xref: /freebsd/contrib/llvm-project/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp (revision 5fb307d29b364982acbde82cbf77db3cae486f8c)
1 //===- HexagonHardwareLoops.cpp - Identify and generate hardware loops ----===//
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 pass identifies loops where we can generate the Hexagon hardware
10 // loop instruction.  The hardware loop can perform loop branches with a
11 // zero-cycle overhead.
12 //
13 // The pattern that defines the induction variable can changed depending on
14 // prior optimizations.  For example, the IndVarSimplify phase run by 'opt'
15 // normalizes induction variables, and the Loop Strength Reduction pass
16 // run by 'llc' may also make changes to the induction variable.
17 // The pattern detected by this phase is due to running Strength Reduction.
18 //
19 // Criteria for hardware loops:
20 //  - Countable loops (w/ ind. var for a trip count)
21 //  - Assumes loops are normalized by IndVarSimplify
22 //  - Try inner-most loops first
23 //  - No function calls in loops.
24 //
25 //===----------------------------------------------------------------------===//
26 
27 #include "HexagonInstrInfo.h"
28 #include "HexagonSubtarget.h"
29 #include "llvm/ADT/ArrayRef.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/StringRef.h"
35 #include "llvm/CodeGen/MachineBasicBlock.h"
36 #include "llvm/CodeGen/MachineDominators.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineFunctionPass.h"
39 #include "llvm/CodeGen/MachineInstr.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineLoopInfo.h"
42 #include "llvm/CodeGen/MachineOperand.h"
43 #include "llvm/CodeGen/MachineRegisterInfo.h"
44 #include "llvm/CodeGen/TargetRegisterInfo.h"
45 #include "llvm/IR/Constants.h"
46 #include "llvm/IR/DebugLoc.h"
47 #include "llvm/InitializePasses.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include <cassert>
55 #include <cstdint>
56 #include <cstdlib>
57 #include <iterator>
58 #include <map>
59 #include <set>
60 #include <string>
61 #include <utility>
62 #include <vector>
63 
64 using namespace llvm;
65 
66 #define DEBUG_TYPE "hwloops"
67 
68 #ifndef NDEBUG
69 static cl::opt<int> HWLoopLimit("hexagon-max-hwloop", cl::Hidden, cl::init(-1));
70 
71 // Option to create preheader only for a specific function.
72 static cl::opt<std::string> PHFn("hexagon-hwloop-phfn", cl::Hidden,
73                                  cl::init(""));
74 #endif
75 
76 // Option to create a preheader if one doesn't exist.
77 static cl::opt<bool> HWCreatePreheader("hexagon-hwloop-preheader",
78     cl::Hidden, cl::init(true),
79     cl::desc("Add a preheader to a hardware loop if one doesn't exist"));
80 
81 // Turn it off by default. If a preheader block is not created here, the
82 // software pipeliner may be unable to find a block suitable to serve as
83 // a preheader. In that case SWP will not run.
84 static cl::opt<bool> SpecPreheader("hwloop-spec-preheader", cl::Hidden,
85                                    cl::desc("Allow speculation of preheader "
86                                             "instructions"));
87 
88 STATISTIC(NumHWLoops, "Number of loops converted to hardware loops");
89 
90 namespace llvm {
91 
92   FunctionPass *createHexagonHardwareLoops();
93   void initializeHexagonHardwareLoopsPass(PassRegistry&);
94 
95 } // end namespace llvm
96 
97 namespace {
98 
99   class CountValue;
100 
101   struct HexagonHardwareLoops : public MachineFunctionPass {
102     MachineLoopInfo            *MLI;
103     MachineRegisterInfo        *MRI;
104     MachineDominatorTree       *MDT;
105     const HexagonInstrInfo     *TII;
106     const HexagonRegisterInfo  *TRI;
107 #ifndef NDEBUG
108     static int Counter;
109 #endif
110 
111   public:
112     static char ID;
113 
114     HexagonHardwareLoops() : MachineFunctionPass(ID) {}
115 
116     bool runOnMachineFunction(MachineFunction &MF) override;
117 
118     StringRef getPassName() const override { return "Hexagon Hardware Loops"; }
119 
120     void getAnalysisUsage(AnalysisUsage &AU) const override {
121       AU.addRequired<MachineDominatorTree>();
122       AU.addRequired<MachineLoopInfo>();
123       MachineFunctionPass::getAnalysisUsage(AU);
124     }
125 
126   private:
127     using LoopFeederMap = std::map<Register, MachineInstr *>;
128 
129     /// Kinds of comparisons in the compare instructions.
