• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===----- HexagonPacketizer.cpp - vliw packetizer ---------------------===//
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 a simple VLIW packetizer using DFA. The packetizer works on
11 // machine basic blocks. For each instruction I in BB, the packetizer consults
12 // the DFA to see if machine resources are available to execute I. If so, the
13 // packetizer checks if I depends on any instruction J in the current packet.
14 // If no dependency is found, I is added to current packet and machine resource
15 // is marked as taken. If any dependency is found, a target API call is made to
16 // prune the dependence.
17 //
18 //===----------------------------------------------------------------------===//
19 #include "HexagonRegisterInfo.h"
20 #include "HexagonSubtarget.h"
21 #include "HexagonTargetMachine.h"
22 #include "HexagonVLIWPacketizer.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineDominators.h"
25 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachineLoopInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/Passes.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 
33 using namespace llvm;
34 
35 #define DEBUG_TYPE "packets"
36 
37 static cl::opt<bool> DisablePacketizer("disable-packetizer", cl::Hidden,
38   cl::ZeroOrMore, cl::init(false),
39   cl::desc("Disable Hexagon packetizer pass"));
40 
41 static cl::opt<bool> PacketizeVolatiles("hexagon-packetize-volatiles",
42   cl::ZeroOrMore, cl::Hidden, cl::init(true),
43   cl::desc("Allow non-solo packetization of volatile memory references"));
44 
45 static cl::opt<bool> EnableGenAllInsnClass("enable-gen-insn", cl::init(false),
46   cl::Hidden, cl::ZeroOrMore, cl::desc("Generate all instruction with TC"));
47 
48 static cl::opt<bool> DisableVecDblNVStores("disable-vecdbl-nv-stores",
49   cl::init(false), cl::Hidden, cl::ZeroOrMore,
50   cl::desc("Disable vector double new-value-stores"));
51 
52 extern cl::opt<bool> ScheduleInlineAsm;
53 
54 namespace llvm {
55   FunctionPass *createHexagonPacketizer();
56   void initializeHexagonPacketizerPass(PassRegistry&);
57 }
58 
59 
60 namespace {
61   class HexagonPacketizer : public MachineFunctionPass {
62   public:
63     static char ID;
HexagonPacketizer()64     HexagonPacketizer() : MachineFunctionPass(ID) {
65       initializeHexagonPacketizerPass(*PassRegistry::getPassRegistry());
66     }
67 
getAnalysisUsage(AnalysisUsage & AU) const68     void getAnalysisUsage(AnalysisUsage &AU) const override {
69       AU.setPreservesCFG();
70       AU.addRequired<AAResultsWrapperPass>();
71       AU.addRequired<MachineBranchProbabilityInfo>();
72       AU.addRequired<MachineDominatorTree>();
73       AU.addRequired<MachineLoopInfo>();
74       AU.addPreserved<MachineDominatorTree>();
75       AU.addPreserved<MachineLoopInfo>();
76       MachineFunctionPass::getAnalysisUsage(AU);
77     }
getPassName() const78     const char *getPassName() const override {
79       return "Hexagon Packetizer";
80     }
81     bool runOnMachineFunction(MachineFunction &Fn) override;
getRequiredProperties() const82     MachineFunctionProperties getRequiredProperties() const override {
83       return MachineFunctionProperties().set(
84           MachineFunctionProperties::Property::AllVRegsAllocated);
85     }
86 
87   private:
88     const HexagonInstrInfo *HII;
89     const HexagonRegisterInfo *HRI;
90   };
91 
92   char HexagonPacketizer::ID = 0;
93 }
94 
95 INITIALIZE_PASS_BEGIN(HexagonPacketizer, "packets", "Hexagon Packetizer",
96                       false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)97 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
98 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
99 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
100 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
101 INITIALIZE_PASS_END(HexagonPacketizer, "packets", "Hexagon Packetizer",
102                     false, false)
103 
104 
105 HexagonPacketizerList::HexagonPacketizerList(MachineFunction &MF,
106       MachineLoopInfo &MLI, AliasAnalysis *AA,
107       const MachineBranchProbabilityInfo *MBPI)
108     : VLIWPacketizerList(MF, MLI, AA), MBPI(MBPI), MLI(&MLI) {
109   HII = MF.getSubtarget<HexagonSubtarget>().getInstrInfo();
110   HRI = MF.getSubtarget<HexagonSubtarget>().getRegisterInfo();
111 }
112 
113 // Check if FirstI modifies a register that SecondI reads.
hasWriteToReadDep(const MachineInstr & FirstI,const MachineInstr & SecondI,const TargetRegisterInfo * TRI)114 static bool hasWriteToReadDep(const MachineInstr &FirstI,
115                               const MachineInstr &SecondI,
116                               const TargetRegisterInfo *TRI) {
117   for (auto &MO : FirstI.operands()) {
118     if (!MO.isReg() || !MO.isDef())
119       continue;
120     unsigned R = MO.getReg();
121     if (SecondI.readsRegister(R, TRI))
122       return true;
123   }
124   return false;
125 }
126 
127 
moveInstrOut(MachineInstr * MI,MachineBasicBlock::iterator BundleIt,bool Before)128 static MachineBasicBlock::iterator moveInstrOut(MachineInstr *MI,
129       MachineBasicBlock::iterator BundleIt, bool Before) {
130   MachineBasicBlock::instr_iterator InsertPt;
131   if (Before)
132     InsertPt = BundleIt.getInstrIterator();
133   else
134     InsertPt = std::next(BundleIt).getInstrIterator();
135 
136   MachineBasicBlock &B = *MI->getParent();
137   // The instruction should at least be bundled with the preceding instruction
138   // (there will always be one, i.e. BUNDLE, if nothing else).
139   assert(MI->isBundledWithPred());
140   if (MI->isBundledWithSucc()) {
141     MI->clearFlag(MachineInstr::BundledSucc);
142     MI->clearFlag(MachineInstr::BundledPred);
143   } else {
144     // If it's not bundled with the successor (i.e. it is the last one
145     // in the bundle), then we can simply unbundle it from the predecessor,
146     // which will take care of updating the predecessor's flag.
147     MI->unbundleFromPred();
148   }
149   B.splice(InsertPt, &B, MI);
150 
151   // Get the size of the bundle without asserting.
152   MachineBasicBlock::const_instr_iterator I = BundleIt.getInstrIterator();
153   MachineBasicBlock::const_instr_iterator E = B.instr_end();
154   unsigned Size = 0;
155   for (++I; I != E && I->isBundledWithPred(); ++I)
156     ++Size;
157 
158   // If there are still two or more instructions, then there is nothing
159   // else to be done.
160   if (Size > 1)
161     return BundleIt;
162 
163   // Otherwise, extract the single instruction out and delete the bundle.
164   MachineBasicBlock::iterator NextIt = std::next(BundleIt);
165   MachineInstr *SingleI = BundleIt->getNextNode();
166   SingleI->unbundleFromPred();
167   assert(!SingleI->isBundledWithSucc());
168   BundleIt->eraseFromParent();
169   return NextIt;
170 }
171 
172 
runOnMachineFunction(MachineFunction & MF)173 bool HexagonPacketizer::runOnMachineFunction(MachineFunction &MF) {
174   if (DisablePacketizer || skipFunction(*MF.getFunction()))
175     return false;
176 
177   HII = MF.getSubtarget<HexagonSubtarget>().getInstrInfo();
178   HRI = MF.getSubtarget<HexagonSubtarget>().getRegisterInfo();
179   auto &MLI = getAnalysis<MachineLoopInfo>();
180   auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
181   auto *MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
182 
183   if (EnableGenAllInsnClass)
184     HII->genAllInsnTimingClasses(MF);
185 
186   // Instantiate the packetizer.
187   HexagonPacketizerList Packetizer(MF, MLI, AA, MBPI);
188 
189   // DFA state table should not be empty.
190   assert(Packetizer.getResourceTracker() && "Empty DFA table!");
191 
192   //
193   // Loop over all basic blocks and remove KILL pseudo-instructions
194   // These instructions confuse the dependence analysis. Consider:
195   // D0 = ...   (Insn 0)
196   // R0 = KILL R0, D0 (Insn 1)
197   // R0 = ... (Insn 2)
198   // Here, Insn 1 will result in the dependence graph not emitting an output
199   // dependence between Insn 0 and Insn 2. This can lead to incorrect
200   // packetization
201   //
202   for (auto &MB : MF) {
203     auto End = MB.end();
204     auto MI = MB.begin();
205     while (MI != End) {
206       auto NextI = std::next(MI);
207       if (MI->isKill()) {
208         MB.erase(MI);
209         End = MB.end();
210       }
211       MI = NextI;
212     }
213   }
214 
215   // Loop over all of the basic blocks.
216   for (auto &MB : MF) {
217     auto Begin = MB.begin(), End = MB.end();
218     while (Begin != End) {
219       // First the first non-boundary starting from the end of the last
220       // scheduling region.
