1 //===- HexagonPacketizer.cpp - VLIW packetizer ----------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements a simple VLIW packetizer using DFA. The packetizer works on
10 // machine basic blocks. For each instruction I in BB, the packetizer consults
11 // the DFA to see if machine resources are available to execute I. If so, the
12 // packetizer checks if I depends on any instruction J in the current packet.
13 // If no dependency is found, I is added to current packet and machine resource
14 // is marked as taken. If any dependency is found, a target API call is made to
15 // prune the dependence.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "HexagonVLIWPacketizer.h"
20 #include "Hexagon.h"
21 #include "HexagonInstrInfo.h"
22 #include "HexagonRegisterInfo.h"
23 #include "HexagonSubtarget.h"
24 #include "llvm/ADT/BitVector.h"
25 #include "llvm/ADT/DenseSet.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/Analysis/AliasAnalysis.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
31 #include "llvm/CodeGen/MachineDominators.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineInstr.h"
36 #include "llvm/CodeGen/MachineInstrBundle.h"
37 #include "llvm/CodeGen/MachineLoopInfo.h"
38 #include "llvm/CodeGen/MachineOperand.h"
39 #include "llvm/CodeGen/ScheduleDAG.h"
40 #include "llvm/CodeGen/TargetRegisterInfo.h"
41 #include "llvm/CodeGen/TargetSubtargetInfo.h"
42 #include "llvm/IR/DebugLoc.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/MC/MCInstrDesc.h"
45 #include "llvm/Pass.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include <cassert>
51 #include <cstdint>
52 #include <iterator>
53 
54 using namespace llvm;
55 
56 #define DEBUG_TYPE "packets"
57 
58 static cl::opt<bool> DisablePacketizer("disable-packetizer", cl::Hidden,
59   cl::ZeroOrMore, cl::init(false),
60   cl::desc("Disable Hexagon packetizer pass"));
61 
62 static cl::opt<bool> Slot1Store("slot1-store-slot0-load", cl::Hidden,
63                                 cl::ZeroOrMore, cl::init(true),
64                                 cl::desc("Allow slot1 store and slot0 load"));
65 
66 static cl::opt<bool> PacketizeVolatiles("hexagon-packetize-volatiles",
67   cl::ZeroOrMore, cl::Hidden, cl::init(true),
68   cl::desc("Allow non-solo packetization of volatile memory references"));
69 
70 static cl::opt<bool> EnableGenAllInsnClass("enable-gen-insn", cl::init(false),
71   cl::Hidden, cl::ZeroOrMore, cl::desc("Generate all instruction with TC"));
72 
73 static cl::opt<bool> DisableVecDblNVStores("disable-vecdbl-nv-stores",
74   cl::init(false), cl::Hidden, cl::ZeroOrMore,
75   cl::desc("Disable vector double new-value-stores"));
76 
77 extern cl::opt<bool> ScheduleInlineAsm;
78 
79 namespace llvm {
80 
81 FunctionPass *createHexagonPacketizer(bool Minimal);
82 void initializeHexagonPacketizerPass(PassRegistry&);
83 
84 } // end namespace llvm
85 
86 namespace {
87 
88   class HexagonPacketizer : public MachineFunctionPass {
89   public:
90     static char ID;
91 
92     HexagonPacketizer(bool Min = false)
93       : MachineFunctionPass(ID), Minimal(Min) {}
94 
95     void getAnalysisUsage(AnalysisUsage &AU) const override {
96       AU.setPreservesCFG();
97       AU.addRequired<AAResultsWrapperPass>();
98       AU.addRequired<MachineBranchProbabilityInfo>();
99       AU.addRequired<MachineDominatorTree>();
100       AU.addRequired<MachineLoopInfo>();
101       AU.addPreserved<MachineDominatorTree>();
102       AU.addPreserved<MachineLoopInfo>();
103       MachineFunctionPass::getAnalysisUsage(AU);
104     }
105 
106     StringRef getPassName() const override { return "Hexagon Packetizer"; }
107     bool runOnMachineFunction(MachineFunction &Fn) override;
108 
109     MachineFunctionProperties getRequiredProperties() const override {
110       return MachineFunctionProperties().set(
111           MachineFunctionProperties::Property::NoVRegs);
112     }
113 
114   private:
115     const HexagonInstrInfo *HII = nullptr;
116     const HexagonRegisterInfo *HRI = nullptr;
117     const bool Minimal = false;
118   };
119 
120 } // end anonymous namespace
121 
122 char HexagonPacketizer::ID = 0;
123 
124 INITIALIZE_PASS_BEGIN(HexagonPacketizer, "hexagon-packetizer",
125                       "Hexagon Packetizer", false, false)
126 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
127 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
128 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
129 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
130 INITIALIZE_PASS_END(HexagonPacketizer, "hexagon-packetizer",
131                     "Hexagon Packetizer", false, false)
132 
133 HexagonPacketizerList::HexagonPacketizerList(MachineFunction &MF,
134       MachineLoopInfo &MLI, AAResults *AA,
135       const MachineBranchProbabilityInfo *MBPI, bool Minimal)
136     : VLIWPacketizerList(MF, MLI, AA), MBPI(MBPI), MLI(&MLI),
137       Minimal(Minimal) {
138   HII = MF.getSubtarget<HexagonSubtarget>().getInstrInfo();
139   HRI = MF.getSubtarget<HexagonSubtarget>().getRegisterInfo();
140 
141   addMutation(std::make_unique<HexagonSubtarget::UsrOverflowMutation>());
142   addMutation(std::make_unique<HexagonSubtarget::HVXMemLatencyMutation>());
143   addMutation(std::make_unique<HexagonSubtarget::BankConflictMutation>());
144 }
145 
146 // Check if FirstI modifies a register that SecondI reads.
147 static bool hasWriteToReadDep(const MachineInstr &FirstI,
148                               const MachineInstr &SecondI,
149                               const TargetRegisterInfo *TRI) {
150   for (auto &MO : FirstI.operands()) {
151     if (!MO.isReg() || !MO.isDef())
152       continue;
153     Register R = MO.getReg();
154     if (SecondI.readsRegister(R, TRI))
155       return true;
156   }
157   return false;
158 }
159 
160 
161 static MachineBasicBlock::iterator moveInstrOut(MachineInstr &MI,
162       MachineBasicBlock::iterator BundleIt, bool Before) {
163   MachineBasicBlock::instr_iterator InsertPt;
164   if (Before)
165     InsertPt = BundleIt.getInstrIterator();
166   else
167     InsertPt = std::next(BundleIt).getInstrIterator();
168 
169   MachineBasicBlock &B = *MI.getParent();
170   // The instruction should at least be bundled with the preceding instruction
171   // (there will always be one, i.e. BUNDLE, if nothing else).
172   assert(MI.isBundledWithPred());
173   if (MI.isBundledWithSucc()) {
174     MI.clearFlag(MachineInstr::BundledSucc);
175     MI.clearFlag(MachineInstr::BundledPred);
176   } else {
177     // If it's not bundled with the successor (i.e. it is the last one
178     // in the bundle), then we can simply unbundle it from the predecessor,
179     // which will take care of updating the predecessor's flag.
180     MI.unbundleFromPred();
181   }
182   B.splice(InsertPt, &B, MI.getIterator());
183 
184   // Get the size of the bundle without asserting.
185   MachineBasicBlock::const_instr_iterator I = BundleIt.getInstrIterator();
186   MachineBasicBlock::const_instr_iterator E = B.instr_end();
187   unsigned Size = 0;
188   for (++I; I != E && I->isBundledWithPred(); ++I)
189     ++Size;
190 
191   // If there are still two or more instructions, then there is nothing
192   // else to be done.
193   if (Size > 1)
194     return BundleIt;
195 
196   // Otherwise, extract the single instruction out and delete the bundle.
197   MachineBasicBlock::iterator NextIt = std::next(BundleIt);
198   MachineInstr &SingleI = *BundleIt->getNextNode();
199   SingleI.unbundleFromPred();
200   assert(!SingleI.isBundledWithSucc());
201   BundleIt->eraseFromParent();
202   return NextIt;
203 }
204 
205 bool HexagonPacketizer::runOnMachineFunction(MachineFunction &MF) {
206   auto &HST = MF.getSubtarget<HexagonSubtarget>();
207   HII = HST.getInstrInfo();
208   HRI = HST.getRegisterInfo();
209   auto &MLI = getAnalysis<MachineLoopInfo>();
210   auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
211   auto *MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
212 
213   if (EnableGenAllInsnClass)
214     HII->genAllInsnTimingClasses(MF);
215 
216   // Instantiate the packetizer.
217   bool MinOnly = Minimal || DisablePacketizer || !HST.usePackets() ||
218                  skipFunction(MF.getFunction());
219   HexagonPacketizerList Packetizer(MF, MLI, AA, MBPI, MinOnly);
220 
221   // DFA state table should not be empty.
222   assert(Packetizer.getResourceTracker() && "Empty DFA table!");
223 
224   // Loop over all basic blocks and remove KILL pseudo-instructions
225   // These instructions confuse the dependence analysis. Consider:
226   // D0 = ...   (Insn 0)
227   // R0 = KILL R0, D0 (Insn 1)
228   // R0 = ... (Insn 2)
229   // Here, Insn 1 will result in the dependence graph not emitting an output
230   // dependence between Insn 0 and Insn 2. This can lead to incorrect
231   // packetization
232   for (MachineBasicBlock &MB : MF) {
233     auto End = MB.end();
234     auto MI = MB.begin();
235     while (MI != End) {
236       auto NextI = std::next(MI);
237       if (MI->isKill()) {
238         MB.erase(MI);
239         End = MB.end();
240       }
241       MI = NextI;
242     }
243   }
244 
245   // TinyCore with Duplexes: Translate to big-instructions.
246   if (HST.isTinyCoreWithDuplex())
247     HII->translateInstrsForDup(MF, true);
248 
249   // Loop over all of the basic blocks.
250   for (auto &MB : MF) {
251     auto Begin = MB.begin(), End = MB.end();
252     while (Begin != End) {
253       // Find the first non-boundary starting from the end of the last
254       // scheduling region.
