0 && "We should have evaluated this pred!")(static_cast <bool> (PredSethiUllman > 0 && "We should have evaluated this pred!" ) ? void (0) : __assert_fail ("PredSethiUllman > 0 && \"We should have evaluated this pred!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1959, __extension__ __PRETTY_FUNCTION__)); |
1960 | if (PredSethiUllman > SethiUllmanNumber) { |
1961 | SethiUllmanNumber = PredSethiUllman; |
1962 | Extra = 0; |
1963 | } else if (PredSethiUllman == SethiUllmanNumber) |
1964 | ++Extra; |
1965 | } |
1966 | |
1967 | SethiUllmanNumber += Extra; |
1968 | if (SethiUllmanNumber == 0) |
1969 | SethiUllmanNumber = 1; |
1970 | SUNumbers[TempSU->NodeNum] = SethiUllmanNumber; |
1971 | WorkList.pop_back(); |
1972 | } |
1973 | |
1974 | assert(SUNumbers[SU->NodeNum] > 0 && "SethiUllman should never be zero!")(static_cast <bool> (SUNumbers[SU->NodeNum] > 0 && "SethiUllman should never be zero!") ? void (0) : __assert_fail ("SUNumbers[SU->NodeNum] > 0 && \"SethiUllman should never be zero!\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 1974, __extension__ __PRETTY_FUNCTION__)); |
1975 | return SUNumbers[SU->NodeNum]; |
1976 | } |
1977 | |
1978 | /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all |
1979 | /// scheduling units. |
1980 | void RegReductionPQBase::CalculateSethiUllmanNumbers() { |
1981 | SethiUllmanNumbers.assign(SUnits->size(), 0); |
1982 | |
1983 | for (const SUnit &SU : *SUnits) |
1984 | CalcNodeSethiUllmanNumber(&SU, SethiUllmanNumbers); |
1985 | } |
1986 | |
1987 | void RegReductionPQBase::addNode(const SUnit *SU) { |
1988 | unsigned SUSize = SethiUllmanNumbers.size(); |
1989 | if (SUnits->size() > SUSize) |
1990 | SethiUllmanNumbers.resize(SUSize*2, 0); |
1991 | CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers); |
1992 | } |
1993 | |
1994 | void RegReductionPQBase::updateNode(const SUnit *SU) { |
1995 | SethiUllmanNumbers[SU->NodeNum] = 0; |
1996 | CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers); |
1997 | } |
1998 | |
1999 | // Lower priority means schedule further down. For bottom-up scheduling, lower |
2000 | // priority SUs are scheduled before higher priority SUs. |
2001 | unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const { |
2002 | assert(SU->NodeNum < SethiUllmanNumbers.size())(static_cast <bool> (SU->NodeNum < SethiUllmanNumbers .size()) ? void (0) : __assert_fail ("SU->NodeNum < SethiUllmanNumbers.size()" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2002, __extension__ __PRETTY_FUNCTION__)); |
2003 | unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0; |
2004 | if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg) |
2005 | // CopyToReg should be close to its uses to facilitate coalescing and |
2006 | // avoid spilling. |
2007 | return 0; |
2008 | if (Opc == TargetOpcode::EXTRACT_SUBREG || |
2009 | Opc == TargetOpcode::SUBREG_TO_REG || |
2010 | Opc == TargetOpcode::INSERT_SUBREG) |
2011 | // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be |
2012 | // close to their uses to facilitate coalescing. |
2013 | return 0; |
2014 | if (SU->NumSuccs == 0 && SU->NumPreds != 0) |
2015 | // If SU does not have a register use, i.e. it doesn't produce a value |
2016 | // that would be consumed (e.g. store), then it terminates a chain of |
2017 | // computation. Give it a large SethiUllman number so it will be |
2018 | // scheduled right before its predecessors that it doesn't lengthen |
2019 | // their live ranges. |
2020 | return 0xffff; |
2021 | if (SU->NumPreds == 0 && SU->NumSuccs != 0) |
2022 | // If SU does not have a register def, schedule it close to its uses |
2023 | // because it does not lengthen any live ranges. |
2024 | return 0; |
2025 | #if 1 |
2026 | return SethiUllmanNumbers[SU->NodeNum]; |
2027 | #else |
2028 | unsigned Priority = SethiUllmanNumbers[SU->NodeNum]; |
2029 | if (SU->isCallOp) { |
2030 | // FIXME: This assumes all of the defs are used as call operands. |
2031 | int NP = (int)Priority - SU->getNode()->getNumValues(); |
2032 | return (NP > 0) ? NP : 0; |
2033 | } |
2034 | return Priority; |
2035 | #endif |
2036 | } |
2037 | |
2038 | //===----------------------------------------------------------------------===// |
2039 | // Register Pressure Tracking |
2040 | //===----------------------------------------------------------------------===// |
2041 | |
2042 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
2043 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegReductionPQBase::dumpRegPressure() const { |
2044 | for (const TargetRegisterClass *RC : TRI->regclasses()) { |
2045 | unsigned Id = RC->getID(); |
2046 | unsigned RP = RegPressure[Id]; |
2047 | if (!RP) continue; |
2048 | DEBUG(dbgs() << TRI->getRegClassName(RC) << ": " << RP << " / "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << TRI->getRegClassName(RC ) << ": " << RP << " / " << RegLimit[ Id] << '\n'; } } while (false) |
2049 | << RegLimit[Id] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << TRI->getRegClassName(RC ) << ": " << RP << " / " << RegLimit[ Id] << '\n'; } } while (false); |
2050 | } |
2051 | } |
2052 | #endif |
2053 | |
2054 | bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const { |
2055 | if (!TLI) |
2056 | return false; |
2057 | |
2058 | for (const SDep &Pred : SU->Preds) { |
2059 | if (Pred.isCtrl()) |
2060 | continue; |
2061 | SUnit *PredSU = Pred.getSUnit(); |
2062 | // NumRegDefsLeft is zero when enough uses of this node have been scheduled |
2063 | // to cover the number of registers defined (they are all live). |
2064 | if (PredSU->NumRegDefsLeft == 0) { |
2065 | continue; |
2066 | } |
2067 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG); |
2068 | RegDefPos.IsValid(); RegDefPos.Advance()) { |
2069 | unsigned RCId, Cost; |
2070 | GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF); |
2071 | |
2072 | if ((RegPressure[RCId] + Cost) >= RegLimit[RCId]) |
2073 | return true; |
2074 | } |
2075 | } |
2076 | return false; |
2077 | } |
2078 | |
2079 | bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const { |
2080 | const SDNode *N = SU->getNode(); |
2081 | |
2082 | if (!N->isMachineOpcode() || !SU->NumSuccs) |
2083 | return false; |
2084 | |
2085 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); |
2086 | for (unsigned i = 0; i != NumDefs; ++i) { |
2087 | MVT VT = N->getSimpleValueType(i); |
2088 | if (!N->hasAnyUseOfValue(i)) |
2089 | continue; |
2090 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); |
2091 | if (RegPressure[RCId] >= RegLimit[RCId]) |
2092 | return true; |
2093 | } |