130     struct Comparison {
131       enum Kind {
132         EQ  = 0x01,
133         NE  = 0x02,
134         L   = 0x04,
135         G   = 0x08,
136         U   = 0x40,
137         LTs = L,
138         LEs = L | EQ,
139         GTs = G,
140         GEs = G | EQ,
141         LTu = L      | U,
142         LEu = L | EQ | U,
143         GTu = G      | U,
144         GEu = G | EQ | U
145       };
146 
147       static Kind getSwappedComparison(Kind Cmp) {
148         assert ((!((Cmp & L) && (Cmp & G))) && "Malformed comparison operator");
149         if ((Cmp & L) || (Cmp & G))
150           return (Kind)(Cmp ^ (L|G));
151         return Cmp;
152       }
153 
154       static Kind getNegatedComparison(Kind Cmp) {
155         if ((Cmp & L) || (Cmp & G))
156           return (Kind)((Cmp ^ (L | G)) ^ EQ);
157         if ((Cmp & NE) || (Cmp & EQ))
158           return (Kind)(Cmp ^ (EQ | NE));
159         return (Kind)0;
160       }
161 
162       static bool isSigned(Kind Cmp) {
163         return (Cmp & (L | G) && !(Cmp & U));
164       }
165 
166       static bool isUnsigned(Kind Cmp) {
167         return (Cmp & U);
168       }
169     };
170 
171     /// Find the register that contains the loop controlling
172     /// induction variable.
173     /// If successful, it will return true and set the \p Reg, \p IVBump
174     /// and \p IVOp arguments.  Otherwise it will return false.
175     /// The returned induction register is the register R that follows the
176     /// following induction pattern:
177     /// loop:
178     ///   R = phi ..., [ R.next, LatchBlock ]
179     ///   R.next = R + #bump
180     ///   if (R.next < #N) goto loop
181     /// IVBump is the immediate value added to R, and IVOp is the instruction
182     /// "R.next = R + #bump".
183     bool findInductionRegister(MachineLoop *L, Register &Reg,
184                                int64_t &IVBump, MachineInstr *&IVOp) const;
185 
186     /// Return the comparison kind for the specified opcode.
187     Comparison::Kind getComparisonKind(unsigned CondOpc,
188                                        MachineOperand *InitialValue,
189                                        const MachineOperand *Endvalue,
190                                        int64_t IVBump) const;
191 
192     /// Analyze the statements in a loop to determine if the loop
193     /// has a computable trip count and, if so, return a value that represents
194     /// the trip count expression.
195     CountValue *getLoopTripCount(MachineLoop *L,
196                                  SmallVectorImpl<MachineInstr *> &OldInsts);
197 
198     /// Return the expression that represents the number of times
199     /// a loop iterates.  The function takes the operands that represent the
200     /// loop start value, loop end value, and induction value.  Based upon
201     /// these operands, the function attempts to compute the trip count.
202     /// If the trip count is not directly available (as an immediate value,
203     /// or a register), the function will attempt to insert computation of it
204     /// to the loop's preheader.
205     CountValue *computeCount(MachineLoop *Loop, const MachineOperand *Start,
206                              const MachineOperand *End, Register IVReg,
207                              int64_t IVBump, Comparison::Kind Cmp) const;
208 
209     /// Return true if the instruction is not valid within a hardware
210     /// loop.
211     bool isInvalidLoopOperation(const MachineInstr *MI,
212                                 bool IsInnerHWLoop) const;
213 
214     /// Return true if the loop contains an instruction that inhibits
215     /// using the hardware loop.