255       MachineBasicBlock::iterator RB = Begin;
256       while (RB != End && HII->isSchedulingBoundary(*RB, &MB, MF))
257         ++RB;
258       // Find the first boundary starting from the beginning of the new
259       // region.
260       MachineBasicBlock::iterator RE = RB;
261       while (RE != End && !HII->isSchedulingBoundary(*RE, &MB, MF))
262         ++RE;
263       // Add the scheduling boundary if it's not block end.
264       if (RE != End)
265         ++RE;
266       // If RB == End, then RE == End.
267       if (RB != End)
268         Packetizer.PacketizeMIs(&MB, RB, RE);
269 
270       Begin = RE;
271     }
272   }
273 
274   // TinyCore with Duplexes: Translate to tiny-instructions.
275   if (HST.isTinyCoreWithDuplex())
276     HII->translateInstrsForDup(MF, false);
277 
278   Packetizer.unpacketizeSoloInstrs(MF);
279   return true;
280 }
281 
282 // Reserve resources for a constant extender. Trigger an assertion if the
283 // reservation fails.
284 void HexagonPacketizerList::reserveResourcesForConstExt() {
285   if (!tryAllocateResourcesForConstExt(true))
286     llvm_unreachable("Resources not available");
287 }
288 
289 bool HexagonPacketizerList::canReserveResourcesForConstExt() {
290   return tryAllocateResourcesForConstExt(false);
291 }
292 
293 // Allocate resources (i.e. 4 bytes) for constant extender. If succeeded,
294 // return true, otherwise, return false.
295 bool HexagonPacketizerList::tryAllocateResourcesForConstExt(bool Reserve) {
296   auto *ExtMI = MF.CreateMachineInstr(HII->get(Hexagon::A4_ext), DebugLoc());
297   bool Avail = ResourceTracker->canReserveResources(*ExtMI);
298   if (Reserve && Avail)
299     ResourceTracker->reserveResources(*ExtMI);
300   MF.DeleteMachineInstr(ExtMI);
301   return Avail;
302 }
303 
304 bool HexagonPacketizerList::isCallDependent(const MachineInstr &MI,
305       SDep::Kind DepType, unsigned DepReg) {
306   // Check for LR dependence.
307   if (DepReg == HRI->getRARegister())
308     return true;
309 
310   if (HII->isDeallocRet(MI))
311     if (DepReg == HRI->getFrameRegister() || DepReg == HRI->getStackRegister())
312       return true;
313 
314   // Call-like instructions can be packetized with preceding instructions
315   // that define registers implicitly used or modified by the call. Explicit
316   // uses are still prohibited, as in the case of indirect calls:
317   //   r0 = ...
318   //   J2_jumpr r0
319   if (DepType == SDep::Data) {
320     for (const MachineOperand &MO : MI.operands())
321       if (MO.isReg() && MO.getReg() == DepReg && !MO.isImplicit())
322         return true;
323   }
324 
325   return false;
326 }
327 
328 static bool isRegDependence(const SDep::Kind DepType) {
329   return DepType == SDep::Data || DepType == SDep::Anti ||
330          DepType == SDep::Output;
331 }
332 
333 static bool isDirectJump(const MachineInstr &MI) {
334   return MI.getOpcode() == Hexagon::J2_jump;
335 }
336 
337 static bool isSchedBarrier(const MachineInstr &MI) {
338   switch (MI.getOpcode()) {
339   case Hexagon::Y2_barrier:
340     return true;
341   }
342   return false;
343 }
344 
345 static bool isControlFlow(const MachineInstr &MI) {
346   return MI.getDesc().isTerminator() || MI.getDesc().isCall();
347 }
348 
349 /// Returns true if the instruction modifies a callee-saved register.
350 static bool doesModifyCalleeSavedReg(const MachineInstr &MI,
351                                      const TargetRegisterInfo *TRI) {
352   const MachineFunction &MF = *MI.getParent()->getParent();
353   for (auto *CSR = TRI->getCalleeSavedRegs(&MF); CSR && *CSR; ++CSR)
354     if (MI.modifiesRegister(*CSR, TRI))
355       return true;
356   return false;
357 }
358 
359 // Returns true if an instruction can be promoted to .new predicate or
360 // new-value store.
361 bool HexagonPacketizerList::isNewifiable(const MachineInstr &MI,
362       const TargetRegisterClass *NewRC) {
363   // Vector stores can be predicated, and can be new-value stores, but
364   // they cannot be predicated on a .new predicate value.
365   if (NewRC == &Hexagon::PredRegsRegClass) {
366     if (HII->isHVXVec(MI) && MI.mayStore())
367       return false;
368     return HII->isPredicated(MI) && HII->getDotNewPredOp(MI, nullptr) > 0;
369   }
370   // If the class is not PredRegs, it could only apply to new-value stores.
371   return HII->mayBeNewStore(MI);
372 }
373 
374 // Promote an instructiont to its .cur form.
375 // At this time, we have already made a call to canPromoteToDotCur and made
376 // sure that it can *indeed* be promoted.
377 bool HexagonPacketizerList::promoteToDotCur(MachineInstr &MI,
378       SDep::Kind DepType, MachineBasicBlock::iterator &MII,
379       const TargetRegisterClass* RC) {
380   assert(DepType == SDep::Data);
381   int CurOpcode = HII->getDotCurOp(MI);
382   MI.setDesc(HII->get(CurOpcode));
383   return true;
384 }
385 
386 void HexagonPacketizerList::cleanUpDotCur() {
387   MachineInstr *MI = nullptr;
388   for (auto BI : CurrentPacketMIs) {
389     LLVM_DEBUG(dbgs() << "Cleanup packet has "; BI->dump(););
390     if (HII->isDotCurInst(*BI)) {
391       MI = BI;
392       continue;
393     }
394     if (MI) {
395       for (auto &MO : BI->operands())
396         if (MO.isReg() && MO.getReg() == MI->getOperand(0).getReg())
397           return;
398     }
399   }
400   if (!MI)
401     return;
402   // We did not find a use of the CUR, so de-cur it.
403   MI->setDesc(HII->get(HII->getNonDotCurOp(*MI)));
404   LLVM_DEBUG(dbgs() << "Demoted CUR "; MI->dump(););
405 }
406 
407 // Check to see if an instruction can be dot cur.
408 bool HexagonPacketizerList::canPromoteToDotCur(const MachineInstr &MI,
409       const SUnit *PacketSU, unsigned DepReg, MachineBasicBlock::iterator &MII,
410       const TargetRegisterClass *RC) {
411   if (!HII->isHVXVec(MI))
412     return false;
413   if (!HII->isHVXVec(*MII))
414     return false;
415 
416   // Already a dot new instruction.
417   if (HII->isDotCurInst(MI) && !HII->mayBeCurLoad(MI))
418     return false;
419 
420   if (!HII->mayBeCurLoad(MI))
421     return false;
422 
423   // The "cur value" cannot come from inline asm.
424   if (PacketSU->getInstr()->isInlineAsm())
425     return false;
426 
427   // Make sure candidate instruction uses cur.
428   LLVM_DEBUG(dbgs() << "Can we DOT Cur Vector MI\n"; MI.dump();
429              dbgs() << "in packet\n";);
430   MachineInstr &MJ = *MII;
431   LLVM_DEBUG({
432     dbgs() << "Checking CUR against ";
433     MJ.dump();
434   });
435   Register DestReg = MI.getOperand(0).getReg();
436   bool FoundMatch = false;
437   for (auto &MO : MJ.operands())
438     if (MO.isReg() && MO.getReg() == DestReg)
439       FoundMatch = true;
440   if (!FoundMatch)
441     return false;
442 
443   // Check for existing uses of a vector register within the packet which
444   // would be affected by converting a vector load into .cur formt.
445   for (auto BI : CurrentPacketMIs) {
446     LLVM_DEBUG(dbgs() << "packet has "; BI->dump(););
447     if (BI->readsRegister(DepReg, MF.getSubtarget().getRegisterInfo()))
448       return false;
449   }
450 
451   LLVM_DEBUG(dbgs() << "Can Dot CUR MI\n"; MI.dump(););
452   // We can convert the opcode into a .cur.
453   return true;
454 }
455 
456 // Promote an instruction to its .new form. At this time, we have already
457 // made a call to canPromoteToDotNew and made sure that it can *indeed* be
458 // promoted.
459 bool HexagonPacketizerList::promoteToDotNew(MachineInstr &MI,
460       SDep::Kind DepType, MachineBasicBlock::iterator &MII,
461       const TargetRegisterClass* RC) {
462   assert(DepType == SDep::Data);
463   int NewOpcode;
464   if (RC == &Hexagon::PredRegsRegClass)
465     NewOpcode = HII->getDotNewPredOp(MI, MBPI);
466   else
467     NewOpcode = HII->getDotNewOp(MI);
468   MI.setDesc(HII->get(NewOpcode));
469   return true;
470 }
471 
472 bool HexagonPacketizerList::demoteToDotOld(MachineInstr &MI) {
473   int NewOpcode = HII->getDotOldOp(MI);
474   MI.setDesc(HII->get(NewOpcode));
475   return true;
476 }
477 
478 bool HexagonPacketizerList::useCallersSP(MachineInstr &MI) {
479   unsigned Opc = MI.getOpcode();
480   switch (Opc) {
481     case Hexagon::S2_storerd_io:
482     case Hexagon::S2_storeri_io:
483     case Hexagon::S2_storerh_io:
484     case Hexagon::S2_storerb_io:
485       break;
486     default:
487       llvm_unreachable("Unexpected instruction");
488   }
489   unsigned FrameSize = MF.getFrameInfo().getStackSize();
490   MachineOperand &Off = MI.getOperand(1);
491   int64_t NewOff = Off.getImm() - (FrameSize + HEXAGON_LRFP_SIZE);
492   if (HII->isValidOffset(Opc, NewOff, HRI)) {
493     Off.setImm(NewOff);
494     return true;
495   }
496   return false;
497 }
498 
499 void HexagonPacketizerList::useCalleesSP(MachineInstr &MI) {
500   unsigned Opc = MI.getOpcode();
501   switch (Opc) {
502     case Hexagon::S2_storerd_io:
503     case Hexagon::S2_storeri_io:
504     case Hexagon::S2_storerh_io:
505     case Hexagon::S2_storerb_io:
506       break;
507     default:
508       llvm_unreachable("Unexpected instruction");
509   }
510   unsigned FrameSize = MF.getFrameInfo().getStackSize();
511   MachineOperand &Off = MI.getOperand(1);
512   Off.setImm(Off.getImm() + FrameSize + HEXAGON_LRFP_SIZE);
513 }
514 
515 /// Return true if we can update the offset in MI so that MI and MJ
516 /// can be packetized together.