221       MachineBasicBlock::iterator RB = Begin;
222       while (RB != End && HII->isSchedulingBoundary(*RB, &MB, MF))
223         ++RB;
224       // First the first boundary starting from the beginning of the new
225       // region.
226       MachineBasicBlock::iterator RE = RB;
227       while (RE != End && !HII->isSchedulingBoundary(*RE, &MB, MF))
228         ++RE;
229       // Add the scheduling boundary if it's not block end.
230       if (RE != End)
231         ++RE;
232       // If RB == End, then RE == End.
233       if (RB != End)
234         Packetizer.PacketizeMIs(&MB, RB, RE);
235 
236       Begin = RE;
237     }
238   }
239 
240   Packetizer.unpacketizeSoloInstrs(MF);
241   return true;
242 }
243 
244 
245 // Reserve resources for a constant extender. Trigger an assertion if the
246 // reservation fails.
reserveResourcesForConstExt()247 void HexagonPacketizerList::reserveResourcesForConstExt() {
248   if (!tryAllocateResourcesForConstExt(true))
249     llvm_unreachable("Resources not available");
250 }
251 
canReserveResourcesForConstExt()252 bool HexagonPacketizerList::canReserveResourcesForConstExt() {
253   return tryAllocateResourcesForConstExt(false);
254 }
255 
256 // Allocate resources (i.e. 4 bytes) for constant extender. If succeeded,
257 // return true, otherwise, return false.
tryAllocateResourcesForConstExt(bool Reserve)258 bool HexagonPacketizerList::tryAllocateResourcesForConstExt(bool Reserve) {
259   auto *ExtMI = MF.CreateMachineInstr(HII->get(Hexagon::A4_ext), DebugLoc());
260   bool Avail = ResourceTracker->canReserveResources(*ExtMI);
261   if (Reserve && Avail)
262     ResourceTracker->reserveResources(*ExtMI);
263   MF.DeleteMachineInstr(ExtMI);
264   return Avail;
265 }
266 
267 
isCallDependent(const MachineInstr * MI,SDep::Kind DepType,unsigned DepReg)268 bool HexagonPacketizerList::isCallDependent(const MachineInstr* MI,
269       SDep::Kind DepType, unsigned DepReg) {
270   // Check for LR dependence.
271   if (DepReg == HRI->getRARegister())
272     return true;
273 
274   if (HII->isDeallocRet(MI))
275     if (DepReg == HRI->getFrameRegister() || DepReg == HRI->getStackRegister())
276       return true;
277 
278   // Check if this is a predicate dependence.
279   const TargetRegisterClass* RC = HRI->getMinimalPhysRegClass(DepReg);
280   if (RC == &Hexagon::PredRegsRegClass)
281     return true;
282 
283   // Assumes that the first operand of the CALLr is the function address.
284   if (HII->isIndirectCall(MI) && (DepType == SDep::Data)) {
285     MachineOperand MO = MI->getOperand(0);
286     if (MO.isReg() && MO.isUse() && (MO.getReg() == DepReg))
287       return true;
288   }
289 
290   return false;
291 }
292 
isRegDependence(const SDep::Kind DepType)293 static bool isRegDependence(const SDep::Kind DepType) {
294   return DepType == SDep::Data || DepType == SDep::Anti ||
295          DepType == SDep::Output;
296 }
297 
isDirectJump(const MachineInstr * MI)298 static bool isDirectJump(const MachineInstr* MI) {
299   return MI->getOpcode() == Hexagon::J2_jump;
300 }
301 
isSchedBarrier(const MachineInstr * MI)302 static bool isSchedBarrier(const MachineInstr* MI) {
303   switch (MI->getOpcode()) {
304   case Hexagon::Y2_barrier:
305     return true;
306   }
307   return false;
308 }
309 
isControlFlow(const MachineInstr * MI)310 static bool isControlFlow(const MachineInstr* MI) {
311   return (MI->getDesc().isTerminator() || MI->getDesc().isCall());
312 }
313 
314 
315 /// Returns true if the instruction modifies a callee-saved register.
doesModifyCalleeSavedReg(const MachineInstr * MI,const TargetRegisterInfo * TRI)316 static bool doesModifyCalleeSavedReg(const MachineInstr *MI,
317                                      const TargetRegisterInfo *TRI) {
318   const MachineFunction &MF = *MI->getParent()->getParent();
319   for (auto *CSR = TRI->getCalleeSavedRegs(&MF); CSR && *CSR; ++CSR)
320     if (MI->modifiesRegister(*CSR, TRI))
321       return true;
322   return false;
323 }
324 
325 // TODO: MI->isIndirectBranch() and IsRegisterJump(MI)
326 // Returns true if an instruction can be promoted to .new predicate or
327 // new-value store.
isNewifiable(const MachineInstr * MI)328 bool HexagonPacketizerList::isNewifiable(const MachineInstr* MI) {
329   return HII->isCondInst(MI) || MI->isReturn() || HII->mayBeNewStore(MI);
330 }
331 
332 // Promote an instructiont to its .cur form.
333 // At this time, we have already made a call to canPromoteToDotCur and made
334 // sure that it can *indeed* be promoted.
promoteToDotCur(MachineInstr * MI,SDep::Kind DepType,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)335 bool HexagonPacketizerList::promoteToDotCur(MachineInstr* MI,
336       SDep::Kind DepType, MachineBasicBlock::iterator &MII,
337       const TargetRegisterClass* RC) {
338   assert(DepType == SDep::Data);
339   int CurOpcode = HII->getDotCurOp(MI);
340   MI->setDesc(HII->get(CurOpcode));
341   return true;
342 }
343 
cleanUpDotCur()344 void HexagonPacketizerList::cleanUpDotCur() {
345   MachineInstr *MI = NULL;
346   for (auto BI : CurrentPacketMIs) {
347     DEBUG(dbgs() << "Cleanup packet has "; BI->dump(););
348     if (BI->getOpcode() == Hexagon::V6_vL32b_cur_ai) {
349       MI = BI;
350       continue;
351     }
352     if (MI) {
353       for (auto &MO : BI->operands())
354         if (MO.isReg() && MO.getReg() == MI->getOperand(0).getReg())
355           return;
356     }
357   }
358   if (!MI)
359     return;
360   // We did not find a use of the CUR, so de-cur it.
361   MI->setDesc(HII->get(Hexagon::V6_vL32b_ai));
362   DEBUG(dbgs() << "Demoted CUR "; MI->dump(););
363 }
364 
365 // Check to see if an instruction can be dot cur.
canPromoteToDotCur(const MachineInstr * MI,const SUnit * PacketSU,unsigned DepReg,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)366 bool HexagonPacketizerList::canPromoteToDotCur(const MachineInstr *MI,
367       const SUnit *PacketSU, unsigned DepReg, MachineBasicBlock::iterator &MII,
368       const TargetRegisterClass *RC) {
369   if (!HII->isV60VectorInstruction(MI))
370     return false;
371   if (!HII->isV60VectorInstruction(&*MII))
372     return false;
373 
374   // Already a dot new instruction.
375   if (HII->isDotCurInst(MI) && !HII->mayBeCurLoad(MI))
376     return false;
377 
378   if (!HII->mayBeCurLoad(MI))
379     return false;
380 
381   // The "cur value" cannot come from inline asm.
382   if (PacketSU->getInstr()->isInlineAsm())
383     return false;
384 
385   // Make sure candidate instruction uses cur.
386   DEBUG(dbgs() << "Can we DOT Cur Vector MI\n";
387         MI->dump();
388         dbgs() << "in packet\n";);
389   MachineInstr &MJ = *MII;
390   DEBUG({
391     dbgs() << "Checking CUR against ";
392     MJ.dump();
393   });
394   unsigned DestReg = MI->getOperand(0).getReg();
395   bool FoundMatch = false;
396   for (auto &MO : MJ.operands())
397     if (MO.isReg() && MO.getReg() == DestReg)
398       FoundMatch = true;
399   if (!FoundMatch)
400     return false;
401 
402   // Check for existing uses of a vector register within the packet which
403   // would be affected by converting a vector load into .cur formt.
404   for (auto BI : CurrentPacketMIs) {
405     DEBUG(dbgs() << "packet has "; BI->dump(););
406     if (BI->readsRegister(DepReg, MF.getSubtarget().getRegisterInfo()))
407       return false;
408   }
409 
410   DEBUG(dbgs() << "Can Dot CUR MI\n"; MI->dump(););
411   // We can convert the opcode into a .cur.