216     bool containsInvalidInstruction(MachineLoop *L, bool IsInnerHWLoop) const;
217 
218     /// Given a loop, check if we can convert it to a hardware loop.
219     /// If so, then perform the conversion and return true.
220     bool convertToHardwareLoop(MachineLoop *L, bool &L0used, bool &L1used);
221 
222     /// Return true if the instruction is now dead.
223     bool isDead(const MachineInstr *MI,
224                 SmallVectorImpl<MachineInstr *> &DeadPhis) const;
225 
226     /// Remove the instruction if it is now dead.
227     void removeIfDead(MachineInstr *MI);
228 
229     /// Make sure that the "bump" instruction executes before the
230     /// compare.  We need that for the IV fixup, so that the compare
231     /// instruction would not use a bumped value that has not yet been
232     /// defined.  If the instructions are out of order, try to reorder them.
233     bool orderBumpCompare(MachineInstr *BumpI, MachineInstr *CmpI);
234 
235     /// Return true if MO and MI pair is visited only once. If visited
236     /// more than once, this indicates there is recursion. In such a case,
237     /// return false.
238     bool isLoopFeeder(MachineLoop *L, MachineBasicBlock *A, MachineInstr *MI,
239                       const MachineOperand *MO,
240                       LoopFeederMap &LoopFeederPhi) const;
241 
242     /// Return true if the Phi may generate a value that may underflow,
243     /// or may wrap.
244     bool phiMayWrapOrUnderflow(MachineInstr *Phi, const MachineOperand *EndVal,
245                                MachineBasicBlock *MBB, MachineLoop *L,
246                                LoopFeederMap &LoopFeederPhi) const;
247 
248     /// Return true if the induction variable may underflow an unsigned
249     /// value in the first iteration.
250     bool loopCountMayWrapOrUnderFlow(const MachineOperand *InitVal,
251                                      const MachineOperand *EndVal,
252                                      MachineBasicBlock *MBB, MachineLoop *L,
253                                      LoopFeederMap &LoopFeederPhi) const;
254 
255     /// Check if the given operand has a compile-time known constant
256     /// value. Return true if yes, and false otherwise. When returning true, set
257     /// Val to the corresponding constant value.
258     bool checkForImmediate(const MachineOperand &MO, int64_t &Val) const;
259 
260     /// Check if the operand has a compile-time known constant value.
261     bool isImmediate(const MachineOperand &MO) const {
262       int64_t V;
263       return checkForImmediate(MO, V);
264     }
265 
266     /// Return the immediate for the specified operand.
267     int64_t getImmediate(const MachineOperand &MO) const {
268       int64_t V;
269       if (!checkForImmediate(MO, V))
270         llvm_unreachable("Invalid operand");
271       return V;
272     }
273 
274     /// Reset the given machine operand to now refer to a new immediate
275     /// value.  Assumes that the operand was already referencing an immediate
276     /// value, either directly, or via a register.
277     void setImmediate(MachineOperand &MO, int64_t Val);
278 
279     /// Fix the data flow of the induction variable.
280     /// The desired flow is: phi ---> bump -+-> comparison-in-latch.
281     ///                                     |
282     ///                                     +-> back to phi
283     /// where "bump" is the increment of the induction variable:
284     ///   iv = iv + #const.
285     /// Due to some prior code transformations, the actual flow may look
286     /// like this:
287     ///   phi -+-> bump ---> back to phi
288     ///        |
289     ///        +-> comparison-in-latch (against upper_bound-bump),
290     /// i.e. the comparison that controls the loop execution may be using
291     /// the value of the induction variable from before the increment.
292     ///
293     /// Return true if the loop's flow is the desired one (i.e. it's
294     /// either been fixed, or no fixing was necessary).
295     /// Otherwise, return false.  This can happen if the induction variable
296     /// couldn't be identified, or if the value in the latch's comparison
297     /// cannot be adjusted to reflect the post-bump value.
298     bool fixupInductionVariable(MachineLoop *L);
299 
300     /// Given a loop, if it does not have a preheader, create one.
301     /// Return the block that is the preheader.