517 bool HexagonPacketizerList::updateOffset(SUnit *SUI, SUnit *SUJ) {
518   assert(SUI->getInstr() && SUJ->getInstr());
519   MachineInstr &MI = *SUI->getInstr();
520   MachineInstr &MJ = *SUJ->getInstr();
521 
522   unsigned BPI, OPI;
523   if (!HII->getBaseAndOffsetPosition(MI, BPI, OPI))
524     return false;
525   unsigned BPJ, OPJ;
526   if (!HII->getBaseAndOffsetPosition(MJ, BPJ, OPJ))
527     return false;
528   Register Reg = MI.getOperand(BPI).getReg();
529   if (Reg != MJ.getOperand(BPJ).getReg())
530     return false;
531   // Make sure that the dependences do not restrict adding MI to the packet.
532   // That is, ignore anti dependences, and make sure the only data dependence
533   // involves the specific register.
534   for (const auto &PI : SUI->Preds)
535     if (PI.getKind() != SDep::Anti &&
536         (PI.getKind() != SDep::Data || PI.getReg() != Reg))
537       return false;
538   int Incr;
539   if (!HII->getIncrementValue(MJ, Incr))
540     return false;
541 
542   int64_t Offset = MI.getOperand(OPI).getImm();
543   if (!HII->isValidOffset(MI.getOpcode(), Offset+Incr, HRI))
544     return false;
545 
546   MI.getOperand(OPI).setImm(Offset + Incr);
547   ChangedOffset = Offset;
548   return true;
549 }
550 
551 /// Undo the changed offset. This is needed if the instruction cannot be
552 /// added to the current packet due to a different instruction.
553 void HexagonPacketizerList::undoChangedOffset(MachineInstr &MI) {
554   unsigned BP, OP;
555   if (!HII->getBaseAndOffsetPosition(MI, BP, OP))
556     llvm_unreachable("Unable to find base and offset operands.");
557   MI.getOperand(OP).setImm(ChangedOffset);
558 }
559 
560 enum PredicateKind {
561   PK_False,
562   PK_True,
563   PK_Unknown
564 };
565 
566 /// Returns true if an instruction is predicated on p0 and false if it's
567 /// predicated on !p0.
568 static PredicateKind getPredicateSense(const MachineInstr &MI,
569                                        const HexagonInstrInfo *HII) {
570   if (!HII->isPredicated(MI))
571     return PK_Unknown;
572   if (HII->isPredicatedTrue(MI))
573     return PK_True;
574   return PK_False;
575 }
576 
577 static const MachineOperand &getPostIncrementOperand(const MachineInstr &MI,
578       const HexagonInstrInfo *HII) {
579   assert(HII->isPostIncrement(MI) && "Not a post increment operation.");
580 #ifndef NDEBUG
581   // Post Increment means duplicates. Use dense map to find duplicates in the
582   // list. Caution: Densemap initializes with the minimum of 64 buckets,
583   // whereas there are at most 5 operands in the post increment.
584   DenseSet<unsigned> DefRegsSet;
585   for (auto &MO : MI.operands())
586     if (MO.isReg() && MO.isDef())
587       DefRegsSet.insert(MO.getReg());
588 
589   for (auto &MO : MI.operands())
590     if (MO.isReg() && MO.isUse() && DefRegsSet.count(MO.getReg()))
591       return MO;
592 #else
593   if (MI.mayLoad()) {
594     const MachineOperand &Op1 = MI.getOperand(1);
595     // The 2nd operand is always the post increment operand in load.
596     assert(Op1.isReg() && "Post increment operand has be to a register.");
597     return Op1;
598   }
599   if (MI.getDesc().mayStore()) {
600     const MachineOperand &Op0 = MI.getOperand(0);
601     // The 1st operand is always the post increment operand in store.
602     assert(Op0.isReg() && "Post increment operand has be to a register.");
603     return Op0;
604   }
605 #endif
606   // we should never come here.
607   llvm_unreachable("mayLoad or mayStore not set for Post Increment operation");
608 }
609 
610 // Get the value being stored.
611 static const MachineOperand& getStoreValueOperand(const MachineInstr &MI) {
612   // value being stored is always the last operand.
613   return MI.getOperand(MI.getNumOperands()-1);
614 }
615 
616 static bool isLoadAbsSet(const MachineInstr &MI) {
617   unsigned Opc = MI.getOpcode();
618   switch (Opc) {
619     case Hexagon::L4_loadrd_ap:
620     case Hexagon::L4_loadrb_ap:
621     case Hexagon::L4_loadrh_ap:
622     case Hexagon::L4_loadrub_ap:
623     case Hexagon::L4_loadruh_ap:
624     case Hexagon::L4_loadri_ap:
625       return true;
626   }
627   return false;
628 }
629 
630 static const MachineOperand &getAbsSetOperand(const MachineInstr &MI) {
631   assert(isLoadAbsSet(MI));
632   return MI.getOperand(1);
633 }
634 
635 // Can be new value store?
636 // Following restrictions are to be respected in convert a store into
637 // a new value store.
638 // 1. If an instruction uses auto-increment, its address register cannot
639 //    be a new-value register. Arch Spec 5.4.2.1
640 // 2. If an instruction uses absolute-set addressing mode, its address
641 //    register cannot be a new-value register. Arch Spec 5.4.2.1.
642 // 3. If an instruction produces a 64-bit result, its registers cannot be used
643 //    as new-value registers. Arch Spec 5.4.2.2.
644 // 4. If the instruction that sets the new-value register is conditional, then
645 //    the instruction that uses the new-value register must also be conditional,
646 //    and both must always have their predicates evaluate identically.
647 //    Arch Spec 5.4.2.3.
648 // 5. There is an implied restriction that a packet cannot have another store,
649 //    if there is a new value store in the packet. Corollary: if there is
650 //    already a store in a packet, there can not be a new value store.
651 //    Arch Spec: 3.4.4.2
652 bool HexagonPacketizerList::canPromoteToNewValueStore(const MachineInstr &MI,
653       const MachineInstr &PacketMI, unsigned DepReg) {
654   // Make sure we are looking at the store, that can be promoted.
655   if (!HII->mayBeNewStore(MI))
656     return false;
657 
658   // Make sure there is dependency and can be new value'd.
659   const MachineOperand &Val = getStoreValueOperand(MI);
660   if (Val.isReg() && Val.getReg() != DepReg)
661     return false;
662 
663   const MCInstrDesc& MCID = PacketMI.getDesc();
664 
665   // First operand is always the result.
666   const TargetRegisterClass *PacketRC = HII->getRegClass(MCID, 0, HRI, MF);
667   // Double regs can not feed into new value store: PRM section: 5.4.2.2.
668   if (PacketRC == &Hexagon::DoubleRegsRegClass)
669     return false;
670 
671   // New-value stores are of class NV (slot 0), dual stores require class ST
672   // in slot 0 (PRM 5.5).
673   for (auto I : CurrentPacketMIs) {
674     SUnit *PacketSU = MIToSUnit.find(I)->second;
675     if (PacketSU->getInstr()->mayStore())
676       return false;
677   }
678 
679   // Make sure it's NOT the post increment register that we are going to
680   // new value.
681   if (HII->isPostIncrement(MI) &&
682       getPostIncrementOperand(MI, HII).getReg() == DepReg) {
683     return false;
684   }
685 
686   if (HII->isPostIncrement(PacketMI) && PacketMI.mayLoad() &&
687       getPostIncrementOperand(PacketMI, HII).getReg() == DepReg) {
688     // If source is post_inc, or absolute-set addressing, it can not feed
689     // into new value store
690     //   r3 = memw(r2++#4)
691     //   memw(r30 + #-1404) = r2.new -> can not be new value store
692     // arch spec section: 5.4.2.1.
693     return false;
694   }
695 
696   if (isLoadAbsSet(PacketMI) && getAbsSetOperand(PacketMI).getReg() == DepReg)
697     return false;
698 
699   // If the source that feeds the store is predicated, new value store must
700   // also be predicated.
701   if (HII->isPredicated(PacketMI)) {
702     if (!HII->isPredicated(MI))
703       return false;
704 
705     // Check to make sure that they both will have their predicates
706     // evaluate identically.
707     unsigned predRegNumSrc = 0;
708     unsigned predRegNumDst = 0;
709     const TargetRegisterClass* predRegClass = nullptr;
710 
711     // Get predicate register used in the source instruction.
712     for (auto &MO : PacketMI.operands()) {
713       if (!MO.isReg())
714         continue;
715       predRegNumSrc = MO.getReg();
716       predRegClass = HRI->getMinimalPhysRegClass(predRegNumSrc);
717       if (predRegClass == &Hexagon::PredRegsRegClass)
718         break;
719     }
720     assert((predRegClass == &Hexagon::PredRegsRegClass) &&
721         "predicate register not found in a predicated PacketMI instruction");
722 
723     // Get predicate register used in new-value store instruction.
724     for (auto &MO : MI.operands()) {
725       if (!MO.isReg())
726         continue;
727       predRegNumDst = MO.getReg();
728       predRegClass = HRI->getMinimalPhysRegClass(predRegNumDst);
729       if (predRegClass == &Hexagon::PredRegsRegClass)
730         break;
731     }
732     assert((predRegClass == &Hexagon::PredRegsRegClass) &&
733            "predicate register not found in a predicated MI instruction");
734 
735     // New-value register producer and user (store) need to satisfy these
736     // constraints:
737     // 1) Both instructions should be predicated on the same register.
738     // 2) If producer of the new-value register is .new predicated then store
739     // should also be .new predicated and if producer is not .new predicated
740     // then store should not be .new predicated.
741     // 3) Both new-value register producer and user should have same predicate
742     // sense, i.e, either both should be negated or both should be non-negated.