2094 | return false; |
2095 | } |
2096 | |
2097 | // Compute the register pressure contribution by this instruction by count up |
2098 | // for uses that are not live and down for defs. Only count register classes |
2099 | // that are already under high pressure. As a side effect, compute the number of |
2100 | // uses of registers that are already live. |
2101 | // |
2102 | // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure |
2103 | // so could probably be factored. |
2104 | int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const { |
2105 | LiveUses = 0; |
2106 | int PDiff = 0; |
2107 | for (const SDep &Pred : SU->Preds) { |
2108 | if (Pred.isCtrl()) |
2109 | continue; |
2110 | SUnit *PredSU = Pred.getSUnit(); |
2111 | // NumRegDefsLeft is zero when enough uses of this node have been scheduled |
2112 | // to cover the number of registers defined (they are all live). |
2113 | if (PredSU->NumRegDefsLeft == 0) { |
2114 | if (PredSU->getNode()->isMachineOpcode()) |
2115 | ++LiveUses; |
2116 | continue; |
2117 | } |
2118 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG); |
2119 | RegDefPos.IsValid(); RegDefPos.Advance()) { |
2120 | MVT VT = RegDefPos.GetValue(); |
2121 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); |
2122 | if (RegPressure[RCId] >= RegLimit[RCId]) |
2123 | ++PDiff; |
2124 | } |
2125 | } |
2126 | const SDNode *N = SU->getNode(); |
2127 | |
2128 | if (!N || !N->isMachineOpcode() || !SU->NumSuccs) |
2129 | return PDiff; |
2130 | |
2131 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); |
2132 | for (unsigned i = 0; i != NumDefs; ++i) { |
2133 | MVT VT = N->getSimpleValueType(i); |
2134 | if (!N->hasAnyUseOfValue(i)) |
2135 | continue; |
2136 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); |
2137 | if (RegPressure[RCId] >= RegLimit[RCId]) |
2138 | --PDiff; |
2139 | } |
2140 | return PDiff; |
2141 | } |
2142 | |
2143 | void RegReductionPQBase::scheduledNode(SUnit *SU) { |
2144 | if (!TracksRegPressure) |
2145 | return; |
2146 | |
2147 | if (!SU->getNode()) |
2148 | return; |
2149 | |
2150 | for (const SDep &Pred : SU->Preds) { |
2151 | if (Pred.isCtrl()) |
2152 | continue; |
2153 | SUnit *PredSU = Pred.getSUnit(); |
2154 | // NumRegDefsLeft is zero when enough uses of this node have been scheduled |
2155 | // to cover the number of registers defined (they are all live). |
2156 | if (PredSU->NumRegDefsLeft == 0) { |
2157 | continue; |
2158 | } |
2159 | // FIXME: The ScheduleDAG currently loses information about which of a |
2160 | // node's values is consumed by each dependence. Consequently, if the node |
2161 | // defines multiple register classes, we don't know which to pressurize |
2162 | // here. Instead the following loop consumes the register defs in an |
2163 | // arbitrary order. At least it handles the common case of clustered loads |
2164 | // to the same class. For precise liveness, each SDep needs to indicate the |
2165 | // result number. But that tightly couples the ScheduleDAG with the |
2166 | // SelectionDAG making updates tricky. A simpler hack would be to attach a |
2167 | // value type or register class to SDep. |
2168 | // |
2169 | // The most important aspect of register tracking is balancing the increase |
2170 | // here with the reduction further below. Note that this SU may use multiple |
2171 | // defs in PredSU. The can't be determined here, but we've already |
2172 | // compensated by reducing NumRegDefsLeft in PredSU during |
2173 | // ScheduleDAGSDNodes::AddSchedEdges. |
2174 | --PredSU->NumRegDefsLeft; |
2175 | unsigned SkipRegDefs = PredSU->NumRegDefsLeft; |
2176 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG); |
2177 | RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) { |
2178 | if (SkipRegDefs) |
2179 | continue; |
2180 | |
2181 | unsigned RCId, Cost; |
2182 | GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF); |
2183 | RegPressure[RCId] += Cost; |
2184 | break; |
2185 | } |
2186 | } |
2187 | |
2188 | // We should have this assert, but there may be dead SDNodes that never |
2189 | // materialize as SUnits, so they don't appear to generate liveness. |
2190 | //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses"); |
2191 | int SkipRegDefs = (int)SU->NumRegDefsLeft; |
2192 | for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG); |
2193 | RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) { |
2194 | if (SkipRegDefs > 0) |
2195 | continue; |
2196 | unsigned RCId, Cost; |
2197 | GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF); |
2198 | if (RegPressure[RCId] < Cost) { |
2199 | // Register pressure tracking is imprecise. This can happen. But we try |
2200 | // hard not to let it happen because it likely results in poor scheduling. |
2201 | 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 (false); |
2202 | RegPressure[RCId] = 0; |
2203 | } |
2204 | else { |
2205 | RegPressure[RCId] -= Cost; |
2206 | } |
2207 | } |
2208 | DEBUG(dumpRegPressure())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dumpRegPressure(); } } while (false); |
2209 | } |
2210 | |
2211 | void RegReductionPQBase::unscheduledNode(SUnit *SU) { |
2212 | if (!TracksRegPressure) |
2213 | return; |
2214 | |
2215 | const SDNode *N = SU->getNode(); |
2216 | if (!N) return; |
2217 | |
2218 | if (!N->isMachineOpcode()) { |
2219 | if (N->getOpcode() != ISD::CopyToReg) |
2220 | return; |
2221 | } else { |
2222 | unsigned Opc = N->getMachineOpcode(); |
2223 | if (Opc == TargetOpcode::EXTRACT_SUBREG || |
2224 | Opc == TargetOpcode::INSERT_SUBREG || |
2225 | Opc == TargetOpcode::SUBREG_TO_REG || |
2226 | Opc == TargetOpcode::REG_SEQUENCE || |
2227 | Opc == TargetOpcode::IMPLICIT_DEF) |
2228 | return; |
2229 | } |
2230 | |
2231 | for (const SDep &Pred : SU->Preds) { |
2232 | if (Pred.isCtrl()) |
2233 | continue; |
2234 | SUnit *PredSU = Pred.getSUnit(); |
2235 | // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only |
2236 | // counts data deps. |
2237 | if (PredSU->NumSuccsLeft != PredSU->Succs.size()) |
2238 | continue; |
2239 | const SDNode *PN = PredSU->getNode(); |
2240 | if (!PN->isMachineOpcode()) { |
2241 | if (PN->getOpcode() == ISD::CopyFromReg) { |
2242 | MVT VT = PN->getSimpleValueType(0); |
2243 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); |
2244 | RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); |
2245 | } |
2246 | continue; |
2247 | } |
2248 | unsigned POpc = PN->getMachineOpcode(); |
2249 | if (POpc == TargetOpcode::IMPLICIT_DEF) |
2250 | continue; |
2251 | if (POpc == TargetOpcode::EXTRACT_SUBREG || |