302     MachineBasicBlock *createPreheaderForLoop(MachineLoop *L);
303   };
304 
305   char HexagonHardwareLoops::ID = 0;
306 #ifndef NDEBUG
307   int HexagonHardwareLoops::Counter = 0;
308 #endif
309 
310   /// Abstraction for a trip count of a loop. A smaller version
311   /// of the MachineOperand class without the concerns of changing the
312   /// operand representation.
313   class CountValue {
314   public:
315     enum CountValueType {
316       CV_Register,
317       CV_Immediate
318     };
319 
320   private:
321     CountValueType Kind;
322     union Values {
323       Values() : R{Register(), 0} {}
324       Values(const Values&) = default;
325       struct {
326         Register Reg;
327         unsigned Sub;
328       } R;
329       unsigned ImmVal;
330     } Contents;
331 
332   public:
333     explicit CountValue(CountValueType t, Register v, unsigned u = 0) {
334       Kind = t;
335       if (Kind == CV_Register) {
336         Contents.R.Reg = v;
337         Contents.R.Sub = u;
338       } else {
339         Contents.ImmVal = v;
340       }
341     }
342 
343     bool isReg() const { return Kind == CV_Register; }
344     bool isImm() const { return Kind == CV_Immediate; }
345 
346     Register getReg() const {
347       assert(isReg() && "Wrong CountValue accessor");
348       return Contents.R.Reg;
349     }
350 
351     unsigned getSubReg() const {
352       assert(isReg() && "Wrong CountValue accessor");
353       return Contents.R.Sub;
354     }
355 
356     unsigned getImm() const {
357       assert(isImm() && "Wrong CountValue accessor");
358       return Contents.ImmVal;
359     }
360 
361     void print(raw_ostream &OS, const TargetRegisterInfo *TRI = nullptr) const {
362       if (isReg()) { OS << printReg(Contents.R.Reg, TRI, Contents.R.Sub); }
363       if (isImm()) { OS << Contents.ImmVal; }
364     }
365   };
366 
367 } // end anonymous namespace
368 
369 INITIALIZE_PASS_BEGIN(HexagonHardwareLoops, "hwloops",
370                       "Hexagon Hardware Loops", false, false)
371 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
372 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
373 INITIALIZE_PASS_END(HexagonHardwareLoops, "hwloops",
374                     "Hexagon Hardware Loops", false, false)
375 
376 FunctionPass *llvm::createHexagonHardwareLoops() {
377   return new HexagonHardwareLoops();
378 }
379 
380 bool HexagonHardwareLoops::runOnMachineFunction(MachineFunction &MF) {
381   LLVM_DEBUG(dbgs() << "********* Hexagon Hardware Loops *********\n");
382   if (skipFunction(MF.getFunction()))
383     return false;
384 
385   bool Changed = false;
386 
387   MLI = &getAnalysis<MachineLoopInfo>();
388   MRI = &MF.getRegInfo();
389   MDT = &getAnalysis<MachineDominatorTree>();
390   const HexagonSubtarget &HST = MF.getSubtarget<HexagonSubtarget>();
391   TII = HST.getInstrInfo();
392   TRI = HST.getRegisterInfo();
393 
394   for (auto &L : *MLI)
395     if (L->isOutermost()) {
396       bool L0Used = false;
397       bool L1Used = false;
398       Changed |= convertToHardwareLoop(L, L0Used, L1Used);
399     }
400 
401   return Changed;
402 }
403 
404 bool HexagonHardwareLoops::findInductionRegister(MachineLoop *L,
405                                                  Register &Reg,
406                                                  int64_t &IVBump,
407                                                  MachineInstr *&IVOp
408                                                  ) const {
409   MachineBasicBlock *Header = L->getHeader();
410   MachineBasicBlock *Preheader = MLI->findLoopPreheader(L, SpecPreheader);
411   MachineBasicBlock *Latch = L->getLoopLatch();
412   MachineBasicBlock *ExitingBlock = L->findLoopControlBlock();
413   if (!Header || !Preheader || !Latch || !ExitingBlock)
414     return false;
415 
416   // This pair represents an induction register together with an immediate
417   // value that will be added to it in each loop iteration.