743     if (predRegNumDst != predRegNumSrc ||
744         HII->isDotNewInst(PacketMI) != HII->isDotNewInst(MI) ||
745         getPredicateSense(MI, HII) != getPredicateSense(PacketMI, HII))
746       return false;
747   }
748 
749   // Make sure that other than the new-value register no other store instruction
750   // register has been modified in the same packet. Predicate registers can be
751   // modified by they should not be modified between the producer and the store
752   // instruction as it will make them both conditional on different values.
753   // We already know this to be true for all the instructions before and
754   // including PacketMI. Howerver, we need to perform the check for the
755   // remaining instructions in the packet.
756 
757   unsigned StartCheck = 0;
758 
759   for (auto I : CurrentPacketMIs) {
760     SUnit *TempSU = MIToSUnit.find(I)->second;
761     MachineInstr &TempMI = *TempSU->getInstr();
762 
763     // Following condition is true for all the instructions until PacketMI is
764     // reached (StartCheck is set to 0 before the for loop).
765     // StartCheck flag is 1 for all the instructions after PacketMI.
766     if (&TempMI != &PacketMI && !StartCheck) // Start processing only after
767       continue;                              // encountering PacketMI.
768 
769     StartCheck = 1;
770     if (&TempMI == &PacketMI) // We don't want to check PacketMI for dependence.
771       continue;
772 
773     for (auto &MO : MI.operands())
774       if (MO.isReg() && TempSU->getInstr()->modifiesRegister(MO.getReg(), HRI))
775         return false;
776   }
777 
778   // Make sure that for non-POST_INC stores:
779   // 1. The only use of reg is DepReg and no other registers.
780   //    This handles base+index registers.
781   //    The following store can not be dot new.
782   //    Eg.   r0 = add(r0, #3)
783   //          memw(r1+r0<<#2) = r0
784   if (!HII->isPostIncrement(MI)) {
785     for (unsigned opNum = 0; opNum < MI.getNumOperands()-1; opNum++) {
786       const MachineOperand &MO = MI.getOperand(opNum);
787       if (MO.isReg() && MO.getReg() == DepReg)
788         return false;
789     }
790   }
791 
792   // If data definition is because of implicit definition of the register,
793   // do not newify the store. Eg.
794   // %r9 = ZXTH %r12, implicit %d6, implicit-def %r12
795   // S2_storerh_io %r8, 2, killed %r12; mem:ST2[%scevgep343]
796   for (auto &MO : PacketMI.operands()) {
797     if (MO.isRegMask() && MO.clobbersPhysReg(DepReg))
798       return false;
799     if (!MO.isReg() || !MO.isDef() || !MO.isImplicit())
800       continue;
801     Register R = MO.getReg();
802     if (R == DepReg || HRI->isSuperRegister(DepReg, R))
803       return false;
804   }
805 
806   // Handle imp-use of super reg case. There is a target independent side
807   // change that should prevent this situation but I am handling it for
808   // just-in-case. For example, we cannot newify R2 in the following case:
809   // %r3 = A2_tfrsi 0;
810   // S2_storeri_io killed %r0, 0, killed %r2, implicit killed %d1;
811   for (auto &MO : MI.operands()) {
812     if (MO.isReg() && MO.isUse() && MO.isImplicit() && MO.getReg() == DepReg)
813       return false;
814   }
815 
816   // Can be dot new store.
817   return true;
818 }
819 
820 // Can this MI to promoted to either new value store or new value jump.
821 bool HexagonPacketizerList::canPromoteToNewValue(const MachineInstr &MI,
822       const SUnit *PacketSU, unsigned DepReg,
823       MachineBasicBlock::iterator &MII) {
824   if (!HII->mayBeNewStore(MI))
825     return false;
826 
827   // Check to see the store can be new value'ed.
828   MachineInstr &PacketMI = *PacketSU->getInstr();
829   if (canPromoteToNewValueStore(MI, PacketMI, DepReg))
830     return true;
831 
832   // Check to see the compare/jump can be new value'ed.
833   // This is done as a pass on its own. Don't need to check it here.
834   return false;
835 }
836 
837 static bool isImplicitDependency(const MachineInstr &I, bool CheckDef,
838       unsigned DepReg) {
839   for (auto &MO : I.operands()) {
840     if (CheckDef && MO.isRegMask() && MO.clobbersPhysReg(DepReg))
841       return true;
842     if (!MO.isReg() || MO.getReg() != DepReg || !MO.isImplicit())
843       continue;
844     if (CheckDef == MO.isDef())
845       return true;
846   }
847   return false;
848 }
849 
850 // Check to see if an instruction can be dot new.
851 bool HexagonPacketizerList::canPromoteToDotNew(const MachineInstr &MI,
852       const SUnit *PacketSU, unsigned DepReg, MachineBasicBlock::iterator &MII,
853       const TargetRegisterClass* RC) {
854   // Already a dot new instruction.
855   if (HII->isDotNewInst(MI) && !HII->mayBeNewStore(MI))
856     return false;
857 
858   if (!isNewifiable(MI, RC))
859     return false;
860 
861   const MachineInstr &PI = *PacketSU->getInstr();
862 
863   // The "new value" cannot come from inline asm.
864   if (PI.isInlineAsm())
865     return false;
866 
867   // IMPLICIT_DEFs won't materialize as real instructions, so .new makes no
868   // sense.
869   if (PI.isImplicitDef())
870     return false;
871 
872   // If dependency is trough an implicitly defined register, we should not
873   // newify the use.
874   if (isImplicitDependency(PI, true, DepReg) ||
875       isImplicitDependency(MI, false, DepReg))
876     return false;
877 
878   const MCInstrDesc& MCID = PI.getDesc();
879   const TargetRegisterClass *VecRC = HII->getRegClass(MCID, 0, HRI, MF);
880   if (DisableVecDblNVStores && VecRC == &Hexagon::HvxWRRegClass)
881     return false;
882 
883   // predicate .new
884   if (RC == &Hexagon::PredRegsRegClass)
885     return HII->predCanBeUsedAsDotNew(PI, DepReg);
886 
887   if (RC != &Hexagon::PredRegsRegClass && !HII->mayBeNewStore(MI))
888     return false;
889 
890   // Create a dot new machine instruction to see if resources can be
891   // allocated. If not, bail out now.
892   int NewOpcode = HII->getDotNewOp(MI);
893   const MCInstrDesc &D = HII->get(NewOpcode);
894   MachineInstr *NewMI = MF.CreateMachineInstr(D, DebugLoc());
895   bool ResourcesAvailable = ResourceTracker->canReserveResources(*NewMI);
896   MF.DeleteMachineInstr(NewMI);
897   if (!ResourcesAvailable)
898     return false;
899 
900   // New Value Store only. New Value Jump generated as a separate pass.
901   if (!canPromoteToNewValue(MI, PacketSU, DepReg, MII))
902     return false;
903 
904   return true;
905 }
906 
907 // Go through the packet instructions and search for an anti dependency between
908 // them and DepReg from MI. Consider this case:
909 // Trying to add
910 // a) %r1 = TFRI_cdNotPt %p3, 2
911 // to this packet:
912 // {
913 //   b) %p0 = C2_or killed %p3, killed %p0
914 //   c) %p3 = C2_tfrrp %r23
915 //   d) %r1 = C2_cmovenewit %p3, 4
916 //  }
917 // The P3 from a) and d) will be complements after
918 // a)'s P3 is converted to .new form
919 // Anti-dep between c) and b) is irrelevant for this case
920 bool HexagonPacketizerList::restrictingDepExistInPacket(MachineInstr &MI,
921                                                         unsigned DepReg) {
922   SUnit *PacketSUDep = MIToSUnit.find(&MI)->second;
923 
924   for (auto I : CurrentPacketMIs) {
925     // We only care for dependencies to predicated instructions
926     if (!HII->isPredicated(*I))
927       continue;
928 
929     // Scheduling Unit for current insn in the packet
930     SUnit *PacketSU = MIToSUnit.find(I)->second;
931 
932     // Look at dependencies between current members of the packet and
933     // predicate defining instruction MI. Make sure that dependency is
934     // on the exact register we care about.
935     if (PacketSU->isSucc(PacketSUDep)) {
936       for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
937         auto &Dep = PacketSU->Succs[i];
938         if (Dep.getSUnit() == PacketSUDep && Dep.getKind() == SDep::Anti &&
939             Dep.getReg() == DepReg)
940           return true;
941       }
942     }
943   }
944 
945   return false;
946 }
947 
948 /// Gets the predicate register of a predicated instruction.
949 static unsigned getPredicatedRegister(MachineInstr &MI,
950                                       const HexagonInstrInfo *QII) {
951   /// We use the following rule: The first predicate register that is a use is
952   /// the predicate register of a predicated instruction.
953   assert(QII->isPredicated(MI) && "Must be predicated instruction");
954 
955   for (auto &Op : MI.operands()) {
956     if (Op.isReg() && Op.getReg() && Op.isUse() &&
957         Hexagon::PredRegsRegClass.contains(Op.getReg()))
958       return Op.getReg();
959   }
960 
961   llvm_unreachable("Unknown instruction operand layout");
962   return 0;
963 }
964 
965 // Given two predicated instructions, this function detects whether
966 // the predicates are complements.
967 bool HexagonPacketizerList::arePredicatesComplements(MachineInstr &MI1,
968                                                      MachineInstr &MI2) {
969   // If we don't know the predicate sense of the instructions bail out early, we
970   // need it later.
971   if (getPredicateSense(MI1, HII) == PK_Unknown ||
972       getPredicateSense(MI2, HII) == PK_Unknown)
973     return false;
974 
975   // Scheduling unit for candidate.
976   SUnit *SU = MIToSUnit[&MI1];
977 
978   // One corner case deals with the following scenario:
979   // Trying to add
980   // a) %r24 = A2_tfrt %p0, %r25
981   // to this packet:
982   // {
983   //   b) %r25 = A2_tfrf %p0, %r24
984   //   c) %p0 = C2_cmpeqi %r26, 1
985   // }
986   //
987   // On general check a) and b) are complements, but presence of c) will
988   // convert a) to .new form, and then it is not a complement.
989   // We attempt to detect it by analyzing existing dependencies in the packet.
990 
991   // Analyze relationships between all existing members of the packet.