2252 | POpc == TargetOpcode::INSERT_SUBREG || |
2253 | POpc == TargetOpcode::SUBREG_TO_REG) { |
2254 | MVT VT = PN->getSimpleValueType(0); |
2255 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); |
2256 | RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); |
2257 | continue; |
2258 | } |
2259 | unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs(); |
2260 | for (unsigned i = 0; i != NumDefs; ++i) { |
2261 | MVT VT = PN->getSimpleValueType(i); |
2262 | if (!PN->hasAnyUseOfValue(i)) |
2263 | continue; |
2264 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); |
2265 | if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT)) |
2266 | // Register pressure tracking is imprecise. This can happen. |
2267 | RegPressure[RCId] = 0; |
2268 | else |
2269 | RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT); |
2270 | } |
2271 | } |
2272 | |
2273 | // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses() |
2274 | // may transfer data dependencies to CopyToReg. |
2275 | if (SU->NumSuccs && N->isMachineOpcode()) { |
2276 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); |
2277 | for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) { |
2278 | MVT VT = N->getSimpleValueType(i); |
2279 | if (VT == MVT::Glue || VT == MVT::Other) |
2280 | continue; |
2281 | if (!N->hasAnyUseOfValue(i)) |
2282 | continue; |
2283 | unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); |
2284 | RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); |
2285 | } |
2286 | } |
2287 | |
2288 | DEBUG(dumpRegPressure())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dumpRegPressure(); } } while (false); |
2289 | } |
2290 | |
2291 | //===----------------------------------------------------------------------===// |
2292 | // Dynamic Node Priority for Register Pressure Reduction |
2293 | //===----------------------------------------------------------------------===// |
2294 | |
2295 | /// closestSucc - Returns the scheduled cycle of the successor which is |
2296 | /// closest to the current cycle. |
2297 | static unsigned closestSucc(const SUnit *SU) { |
2298 | unsigned MaxHeight = 0; |
2299 | for (const SDep &Succ : SU->Succs) { |
2300 | if (Succ.isCtrl()) continue; // ignore chain succs |
2301 | unsigned Height = Succ.getSUnit()->getHeight(); |
2302 | // If there are bunch of CopyToRegs stacked up, they should be considered |
2303 | // to be at the same position. |
2304 | if (Succ.getSUnit()->getNode() && |
2305 | Succ.getSUnit()->getNode()->getOpcode() == ISD::CopyToReg) |
2306 | Height = closestSucc(Succ.getSUnit())+1; |
2307 | if (Height > MaxHeight) |
2308 | MaxHeight = Height; |
2309 | } |
2310 | return MaxHeight; |
2311 | } |
2312 | |
2313 | /// calcMaxScratches - Returns an cost estimate of the worse case requirement |
2314 | /// for scratch registers, i.e. number of data dependencies. |
2315 | static unsigned calcMaxScratches(const SUnit *SU) { |
2316 | unsigned Scratches = 0; |
2317 | for (const SDep &Pred : SU->Preds) { |
2318 | if (Pred.isCtrl()) continue; // ignore chain preds |
2319 | Scratches++; |
2320 | } |
2321 | return Scratches; |
2322 | } |
2323 | |
2324 | /// hasOnlyLiveInOpers - Return true if SU has only value predecessors that are |
2325 | /// CopyFromReg from a virtual register. |
2326 | static bool hasOnlyLiveInOpers(const SUnit *SU) { |
2327 | bool RetVal = false; |
2328 | for (const SDep &Pred : SU->Preds) { |
2329 | if (Pred.isCtrl()) continue; |
2330 | const SUnit *PredSU = Pred.getSUnit(); |
2331 | if (PredSU->getNode() && |
2332 | PredSU->getNode()->getOpcode() == ISD::CopyFromReg) { |
2333 | unsigned Reg = |
2334 | cast<RegisterSDNode>(PredSU->getNode()->getOperand(1))->getReg(); |
2335 | if (TargetRegisterInfo::isVirtualRegister(Reg)) { |
2336 | RetVal = true; |
2337 | continue; |
2338 | } |
2339 | } |
2340 | return false; |
2341 | } |
2342 | return RetVal; |
2343 | } |
2344 | |
2345 | /// hasOnlyLiveOutUses - Return true if SU has only value successors that are |
2346 | /// CopyToReg to a virtual register. This SU def is probably a liveout and |
2347 | /// it has no other use. It should be scheduled closer to the terminator. |
2348 | static bool hasOnlyLiveOutUses(const SUnit *SU) { |
2349 | bool RetVal = false; |
2350 | for (const SDep &Succ : SU->Succs) { |
2351 | if (Succ.isCtrl()) continue; |
2352 | const SUnit *SuccSU = Succ.getSUnit(); |
2353 | if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) { |
2354 | unsigned Reg = |
2355 | cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg(); |
2356 | if (TargetRegisterInfo::isVirtualRegister(Reg)) { |
2357 | RetVal = true; |
2358 | continue; |
2359 | } |
2360 | } |
2361 | return false; |
2362 | } |
2363 | return RetVal; |
2364 | } |
2365 | |
2366 | // Set isVRegCycle for a node with only live in opers and live out uses. Also |
2367 | // set isVRegCycle for its CopyFromReg operands. |
2368 | // |
2369 | // This is only relevant for single-block loops, in which case the VRegCycle |
2370 | // node is likely an induction variable in which the operand and target virtual |
2371 | // registers should be coalesced (e.g. pre/post increment values). Setting the |
2372 | // isVRegCycle flag helps the scheduler prioritize other uses of the same |
2373 | // CopyFromReg so that this node becomes the virtual register "kill". This |
2374 | // avoids interference between the values live in and out of the block and |
2375 | // eliminates a copy inside the loop. |
2376 | static void initVRegCycle(SUnit *SU) { |
2377 | if (DisableSchedVRegCycle) |
2378 | return; |
2379 | |
2380 | if (!hasOnlyLiveInOpers(SU) || !hasOnlyLiveOutUses(SU)) |
2381 | return; |
2382 | |
2383 | DEBUG(dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n"; } } while (false); |
2384 | |
2385 | SU->isVRegCycle = true; |
2386 | |
2387 | for (const SDep &Pred : SU->Preds) { |
2388 | if (Pred.isCtrl()) continue; |
2389 | Pred.getSUnit()->isVRegCycle = true; |
2390 | } |
2391 | } |
2392 | |
2393 | // After scheduling the definition of a VRegCycle, clear the isVRegCycle flag of |
2394 | // CopyFromReg operands. We should no longer penalize other uses of this VReg. |
2395 | static void resetVRegCycle(SUnit *SU) { |
2396 | if (!SU->isVRegCycle) |
2397 | return; |
2398 | |
2399 | for (const SDep &Pred : SU->Preds) { |
2400 | if (Pred.isCtrl()) continue; // ignore chain preds |
2401 | SUnit *PredSU = Pred.getSUnit(); |
2402 | if (PredSU->isVRegCycle) { |