418   using RegisterBump = std::pair<Register, int64_t>;
419 
420   // Mapping:  R.next -> (R, bump), where R, R.next and bump are derived
421   // from an induction operation
422   //   R.next = R + bump
423   // where bump is an immediate value.
424   using InductionMap = std::map<Register, RegisterBump>;
425 
426   InductionMap IndMap;
427 
428   using instr_iterator = MachineBasicBlock::instr_iterator;
429 
430   for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
431        I != E && I->isPHI(); ++I) {
432     MachineInstr *Phi = &*I;
433 
434     // Have a PHI instruction.  Get the operand that corresponds to the
435     // latch block, and see if is a result of an addition of form "reg+imm",
436     // where the "reg" is defined by the PHI node we are looking at.
437     for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) {
438       if (Phi->getOperand(i+1).getMBB() != Latch)
439         continue;
440 
441       Register PhiOpReg = Phi->getOperand(i).getReg();
442       MachineInstr *DI = MRI->getVRegDef(PhiOpReg);
443 
444       if (DI->getDesc().isAdd()) {
445         // If the register operand to the add is the PHI we're looking at, this
446         // meets the induction pattern.
447         Register IndReg = DI->getOperand(1).getReg();
448         MachineOperand &Opnd2 = DI->getOperand(2);
449         int64_t V;
450         if (MRI->getVRegDef(IndReg) == Phi && checkForImmediate(Opnd2, V)) {
451           Register UpdReg = DI->getOperand(0).getReg();
452           IndMap.insert(std::make_pair(UpdReg, std::make_pair(IndReg, V)));
453         }
454       }
455     }  // for (i)
456   }  // for (instr)
457 
458   SmallVector<MachineOperand,2> Cond;
459   MachineBasicBlock *TB = nullptr, *FB = nullptr;
460   bool NotAnalyzed = TII->analyzeBranch(*ExitingBlock, TB, FB, Cond, false);
461   if (NotAnalyzed)
462     return false;
463 
464   Register PredR;
465   unsigned PredPos, PredRegFlags;
466   if (!TII->getPredReg(Cond, PredR, PredPos, PredRegFlags))
467     return false;
468 
469   MachineInstr *PredI = MRI->getVRegDef(PredR);
470   if (!PredI->isCompare())
471     return false;
472 
473   Register CmpReg1, CmpReg2;
474   int64_t CmpImm = 0, CmpMask = 0;
475   bool CmpAnalyzed =
476       TII->analyzeCompare(*PredI, CmpReg1, CmpReg2, CmpMask, CmpImm);
477   // Fail if the compare was not analyzed, or it's not comparing a register
478   // with an immediate value.  Not checking the mask here, since we handle
479   // the individual compare opcodes (including A4_cmpb*) later on.
480   if (!CmpAnalyzed)
481     return false;
482 
483   // Exactly one of the input registers to the comparison should be among
484   // the induction registers.
485   InductionMap::iterator IndMapEnd = IndMap.end();
486   InductionMap::iterator F = IndMapEnd;
487   if (CmpReg1 != 0) {
488     InductionMap::iterator F1 = IndMap.find(CmpReg1);
489     if (F1 != IndMapEnd)
490       F = F1;
491   }
492   if (CmpReg2 != 0) {
493     InductionMap::iterator F2 = IndMap.find(CmpReg2);
494     if (F2 != IndMapEnd) {
495       if (F != IndMapEnd)
496         return false;
497       F = F2;
498     }
499   }
500   if (F == IndMapEnd)
501     return false;
502 
503   Reg = F->second.first;
504   IVBump = F->second.second;
505   IVOp = MRI->getVRegDef(F->first);
506   return true;
507 }
508 
509 // Return the comparison kind for the specified opcode.