992   // Look for Anti dependecy on the same predicate reg as used in the
993   // candidate.
994   for (auto I : CurrentPacketMIs) {
995     // Scheduling Unit for current insn in the packet.
996     SUnit *PacketSU = MIToSUnit.find(I)->second;
997 
998     // If this instruction in the packet is succeeded by the candidate...
999     if (PacketSU->isSucc(SU)) {
1000       for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
1001         auto Dep = PacketSU->Succs[i];
1002         // The corner case exist when there is true data dependency between
1003         // candidate and one of current packet members, this dep is on
1004         // predicate reg, and there already exist anti dep on the same pred in
1005         // the packet.
1006         if (Dep.getSUnit() == SU && Dep.getKind() == SDep::Data &&
1007             Hexagon::PredRegsRegClass.contains(Dep.getReg())) {
1008           // Here I know that I is predicate setting instruction with true
1009           // data dep to candidate on the register we care about - c) in the
1010           // above example. Now I need to see if there is an anti dependency
1011           // from c) to any other instruction in the same packet on the pred
1012           // reg of interest.
1013           if (restrictingDepExistInPacket(*I, Dep.getReg()))
1014             return false;
1015         }
1016       }
1017     }
1018   }
1019 
1020   // If the above case does not apply, check regular complement condition.
1021   // Check that the predicate register is the same and that the predicate
1022   // sense is different We also need to differentiate .old vs. .new: !p0
1023   // is not complementary to p0.new.
1024   unsigned PReg1 = getPredicatedRegister(MI1, HII);
1025   unsigned PReg2 = getPredicatedRegister(MI2, HII);
1026   return PReg1 == PReg2 &&
1027          Hexagon::PredRegsRegClass.contains(PReg1) &&
1028          Hexagon::PredRegsRegClass.contains(PReg2) &&
1029          getPredicateSense(MI1, HII) != getPredicateSense(MI2, HII) &&
1030          HII->isDotNewInst(MI1) == HII->isDotNewInst(MI2);
1031 }
1032 
1033 // Initialize packetizer flags.
1034 void HexagonPacketizerList::initPacketizerState() {
1035   Dependence = false;
1036   PromotedToDotNew = false;
1037   GlueToNewValueJump = false;
1038   GlueAllocframeStore = false;
1039   FoundSequentialDependence = false;
1040   ChangedOffset = INT64_MAX;
1041 }
1042 
1043 // Ignore bundling of pseudo instructions.
1044 bool HexagonPacketizerList::ignorePseudoInstruction(const MachineInstr &MI,
1045                                                     const MachineBasicBlock *) {
1046   if (MI.isDebugInstr())
1047     return true;
1048 
1049   if (MI.isCFIInstruction())
1050     return false;
1051 
1052   // We must print out inline assembly.
1053   if (MI.isInlineAsm())
1054     return false;
1055 
1056   if (MI.isImplicitDef())
1057     return false;
1058 
1059   // We check if MI has any functional units mapped to it. If it doesn't,
1060   // we ignore the instruction.
1061   const MCInstrDesc& TID = MI.getDesc();
1062   auto *IS = ResourceTracker->getInstrItins()->beginStage(TID.getSchedClass());
1063   return !IS->getUnits();
1064 }
1065 
1066 bool HexagonPacketizerList::isSoloInstruction(const MachineInstr &MI) {
1067   // Ensure any bundles created by gather packetize remain separate.
1068   if (MI.isBundle())
1069     return true;
1070 
1071   if (MI.isEHLabel() || MI.isCFIInstruction())
1072     return true;
1073 
1074   // Consider inline asm to not be a solo instruction by default.
1075   // Inline asm will be put in a packet temporarily, but then it will be
1076   // removed, and placed outside of the packet (before or after, depending
1077   // on dependencies).  This is to reduce the impact of inline asm as a
1078   // "packet splitting" instruction.
1079   if (MI.isInlineAsm() && !ScheduleInlineAsm)
1080     return true;
1081 
1082   if (isSchedBarrier(MI))
1083     return true;
1084 
1085   if (HII->isSolo(MI))
1086     return true;
1087 
1088   if (MI.getOpcode() == Hexagon::A2_nop)
1089     return true;
1090 
1091   return false;
1092 }
1093 
1094 // Quick check if instructions MI and MJ cannot coexist in the same packet.
1095 // Limit the tests to be "one-way", e.g.  "if MI->isBranch and MJ->isInlineAsm",
1096 // but not the symmetric case: "if MJ->isBranch and MI->isInlineAsm".
1097 // For full test call this function twice:
1098 //   cannotCoexistAsymm(MI, MJ) || cannotCoexistAsymm(MJ, MI)
1099 // Doing the test only one way saves the amount of code in this function,
1100 // since every test would need to be repeated with the MI and MJ reversed.
1101 static bool cannotCoexistAsymm(const MachineInstr &MI, const MachineInstr &MJ,
1102       const HexagonInstrInfo &HII) {
1103   const MachineFunction *MF = MI.getParent()->getParent();
1104   if (MF->getSubtarget<HexagonSubtarget>().hasV60OpsOnly() &&
1105       HII.isHVXMemWithAIndirect(MI, MJ))
1106     return true;
1107 
1108   // An inline asm cannot be together with a branch, because we may not be
1109   // able to remove the asm out after packetizing (i.e. if the asm must be
1110   // moved past the bundle).  Similarly, two asms cannot be together to avoid
1111   // complications when determining their relative order outside of a bundle.
1112   if (MI.isInlineAsm())
1113     return MJ.isInlineAsm() || MJ.isBranch() || MJ.isBarrier() ||
1114            MJ.isCall() || MJ.isTerminator();
1115 
1116   // New-value stores cannot coexist with any other stores.
1117   if (HII.isNewValueStore(MI) && MJ.mayStore())
1118     return true;
1119 
1120   switch (MI.getOpcode()) {
1121   case Hexagon::S2_storew_locked:
1122   case Hexagon::S4_stored_locked:
1123   case Hexagon::L2_loadw_locked:
1124   case Hexagon::L4_loadd_locked:
1125   case Hexagon::Y2_dccleana:
1126   case Hexagon::Y2_dccleaninva:
1127   case Hexagon::Y2_dcinva:
1128   case Hexagon::Y2_dczeroa:
1129   case Hexagon::Y4_l2fetch:
1130   case Hexagon::Y5_l2fetch: {
1131     // These instructions can only be grouped with ALU32 or non-floating-point
1132     // XTYPE instructions.  Since there is no convenient way of identifying fp
1133     // XTYPE instructions, only allow grouping with ALU32 for now.
1134     unsigned TJ = HII.getType(MJ);
1135     if (TJ != HexagonII::TypeALU32_2op &&
1136         TJ != HexagonII::TypeALU32_3op &&
1137         TJ != HexagonII::TypeALU32_ADDI)
1138       return true;
1139     break;
1140   }
1141   default:
1142     break;
1143   }
1144 
1145   // "False" really means that the quick check failed to determine if
1146   // I and J cannot coexist.
1147   return false;
1148 }
1149 
1150 // Full, symmetric check.
1151 bool HexagonPacketizerList::cannotCoexist(const MachineInstr &MI,
1152       const MachineInstr &MJ) {
1153   return cannotCoexistAsymm(MI, MJ, *HII) || cannotCoexistAsymm(MJ, MI, *HII);
1154 }
1155 
1156 void HexagonPacketizerList::unpacketizeSoloInstrs(MachineFunction &MF) {
1157   for (auto &B : MF) {
1158     MachineBasicBlock::iterator BundleIt;
1159     MachineBasicBlock::instr_iterator NextI;
1160     for (auto I = B.instr_begin(), E = B.instr_end(); I != E; I = NextI) {
1161       NextI = std::next(I);
1162       MachineInstr &MI = *I;
1163       if (MI.isBundle())
1164         BundleIt = I;
1165       if (!MI.isInsideBundle())
1166         continue;
1167 
1168       // Decide on where to insert the instruction that we are pulling out.
1169       // Debug instructions always go before the bundle, but the placement of
1170       // INLINE_ASM depends on potential dependencies.  By default, try to
1171       // put it before the bundle, but if the asm writes to a register that
1172       // other instructions in the bundle read, then we need to place it
1173       // after the bundle (to preserve the bundle semantics).
1174       bool InsertBeforeBundle;
1175       if (MI.isInlineAsm())
1176         InsertBeforeBundle = !hasWriteToReadDep(MI, *BundleIt, HRI);
1177       else if (MI.isDebugValue())
1178         InsertBeforeBundle = true;
1179       else
1180         continue;
1181 
1182       BundleIt = moveInstrOut(MI, BundleIt, InsertBeforeBundle);
1183     }
1184   }
1185 }
1186 
1187 // Check if a given instruction is of class "system".
1188 static bool isSystemInstr(const MachineInstr &MI) {
1189   unsigned Opc = MI.getOpcode();
1190   switch (Opc) {
1191     case Hexagon::Y2_barrier:
1192     case Hexagon::Y2_dcfetchbo:
1193     case Hexagon::Y4_l2fetch:
1194     case Hexagon::Y5_l2fetch:
1195       return true;
1196   }
1197   return false;
1198 }
1199 
1200 bool HexagonPacketizerList::hasDeadDependence(const MachineInstr &I,
1201                                               const MachineInstr &J) {
1202   // The dependence graph may not include edges between dead definitions,
1203   // so without extra checks, we could end up packetizing two instruction
1204   // defining the same (dead) register.
1205   if (I.isCall() || J.isCall())
1206     return false;
1207   if (HII->isPredicated(I) || HII->isPredicated(J))
1208     return false;
1209 
1210   BitVector DeadDefs(Hexagon::NUM_TARGET_REGS);
1211   for (auto &MO : I.operands()) {
1212     if (!MO.isReg() || !MO.isDef() || !MO.isDead())
1213       continue;
1214     DeadDefs[MO.getReg()] = true;
1215   }
1216 
1217   for (auto &MO : J.operands()) {
1218     if (!MO.isReg() || !MO.isDef() || !MO.isDead())
1219       continue;
1220     Register R = MO.getReg();
1221     if (R != Hexagon::USR_OVF && DeadDefs[R])
1222       return true;
1223   }
1224   return false;
1225 }
1226 
1227 bool HexagonPacketizerList::hasControlDependence(const MachineInstr &I,
1228                                                  const MachineInstr &J) {
1229   // A save callee-save register function call can only be in a packet
1230   // with instructions that don't write to the callee-save registers.