2403 | assert(PredSU->getNode()->getOpcode() == ISD::CopyFromReg &&(static_cast <bool> (PredSU->getNode()->getOpcode () == ISD::CopyFromReg && "VRegCycle def must be CopyFromReg" ) ? void (0) : __assert_fail ("PredSU->getNode()->getOpcode() == ISD::CopyFromReg && \"VRegCycle def must be CopyFromReg\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2404, __extension__ __PRETTY_FUNCTION__)) |
2404 | "VRegCycle def must be CopyFromReg")(static_cast <bool> (PredSU->getNode()->getOpcode () == ISD::CopyFromReg && "VRegCycle def must be CopyFromReg" ) ? void (0) : __assert_fail ("PredSU->getNode()->getOpcode() == ISD::CopyFromReg && \"VRegCycle def must be CopyFromReg\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2404, __extension__ __PRETTY_FUNCTION__)); |
2405 | Pred.getSUnit()->isVRegCycle = false; |
2406 | } |
2407 | } |
2408 | } |
2409 | |
2410 | // Return true if this SUnit uses a CopyFromReg node marked as a VRegCycle. This |
2411 | // means a node that defines the VRegCycle has not been scheduled yet. |
2412 | static bool hasVRegCycleUse(const SUnit *SU) { |
2413 | // If this SU also defines the VReg, don't hoist it as a "use". |
2414 | if (SU->isVRegCycle) |
2415 | return false; |
2416 | |
2417 | for (const SDep &Pred : SU->Preds) { |
2418 | if (Pred.isCtrl()) continue; // ignore chain preds |
2419 | if (Pred.getSUnit()->isVRegCycle && |
2420 | Pred.getSUnit()->getNode()->getOpcode() == ISD::CopyFromReg) { |
2421 | 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 (false); |
2422 | return true; |
2423 | } |
2424 | } |
2425 | return false; |
2426 | } |
2427 | |
2428 | // Check for either a dependence (latency) or resource (hazard) stall. |
2429 | // |
2430 | // Note: The ScheduleHazardRecognizer interface requires a non-const SU. |
2431 | static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) { |
2432 | if ((int)SPQ->getCurCycle() < Height) return true; |
2433 | if (SPQ->getHazardRec()->getHazardType(SU, 0) |
2434 | != ScheduleHazardRecognizer::NoHazard) |
2435 | return true; |
2436 | return false; |
2437 | } |
2438 | |
2439 | // Return -1 if left has higher priority, 1 if right has higher priority. |
2440 | // Return 0 if latency-based priority is equivalent. |
2441 | static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref, |
2442 | RegReductionPQBase *SPQ) { |
2443 | // Scheduling an instruction that uses a VReg whose postincrement has not yet |
2444 | // been scheduled will induce a copy. Model this as an extra cycle of latency. |
2445 | int LPenalty = hasVRegCycleUse(left) ? 1 : 0; |
2446 | int RPenalty = hasVRegCycleUse(right) ? 1 : 0; |
2447 | int LHeight = (int)left->getHeight() + LPenalty; |
2448 | int RHeight = (int)right->getHeight() + RPenalty; |
2449 | |
2450 | bool LStall = (!checkPref || left->SchedulingPref == Sched::ILP) && |
2451 | BUHasStall(left, LHeight, SPQ); |
2452 | bool RStall = (!checkPref || right->SchedulingPref == Sched::ILP) && |
2453 | BUHasStall(right, RHeight, SPQ); |
2454 | |
2455 | // If scheduling one of the node will cause a pipeline stall, delay it. |
2456 | // If scheduling either one of the node will cause a pipeline stall, sort |
2457 | // them according to their height. |
2458 | if (LStall) { |
2459 | if (!RStall) |
2460 | return 1; |
2461 | if (LHeight != RHeight) |
2462 | return LHeight > RHeight ? 1 : -1; |
2463 | } else if (RStall) |
2464 | return -1; |
2465 | |
2466 | // If either node is scheduling for latency, sort them by height/depth |
2467 | // and latency. |
2468 | if (!checkPref || (left->SchedulingPref == Sched::ILP || |
2469 | right->SchedulingPref == Sched::ILP)) { |
2470 | // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer |
2471 | // is enabled, grouping instructions by cycle, then its height is already |
2472 | // covered so only its depth matters. We also reach this point if both stall |
2473 | // but have the same height. |
2474 | if (!SPQ->getHazardRec()->isEnabled()) { |
2475 | if (LHeight != RHeight) |
2476 | return LHeight > RHeight ? 1 : -1; |
2477 | } |
2478 | int LDepth = left->getDepth() - LPenalty; |
2479 | int RDepth = right->getDepth() - RPenalty; |
2480 | if (LDepth != RDepth) { |
2481 | 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 (false) |
2482 | << ") 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 (false) |
2483 | << ") 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 (false); |
2484 | return LDepth < RDepth ? 1 : -1; |
2485 | } |
2486 | if (left->Latency != right->Latency) |
2487 | return left->Latency > right->Latency ? 1 : -1; |
2488 | } |
2489 | return 0; |
2490 | } |
2491 | |
2492 | static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) { |
2493 | // Schedule physical register definitions close to their use. This is |
2494 | // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as |
2495 | // long as shortening physreg live ranges is generally good, we can defer |
2496 | // creating a subtarget hook. |
2497 | if (!DisableSchedPhysRegJoin) { |
2498 | bool LHasPhysReg = left->hasPhysRegDefs; |
2499 | bool RHasPhysReg = right->hasPhysRegDefs; |
2500 | if (LHasPhysReg != RHasPhysReg) { |
2501 | #ifndef NDEBUG |
2502 | static const char *const PhysRegMsg[] = { " has no physreg", |
2503 | " defines a physreg" }; |
2504 | #endif |
2505 | 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 (false) |
2506 | << 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 (false) |
2507 | << 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 (false); |
2508 | return LHasPhysReg < RHasPhysReg; |
2509 | } |
2510 | } |
2511 | |
2512 | // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down. |
2513 | unsigned LPriority = SPQ->getNodePriority(left); |
2514 | unsigned RPriority = SPQ->getNodePriority(right); |
2515 | |
2516 | // Be really careful about hoisting call operands above previous calls. |
2517 | // Only allows it if it would reduce register pressure. |
2518 | if (left->isCall && right->isCallOp) { |
2519 | unsigned RNumVals = right->getNode()->getNumValues(); |
2520 | RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0; |
2521 | } |
2522 | if (right->isCall && left->isCallOp) { |
2523 | unsigned LNumVals = left->getNode()->getNumValues(); |
2524 | LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0; |
2525 | } |
2526 | |
2527 | if (LPriority != RPriority) |
2528 | return LPriority > RPriority; |
2529 | |
2530 | // One or both of the nodes are calls and their sethi-ullman numbers are the |
2531 | // same, then keep source order. |
2532 | if (left->isCall || right->isCall) { |
2533 | unsigned LOrder = SPQ->getNodeOrdering(left); |
2534 | unsigned ROrder = SPQ->getNodeOrdering(right); |
2535 | |