412   return true;
413 }
414 
415 // Promote an instruction to its .new form. At this time, we have already
416 // made a call to canPromoteToDotNew and made sure that it can *indeed* be
417 // promoted.
promoteToDotNew(MachineInstr * MI,SDep::Kind DepType,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)418 bool HexagonPacketizerList::promoteToDotNew(MachineInstr* MI,
419       SDep::Kind DepType, MachineBasicBlock::iterator &MII,
420       const TargetRegisterClass* RC) {
421   assert (DepType == SDep::Data);
422   int NewOpcode;
423   if (RC == &Hexagon::PredRegsRegClass)
424     NewOpcode = HII->getDotNewPredOp(MI, MBPI);
425   else
426     NewOpcode = HII->getDotNewOp(MI);
427   MI->setDesc(HII->get(NewOpcode));
428   return true;
429 }
430 
demoteToDotOld(MachineInstr * MI)431 bool HexagonPacketizerList::demoteToDotOld(MachineInstr* MI) {
432   int NewOpcode = HII->getDotOldOp(MI->getOpcode());
433   MI->setDesc(HII->get(NewOpcode));
434   return true;
435 }
436 
437 enum PredicateKind {
438   PK_False,
439   PK_True,
440   PK_Unknown
441 };
442 
443 /// Returns true if an instruction is predicated on p0 and false if it's
444 /// predicated on !p0.
getPredicateSense(const MachineInstr & MI,const HexagonInstrInfo * HII)445 static PredicateKind getPredicateSense(const MachineInstr &MI,
446                                        const HexagonInstrInfo *HII) {
447   if (!HII->isPredicated(MI))
448     return PK_Unknown;
449   if (HII->isPredicatedTrue(MI))
450     return PK_True;
451   return PK_False;
452 }
453 
getPostIncrementOperand(const MachineInstr * MI,const HexagonInstrInfo * HII)454 static const MachineOperand &getPostIncrementOperand(const MachineInstr *MI,
455       const HexagonInstrInfo *HII) {
456   assert(HII->isPostIncrement(MI) && "Not a post increment operation.");
457 #ifndef NDEBUG
458   // Post Increment means duplicates. Use dense map to find duplicates in the
459   // list. Caution: Densemap initializes with the minimum of 64 buckets,
460   // whereas there are at most 5 operands in the post increment.
461   DenseSet<unsigned> DefRegsSet;
462   for (auto &MO : MI->operands())
463     if (MO.isReg() && MO.isDef())
464       DefRegsSet.insert(MO.getReg());
465 
466   for (auto &MO : MI->operands())
467     if (MO.isReg() && MO.isUse() && DefRegsSet.count(MO.getReg()))
468       return MO;
469 #else
470   if (MI->mayLoad()) {
471     const MachineOperand &Op1 = MI->getOperand(1);
472     // The 2nd operand is always the post increment operand in load.
473     assert(Op1.isReg() && "Post increment operand has be to a register.");
474     return Op1;
475   }
476   if (MI->getDesc().mayStore()) {
477     const MachineOperand &Op0 = MI->getOperand(0);
478     // The 1st operand is always the post increment operand in store.
479     assert(Op0.isReg() && "Post increment operand has be to a register.");
480     return Op0;
481   }
482 #endif
483   // we should never come here.
484   llvm_unreachable("mayLoad or mayStore not set for Post Increment operation");
485 }
486 
487 // Get the value being stored.
getStoreValueOperand(const MachineInstr * MI)488 static const MachineOperand& getStoreValueOperand(const MachineInstr *MI) {
489   // value being stored is always the last operand.
490   return MI->getOperand(MI->getNumOperands()-1);
491 }
492 
isLoadAbsSet(const MachineInstr * MI)493 static bool isLoadAbsSet(const MachineInstr *MI) {
494   unsigned Opc = MI->getOpcode();
495   switch (Opc) {
496     case Hexagon::L4_loadrd_ap:
497     case Hexagon::L4_loadrb_ap:
498     case Hexagon::L4_loadrh_ap:
499     case Hexagon::L4_loadrub_ap:
500     case Hexagon::L4_loadruh_ap:
501     case Hexagon::L4_loadri_ap:
502       return true;
503   }
504   return false;
505 }
506 
getAbsSetOperand(const MachineInstr * MI)507 static const MachineOperand &getAbsSetOperand(const MachineInstr *MI) {
508   assert(isLoadAbsSet(MI));
509   return MI->getOperand(1);
510 }
511 
512 
513 // Can be new value store?
514 // Following restrictions are to be respected in convert a store into
515 // a new value store.
516 // 1. If an instruction uses auto-increment, its address register cannot
517 //    be a new-value register. Arch Spec 5.4.2.1
518 // 2. If an instruction uses absolute-set addressing mode, its address
519 //    register cannot be a new-value register. Arch Spec 5.4.2.1.
520 // 3. If an instruction produces a 64-bit result, its registers cannot be used
521 //    as new-value registers. Arch Spec 5.4.2.2.
522 // 4. If the instruction that sets the new-value register is conditional, then
523 //    the instruction that uses the new-value register must also be conditional,
524 //    and both must always have their predicates evaluate identically.
525 //    Arch Spec 5.4.2.3.
526 // 5. There is an implied restriction that a packet cannot have another store,
527 //    if there is a new value store in the packet. Corollary: if there is
528 //    already a store in a packet, there can not be a new value store.
529 //    Arch Spec: 3.4.4.2
canPromoteToNewValueStore(const MachineInstr * MI,const MachineInstr * PacketMI,unsigned DepReg)530 bool HexagonPacketizerList::canPromoteToNewValueStore(const MachineInstr *MI,
531       const MachineInstr *PacketMI, unsigned DepReg) {
532   // Make sure we are looking at the store, that can be promoted.
533   if (!HII->mayBeNewStore(MI))
534     return false;
535 
536   // Make sure there is dependency and can be new value'd.
537   const MachineOperand &Val = getStoreValueOperand(MI);
538   if (Val.isReg() && Val.getReg() != DepReg)
539     return false;
540 
541   const MCInstrDesc& MCID = PacketMI->getDesc();
542 
543   // First operand is always the result.
544   const TargetRegisterClass *PacketRC = HII->getRegClass(MCID, 0, HRI, MF);
545   // Double regs can not feed into new value store: PRM section: 5.4.2.2.
546   if (PacketRC == &Hexagon::DoubleRegsRegClass)
547     return false;
548 
549   // New-value stores are of class NV (slot 0), dual stores require class ST
550   // in slot 0 (PRM 5.5).
551   for (auto I : CurrentPacketMIs) {
552     SUnit *PacketSU = MIToSUnit.find(I)->second;
553     if (PacketSU->getInstr()->mayStore())
554       return false;
555   }
556 
557   // Make sure it's NOT the post increment register that we are going to
558   // new value.
559   if (HII->isPostIncrement(MI) &&
560       getPostIncrementOperand(MI, HII).getReg() == DepReg) {
561     return false;
562   }
563 
564   if (HII->isPostIncrement(PacketMI) && PacketMI->mayLoad() &&
565       getPostIncrementOperand(PacketMI, HII).getReg() == DepReg) {
566     // If source is post_inc, or absolute-set addressing, it can not feed
567     // into new value store
568     //   r3 = memw(r2++#4)
569     //   memw(r30 + #-1404) = r2.new -> can not be new value store
570     // arch spec section: 5.4.2.1.
571     return false;
572   }
573 
574   if (isLoadAbsSet(PacketMI) && getAbsSetOperand(PacketMI).getReg() == DepReg)
575     return false;
576 
577   // If the source that feeds the store is predicated, new value store must
578   // also be predicated.
579   if (HII->isPredicated(*PacketMI)) {
580     if (!HII->isPredicated(*MI))
581       return false;
582 
583     // Check to make sure that they both will have their predicates
584     // evaluate identically.
585     unsigned predRegNumSrc = 0;
586     unsigned predRegNumDst = 0;
587     const TargetRegisterClass* predRegClass = nullptr;
588 
589     // Get predicate register used in the source instruction.
590     for (auto &MO : PacketMI->operands()) {
591       if (!MO.isReg())
592         continue;
593       predRegNumSrc = MO.getReg();
594       predRegClass = HRI->getMinimalPhysRegClass(predRegNumSrc);
595       if (predRegClass == &Hexagon::PredRegsRegClass)
596         break;
597     }
598     assert((predRegClass == &Hexagon::PredRegsRegClass) &&
599         "predicate register not found in a predicated PacketMI instruction");
600 
601     // Get predicate register used in new-value store instruction.
602     for (auto &MO : MI->operands()) {
603       if (!MO.isReg())
604         continue;
605       predRegNumDst = MO.getReg();
606       predRegClass = HRI->getMinimalPhysRegClass(predRegNumDst);
607       if (predRegClass == &Hexagon::PredRegsRegClass)
608         break;
609     }
610     assert((predRegClass == &Hexagon::PredRegsRegClass) &&
611            "predicate register not found in a predicated MI instruction");
612 
613     // New-value register producer and user (store) need to satisfy these
614     // constraints:
615     // 1) Both instructions should be predicated on the same register.