1231   if ((HII->isSaveCalleeSavedRegsCall(I) &&
1232        doesModifyCalleeSavedReg(J, HRI)) ||
1233       (HII->isSaveCalleeSavedRegsCall(J) &&
1234        doesModifyCalleeSavedReg(I, HRI)))
1235     return true;
1236 
1237   // Two control flow instructions cannot go in the same packet.
1238   if (isControlFlow(I) && isControlFlow(J))
1239     return true;
1240 
1241   // \ref-manual (7.3.4) A loop setup packet in loopN or spNloop0 cannot
1242   // contain a speculative indirect jump,
1243   // a new-value compare jump or a dealloc_return.
1244   auto isBadForLoopN = [this] (const MachineInstr &MI) -> bool {
1245     if (MI.isCall() || HII->isDeallocRet(MI) || HII->isNewValueJump(MI))
1246       return true;
1247     if (HII->isPredicated(MI) && HII->isPredicatedNew(MI) && HII->isJumpR(MI))
1248       return true;
1249     return false;
1250   };
1251 
1252   if (HII->isLoopN(I) && isBadForLoopN(J))
1253     return true;
1254   if (HII->isLoopN(J) && isBadForLoopN(I))
1255     return true;
1256 
1257   // dealloc_return cannot appear in the same packet as a conditional or
1258   // unconditional jump.
1259   return HII->isDeallocRet(I) &&
1260          (J.isBranch() || J.isCall() || J.isBarrier());
1261 }
1262 
1263 bool HexagonPacketizerList::hasRegMaskDependence(const MachineInstr &I,
1264                                                  const MachineInstr &J) {
1265   // Adding I to a packet that has J.
1266 
1267   // Regmasks are not reflected in the scheduling dependency graph, so
1268   // we need to check them manually. This code assumes that regmasks only
1269   // occur on calls, and the problematic case is when we add an instruction
1270   // defining a register R to a packet that has a call that clobbers R via
1271   // a regmask. Those cannot be packetized together, because the call will
1272   // be executed last. That's also a reson why it is ok to add a call
1273   // clobbering R to a packet that defines R.
1274 
1275   // Look for regmasks in J.
1276   for (const MachineOperand &OpJ : J.operands()) {
1277     if (!OpJ.isRegMask())
1278       continue;
1279     assert((J.isCall() || HII->isTailCall(J)) && "Regmask on a non-call");
1280     for (const MachineOperand &OpI : I.operands()) {
1281       if (OpI.isReg()) {
1282         if (OpJ.clobbersPhysReg(OpI.getReg()))
1283           return true;
1284       } else if (OpI.isRegMask()) {
1285         // Both are regmasks. Assume that they intersect.
1286         return true;
1287       }
1288     }
1289   }
1290   return false;
1291 }
1292 
1293 bool HexagonPacketizerList::hasDualStoreDependence(const MachineInstr &I,
1294                                                    const MachineInstr &J) {
1295   bool SysI = isSystemInstr(I), SysJ = isSystemInstr(J);
1296   bool StoreI = I.mayStore(), StoreJ = J.mayStore();
1297   if ((SysI && StoreJ) || (SysJ && StoreI))
1298     return true;
1299 
1300   if (StoreI && StoreJ) {
1301     if (HII->isNewValueInst(J) || HII->isMemOp(J) || HII->isMemOp(I))
1302       return true;
1303   } else {
1304     // A memop cannot be in the same packet with another memop or a store.
1305     // Two stores can be together, but here I and J cannot both be stores.
1306     bool MopStI = HII->isMemOp(I) || StoreI;
1307     bool MopStJ = HII->isMemOp(J) || StoreJ;
1308     if (MopStI && MopStJ)
1309       return true;
1310   }
1311 
1312   return (StoreJ && HII->isDeallocRet(I)) || (StoreI && HII->isDeallocRet(J));
1313 }
1314 
1315 // SUI is the current instruction that is out side of the current packet.
1316 // SUJ is the current instruction inside the current packet against which that
1317 // SUI will be packetized.
1318 bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) {
1319   assert(SUI->getInstr() && SUJ->getInstr());
1320   MachineInstr &I = *SUI->getInstr();
1321   MachineInstr &J = *SUJ->getInstr();
1322 
1323   // Clear IgnoreDepMIs when Packet starts.
1324   if (CurrentPacketMIs.size() == 1)
1325     IgnoreDepMIs.clear();
1326 
1327   MachineBasicBlock::iterator II = I.getIterator();
1328 
1329   // Solo instructions cannot go in the packet.
1330   assert(!isSoloInstruction(I) && "Unexpected solo instr!");
1331 
1332   if (cannotCoexist(I, J))
1333     return false;
1334 
1335   Dependence = hasDeadDependence(I, J) || hasControlDependence(I, J);
1336   if (Dependence)
1337     return false;
1338 
1339   // Regmasks are not accounted for in the scheduling graph, so we need
1340   // to explicitly check for dependencies caused by them. They should only
1341   // appear on calls, so it's not too pessimistic to reject all regmask
1342   // dependencies.
1343   Dependence = hasRegMaskDependence(I, J);
1344   if (Dependence)
1345     return false;
1346 
1347   // Dual-store does not allow second store, if the first store is not
1348   // in SLOT0. New value store, new value jump, dealloc_return and memop
1349   // always take SLOT0. Arch spec 3.4.4.2.
1350   Dependence = hasDualStoreDependence(I, J);
1351   if (Dependence)
1352     return false;
1353 
1354   // If an instruction feeds new value jump, glue it.
1355   MachineBasicBlock::iterator NextMII = I.getIterator();
1356   ++NextMII;
1357   if (NextMII != I.getParent()->end() && HII->isNewValueJump(*NextMII)) {
1358     MachineInstr &NextMI = *NextMII;
1359 
1360     bool secondRegMatch = false;
1361     const MachineOperand &NOp0 = NextMI.getOperand(0);
1362     const MachineOperand &NOp1 = NextMI.getOperand(1);
1363 
1364     if (NOp1.isReg() && I.getOperand(0).getReg() == NOp1.getReg())
1365       secondRegMatch = true;
1366 
1367     for (MachineInstr *PI : CurrentPacketMIs) {
1368       // NVJ can not be part of the dual jump - Arch Spec: section 7.8.
1369       if (PI->isCall()) {
1370         Dependence = true;
1371         break;
1372       }
1373       // Validate:
1374       // 1. Packet does not have a store in it.
1375       // 2. If the first operand of the nvj is newified, and the second
1376       //    operand is also a reg, it (second reg) is not defined in
1377       //    the same packet.
1378       // 3. If the second operand of the nvj is newified, (which means
1379       //    first operand is also a reg), first reg is not defined in
1380       //    the same packet.
1381       if (PI->getOpcode() == Hexagon::S2_allocframe || PI->mayStore() ||
1382           HII->isLoopN(*PI)) {
1383         Dependence = true;
1384         break;
1385       }
1386       // Check #2/#3.
1387       const MachineOperand &OpR = secondRegMatch ? NOp0 : NOp1;
1388       if (OpR.isReg() && PI->modifiesRegister(OpR.getReg(), HRI)) {
1389         Dependence = true;
1390         break;
1391       }
1392     }
1393 
1394     GlueToNewValueJump = true;
1395     if (Dependence)
1396       return false;
1397   }
1398 
1399   // There no dependency between a prolog instruction and its successor.
1400   if (!SUJ->isSucc(SUI))
1401     return true;
1402 
1403   for (unsigned i = 0; i < SUJ->Succs.size(); ++i) {
1404     if (FoundSequentialDependence)
1405       break;
1406 
1407     if (SUJ->Succs[i].getSUnit() != SUI)
1408       continue;
1409 
1410     SDep::Kind DepType = SUJ->Succs[i].getKind();
1411     // For direct calls:
1412     // Ignore register dependences for call instructions for packetization
1413     // purposes except for those due to r31 and predicate registers.
1414     //
1415     // For indirect calls:
1416     // Same as direct calls + check for true dependences to the register
1417     // used in the indirect call.
1418     //
1419     // We completely ignore Order dependences for call instructions.
1420     //
1421     // For returns:
1422     // Ignore register dependences for return instructions like jumpr,
1423     // dealloc return unless we have dependencies on the explicit uses
1424     // of the registers used by jumpr (like r31) or dealloc return
1425     // (like r29 or r30).
1426     unsigned DepReg = 0;
1427     const TargetRegisterClass *RC = nullptr;
1428     if (DepType == SDep::Data) {
1429       DepReg = SUJ->Succs[i].getReg();
1430       RC = HRI->getMinimalPhysRegClass(DepReg);
1431     }
1432 
1433     if (I.isCall() || HII->isJumpR(I) || I.isReturn() || HII->isTailCall(I)) {
1434       if (!isRegDependence(DepType))
1435         continue;
1436       if (!isCallDependent(I, DepType, SUJ->Succs[i].getReg()))
1437         continue;
1438     }
1439 
1440     if (DepType == SDep::Data) {
1441       if (canPromoteToDotCur(J, SUJ, DepReg, II, RC))
1442         if (promoteToDotCur(J, DepType, II, RC))
1443           continue;
1444     }
1445 
1446     // Data dpendence ok if we have load.cur.
1447     if (DepType == SDep::Data && HII->isDotCurInst(J)) {
1448       if (HII->isHVXVec(I))
1449         continue;
1450     }
1451 
1452     // For instructions that can be promoted to dot-new, try to promote.
1453     if (DepType == SDep::Data) {
1454       if (canPromoteToDotNew(I, SUJ, DepReg, II, RC)) {
1455         if (promoteToDotNew(I, DepType, II, RC)) {
1456           PromotedToDotNew = true;
1457           if (cannotCoexist(I, J))
1458             FoundSequentialDependence = true;
1459           continue;
1460         }
1461       }
1462       if (HII->isNewValueJump(I))
1463         continue;
1464     }
1465 
1466     // For predicated instructions, if the predicates are complements then
1467     // there can be no dependence.