2536 | // Prefer an ordering where the lower the non-zero order number, the higher |
2537 | // the preference. |
2538 | if ((LOrder || ROrder) && LOrder != ROrder) |
2539 | return LOrder != 0 && (LOrder < ROrder || ROrder == 0); |
2540 | } |
2541 | |
2542 | // Try schedule def + use closer when Sethi-Ullman numbers are the same. |
2543 | // e.g. |
2544 | // t1 = op t2, c1 |
2545 | // t3 = op t4, c2 |
2546 | // |
2547 | // and the following instructions are both ready. |
2548 | // t2 = op c3 |
2549 | // t4 = op c4 |
2550 | // |
2551 | // Then schedule t2 = op first. |
2552 | // i.e. |
2553 | // t4 = op c4 |
2554 | // t2 = op c3 |
2555 | // t1 = op t2, c1 |
2556 | // t3 = op t4, c2 |
2557 | // |
2558 | // This creates more short live intervals. |
2559 | unsigned LDist = closestSucc(left); |
2560 | unsigned RDist = closestSucc(right); |
2561 | if (LDist != RDist) |
2562 | return LDist < RDist; |
2563 | |
2564 | // How many registers becomes live when the node is scheduled. |
2565 | unsigned LScratch = calcMaxScratches(left); |
2566 | unsigned RScratch = calcMaxScratches(right); |
2567 | if (LScratch != RScratch) |
2568 | return LScratch > RScratch; |
2569 | |
2570 | // Comparing latency against a call makes little sense unless the node |
2571 | // is register pressure-neutral. |
2572 | if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0)) |
2573 | return (left->NodeQueueId > right->NodeQueueId); |
2574 | |
2575 | // Do not compare latencies when one or both of the nodes are calls. |
2576 | if (!DisableSchedCycles && |
2577 | !(left->isCall || right->isCall)) { |
2578 | int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ); |
2579 | if (result != 0) |
2580 | return result > 0; |
2581 | } |
2582 | else { |
2583 | if (left->getHeight() != right->getHeight()) |
2584 | return left->getHeight() > right->getHeight(); |
2585 | |
2586 | if (left->getDepth() != right->getDepth()) |
2587 | return left->getDepth() < right->getDepth(); |
2588 | } |
2589 | |
2590 | assert(left->NodeQueueId && right->NodeQueueId &&(static_cast <bool> (left->NodeQueueId && right ->NodeQueueId && "NodeQueueId cannot be zero") ? void (0) : __assert_fail ("left->NodeQueueId && right->NodeQueueId && \"NodeQueueId cannot be zero\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2591, __extension__ __PRETTY_FUNCTION__)) |
2591 | "NodeQueueId cannot be zero")(static_cast <bool> (left->NodeQueueId && right ->NodeQueueId && "NodeQueueId cannot be zero") ? void (0) : __assert_fail ("left->NodeQueueId && right->NodeQueueId && \"NodeQueueId cannot be zero\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2591, __extension__ __PRETTY_FUNCTION__)); |
2592 | return (left->NodeQueueId > right->NodeQueueId); |
2593 | } |
2594 | |
2595 | // Bottom up |
2596 | bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { |
2597 | if (int res = checkSpecialNodes(left, right)) |
2598 | return res > 0; |
2599 | |
2600 | return BURRSort(left, right, SPQ); |
2601 | } |
2602 | |
2603 | // Source order, otherwise bottom up. |
2604 | bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { |
2605 | if (int res = checkSpecialNodes(left, right)) |
2606 | return res > 0; |
2607 | |
2608 | unsigned LOrder = SPQ->getNodeOrdering(left); |
2609 | unsigned ROrder = SPQ->getNodeOrdering(right); |
2610 | |
2611 | // Prefer an ordering where the lower the non-zero order number, the higher |
2612 | // the preference. |
2613 | if ((LOrder || ROrder) && LOrder != ROrder) |
2614 | return LOrder != 0 && (LOrder < ROrder || ROrder == 0); |
2615 | |
2616 | return BURRSort(left, right, SPQ); |
2617 | } |
2618 | |
2619 | // If the time between now and when the instruction will be ready can cover |
2620 | // the spill code, then avoid adding it to the ready queue. This gives long |
2621 | // stalls highest priority and allows hoisting across calls. It should also |
2622 | // speed up processing the available queue. |
2623 | bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const { |
2624 | static const unsigned ReadyDelay = 3; |
2625 | |
2626 | if (SPQ->MayReduceRegPressure(SU)) return true; |
2627 | |
2628 | if (SU->getHeight() > (CurCycle + ReadyDelay)) return false; |
2629 | |
2630 | if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay) |
2631 | != ScheduleHazardRecognizer::NoHazard) |
2632 | return false; |
2633 | |
2634 | return true; |
2635 | } |
2636 | |
2637 | // Return true if right should be scheduled with higher priority than left. |
2638 | bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { |
2639 | if (int res = checkSpecialNodes(left, right)) |
2640 | return res > 0; |
2641 | |
2642 | if (left->isCall || right->isCall) |
2643 | // No way to compute latency of calls. |
2644 | return BURRSort(left, right, SPQ); |
2645 | |
2646 | bool LHigh = SPQ->HighRegPressure(left); |
2647 | bool RHigh = SPQ->HighRegPressure(right); |
2648 | // Avoid causing spills. If register pressure is high, schedule for |
2649 | // register pressure reduction. |
2650 | if (LHigh && !RHigh) { |
2651 | 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 (false) |
2652 | << right->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " pressure SU(" << left->NodeNum << ") > SU(" << right->NodeNum << ")\n"; } } while (false); |
2653 | return true; |
2654 | } |
2655 | else if (!LHigh && RHigh) { |
2656 | 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 (false) |
2657 | << left->NodeNum << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pre-RA-sched")) { dbgs() << " pressure SU(" << right->NodeNum << ") > SU(" << left->NodeNum << ")\n"; } } while (false); |
2658 | return false; |
2659 | } |
2660 | if (!LHigh && !RHigh) { |
2661 | int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ); |
2662 | if (result != 0) |
2663 | return result > 0; |
2664 | } |
2665 | return BURRSort(left, right, SPQ); |
2666 | } |
2667 | |
2668 | // Schedule as many instructions in each cycle as possible. So don't make an |
2669 | // instruction available unless it is ready in the current cycle. |
2670 | bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const { |
2671 | if (SU->getHeight() > CurCycle) return false; |
2672 | |
2673 | if (SPQ->getHazardRec()->getHazardType(SU, 0) |
2674 | != ScheduleHazardRecognizer::NoHazard) |
2675 | return false; |
2676 | |
2677 | return true; |
2678 | } |
2679 | |
2680 | static bool canEnableCoalescing(SUnit *SU) { |
2681 | unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0; |
2682 | if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg) |
2683 | // CopyToReg should be close to its uses to facilitate coalescing and |