510 HexagonHardwareLoops::Comparison::Kind
511 HexagonHardwareLoops::getComparisonKind(unsigned CondOpc,
512                                         MachineOperand *InitialValue,
513                                         const MachineOperand *EndValue,
514                                         int64_t IVBump) const {
515   Comparison::Kind Cmp = (Comparison::Kind)0;
516   switch (CondOpc) {
517   case Hexagon::C2_cmpeq:
518   case Hexagon::C2_cmpeqi:
519   case Hexagon::C2_cmpeqp:
520     Cmp = Comparison::EQ;
521     break;
522   case Hexagon::C4_cmpneq:
523   case Hexagon::C4_cmpneqi:
524     Cmp = Comparison::NE;
525     break;
526   case Hexagon::C2_cmplt:
527     Cmp = Comparison::LTs;
528     break;
529   case Hexagon::C2_cmpltu:
530     Cmp = Comparison::LTu;
531     break;
532   case Hexagon::C4_cmplte:
533   case Hexagon::C4_cmpltei:
534     Cmp = Comparison::LEs;
535     break;
536   case Hexagon::C4_cmplteu:
537   case Hexagon::C4_cmplteui:
538     Cmp = Comparison::LEu;
539     break;
540   case Hexagon::C2_cmpgt:
541   case Hexagon::C2_cmpgti:
542   case Hexagon::C2_cmpgtp:
543     Cmp = Comparison::GTs;
544     break;
545   case Hexagon::C2_cmpgtu:
546   case Hexagon::C2_cmpgtui:
547   case Hexagon::C2_cmpgtup:
548     Cmp = Comparison::GTu;
549     break;
550   case Hexagon::C2_cmpgei:
551     Cmp = Comparison::GEs;
552     break;
553   case Hexagon::C2_cmpgeui:
554     Cmp = Comparison::GEs;
555     break;
556   default:
557     return (Comparison::Kind)0;
558   }
559   return Cmp;
560 }
561 
562 /// Analyze the statements in a loop to determine if the loop has
563 /// a computable trip count and, if so, return a value that represents
564 /// the trip count expression.
565 ///
566 /// This function iterates over the phi nodes in the loop to check for
567 /// induction variable patterns that are used in the calculation for
568 /// the number of time the loop is executed.
569 CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L,
570     SmallVectorImpl<MachineInstr *> &OldInsts) {
571   MachineBasicBlock *TopMBB = L->getTopBlock();
572   MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin();
573   assert(PI != TopMBB->pred_end() &&
574          "Loop must have more than one incoming edge!");
575   MachineBasicBlock *Backedge = *PI++;
576   if (PI == TopMBB->pred_end())  // dead loop?
577     return nullptr;
578   MachineBasicBlock *Incoming = *PI++;
579   if (PI != TopMBB->pred_end())  // multiple backedges?
580     return nullptr;
581 
582   // Make sure there is one incoming and one backedge and determine which
583   // is which.
584   if (L->contains(Incoming)) {
585     if (L->contains(Backedge))
586       return nullptr;
587     std::swap(Incoming, Backedge);
588   } else if (!L->contains(Backedge))
589     return nullptr;
590 
591   // Look for the cmp instruction to determine if we can get a useful trip
592   // count.  The trip count can be either a register or an immediate.  The
593   // location of the value depends upon the type (reg or imm).
594   MachineBasicBlock *ExitingBlock = L->findLoopControlBlock();
595   if (!ExitingBlock)
596     return nullptr;
597 
598   Register IVReg = 0;
599   int64_t IVBump = 0;
600   MachineInstr *IVOp;
601   bool FoundIV = findInductionRegister(L, IVReg, IVBump, IVOp);
602   if (!FoundIV)
603     return nullptr;
604 
605   MachineBasicBlock *Preheader = MLI->findLoopPreheader(L, SpecPreheader);
606 
607   MachineOperand *InitialValue = nullptr;
608   MachineInstr *IV_Phi = MRI->getVRegDef(IVReg);
609   MachineBasicBlock *Latch = L->getLoopLatch();
610   for (unsigned i = 1, n = IV_Phi->getNumOperands(); i < n; i += 2) {
611     MachineBasicBlock *MBB = IV_Phi->getOperand(i+1).getMBB();
612     if (MBB == Preheader)
613       InitialValue = &IV_Phi->getOperand(i);
614     else if (MBB == Latch)