1468     if (HII->isPredicated(I) && HII->isPredicated(J) &&
1469         arePredicatesComplements(I, J)) {
1470       // Not always safe to do this translation.
1471       // DAG Builder attempts to reduce dependence edges using transitive
1472       // nature of dependencies. Here is an example:
1473       //
1474       // r0 = tfr_pt ... (1)
1475       // r0 = tfr_pf ... (2)
1476       // r0 = tfr_pt ... (3)
1477       //
1478       // There will be an output dependence between (1)->(2) and (2)->(3).
1479       // However, there is no dependence edge between (1)->(3). This results
1480       // in all 3 instructions going in the same packet. We ignore dependce
1481       // only once to avoid this situation.
1482       auto Itr = find(IgnoreDepMIs, &J);
1483       if (Itr != IgnoreDepMIs.end()) {
1484         Dependence = true;
1485         return false;
1486       }
1487       IgnoreDepMIs.push_back(&I);
1488       continue;
1489     }
1490 
1491     // Ignore Order dependences between unconditional direct branches
1492     // and non-control-flow instructions.
1493     if (isDirectJump(I) && !J.isBranch() && !J.isCall() &&
1494         DepType == SDep::Order)
1495       continue;
1496 
1497     // Ignore all dependences for jumps except for true and output
1498     // dependences.
1499     if (I.isConditionalBranch() && DepType != SDep::Data &&
1500         DepType != SDep::Output)
1501       continue;
1502 
1503     if (DepType == SDep::Output) {
1504       FoundSequentialDependence = true;
1505       break;
1506     }
1507 
1508     // For Order dependences:
1509     // 1. Volatile loads/stores can be packetized together, unless other
1510     //    rules prevent is.
1511     // 2. Store followed by a load is not allowed.
1512     // 3. Store followed by a store is valid.
1513     // 4. Load followed by any memory operation is allowed.
1514     if (DepType == SDep::Order) {
1515       if (!PacketizeVolatiles) {
1516         bool OrdRefs = I.hasOrderedMemoryRef() || J.hasOrderedMemoryRef();
1517         if (OrdRefs) {
1518           FoundSequentialDependence = true;
1519           break;
1520         }
1521       }
1522       // J is first, I is second.
1523       bool LoadJ = J.mayLoad(), StoreJ = J.mayStore();
1524       bool LoadI = I.mayLoad(), StoreI = I.mayStore();
1525       bool NVStoreJ = HII->isNewValueStore(J);
1526       bool NVStoreI = HII->isNewValueStore(I);
1527       bool IsVecJ = HII->isHVXVec(J);
1528       bool IsVecI = HII->isHVXVec(I);
1529 
1530       if (Slot1Store && MF.getSubtarget<HexagonSubtarget>().hasV65Ops() &&
1531           ((LoadJ && StoreI && !NVStoreI) ||
1532            (StoreJ && LoadI && !NVStoreJ)) &&
1533           (J.getOpcode() != Hexagon::S2_allocframe &&
1534            I.getOpcode() != Hexagon::S2_allocframe) &&
1535           (J.getOpcode() != Hexagon::L2_deallocframe &&
1536            I.getOpcode() != Hexagon::L2_deallocframe) &&
1537           (!HII->isMemOp(J) && !HII->isMemOp(I)) && (!IsVecJ && !IsVecI))
1538         setmemShufDisabled(true);
1539       else
1540         if (StoreJ && LoadI && alias(J, I)) {
1541           FoundSequentialDependence = true;
1542           break;
1543         }
1544 
1545       if (!StoreJ)
1546         if (!LoadJ || (!LoadI && !StoreI)) {
1547           // If J is neither load nor store, assume a dependency.
1548           // If J is a load, but I is neither, also assume a dependency.
1549           FoundSequentialDependence = true;
1550           break;
1551         }
1552       // Store followed by store: not OK on V2.
1553       // Store followed by load: not OK on all.
1554       // Load followed by store: OK on all.
1555       // Load followed by load: OK on all.
1556       continue;
1557     }
1558 
1559     // Special case for ALLOCFRAME: even though there is dependency
1560     // between ALLOCFRAME and subsequent store, allow it to be packetized
1561     // in a same packet. This implies that the store is using the caller's
1562     // SP. Hence, offset needs to be updated accordingly.
1563     if (DepType == SDep::Data && J.getOpcode() == Hexagon::S2_allocframe) {
1564       unsigned Opc = I.getOpcode();
1565       switch (Opc) {
1566         case Hexagon::S2_storerd_io:
1567         case Hexagon::S2_storeri_io:
1568         case Hexagon::S2_storerh_io:
1569         case Hexagon::S2_storerb_io:
1570           if (I.getOperand(0).getReg() == HRI->getStackRegister()) {
1571             // Since this store is to be glued with allocframe in the same
1572             // packet, it will use SP of the previous stack frame, i.e.
1573             // caller's SP. Therefore, we need to recalculate offset
1574             // according to this change.
1575             GlueAllocframeStore = useCallersSP(I);
1576             if (GlueAllocframeStore)
1577               continue;
1578           }
1579           break;
1580         default:
1581           break;
1582       }
1583     }
1584 
1585     // There are certain anti-dependencies that cannot be ignored.
1586     // Specifically:
1587     //   J2_call ... implicit-def %r0   ; SUJ
1588     //   R0 = ...                   ; SUI
1589     // Those cannot be packetized together, since the call will observe
1590     // the effect of the assignment to R0.
1591     if ((DepType == SDep::Anti || DepType == SDep::Output) && J.isCall()) {
1592       // Check if I defines any volatile register. We should also check
1593       // registers that the call may read, but these happen to be a
1594       // subset of the volatile register set.
1595       for (const MachineOperand &Op : I.operands()) {
1596         if (Op.isReg() && Op.isDef()) {
1597           Register R = Op.getReg();
1598           if (!J.readsRegister(R, HRI) && !J.modifiesRegister(R, HRI))
1599             continue;
1600         } else if (!Op.isRegMask()) {
1601           // If I has a regmask assume dependency.
1602           continue;
1603         }
1604         FoundSequentialDependence = true;
1605         break;
1606       }
1607     }
1608 
1609     // Skip over remaining anti-dependences. Two instructions that are
1610     // anti-dependent can share a packet, since in most such cases all
1611     // operands are read before any modifications take place.
1612     // The exceptions are branch and call instructions, since they are
1613     // executed after all other instructions have completed (at least
1614     // conceptually).
1615     if (DepType != SDep::Anti) {
1616       FoundSequentialDependence = true;
1617       break;
1618     }
1619   }
1620 
1621   if (FoundSequentialDependence) {
1622     Dependence = true;
1623     return false;
1624   }
1625 
1626   return true;
1627 }
1628 
1629 bool HexagonPacketizerList::isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) {
1630   assert(SUI->getInstr() && SUJ->getInstr());
1631   MachineInstr &I = *SUI->getInstr();
1632   MachineInstr &J = *SUJ->getInstr();
1633 
1634   bool Coexist = !cannotCoexist(I, J);
1635 
1636   if (Coexist && !Dependence)
1637     return true;
1638 
1639   // Check if the instruction was promoted to a dot-new. If so, demote it
1640   // back into a dot-old.
1641   if (PromotedToDotNew)
1642     demoteToDotOld(I);
1643 
1644   cleanUpDotCur();
1645   // Check if the instruction (must be a store) was glued with an allocframe
1646   // instruction. If so, restore its offset to its original value, i.e. use
1647   // current SP instead of caller's SP.
1648   if (GlueAllocframeStore) {
1649     useCalleesSP(I);
1650     GlueAllocframeStore = false;
1651   }
1652 
1653   if (ChangedOffset != INT64_MAX)
1654     undoChangedOffset(I);
1655 
1656   if (GlueToNewValueJump) {
1657     // Putting I and J together would prevent the new-value jump from being
1658     // packetized with the producer. In that case I and J must be separated.
1659     GlueToNewValueJump = false;
1660     return false;
1661   }
1662 
1663   if (!Coexist)
1664     return false;
1665 
1666   if (ChangedOffset == INT64_MAX && updateOffset(SUI, SUJ)) {
1667     FoundSequentialDependence = false;
1668     Dependence = false;
1669     return true;
1670   }
1671 
1672   return false;
1673 }
1674 
1675 
1676 bool HexagonPacketizerList::foundLSInPacket() {
1677   bool FoundLoad = false;
1678   bool FoundStore = false;
1679 
1680   for (auto MJ : CurrentPacketMIs) {
1681     unsigned Opc = MJ->getOpcode();
1682     if (Opc == Hexagon::S2_allocframe || Opc == Hexagon::L2_deallocframe)
1683       continue;
1684     if (HII->isMemOp(*MJ))
1685       continue;
1686     if (MJ->mayLoad())
1687       FoundLoad = true;
1688     if (MJ->mayStore() && !HII->isNewValueStore(*MJ))
1689       FoundStore = true;
1690   }
1691   return FoundLoad && FoundStore;
1692 }
1693 
1694 
1695 MachineBasicBlock::iterator
1696 HexagonPacketizerList::addToPacket(MachineInstr &MI) {
1697   MachineBasicBlock::iterator MII = MI.getIterator();
1698   MachineBasicBlock *MBB = MI.getParent();
1699 
1700   if (CurrentPacketMIs.empty())
1701     PacketStalls = false;
1702   PacketStalls |= producesStall(MI);
1703 
1704   if (MI.isImplicitDef()) {
1705     // Add to the packet to allow subsequent instructions to be checked
1706     // properly.
1707     CurrentPacketMIs.push_back(&MI);
1708     return MII;
1709   }
1710   assert(ResourceTracker->canReserveResources(MI));
1711 
1712   bool ExtMI = HII->isExtended(MI) || HII->isConstExtended(MI);
1713   bool Good = true;
1714 
1715   if (GlueToNewValueJump) {
1716     MachineInstr &NvjMI = *++MII;
1717     // We need to put both instructions in the same packet: MI and NvjMI.
1718     // Either of them can require a constant extender. Try to add both to
1719     // the current packet, and if that fails, end the packet and start a
1720     // new one.