2684 | // avoid spilling. |
2685 | return true; |
2686 | |
2687 | if (Opc == TargetOpcode::EXTRACT_SUBREG || |
2688 | Opc == TargetOpcode::SUBREG_TO_REG || |
2689 | Opc == TargetOpcode::INSERT_SUBREG) |
2690 | // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be |
2691 | // close to their uses to facilitate coalescing. |
2692 | return true; |
2693 | |
2694 | if (SU->NumPreds == 0 && SU->NumSuccs != 0) |
2695 | // If SU does not have a register def, schedule it close to its uses |
2696 | // because it does not lengthen any live ranges. |
2697 | return true; |
2698 | |
2699 | return false; |
2700 | } |
2701 | |
2702 | // list-ilp is currently an experimental scheduler that allows various |
2703 | // heuristics to be enabled prior to the normal register reduction logic. |
2704 | bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const { |
2705 | if (int res = checkSpecialNodes(left, right)) |
2706 | return res > 0; |
2707 | |
2708 | if (left->isCall || right->isCall) |
2709 | // No way to compute latency of calls. |
2710 | return BURRSort(left, right, SPQ); |
2711 | |
2712 | unsigned LLiveUses = 0, RLiveUses = 0; |
2713 | int LPDiff = 0, RPDiff = 0; |
2714 | if (!DisableSchedRegPressure || !DisableSchedLiveUses) { |
2715 | LPDiff = SPQ->RegPressureDiff(left, LLiveUses); |
2716 | RPDiff = SPQ->RegPressureDiff(right, RLiveUses); |
2717 | } |
2718 | if (!DisableSchedRegPressure && LPDiff != RPDiff) { |
2719 | 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 (false) |
2720 | << " != 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 (false); |
2721 | return LPDiff > RPDiff; |
2722 | } |
2723 | |
2724 | if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) { |
2725 | bool LReduce = canEnableCoalescing(left); |
2726 | bool RReduce = canEnableCoalescing(right); |
2727 | if (LReduce && !RReduce) return false; |
2728 | if (RReduce && !LReduce) return true; |
2729 | } |
2730 | |
2731 | if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) { |
2732 | 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 (false) |
2733 | << " != 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 (false); |
2734 | return LLiveUses < RLiveUses; |
2735 | } |
2736 | |
2737 | if (!DisableSchedStalls) { |
2738 | bool LStall = BUHasStall(left, left->getHeight(), SPQ); |
2739 | bool RStall = BUHasStall(right, right->getHeight(), SPQ); |
2740 | if (LStall != RStall) |
2741 | return left->getHeight() > right->getHeight(); |
2742 | } |
2743 | |
2744 | if (!DisableSchedCriticalPath) { |
2745 | int spread = (int)left->getDepth() - (int)right->getDepth(); |
2746 | if (std::abs(spread) > MaxReorderWindow) { |
2747 | 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 (false) |
2748 | << 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 (false) |
2749 | << 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 (false); |
2750 | return left->getDepth() < right->getDepth(); |
2751 | } |
2752 | } |
2753 | |
2754 | if (!DisableSchedHeight && left->getHeight() != right->getHeight()) { |
2755 | int spread = (int)left->getHeight() - (int)right->getHeight(); |
2756 | if (std::abs(spread) > MaxReorderWindow) |
2757 | return left->getHeight() > right->getHeight(); |
2758 | } |
2759 | |
2760 | return BURRSort(left, right, SPQ); |
2761 | } |
2762 | |
2763 | void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) { |
2764 | SUnits = &sunits; |
2765 | // Add pseudo dependency edges for two-address nodes. |
2766 | if (!Disable2AddrHack) |
2767 | AddPseudoTwoAddrDeps(); |
2768 | // Reroute edges to nodes with multiple uses. |
2769 | if (!TracksRegPressure && !SrcOrder) |
2770 | PrescheduleNodesWithMultipleUses(); |
2771 | // Calculate node priorities. |
2772 | CalculateSethiUllmanNumbers(); |
2773 | |
2774 | // For single block loops, mark nodes that look like canonical IV increments. |
2775 | if (scheduleDAG->BB->isSuccessor(scheduleDAG->BB)) |
2776 | for (SUnit &SU : sunits) |
2777 | initVRegCycle(&SU); |
2778 | } |
2779 | |
2780 | //===----------------------------------------------------------------------===// |
2781 | // Preschedule for Register Pressure |
2782 | //===----------------------------------------------------------------------===// |
2783 | |
2784 | bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) { |
2785 | if (SU->isTwoAddress) { |
2786 | unsigned Opc = SU->getNode()->getMachineOpcode(); |
2787 | const MCInstrDesc &MCID = TII->get(Opc); |
2788 | unsigned NumRes = MCID.getNumDefs(); |
2789 | unsigned NumOps = MCID.getNumOperands() - NumRes; |
2790 | for (unsigned i = 0; i != NumOps; ++i) { |
2791 | if (MCID.getOperandConstraint(i+NumRes, MCOI::TIED_TO) != -1) { |
2792 | SDNode *DU = SU->getNode()->getOperand(i).getNode(); |
2793 | if (DU->getNodeId() != -1 && |
2794 | Op->OrigNode == &(*SUnits)[DU->getNodeId()]) |
2795 | return true; |
2796 | } |
2797 | } |
2798 | } |
2799 | return false; |
2800 | } |
2801 | |
2802 | /// canClobberReachingPhysRegUse - True if SU would clobber one of it's |
2803 | /// successor's explicit physregs whose definition can reach DepSU. |
2804 | /// i.e. DepSU should not be scheduled above SU. |
2805 | static bool canClobberReachingPhysRegUse(const SUnit *DepSU, const SUnit *SU, |
2806 | ScheduleDAGRRList *scheduleDAG, |
2807 | const TargetInstrInfo *TII, |
2808 | const TargetRegisterInfo *TRI) { |
2809 | const MCPhysReg *ImpDefs |
2810 | = TII->get(SU->getNode()->getMachineOpcode()).getImplicitDefs(); |
2811 | const uint32_t *RegMask = getNodeRegMask(SU->getNode()); |
2812 | if(!ImpDefs && !RegMask) |
2813 | return false; |
2814 | |
2815 | for (const SDep &Succ : SU->Succs) { |
2816 | SUnit *SuccSU = Succ.getSUnit(); |
2817 | for (const SDep &SuccPred : SuccSU->Preds) { |
2818 | if (!SuccPred.isAssignedRegDep()) |
2819 | continue; |
2820 | |
2821 | if (RegMask && |
2822 | MachineOperand::clobbersPhysReg(RegMask, SuccPred.getReg()) && |
2823 | scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit())) |
2824 | return true; |
2825 | |
2826 | if (ImpDefs) |
2827 | for (const MCPhysReg *ImpDef = ImpDefs; *ImpDef; ++ImpDef) |
2828 | // Return true if SU clobbers this physical register use and the |
2829 | // definition of the register reaches from DepSU. IsReachable queries |
2830 | // a topological forward sort of the DAG (following the successors). |
2831 | if (TRI->regsOverlap(*ImpDef, SuccPred.getReg()) && |
2832 | scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit())) |
2833 | return true; |
2834 | } |
2835 | } |
2836 | return false; |
2837 | } |
2838 | |