616     // 2) If producer of the new-value register is .new predicated then store
617     // should also be .new predicated and if producer is not .new predicated
618     // then store should not be .new predicated.
619     // 3) Both new-value register producer and user should have same predicate
620     // sense, i.e, either both should be negated or both should be non-negated.
621     if (predRegNumDst != predRegNumSrc ||
622         HII->isDotNewInst(PacketMI) != HII->isDotNewInst(MI) ||
623         getPredicateSense(*MI, HII) != getPredicateSense(*PacketMI, HII))
624       return false;
625   }
626 
627   // Make sure that other than the new-value register no other store instruction
628   // register has been modified in the same packet. Predicate registers can be
629   // modified by they should not be modified between the producer and the store
630   // instruction as it will make them both conditional on different values.
631   // We already know this to be true for all the instructions before and
632   // including PacketMI. Howerver, we need to perform the check for the
633   // remaining instructions in the packet.
634 
635   unsigned StartCheck = 0;
636 
637   for (auto I : CurrentPacketMIs) {
638     SUnit *TempSU = MIToSUnit.find(I)->second;
639     MachineInstr* TempMI = TempSU->getInstr();
640 
641     // Following condition is true for all the instructions until PacketMI is
642     // reached (StartCheck is set to 0 before the for loop).
643     // StartCheck flag is 1 for all the instructions after PacketMI.
644     if (TempMI != PacketMI && !StartCheck) // Start processing only after
645       continue;                            // encountering PacketMI.
646 
647     StartCheck = 1;
648     if (TempMI == PacketMI) // We don't want to check PacketMI for dependence.
649       continue;
650 
651     for (auto &MO : MI->operands())
652       if (MO.isReg() && TempSU->getInstr()->modifiesRegister(MO.getReg(), HRI))
653         return false;
654   }
655 
656   // Make sure that for non-POST_INC stores:
657   // 1. The only use of reg is DepReg and no other registers.
658   //    This handles V4 base+index registers.
659   //    The following store can not be dot new.
660   //    Eg.   r0 = add(r0, #3)
661   //          memw(r1+r0<<#2) = r0
662   if (!HII->isPostIncrement(MI)) {
663     for (unsigned opNum = 0; opNum < MI->getNumOperands()-1; opNum++) {
664       const MachineOperand &MO = MI->getOperand(opNum);
665       if (MO.isReg() && MO.getReg() == DepReg)
666         return false;
667     }
668   }
669 
670   // If data definition is because of implicit definition of the register,
671   // do not newify the store. Eg.
672   // %R9<def> = ZXTH %R12, %D6<imp-use>, %R12<imp-def>
673   // S2_storerh_io %R8, 2, %R12<kill>; mem:ST2[%scevgep343]
674   for (auto &MO : PacketMI->operands()) {
675     if (!MO.isReg() || !MO.isDef() || !MO.isImplicit())
676       continue;
677     unsigned R = MO.getReg();
678     if (R == DepReg || HRI->isSuperRegister(DepReg, R))
679       return false;
680   }
681 
682   // Handle imp-use of super reg case. There is a target independent side
683   // change that should prevent this situation but I am handling it for
684   // just-in-case. For example, we cannot newify R2 in the following case:
685   // %R3<def> = A2_tfrsi 0;
686   // S2_storeri_io %R0<kill>, 0, %R2<kill>, %D1<imp-use,kill>;
687   for (auto &MO : MI->operands()) {
688     if (MO.isReg() && MO.isUse() && MO.isImplicit() && MO.getReg() == DepReg)
689       return false;
690   }
691 
692   // Can be dot new store.
693   return true;
694 }
695 
696 // Can this MI to promoted to either new value store or new value jump.
canPromoteToNewValue(const MachineInstr * MI,const SUnit * PacketSU,unsigned DepReg,MachineBasicBlock::iterator & MII)697 bool HexagonPacketizerList::canPromoteToNewValue(const MachineInstr *MI,
698       const SUnit *PacketSU, unsigned DepReg,
699       MachineBasicBlock::iterator &MII) {
700   if (!HII->mayBeNewStore(MI))
701     return false;
702 
703   // Check to see the store can be new value'ed.
704   MachineInstr *PacketMI = PacketSU->getInstr();
705   if (canPromoteToNewValueStore(MI, PacketMI, DepReg))
706     return true;
707 
708   // Check to see the compare/jump can be new value'ed.
709   // This is done as a pass on its own. Don't need to check it here.
710   return false;
711 }
712 
isImplicitDependency(const MachineInstr * I,unsigned DepReg)713 static bool isImplicitDependency(const MachineInstr *I, unsigned DepReg) {
714   for (auto &MO : I->operands())
715     if (MO.isReg() && MO.isDef() && (MO.getReg() == DepReg) && MO.isImplicit())
716       return true;
717   return false;
718 }
719 
720 // Check to see if an instruction can be dot new
721 // There are three kinds.
722 // 1. dot new on predicate - V2/V3/V4
723 // 2. dot new on stores NV/ST - V4
724 // 3. dot new on jump NV/J - V4 -- This is generated in a pass.
canPromoteToDotNew(const MachineInstr * MI,const SUnit * PacketSU,unsigned DepReg,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)725 bool HexagonPacketizerList::canPromoteToDotNew(const MachineInstr *MI,
726       const SUnit *PacketSU, unsigned DepReg, MachineBasicBlock::iterator &MII,
727       const TargetRegisterClass* RC) {
728   // Already a dot new instruction.
729   if (HII->isDotNewInst(MI) && !HII->mayBeNewStore(MI))
730     return false;
731 
732   if (!isNewifiable(MI))
733     return false;
734 
735   const MachineInstr *PI = PacketSU->getInstr();
736 
737   // The "new value" cannot come from inline asm.
738   if (PI->isInlineAsm())
739     return false;
740 
741   // IMPLICIT_DEFs won't materialize as real instructions, so .new makes no
742   // sense.
743   if (PI->isImplicitDef())
744     return false;
745 
746   // If dependency is trough an implicitly defined register, we should not
747   // newify the use.
748   if (isImplicitDependency(PI, DepReg))
749     return false;
750 
751   const MCInstrDesc& MCID = PI->getDesc();
752   const TargetRegisterClass *VecRC = HII->getRegClass(MCID, 0, HRI, MF);
753   if (DisableVecDblNVStores && VecRC == &Hexagon::VecDblRegsRegClass)
754     return false;
755 
756   // predicate .new
757   // bug 5670: until that is fixed
758   // TODO: MI->isIndirectBranch() and IsRegisterJump(MI)
759   if (RC == &Hexagon::PredRegsRegClass)
760     if (HII->isCondInst(MI) || MI->isReturn())
761       return HII->predCanBeUsedAsDotNew(PI, DepReg);
762 
763   if (RC != &Hexagon::PredRegsRegClass && !HII->mayBeNewStore(MI))
764     return false;
765 
766   // Create a dot new machine instruction to see if resources can be
767   // allocated. If not, bail out now.
768   int NewOpcode = HII->getDotNewOp(MI);
769   const MCInstrDesc &D = HII->get(NewOpcode);
770   MachineInstr *NewMI = MF.CreateMachineInstr(D, DebugLoc());
771   bool ResourcesAvailable = ResourceTracker->canReserveResources(*NewMI);
772   MF.DeleteMachineInstr(NewMI);
773   if (!ResourcesAvailable)
774     return false;
775 
776   // New Value Store only. New Value Jump generated as a separate pass.
777   if (!canPromoteToNewValue(MI, PacketSU, DepReg, MII))
778     return false;
779 
780   return true;
781 }
782 
783 // Go through the packet instructions and search for an anti dependency between
784 // them and DepReg from MI. Consider this case:
785 // Trying to add
786 // a) %R1<def> = TFRI_cdNotPt %P3, 2
787 // to this packet:
788 // {
789 //   b) %P0<def> = C2_or %P3<kill>, %P0<kill>
790 //   c) %P3<def> = C2_tfrrp %R23
791 //   d) %R1<def> = C2_cmovenewit %P3, 4
792 //  }
793 // The P3 from a) and d) will be complements after
794 // a)'s P3 is converted to .new form
795 // Anti-dep between c) and b) is irrelevant for this case
restrictingDepExistInPacket(MachineInstr * MI,unsigned DepReg)796 bool HexagonPacketizerList::restrictingDepExistInPacket(MachineInstr* MI,
797                                                         unsigned DepReg) {
798   SUnit *PacketSUDep = MIToSUnit.find(MI)->second;
799 
800   for (auto I : CurrentPacketMIs) {
801     // We only care for dependencies to predicated instructions
802     if (!HII->isPredicated(*I))
803       continue;
804 
805     // Scheduling Unit for current insn in the packet
806     SUnit *PacketSU = MIToSUnit.find(I)->second;
807 
808     // Look at dependencies between current members of the packet and
809     // predicate defining instruction MI. Make sure that dependency is
810     // on the exact register we care about.