1721     ResourceTracker->reserveResources(MI);
1722     if (ExtMI)
1723       Good = tryAllocateResourcesForConstExt(true);
1724 
1725     bool ExtNvjMI = HII->isExtended(NvjMI) || HII->isConstExtended(NvjMI);
1726     if (Good) {
1727       if (ResourceTracker->canReserveResources(NvjMI))
1728         ResourceTracker->reserveResources(NvjMI);
1729       else
1730         Good = false;
1731     }
1732     if (Good && ExtNvjMI)
1733       Good = tryAllocateResourcesForConstExt(true);
1734 
1735     if (!Good) {
1736       endPacket(MBB, MI);
1737       assert(ResourceTracker->canReserveResources(MI));
1738       ResourceTracker->reserveResources(MI);
1739       if (ExtMI) {
1740         assert(canReserveResourcesForConstExt());
1741         tryAllocateResourcesForConstExt(true);
1742       }
1743       assert(ResourceTracker->canReserveResources(NvjMI));
1744       ResourceTracker->reserveResources(NvjMI);
1745       if (ExtNvjMI) {
1746         assert(canReserveResourcesForConstExt());
1747         reserveResourcesForConstExt();
1748       }
1749     }
1750     CurrentPacketMIs.push_back(&MI);
1751     CurrentPacketMIs.push_back(&NvjMI);
1752     return MII;
1753   }
1754 
1755   ResourceTracker->reserveResources(MI);
1756   if (ExtMI && !tryAllocateResourcesForConstExt(true)) {
1757     endPacket(MBB, MI);
1758     if (PromotedToDotNew)
1759       demoteToDotOld(MI);
1760     if (GlueAllocframeStore) {
1761       useCalleesSP(MI);
1762       GlueAllocframeStore = false;
1763     }
1764     ResourceTracker->reserveResources(MI);
1765     reserveResourcesForConstExt();
1766   }
1767 
1768   CurrentPacketMIs.push_back(&MI);
1769   return MII;
1770 }
1771 
1772 void HexagonPacketizerList::endPacket(MachineBasicBlock *MBB,
1773                                       MachineBasicBlock::iterator EndMI) {
1774   // Replace VLIWPacketizerList::endPacket(MBB, EndMI).
1775   LLVM_DEBUG({
1776     if (!CurrentPacketMIs.empty()) {
1777       dbgs() << "Finalizing packet:\n";
1778       unsigned Idx = 0;
1779       for (MachineInstr *MI : CurrentPacketMIs) {
1780         unsigned R = ResourceTracker->getUsedResources(Idx++);
1781         dbgs() << " * [res:0x" << utohexstr(R) << "] " << *MI;
1782       }
1783     }
1784   });
1785 
1786   bool memShufDisabled = getmemShufDisabled();
1787   if (memShufDisabled && !foundLSInPacket()) {
1788     setmemShufDisabled(false);
1789     LLVM_DEBUG(dbgs() << "  Not added to NoShufPacket\n");
1790   }
1791   memShufDisabled = getmemShufDisabled();
1792 
1793   OldPacketMIs.clear();
1794   for (MachineInstr *MI : CurrentPacketMIs) {
1795     MachineBasicBlock::instr_iterator NextMI = std::next(MI->getIterator());
1796     for (auto &I : make_range(HII->expandVGatherPseudo(*MI), NextMI))
1797       OldPacketMIs.push_back(&I);
1798   }
1799   CurrentPacketMIs.clear();
1800 
1801   if (OldPacketMIs.size() > 1) {
1802     MachineBasicBlock::instr_iterator FirstMI(OldPacketMIs.front());
1803     MachineBasicBlock::instr_iterator LastMI(EndMI.getInstrIterator());
1804     finalizeBundle(*MBB, FirstMI, LastMI);
1805     auto BundleMII = std::prev(FirstMI);
1806     if (memShufDisabled)
1807       HII->setBundleNoShuf(BundleMII);
1808 
1809     setmemShufDisabled(false);
1810   }
1811 
1812   PacketHasDuplex = false;
1813   PacketHasSLOT0OnlyInsn = false;
1814   ResourceTracker->clearResources();
1815   LLVM_DEBUG(dbgs() << "End packet\n");
1816 }
1817 
1818 bool HexagonPacketizerList::shouldAddToPacket(const MachineInstr &MI) {
1819   if (Minimal)
1820     return false;
1821 
1822   // Constrainst for not packetizing this MI with existing instructions in a
1823   // packet.
1824   //	MI is a store instruction.
1825   //	CurrentPacketMIs has a SLOT0 only instruction with constraint
1826   //    A_RESTRICT_NOSLOT1_STORE/isRestrictNoSlot1Store.
1827   if (MI.mayStore() && isPureSlot0InsnWithNoSlot1Store(MI))
1828     return false;
1829 
1830   if (producesStall(MI))
1831     return false;
1832 
1833   // If TinyCore with Duplexes is enabled, check if this MI can form a Duplex
1834   // with any other instruction in the existing packet.
1835   auto &HST = MI.getParent()->getParent()->getSubtarget<HexagonSubtarget>();
1836   // Constraint 1: Only one duplex allowed per packet.
1837   // Constraint 2: Consider duplex checks only if there is atleast one
1838   // instruction in a packet.
1839   // Constraint 3: If one of the existing instructions in the packet has a
1840   // SLOT0 only instruction that can not be duplexed, do not attempt to form
1841   // duplexes. (TODO: This will invalidate the L4_return* instructions to form a
1842   // duplex)
1843   if (HST.isTinyCoreWithDuplex() && CurrentPacketMIs.size() > 0 &&
1844       !PacketHasDuplex) {
1845     // Check for SLOT0 only non-duplexable instruction in packet.
1846     for (auto &MJ : CurrentPacketMIs)
1847       PacketHasSLOT0OnlyInsn |= HII->isPureSlot0(*MJ);
1848     // Get the Big Core Opcode (dup_*).
1849     int Opcode = HII->getDuplexOpcode(MI, false);
1850     if (Opcode >= 0) {
1851       // We now have an instruction that can be duplexed.
1852       for (auto &MJ : CurrentPacketMIs) {
1853         if (HII->isDuplexPair(MI, *MJ) && !PacketHasSLOT0OnlyInsn) {
1854           PacketHasDuplex = true;
1855           return true;
1856         }
1857       }
1858       // If it can not be duplexed, check if there is a valid transition in DFA
1859       // with the original opcode.
1860       MachineInstr &MIRef = const_cast<MachineInstr &>(MI);
1861       MIRef.setDesc(HII->get(Opcode));
1862       return ResourceTracker->canReserveResources(MIRef);
1863     }
1864   }
1865 
1866   return true;
1867 }
1868 
1869 bool HexagonPacketizerList::isPureSlot0InsnWithNoSlot1Store(
1870     const MachineInstr &MI) {
1871   bool noSlot1Store = false;
1872   bool isSlot0Only = false;
1873   for (auto J : CurrentPacketMIs) {
1874     noSlot1Store |= HII->isRestrictNoSlot1Store(*J);
1875     isSlot0Only |= HII->isPureSlot0(*J);
1876   }
1877 
1878   return (noSlot1Store && isSlot0Only);
1879 }
1880 
1881 // V60 forward scheduling.
1882 bool HexagonPacketizerList::producesStall(const MachineInstr &I) {
1883   // If the packet already stalls, then ignore the stall from a subsequent
1884   // instruction in the same packet.
1885   if (PacketStalls)
1886     return false;
1887 
1888   // Check whether the previous packet is in a different loop. If this is the
1889   // case, there is little point in trying to avoid a stall because that would
1890   // favor the rare case (loop entry) over the common case (loop iteration).
1891   //
1892   // TODO: We should really be able to check all the incoming edges if this is
1893   // the first packet in a basic block, so we can avoid stalls from the loop
1894   // backedge.
1895   if (!OldPacketMIs.empty()) {
1896     auto *OldBB = OldPacketMIs.front()->getParent();
1897     auto *ThisBB = I.getParent();
1898     if (MLI->getLoopFor(OldBB) != MLI->getLoopFor(ThisBB))
1899       return false;
1900   }
1901 
1902   SUnit *SUI = MIToSUnit[const_cast<MachineInstr *>(&I)];
1903 
1904   // If the latency is 0 and there is a data dependence between this
1905   // instruction and any instruction in the current packet, we disregard any
1906   // potential stalls due to the instructions in the previous packet. Most of
1907   // the instruction pairs that can go together in the same packet have 0
1908   // latency between them. The exceptions are
1909   // 1. NewValueJumps as they're generated much later and the latencies can't
1910   // be changed at that point.
1911   // 2. .cur instructions, if its consumer has a 0 latency successor (such as
1912   // .new). In this case, the latency between .cur and the consumer stays
1913   // non-zero even though we can have  both .cur and .new in the same packet.
1914   // Changing the latency to 0 is not an option as it causes software pipeliner
1915   // to not pipeline in some cases.
1916 
1917   // For Example:
1918   // {
1919   //   I1:  v6.cur = vmem(r0++#1)
1920   //   I2:  v7 = valign(v6,v4,r2)
1921   //   I3:  vmem(r5++#1) = v7.new
1922   // }
1923   // Here I2 and I3 has 0 cycle latency, but I1 and I2 has 2.
1924 
1925   for (auto J : CurrentPacketMIs) {
1926     SUnit *SUJ = MIToSUnit[J];
1927     for (auto &Pred : SUI->Preds)
1928       if (Pred.getSUnit() == SUJ)
1929         if ((Pred.getLatency() == 0 && Pred.isAssignedRegDep()) ||
1930             HII->isNewValueJump(I) || HII->isToBeScheduledASAP(*J, I))
1931           return false;
1932   }
1933 
1934   // Check if the latency is greater than one between this instruction and any
1935   // instruction in the previous packet.
1936   for (auto J : OldPacketMIs) {
1937     SUnit *SUJ = MIToSUnit[J];
1938     for (auto &Pred : SUI->Preds)
1939       if (Pred.getSUnit() == SUJ && Pred.getLatency() > 1)
1940         return true;
1941   }
1942 
1943   return false;
1944 }
1945 
1946 //===----------------------------------------------------------------------===//
1947 //                         Public Constructor Functions
1948 //===----------------------------------------------------------------------===//
1949 
1950 FunctionPass *llvm::createHexagonPacketizer(bool Minimal) {
1951   return new HexagonPacketizer(Minimal);
1952 }
1953