2839 | /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's |
2840 | /// physical register defs. |
2841 | static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU, |
2842 | const TargetInstrInfo *TII, |
2843 | const TargetRegisterInfo *TRI) { |
2844 | SDNode *N = SuccSU->getNode(); |
2845 | unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); |
2846 | const MCPhysReg *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs(); |
2847 | assert(ImpDefs && "Caller should check hasPhysRegDefs")(static_cast <bool> (ImpDefs && "Caller should check hasPhysRegDefs" ) ? void (0) : __assert_fail ("ImpDefs && \"Caller should check hasPhysRegDefs\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2847, __extension__ __PRETTY_FUNCTION__)); |
2848 | for (const SDNode *SUNode = SU->getNode(); SUNode; |
2849 | SUNode = SUNode->getGluedNode()) { |
2850 | if (!SUNode->isMachineOpcode()) |
2851 | continue; |
2852 | const MCPhysReg *SUImpDefs = |
2853 | TII->get(SUNode->getMachineOpcode()).getImplicitDefs(); |
2854 | const uint32_t *SURegMask = getNodeRegMask(SUNode); |
2855 | if (!SUImpDefs && !SURegMask) |
2856 | continue; |
2857 | for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) { |
2858 | MVT VT = N->getSimpleValueType(i); |
2859 | if (VT == MVT::Glue || VT == MVT::Other) |
2860 | continue; |
2861 | if (!N->hasAnyUseOfValue(i)) |
2862 | continue; |
2863 | unsigned Reg = ImpDefs[i - NumDefs]; |
2864 | if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg)) |
2865 | return true; |
2866 | if (!SUImpDefs) |
2867 | continue; |
2868 | for (;*SUImpDefs; ++SUImpDefs) { |
2869 | unsigned SUReg = *SUImpDefs; |
2870 | if (TRI->regsOverlap(Reg, SUReg)) |
2871 | return true; |
2872 | } |
2873 | } |
2874 | } |
2875 | return false; |
2876 | } |
2877 | |
2878 | /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses |
2879 | /// are not handled well by the general register pressure reduction |
2880 | /// heuristics. When presented with code like this: |
2881 | /// |
2882 | /// N |
2883 | /// / | |
2884 | /// / | |
2885 | /// U store |
2886 | /// | |
2887 | /// ... |
2888 | /// |
2889 | /// the heuristics tend to push the store up, but since the |
2890 | /// operand of the store has another use (U), this would increase |
2891 | /// the length of that other use (the U->N edge). |
2892 | /// |
2893 | /// This function transforms code like the above to route U's |
2894 | /// dependence through the store when possible, like this: |
2895 | /// |
2896 | /// N |
2897 | /// || |
2898 | /// || |
2899 | /// store |
2900 | /// | |
2901 | /// U |
2902 | /// | |
2903 | /// ... |
2904 | /// |
2905 | /// This results in the store being scheduled immediately |
2906 | /// after N, which shortens the U->N live range, reducing |
2907 | /// register pressure. |
2908 | void RegReductionPQBase::PrescheduleNodesWithMultipleUses() { |
2909 | // Visit all the nodes in topological order, working top-down. |
2910 | for (SUnit &SU : *SUnits) { |
2911 | // For now, only look at nodes with no data successors, such as stores. |
2912 | // These are especially important, due to the heuristics in |
2913 | // getNodePriority for nodes with no data successors. |
2914 | if (SU.NumSuccs != 0) |
2915 | continue; |
2916 | // For now, only look at nodes with exactly one data predecessor. |
2917 | if (SU.NumPreds != 1) |
2918 | continue; |
2919 | // Avoid prescheduling copies to virtual registers, which don't behave |
2920 | // like other nodes from the perspective of scheduling heuristics. |
2921 | if (SDNode *N = SU.getNode()) |
2922 | if (N->getOpcode() == ISD::CopyToReg && |
2923 | TargetRegisterInfo::isVirtualRegister |
2924 | (cast<RegisterSDNode>(N->getOperand(1))->getReg())) |
2925 | continue; |
2926 | |
2927 | // Locate the single data predecessor. |
2928 | SUnit *PredSU = nullptr; |
2929 | for (const SDep &Pred : SU.Preds) |
2930 | if (!Pred.isCtrl()) { |
2931 | PredSU = Pred.getSUnit(); |
2932 | break; |
2933 | } |
2934 | assert(PredSU)(static_cast <bool> (PredSU) ? void (0) : __assert_fail ("PredSU", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2934, __extension__ __PRETTY_FUNCTION__)); |
2935 | |
2936 | // Don't rewrite edges that carry physregs, because that requires additional |
2937 | // support infrastructure. |
2938 | if (PredSU->hasPhysRegDefs) |
2939 | continue; |
2940 | // Short-circuit the case where SU is PredSU's only data successor. |
2941 | if (PredSU->NumSuccs == 1) |
2942 | continue; |
2943 | // Avoid prescheduling to copies from virtual registers, which don't behave |
2944 | // like other nodes from the perspective of scheduling heuristics. |
2945 | if (SDNode *N = SU.getNode()) |
2946 | if (N->getOpcode() == ISD::CopyFromReg && |
2947 | TargetRegisterInfo::isVirtualRegister |
2948 | (cast<RegisterSDNode>(N->getOperand(1))->getReg())) |
2949 | continue; |
2950 | |
2951 | // Perform checks on the successors of PredSU. |
2952 | for (const SDep &PredSucc : PredSU->Succs) { |
2953 | SUnit *PredSuccSU = PredSucc.getSUnit(); |
2954 | if (PredSuccSU == &SU) continue; |
2955 | // If PredSU has another successor with no data successors, for |
2956 | // now don't attempt to choose either over the other. |
2957 | if (PredSuccSU->NumSuccs == 0) |
2958 | goto outer_loop_continue; |
2959 | // Don't break physical register dependencies. |
2960 | if (SU.hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs) |
2961 | if (canClobberPhysRegDefs(PredSuccSU, &SU, TII, TRI)) |
2962 | goto outer_loop_continue; |
2963 | // Don't introduce graph cycles. |
2964 | if (scheduleDAG->IsReachable(&SU, PredSuccSU)) |
2965 | goto outer_loop_continue; |
2966 | } |
2967 | |
2968 | // Ok, the transformation is safe and the heuristics suggest it is |
2969 | // profitable. Update the graph. |
2970 | 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 (false) |
2971 | << " 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 (false) |
2972 | << " 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 (false); |
2973 | for (unsigned i = 0; i != PredSU->Succs.size(); ++i) { |
2974 | SDep Edge = PredSU->Succs[i]; |
2975 | assert(!Edge.isAssignedRegDep())(static_cast <bool> (!Edge.isAssignedRegDep()) ? void ( 0) : __assert_fail ("!Edge.isAssignedRegDep()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp" , 2975, __extension__ __PRETTY_FUNCTION__)); |
2976 | SUnit *SuccSU = Edge.getSUnit(); |
2977 | if (SuccSU != &SU) { |
2978 | Edge.setSUnit(PredSU); |
2979 | scheduleDAG->RemovePred(SuccSU, Edge); |
2980 | scheduleDAG->AddPred(&SU, Edge); |
2981 | Edge.setSUnit(&SU); |