811     if (PacketSU->isSucc(PacketSUDep)) {
812       for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
813         auto &Dep = PacketSU->Succs[i];
814         if (Dep.getSUnit() == PacketSUDep && Dep.getKind() == SDep::Anti &&
815             Dep.getReg() == DepReg)
816           return true;
817       }
818     }
819   }
820 
821   return false;
822 }
823 
824 
825 /// Gets the predicate register of a predicated instruction.
getPredicatedRegister(MachineInstr & MI,const HexagonInstrInfo * QII)826 static unsigned getPredicatedRegister(MachineInstr &MI,
827                                       const HexagonInstrInfo *QII) {
828   /// We use the following rule: The first predicate register that is a use is
829   /// the predicate register of a predicated instruction.
830   assert(QII->isPredicated(MI) && "Must be predicated instruction");
831 
832   for (auto &Op : MI.operands()) {
833     if (Op.isReg() && Op.getReg() && Op.isUse() &&
834         Hexagon::PredRegsRegClass.contains(Op.getReg()))
835       return Op.getReg();
836   }
837 
838   llvm_unreachable("Unknown instruction operand layout");
839   return 0;
840 }
841 
842 // Given two predicated instructions, this function detects whether
843 // the predicates are complements.
arePredicatesComplements(MachineInstr & MI1,MachineInstr & MI2)844 bool HexagonPacketizerList::arePredicatesComplements(MachineInstr &MI1,
845                                                      MachineInstr &MI2) {
846   // If we don't know the predicate sense of the instructions bail out early, we
847   // need it later.
848   if (getPredicateSense(MI1, HII) == PK_Unknown ||
849       getPredicateSense(MI2, HII) == PK_Unknown)
850     return false;
851 
852   // Scheduling unit for candidate.
853   SUnit *SU = MIToSUnit[&MI1];
854 
855   // One corner case deals with the following scenario:
856   // Trying to add
857   // a) %R24<def> = A2_tfrt %P0, %R25
858   // to this packet:
859   // {
860   //   b) %R25<def> = A2_tfrf %P0, %R24
861   //   c) %P0<def> = C2_cmpeqi %R26, 1
862   // }
863   //
864   // On general check a) and b) are complements, but presence of c) will
865   // convert a) to .new form, and then it is not a complement.
866   // We attempt to detect it by analyzing existing dependencies in the packet.
867 
868   // Analyze relationships between all existing members of the packet.
869   // Look for Anti dependecy on the same predicate reg as used in the
870   // candidate.
871   for (auto I : CurrentPacketMIs) {
872     // Scheduling Unit for current insn in the packet.
873     SUnit *PacketSU = MIToSUnit.find(I)->second;
874 
875     // If this instruction in the packet is succeeded by the candidate...
876     if (PacketSU->isSucc(SU)) {
877       for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
878         auto Dep = PacketSU->Succs[i];
879         // The corner case exist when there is true data dependency between
880         // candidate and one of current packet members, this dep is on
881         // predicate reg, and there already exist anti dep on the same pred in
882         // the packet.
883         if (Dep.getSUnit() == SU && Dep.getKind() == SDep::Data &&
884             Hexagon::PredRegsRegClass.contains(Dep.getReg())) {
885           // Here I know that I is predicate setting instruction with true
886           // data dep to candidate on the register we care about - c) in the
887           // above example. Now I need to see if there is an anti dependency
888           // from c) to any other instruction in the same packet on the pred
889           // reg of interest.
890           if (restrictingDepExistInPacket(I, Dep.getReg()))
891             return false;
892         }
893       }
894     }
895   }
896 
897   // If the above case does not apply, check regular complement condition.
898   // Check that the predicate register is the same and that the predicate
899   // sense is different We also need to differentiate .old vs. .new: !p0
900   // is not complementary to p0.new.
901   unsigned PReg1 = getPredicatedRegister(MI1, HII);
902   unsigned PReg2 = getPredicatedRegister(MI2, HII);
903   return PReg1 == PReg2 &&
904          Hexagon::PredRegsRegClass.contains(PReg1) &&
905          Hexagon::PredRegsRegClass.contains(PReg2) &&
906          getPredicateSense(MI1, HII) != getPredicateSense(MI2, HII) &&
907          HII->isDotNewInst(&MI1) == HII->isDotNewInst(&MI2);
908 }
909 
910 // Initialize packetizer flags.
initPacketizerState()911 void HexagonPacketizerList::initPacketizerState() {
912   Dependence = false;
913   PromotedToDotNew = false;
914   GlueToNewValueJump = false;
915   GlueAllocframeStore = false;
916   FoundSequentialDependence = false;
917 }
918 
919 // Ignore bundling of pseudo instructions.
ignorePseudoInstruction(const MachineInstr & MI,const MachineBasicBlock *)920 bool HexagonPacketizerList::ignorePseudoInstruction(const MachineInstr &MI,
921                                                     const MachineBasicBlock *) {
922   if (MI.isDebugValue())
923     return true;
924 
925   if (MI.isCFIInstruction())
926     return false;
927 
928   // We must print out inline assembly.
929   if (MI.isInlineAsm())
930     return false;
931 
932   if (MI.isImplicitDef())
933     return false;
934 
935   // We check if MI has any functional units mapped to it. If it doesn't,
936   // we ignore the instruction.
937   const MCInstrDesc& TID = MI.getDesc();
938   auto *IS = ResourceTracker->getInstrItins()->beginStage(TID.getSchedClass());
939   unsigned FuncUnits = IS->getUnits();
940   return !FuncUnits;
941 }
942 
isSoloInstruction(const MachineInstr & MI)943 bool HexagonPacketizerList::isSoloInstruction(const MachineInstr &MI) {
944   if (MI.isEHLabel() || MI.isCFIInstruction())
945     return true;
946 
947   // Consider inline asm to not be a solo instruction by default.
948   // Inline asm will be put in a packet temporarily, but then it will be
949   // removed, and placed outside of the packet (before or after, depending
950   // on dependencies).  This is to reduce the impact of inline asm as a
951   // "packet splitting" instruction.
952   if (MI.isInlineAsm() && !ScheduleInlineAsm)
953     return true;
954 
955   // From Hexagon V4 Programmer's Reference Manual 3.4.4 Grouping constraints:
956   // trap, pause, barrier, icinva, isync, and syncht are solo instructions.
957   // They must not be grouped with other instructions in a packet.
958   if (isSchedBarrier(&MI))
959     return true;
960 
961   if (HII->isSolo(&MI))
962     return true;
963 
964   if (MI.getOpcode() == Hexagon::A2_nop)
965     return true;
966 
967   return false;
968 }
969 
970 
971 // Quick check if instructions MI and MJ cannot coexist in the same packet.
972 // Limit the tests to be "one-way", e.g.  "if MI->isBranch and MJ->isInlineAsm",
973 // but not the symmetric case: "if MJ->isBranch and MI->isInlineAsm".
974 // For full test call this function twice:
975 //   cannotCoexistAsymm(MI, MJ) || cannotCoexistAsymm(MJ, MI)
976 // Doing the test only one way saves the amount of code in this function,
977 // since every test would need to be repeated with the MI and MJ reversed.
cannotCoexistAsymm(const MachineInstr * MI,const MachineInstr * MJ,const HexagonInstrInfo & HII)978 static bool cannotCoexistAsymm(const MachineInstr *MI, const MachineInstr *MJ,
979       const HexagonInstrInfo &HII) {
980   const MachineFunction *MF = MI->getParent()->getParent();
981   if (MF->getSubtarget<HexagonSubtarget>().hasV60TOpsOnly() &&
982       HII.isHVXMemWithAIndirect(MI, MJ))
983     return true;
984 
985   // An inline asm cannot be together with a branch, because we may not be
986   // able to remove the asm out after packetizing (i.e. if the asm must be
987   // moved past the bundle).  Similarly, two asms cannot be together to avoid
988   // complications when determining their relative order outside of a bundle.
989   if (MI->isInlineAsm())
990     return MJ->isInlineAsm() || MJ->isBranch() || MJ->isBarrier() ||
991            MJ->isCall() || MJ->isTerminator();
992 
993   // "False" really means that the quick check failed to determine if
994   // I and J cannot coexist.
995   return false;
996 }
997 
998 
999 // Full, symmetric check.
cannotCoexist(const MachineInstr * MI,const MachineInstr * MJ)1000 bool HexagonPacketizerList::cannotCoexist(const MachineInstr *MI,
1001       const MachineInstr *MJ) {
1002   return cannotCoexistAsymm(MI, MJ, *HII) || cannotCoexistAsymm(MJ, MI, *HII);
1003 }
1004 
unpacketizeSoloInstrs(MachineFunction & MF)1005 void HexagonPacketizerList::unpacketizeSoloInstrs(MachineFunction &MF) {
1006   for (auto &B : MF) {
1007     MachineBasicBlock::iterator BundleIt;
1008     MachineBasicBlock::instr_iterator NextI;
1009     for (auto I = B.instr_begin(), E = B.instr_end(); I != E; I = NextI) {
1010       NextI = std::next(I);
1011       MachineInstr *MI = &*I;
1012       if (MI->isBundle())
1013         BundleIt = I;
1014       if (!MI->isInsideBundle())
1015         continue;
1016 
1017       // Decide on where to insert the instruction that we are pulling out.
1018       // Debug instructions always go before the bundle, but the placement of
1019       // INLINE_ASM depends on potential dependencies.  By default, try to
1020       // put it before the bundle, but if the asm writes to a register that
1021       // other instructions in the bundle read, then we need to place it
1022       // after the bundle (to preserve the bundle semantics).
1023       bool InsertBeforeBundle;
1024       if (MI->isInlineAsm())
1025         InsertBeforeBundle = !hasWriteToReadDep(*MI, *BundleIt, HRI);
1026       else if (MI->isDebugValue())
1027         InsertBeforeBundle = true;
1028       else
1029         continue;
1030 
1031       BundleIt = moveInstrOut(MI, BundleIt, InsertBeforeBundle);
1032     }
1033   }
1034 }
1035 
1036 // Check if a given instruction is of class "system".
isSystemInstr(const MachineInstr * MI)1037 static bool isSystemInstr(const MachineInstr *MI) {
1038   unsigned Opc = MI->getOpcode();
1039   switch (Opc) {
1040     case Hexagon::Y2_barrier:
1041     case Hexagon::Y2_dcfetchbo:
1042       return true;
1043   }
1044   return false;
1045 }
1046 
hasDeadDependence(const MachineInstr * I,const MachineInstr * J)1047 bool HexagonPacketizerList::hasDeadDependence(const MachineInstr *I,
1048                                               const MachineInstr *J) {
1049   // The dependence graph may not include edges between dead definitions,
1050   // so without extra checks, we could end up packetizing two instruction
1051   // defining the same (dead) register.
1052   if (I->isCall() || J->isCall())
1053     return false;
1054   if (HII->isPredicated(*I) || HII->isPredicated(*J))
1055     return false;
1056 
1057   BitVector DeadDefs(Hexagon::NUM_TARGET_REGS);
1058   for (auto &MO : I->operands()) {
1059     if (!MO.isReg() || !MO.isDef() || !MO.isDead())
1060       continue;
1061     DeadDefs[MO.getReg()] = true;
1062   }
1063 
1064   for (auto &MO : J->operands()) {
1065     if (!MO.isReg() || !MO.isDef() || !MO.isDead())
1066       continue;
1067     unsigned R = MO.getReg();
1068     if (R != Hexagon::USR_OVF && DeadDefs[R])
1069       return true;
1070   }
1071   return false;
1072 }
1073 
hasControlDependence(const MachineInstr * I,const MachineInstr * J)1074 bool HexagonPacketizerList::hasControlDependence(const MachineInstr *I,
1075                                                  const MachineInstr *J) {
1076   // A save callee-save register function call can only be in a packet
1077   // with instructions that don't write to the callee-save registers.
1078   if ((HII->isSaveCalleeSavedRegsCall(I) &&
1079        doesModifyCalleeSavedReg(J, HRI)) ||
1080       (HII->isSaveCalleeSavedRegsCall(J) &&
1081        doesModifyCalleeSavedReg(I, HRI)))
1082     return true;
1083 
1084   // Two control flow instructions cannot go in the same packet.
1085   if (isControlFlow(I) && isControlFlow(J))
1086     return true;
1087 
1088   // \ref-manual (7.3.4) A loop setup packet in loopN or spNloop0 cannot
1089   // contain a speculative indirect jump,
1090   // a new-value compare jump or a dealloc_return.
1091   auto isBadForLoopN = [this] (const MachineInstr *MI) -> bool {
1092     if (MI->isCall() || HII->isDeallocRet(MI) || HII->isNewValueJump(MI))
1093       return true;
1094     if (HII->isPredicated(*MI) && HII->isPredicatedNew(*MI) && HII->isJumpR(MI))
1095       return true;
1096     return false;
1097   };
1098 
1099   if (HII->isLoopN(I) && isBadForLoopN(J))
1100     return true;
1101   if (HII->isLoopN(J) && isBadForLoopN(I))
1102     return true;
1103 
1104   // dealloc_return cannot appear in the same packet as a conditional or
1105   // unconditional jump.
1106   return HII->isDeallocRet(I) &&
1107          (J->isBranch() || J->isCall() || J->isBarrier());
1108 }
1109 
hasV4SpecificDependence(const MachineInstr * I,const MachineInstr * J)1110 bool HexagonPacketizerList::hasV4SpecificDependence(const MachineInstr *I,
1111                                                     const MachineInstr *J) {
1112   bool SysI = isSystemInstr(I), SysJ = isSystemInstr(J);
1113   bool StoreI = I->mayStore(), StoreJ = J->mayStore();
1114   if ((SysI && StoreJ) || (SysJ && StoreI))
1115     return true;
1116 
1117   if (StoreI && StoreJ) {
1118     if (HII->isNewValueInst(J) || HII->isMemOp(J) || HII->isMemOp(I))
1119       return true;
1120   } else {
1121     // A memop cannot be in the same packet with another memop or a store.
1122     // Two stores can be together, but here I and J cannot both be stores.
1123     bool MopStI = HII->isMemOp(I) || StoreI;
1124     bool MopStJ = HII->isMemOp(J) || StoreJ;
1125     if (MopStI && MopStJ)
1126       return true;
1127   }
1128 
1129   return (StoreJ && HII->isDeallocRet(I)) || (StoreI && HII->isDeallocRet(J));
1130 }
1131 
1132 // SUI is the current instruction that is out side of the current packet.
1133 // SUJ is the current instruction inside the current packet against which that
1134 // SUI will be packetized.
isLegalToPacketizeTogether(SUnit * SUI,SUnit * SUJ)1135 bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) {
1136   MachineInstr *I = SUI->getInstr();
1137   MachineInstr *J = SUJ->getInstr();
1138   assert(I && J && "Unable to packetize null instruction!");
1139 
1140   // Clear IgnoreDepMIs when Packet starts.
1141   if (CurrentPacketMIs.size() == 1)
1142     IgnoreDepMIs.clear();
1143 
1144   MachineBasicBlock::iterator II = I;
1145   const unsigned FrameSize = MF.getFrameInfo()->getStackSize();
1146 
1147   // Solo instructions cannot go in the packet.
1148   assert(!isSoloInstruction(*I) && "Unexpected solo instr!");
1149 
1150   if (cannotCoexist(I, J))
1151     return false;
1152 
1153   Dependence = hasDeadDependence(I, J) || hasControlDependence(I, J);
1154   if (Dependence)
1155     return false;
1156 
1157   // V4 allows dual stores. It does not allow second store, if the first
1158   // store is not in SLOT0. New value store, new value jump, dealloc_return
1159   // and memop always take SLOT0. Arch spec 3.4.4.2.
1160   Dependence = hasV4SpecificDependence(I, J);
1161   if (Dependence)
1162     return false;
1163 
1164   // If an instruction feeds new value jump, glue it.
1165   MachineBasicBlock::iterator NextMII = I;
1166   ++NextMII;
1167   if (NextMII != I->getParent()->end() && HII->isNewValueJump(&*NextMII)) {
1168     MachineInstr &NextMI = *NextMII;
1169 
1170     bool secondRegMatch = false;
1171     const MachineOperand &NOp0 = NextMI.getOperand(0);
1172     const MachineOperand &NOp1 = NextMI.getOperand(1);
1173 
1174     if (NOp1.isReg() && I->getOperand(0).getReg() == NOp1.getReg())
1175       secondRegMatch = true;
1176 
1177     for (auto I : CurrentPacketMIs) {
1178       SUnit *PacketSU = MIToSUnit.find(I)->second;
1179       MachineInstr *PI = PacketSU->getInstr();
1180       // NVJ can not be part of the dual jump - Arch Spec: section 7.8.
1181       if (PI->isCall()) {
1182         Dependence = true;
1183         break;
1184       }
1185       // Validate:
1186       // 1. Packet does not have a store in it.
1187       // 2. If the first operand of the nvj is newified, and the second
1188       //    operand is also a reg, it (second reg) is not defined in
1189       //    the same packet.
1190       // 3. If the second operand of the nvj is newified, (which means
1191       //    first operand is also a reg), first reg is not defined in
1192       //    the same packet.
1193       if (PI->getOpcode() == Hexagon::S2_allocframe || PI->mayStore() ||
1194           HII->isLoopN(PI)) {
1195         Dependence = true;
1196         break;
1197       }
1198       // Check #2/#3.