2982 | scheduleDAG->AddPred(SuccSU, Edge); |
2983 | --i; |
2984 | } |
2985 | } |
2986 | outer_loop_continue:; |
2987 | } |
2988 | } |
2989 | |
2990 | /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses |
2991 | /// it as a def&use operand. Add a pseudo control edge from it to the other |
2992 | /// node (if it won't create a cycle) so the two-address one will be scheduled |
2993 | /// first (lower in the schedule). If both nodes are two-address, favor the |
2994 | /// one that has a CopyToReg use (more likely to be a loop induction update). |
2995 | /// If both are two-address, but one is commutable while the other is not |
2996 | /// commutable, favor the one that's not commutable. |
2997 | void RegReductionPQBase::AddPseudoTwoAddrDeps() { |
2998 | for (SUnit &SU : *SUnits) { |
2999 | if (!SU.isTwoAddress) |
3000 | continue; |
3001 | |
3002 | SDNode *Node = SU.getNode(); |
3003 | if (!Node || !Node->isMachineOpcode() || SU.getNode()->getGluedNode()) |
3004 | continue; |
3005 | |
3006 | bool isLiveOut = hasOnlyLiveOutUses(&SU); |
3007 | unsigned Opc = Node->getMachineOpcode(); |
3008 | const MCInstrDesc &MCID = TII->get(Opc); |
3009 | unsigned NumRes = MCID.getNumDefs(); |
3010 | unsigned NumOps = MCID.getNumOperands() - NumRes; |
3011 | for (unsigned j = 0; j != NumOps; ++j) { |
3012 | if (MCID.getOperandConstraint(j+NumRes, MCOI::TIED_TO) == -1) |
3013 | continue; |
3014 | SDNode *DU = SU.getNode()->getOperand(j).getNode(); |
3015 | if (DU->getNodeId() == -1) |
3016 | continue; |
3017 | const SUnit *DUSU = &(*SUnits)[DU->getNodeId()]; |
3018 | if (!DUSU) |
3019 | continue; |
3020 | for (const SDep &Succ : DUSU->Succs) { |
3021 | if (Succ.isCtrl()) |
3022 | continue; |
3023 | SUnit *SuccSU = Succ.getSUnit(); |
3024 | if (SuccSU == &SU) |
3025 | continue; |
3026 | // Be conservative. Ignore if nodes aren't at roughly the same |
3027 | // depth and height. |
3028 | if (SuccSU->getHeight() < SU.getHeight() && |
3029 | (SU.getHeight() - SuccSU->getHeight()) > 1) |
3030 | continue; |
3031 | // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge |
3032 | // constrains whatever is using the copy, instead of the copy |
3033 | // itself. In the case that the copy is coalesced, this |
3034 | // preserves the intent of the pseudo two-address heurietics. |
3035 | while (SuccSU->Succs.size() == 1 && |
3036 | SuccSU->getNode()->isMachineOpcode() && |
3037 | SuccSU->getNode()->getMachineOpcode() == |
3038 | TargetOpcode::COPY_TO_REGCLASS) |
3039 | SuccSU = SuccSU->Succs.front().getSUnit(); |
3040 | // Don't constrain non-instruction nodes. |
3041 | if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode()) |
3042 | continue; |
3043 | // Don't constrain nodes with physical register defs if the |
3044 | // predecessor can clobber them. |
3045 | if (SuccSU->hasPhysRegDefs && SU.hasPhysRegClobbers) { |
3046 | if (canClobberPhysRegDefs(SuccSU, &SU, TII, TRI)) |
3047 | continue; |
3048 | } |
3049 | // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG; |
3050 | // these may be coalesced away. We want them close to their uses. |
3051 | unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode(); |
3052 | if (SuccOpc == TargetOpcode::EXTRACT_SUBREG || |
3053 | SuccOpc == TargetOpcode::INSERT_SUBREG || |
3054 | SuccOpc == TargetOpcode::SUBREG_TO_REG) |
3055 | continue; |
3056 | if (!canClobberReachingPhysRegUse(SuccSU, &SU, scheduleDAG, TII, TRI) && |
3057 | (!canClobber(SuccSU, DUSU) || |
3058 | (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) || |
3059 | (!SU.isCommutable && SuccSU->isCommutable)) && |
3060 | !scheduleDAG->IsReachable(SuccSU, &SU)) { |
3061 | 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 (false) |
3062 | << 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 (false); |
3063 | scheduleDAG->AddPred(&SU, SDep(SuccSU, SDep::Artificial)); |
3064 | } |
3065 | } |
3066 | } |
3067 | } |
3068 | } |
3069 | |
3070 | //===----------------------------------------------------------------------===// |
3071 | // Public Constructor Functions |
3072 | //===----------------------------------------------------------------------===// |
3073 | |
3074 | ScheduleDAGSDNodes * |
3075 | llvm::createBURRListDAGScheduler(SelectionDAGISel *IS, |
3076 | CodeGenOpt::Level OptLevel) { |
3077 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); |
3078 | const TargetInstrInfo *TII = STI.getInstrInfo(); |
3079 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); |
3080 | |
3081 | BURegReductionPriorityQueue *PQ = |
3082 | new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, nullptr); |
3083 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel); |
3084 | PQ->setScheduleDAG(SD); |
3085 | return SD; |
3086 | } |
3087 | |
3088 | ScheduleDAGSDNodes * |
3089 | llvm::createSourceListDAGScheduler(SelectionDAGISel *IS, |
3090 | CodeGenOpt::Level OptLevel) { |
3091 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); |
3092 | const TargetInstrInfo *TII = STI.getInstrInfo(); |
3093 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); |
3094 | |
3095 | SrcRegReductionPriorityQueue *PQ = |
3096 | new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, nullptr); |
3097 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel); |
3098 | PQ->setScheduleDAG(SD); |
3099 | return SD; |
3100 | } |
3101 | |
3102 | ScheduleDAGSDNodes * |
3103 | llvm::createHybridListDAGScheduler(SelectionDAGISel *IS, |
3104 | CodeGenOpt::Level OptLevel) { |
3105 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); |
3106 | const TargetInstrInfo *TII = STI.getInstrInfo(); |
3107 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); |
3108 | const TargetLowering *TLI = IS->TLI; |
3109 | |
3110 | HybridBURRPriorityQueue *PQ = |
3111 | new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI); |
3112 | |
3113 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel); |
3114 | PQ->setScheduleDAG(SD); |
3115 | return SD; |
3116 | } |
3117 | |
3118 | ScheduleDAGSDNodes * |
3119 | llvm::createILPListDAGScheduler(SelectionDAGISel *IS, |
3120 | CodeGenOpt::Level OptLevel) { |
3121 | const TargetSubtargetInfo &STI = IS->MF->getSubtarget(); |
3122 | const TargetInstrInfo *TII = STI.getInstrInfo(); |
3123 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); |
3124 | const TargetLowering *TLI = IS->TLI; |
3125 | |
3126 | ILPBURRPriorityQueue *PQ = |
3127 | new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI); |
3128 | ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel); |
3129 | PQ->setScheduleDAG(SD); |
3130 | return SD; |
3131 | } |