1199       const MachineOperand &OpR = secondRegMatch ? NOp0 : NOp1;
1200       if (OpR.isReg() && PI->modifiesRegister(OpR.getReg(), HRI)) {
1201         Dependence = true;
1202         break;
1203       }
1204     }
1205 
1206     if (Dependence)
1207       return false;
1208     GlueToNewValueJump = true;
1209   }
1210 
1211   // There no dependency between a prolog instruction and its successor.
1212   if (!SUJ->isSucc(SUI))
1213     return true;
1214 
1215   for (unsigned i = 0; i < SUJ->Succs.size(); ++i) {
1216     if (FoundSequentialDependence)
1217       break;
1218 
1219     if (SUJ->Succs[i].getSUnit() != SUI)
1220       continue;
1221 
1222     SDep::Kind DepType = SUJ->Succs[i].getKind();
1223     // For direct calls:
1224     // Ignore register dependences for call instructions for packetization
1225     // purposes except for those due to r31 and predicate registers.
1226     //
1227     // For indirect calls:
1228     // Same as direct calls + check for true dependences to the register
1229     // used in the indirect call.
1230     //
1231     // We completely ignore Order dependences for call instructions.
1232     //
1233     // For returns:
1234     // Ignore register dependences for return instructions like jumpr,
1235     // dealloc return unless we have dependencies on the explicit uses
1236     // of the registers used by jumpr (like r31) or dealloc return
1237     // (like r29 or r30).
1238     //
1239     // TODO: Currently, jumpr is handling only return of r31. So, the
1240     // following logic (specificaly isCallDependent) is working fine.
1241     // We need to enable jumpr for register other than r31 and then,
1242     // we need to rework the last part, where it handles indirect call
1243     // of that (isCallDependent) function. Bug 6216 is opened for this.
1244     unsigned DepReg = 0;
1245     const TargetRegisterClass *RC = nullptr;
1246     if (DepType == SDep::Data) {
1247       DepReg = SUJ->Succs[i].getReg();
1248       RC = HRI->getMinimalPhysRegClass(DepReg);
1249     }
1250 
1251     if (I->isCall() || I->isReturn() || HII->isTailCall(I)) {
1252       if (!isRegDependence(DepType))
1253         continue;
1254       if (!isCallDependent(I, DepType, SUJ->Succs[i].getReg()))
1255         continue;
1256     }
1257 
1258     if (DepType == SDep::Data) {
1259       if (canPromoteToDotCur(J, SUJ, DepReg, II, RC))
1260         if (promoteToDotCur(J, DepType, II, RC))
1261           continue;
1262     }
1263 
1264     // Data dpendence ok if we have load.cur.
1265     if (DepType == SDep::Data && HII->isDotCurInst(J)) {
1266       if (HII->isV60VectorInstruction(I))
1267         continue;
1268     }
1269 
1270     // For instructions that can be promoted to dot-new, try to promote.
1271     if (DepType == SDep::Data) {
1272       if (canPromoteToDotNew(I, SUJ, DepReg, II, RC)) {
1273         if (promoteToDotNew(I, DepType, II, RC)) {
1274           PromotedToDotNew = true;
1275           continue;
1276         }
1277       }
1278       if (HII->isNewValueJump(I))
1279         continue;
1280     }
1281 
1282     // For predicated instructions, if the predicates are complements then
1283     // there can be no dependence.
1284     if (HII->isPredicated(*I) && HII->isPredicated(*J) &&
1285         arePredicatesComplements(*I, *J)) {
1286       // Not always safe to do this translation.
1287       // DAG Builder attempts to reduce dependence edges using transitive
1288       // nature of dependencies. Here is an example:
1289       //
1290       // r0 = tfr_pt ... (1)
1291       // r0 = tfr_pf ... (2)
1292       // r0 = tfr_pt ... (3)
1293       //
1294       // There will be an output dependence between (1)->(2) and (2)->(3).
1295       // However, there is no dependence edge between (1)->(3). This results
1296       // in all 3 instructions going in the same packet. We ignore dependce
1297       // only once to avoid this situation.
1298       auto Itr = std::find(IgnoreDepMIs.begin(), IgnoreDepMIs.end(), J);
1299       if (Itr != IgnoreDepMIs.end()) {
1300         Dependence = true;
1301         return false;
1302       }
1303       IgnoreDepMIs.push_back(I);
1304       continue;
1305     }
1306 
1307     // Ignore Order dependences between unconditional direct branches
1308     // and non-control-flow instructions.
1309     if (isDirectJump(I) && !J->isBranch() && !J->isCall() &&
1310         DepType == SDep::Order)
1311       continue;
1312 
1313     // Ignore all dependences for jumps except for true and output
1314     // dependences.
1315     if (I->isConditionalBranch() && DepType != SDep::Data &&
1316         DepType != SDep::Output)
1317       continue;
1318 
1319     // Ignore output dependences due to superregs. We can write to two
1320     // different subregisters of R1:0 for instance in the same cycle.
1321 
1322     // If neither I nor J defines DepReg, then this is a superfluous output
1323     // dependence. The dependence must be of the form:
1324     //   R0 = ...
1325     //   R1 = ...
1326     // and there is an output dependence between the two instructions with
1327     // DepReg = D0.
1328     // We want to ignore these dependences. Ideally, the dependence
1329     // constructor should annotate such dependences. We can then avoid this
1330     // relatively expensive check.
1331     //
1332     if (DepType == SDep::Output) {
1333       // DepReg is the register that's responsible for the dependence.
1334       unsigned DepReg = SUJ->Succs[i].getReg();
1335 
1336       // Check if I and J really defines DepReg.
1337       if (!I->definesRegister(DepReg) && !J->definesRegister(DepReg))
1338         continue;
1339       FoundSequentialDependence = true;
1340       break;
1341     }
1342 
1343     // For Order dependences:
1344     // 1. On V4 or later, volatile loads/stores can be packetized together,
1345     //    unless other rules prevent is.
1346     // 2. Store followed by a load is not allowed.
1347     // 3. Store followed by a store is only valid on V4 or later.
1348     // 4. Load followed by any memory operation is allowed.
1349     if (DepType == SDep::Order) {
1350       if (!PacketizeVolatiles) {
1351         bool OrdRefs = I->hasOrderedMemoryRef() || J->hasOrderedMemoryRef();
1352         if (OrdRefs) {
1353           FoundSequentialDependence = true;
1354           break;
1355         }
1356       }
1357       // J is first, I is second.
1358       bool LoadJ = J->mayLoad(), StoreJ = J->mayStore();
1359       bool LoadI = I->mayLoad(), StoreI = I->mayStore();
1360       if (StoreJ) {
1361         // Two stores are only allowed on V4+. Load following store is never
1362         // allowed.
1363         if (LoadI) {
1364           FoundSequentialDependence = true;
1365           break;
1366         }
1367       } else if (!LoadJ || (!LoadI && !StoreI)) {
1368         // If J is neither load nor store, assume a dependency.
1369         // If J is a load, but I is neither, also assume a dependency.
1370         FoundSequentialDependence = true;
1371         break;
1372       }
1373       // Store followed by store: not OK on V2.
1374       // Store followed by load: not OK on all.
1375       // Load followed by store: OK on all.
1376       // Load followed by load: OK on all.
1377       continue;
1378     }
1379 
1380     // For V4, special case ALLOCFRAME. Even though there is dependency
1381     // between ALLOCFRAME and subsequent store, allow it to be packetized
1382     // in a same packet. This implies that the store is using the caller's
1383     // SP. Hence, offset needs to be updated accordingly.
1384     if (DepType == SDep::Data && J->getOpcode() == Hexagon::S2_allocframe) {
1385       unsigned Opc = I->getOpcode();
1386       switch (Opc) {
1387         case Hexagon::S2_storerd_io:
1388         case Hexagon::S2_storeri_io:
1389         case Hexagon::S2_storerh_io:
1390         case Hexagon::S2_storerb_io:
1391           if (I->getOperand(0).getReg() == HRI->getStackRegister()) {
1392             int64_t Imm = I->getOperand(1).getImm();
1393             int64_t NewOff = Imm - (FrameSize + HEXAGON_LRFP_SIZE);
1394             if (HII->isValidOffset(Opc, NewOff)) {
1395               GlueAllocframeStore = true;
1396               // Since this store is to be glued with allocframe in the same
1397               // packet, it will use SP of the previous stack frame, i.e.
1398               // caller's SP. Therefore, we need to recalculate offset
1399               // according to this change.
1400               I->getOperand(1).setImm(NewOff);
1401               continue;
1402             }
1403           }
1404         default:
1405           break;
1406       }
1407     }
1408 
1409     // There are certain anti-dependencies that cannot be ignored.
1410     // Specifically:
1411     //   J2_call ... %R0<imp-def>   ; SUJ
1412     //   R0 = ...                   ; SUI
1413     // Those cannot be packetized together, since the call will observe
1414     // the effect of the assignment to R0.
1415     if (DepType == SDep::Anti && J->isCall()) {
1416       // Check if I defines any volatile register. We should also check
1417       // registers that the call may read, but these happen to be a
1418       // subset of the volatile register set.
1419       for (const MCPhysReg *P = J->getDesc().ImplicitDefs; P && *P; ++P) {
1420         if (!I->modifiesRegister(*P, HRI))
1421           continue;
1422         FoundSequentialDependence = true;
1423         break;
1424       }
1425     }
1426 
1427     // Skip over remaining anti-dependences. Two instructions that are
1428     // anti-dependent can share a packet, since in most such cases all
1429     // operands are read before any modifications take place.
1430     // The exceptions are branch and call instructions, since they are
1431     // executed after all other instructions have completed (at least
1432     // conceptually).
1433     if (DepType != SDep::Anti) {
1434       FoundSequentialDependence = true;
1435       break;
1436     }
1437   }
1438 
1439   if (FoundSequentialDependence) {
1440     Dependence = true;
1441     return false;
1442   }
1443 
1444   return true;
1445 }
1446 
isLegalToPruneDependencies(SUnit * SUI,SUnit * SUJ)1447 bool HexagonPacketizerList::isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) {
1448   MachineInstr *I = SUI->getInstr();
1449   MachineInstr *J = SUJ->getInstr();
1450   assert(I && J && "Unable to packetize null instruction!");
1451 
1452   if (cannotCoexist(I, J))
1453     return false;
1454 
1455   if (!Dependence)
1456     return true;
1457 
1458   // Check if the instruction was promoted to a dot-new. If so, demote it
1459   // back into a dot-old.
1460   if (PromotedToDotNew)
1461     demoteToDotOld(I);
1462 
1463   cleanUpDotCur();
1464   // Check if the instruction (must be a store) was glued with an allocframe
1465   // instruction. If so, restore its offset to its original value, i.e. use
1466   // current SP instead of caller's SP.
1467   if (GlueAllocframeStore) {
1468     unsigned FrameSize = MF.getFrameInfo()->getStackSize();
1469     MachineOperand &MOff = I->getOperand(1);
1470     MOff.setImm(MOff.getImm() + FrameSize + HEXAGON_LRFP_SIZE);
1471   }
1472   return false;
1473 }
1474 
1475 MachineBasicBlock::iterator
addToPacket(MachineInstr & MI)1476 HexagonPacketizerList::addToPacket(MachineInstr &MI) {
1477   MachineBasicBlock::iterator MII = MI;
1478   MachineBasicBlock *MBB = MI.getParent();
1479   if (MI.isImplicitDef()) {
1480     unsigned R = MI.getOperand(0).getReg();
1481     if (Hexagon::IntRegsRegClass.contains(R)) {
1482       MCSuperRegIterator S(R, HRI, false);
1483       MI.addOperand(MachineOperand::CreateReg(*S, true, true));
1484     }
1485     return MII;
1486   }
1487   assert(ResourceTracker->canReserveResources(MI));
1488 
1489   bool ExtMI = HII->isExtended(&MI) || HII->isConstExtended(&MI);
1490   bool Good = true;
1491 
1492   if (GlueToNewValueJump) {
1493     MachineInstr &NvjMI = *++MII;
1494     // We need to put both instructions in the same packet: MI and NvjMI.
1495     // Either of them can require a constant extender. Try to add both to
1496     // the current packet, and if that fails, end the packet and start a
1497     // new one.
1498     ResourceTracker->reserveResources(MI);
1499     if (ExtMI)
1500       Good = tryAllocateResourcesForConstExt(true);
1501 
1502     bool ExtNvjMI = HII->isExtended(&NvjMI) || HII->isConstExtended(&NvjMI);
1503     if (Good) {
1504       if (ResourceTracker->canReserveResources(NvjMI))
1505         ResourceTracker->reserveResources(NvjMI);
1506       else
1507         Good = false;
1508     }
1509     if (Good && ExtNvjMI)
1510       Good = tryAllocateResourcesForConstExt(true);
1511 
1512     if (!Good) {
1513       endPacket(MBB, MI);
1514       assert(ResourceTracker->canReserveResources(MI));
1515       ResourceTracker->reserveResources(MI);
1516       if (ExtMI) {
1517         assert(canReserveResourcesForConstExt());
1518         tryAllocateResourcesForConstExt(true);
1519       }
1520       assert(ResourceTracker->canReserveResources(NvjMI));
1521       ResourceTracker->reserveResources(NvjMI);
1522       if (ExtNvjMI) {
1523         assert(canReserveResourcesForConstExt());
1524         reserveResourcesForConstExt();
1525       }
1526     }
1527     CurrentPacketMIs.push_back(&MI);
1528     CurrentPacketMIs.push_back(&NvjMI);
1529     return MII;
1530   }
1531 
1532   ResourceTracker->reserveResources(MI);
1533   if (ExtMI && !tryAllocateResourcesForConstExt(true)) {
1534     endPacket(MBB, MI);
1535     if (PromotedToDotNew)
1536       demoteToDotOld(&MI);
1537     ResourceTracker->reserveResources(MI);
1538     reserveResourcesForConstExt();
1539   }
1540 
1541   CurrentPacketMIs.push_back(&MI);
1542   return MII;
1543 }
1544 
endPacket(MachineBasicBlock * MBB,MachineBasicBlock::iterator MI)1545 void HexagonPacketizerList::endPacket(MachineBasicBlock *MBB,
1546                                       MachineBasicBlock::iterator MI) {
1547   OldPacketMIs = CurrentPacketMIs;
1548   VLIWPacketizerList::endPacket(MBB, MI);
1549 }
1550 
shouldAddToPacket(const MachineInstr & MI)1551 bool HexagonPacketizerList::shouldAddToPacket(const MachineInstr &MI) {
1552   return !producesStall(&MI);
1553 }
1554 
1555 
1556 // Return true when ConsMI uses a register defined by ProdMI.
isDependent(const MachineInstr * ProdMI,const MachineInstr * ConsMI)1557 static bool isDependent(const MachineInstr *ProdMI,
1558       const MachineInstr *ConsMI) {
1559   if (!ProdMI->getOperand(0).isReg())
1560     return false;
1561   unsigned DstReg = ProdMI->getOperand(0).getReg();
1562 
1563   for (auto &Op : ConsMI->operands())
1564     if (Op.isReg() && Op.isUse() && Op.getReg() == DstReg)
1565       // The MIs depend on each other.
1566       return true;
1567 
1568   return false;
1569 }
1570 
1571 // V60 forward scheduling.
producesStall(const MachineInstr * I)1572 bool HexagonPacketizerList::producesStall(const MachineInstr *I) {
1573   // Check whether the previous packet is in a different loop. If this is the
1574   // case, there is little point in trying to avoid a stall because that would
1575   // favor the rare case (loop entry) over the common case (loop iteration).
1576   //
1577   // TODO: We should really be able to check all the incoming edges if this is
1578   // the first packet in a basic block, so we can avoid stalls from the loop
1579   // backedge.
1580   if (!OldPacketMIs.empty()) {
1581     auto *OldBB = OldPacketMIs.front()->getParent();
1582     auto *ThisBB = I->getParent();
1583     if (MLI->getLoopFor(OldBB) != MLI->getLoopFor(ThisBB))
1584       return false;
1585   }
1586 
1587   // Check for stall between two vector instructions.
1588   if (HII->isV60VectorInstruction(I)) {
1589     for (auto J : OldPacketMIs) {
1590       if (!HII->isV60VectorInstruction(J))
1591         continue;
1592       if (isDependent(J, I) && !HII->isVecUsableNextPacket(J, I))
1593         return true;
1594     }
1595     return false;
1596   }
1597 
1598   // Check for stall between two scalar instructions. First, check that
1599   // there is no definition of a use in the current packet, because it
1600   // may be a candidate for .new.
1601   for (auto J : CurrentPacketMIs)
1602     if (!HII->isV60VectorInstruction(J) && isDependent(J, I))
1603       return false;
1604 
1605   // Check for stall between I and instructions in the previous packet.
1606   if (MF.getSubtarget<HexagonSubtarget>().useBSBScheduling()) {
1607     for (auto J : OldPacketMIs) {
1608       if (HII->isV60VectorInstruction(J))
1609         continue;
1610       if (!HII->isLateInstrFeedsEarlyInstr(J, I))
1611         continue;
1612       if (isDependent(J, I) && !HII->canExecuteInBundle(J, I))
1613         return true;
1614     }
1615   }
1616 
1617   return false;
1618 }
1619 
1620 
1621 //===----------------------------------------------------------------------===//
1622 //                         Public Constructor Functions
1623 //===----------------------------------------------------------------------===//
1624 
createHexagonPacketizer()1625 FunctionPass *llvm::createHexagonPacketizer() {
1626   return new HexagonPacketizer();
1627 }
1628