615       IVReg = IV_Phi->getOperand(i).getReg();  // Want IV reg after bump.
616   }
617   if (!InitialValue)
618     return nullptr;
619 
620   SmallVector<MachineOperand,2> Cond;
621   MachineBasicBlock *TB = nullptr, *FB = nullptr;
622   bool NotAnalyzed = TII->analyzeBranch(*ExitingBlock, TB, FB, Cond, false);
623   if (NotAnalyzed)
624     return nullptr;
625 
626   MachineBasicBlock *Header = L->getHeader();
627   // TB must be non-null.  If FB is also non-null, one of them must be
628   // the header.  Otherwise, branch to TB could be exiting the loop, and
629   // the fall through can go to the header.
630   assert (TB && "Exit block without a branch?");
631   if (ExitingBlock != Latch && (TB == Latch || FB == Latch)) {
632     MachineBasicBlock *LTB = nullptr, *LFB = nullptr;
633     SmallVector<MachineOperand,2> LCond;
634     bool NotAnalyzed = TII->analyzeBranch(*Latch, LTB, LFB, LCond, false);
635     if (NotAnalyzed)
636       return nullptr;
637     if (TB == Latch)
638       TB = (LTB == Header) ? LTB : LFB;
639     else
640       FB = (LTB == Header) ? LTB: LFB;
641   }
642   assert ((!FB || TB == Header || FB == Header) && "Branches not to header?");
643   if (!TB || (FB && TB != Header && FB != Header))
644     return nullptr;
645 
646   // Branches of form "if (!P) ..." cause HexagonInstrInfo::analyzeBranch
647   // to put imm(0), followed by P in the vector Cond.
648   // If TB is not the header, it means that the "not-taken" path must lead
649   // to the header.
650   bool Negated = TII->predOpcodeHasNot(Cond) ^ (TB != Header);
651   Register PredReg;
652   unsigned PredPos, PredRegFlags;
653   if (!TII->getPredReg(Cond, PredReg, PredPos, PredRegFlags))
654     return nullptr;
655   MachineInstr *CondI = MRI->getVRegDef(PredReg);
656   unsigned CondOpc = CondI->getOpcode();
657 
658   Register CmpReg1, CmpReg2;
659   int64_t Mask = 0, ImmValue = 0;
660   bool AnalyzedCmp =
661       TII->analyzeCompare(*CondI, CmpReg1, CmpReg2, Mask, ImmValue);
662   if (!AnalyzedCmp)
663     return nullptr;
664 
665   // The comparison operator type determines how we compute the loop
666   // trip count.
667   OldInsts.push_back(CondI);
668   OldInsts.push_back(IVOp);
669 
670   // Sadly, the following code gets information based on the position
671   // of the operands in the compare instruction.  This has to be done
672   // this way, because the comparisons check for a specific relationship
673   // between the operands (e.g. is-less-than), rather than to find out
674   // what relationship the operands are in (as on PPC).
675   Comparison::Kind Cmp;
676   bool isSwapped = false;
677   const MachineOperand &Op1 = CondI->getOperand(1);
678   const MachineOperand &Op2 = CondI->getOperand(2);
679   const MachineOperand *EndValue = nullptr;
680 
681   if (Op1.isReg()) {
682     if (Op2.isImm() || Op1.getReg() == IVReg)
683       EndValue = &Op2;
684     else {
685       EndValue = &Op1;
686       isSwapped = true;
687     }
688   }
689 
690   if (!EndValue)
691     return nullptr;
692 
693   Cmp = getComparisonKind(CondOpc, InitialValue, EndValue, IVBump);
694   if (!Cmp)
695     return nullptr;
696   if (Negated)
697     Cmp = Comparison::getNegatedComparison(Cmp);
698   if (isSwapped)
699     Cmp = Comparison::getSwappedComparison(Cmp);
700 
701   if (InitialValue->isReg()) {
702     Register R = InitialValue->getReg();
703     MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent();
704     if (!MDT->properlyDominates(DefBB, Header)) {
705       int64_t V;
706       if (!checkForImmediate(*InitialValue, V))
707         return nullptr;
708     }
709     OldInsts.push_back(MRI->getVRegDef(R));
710   }
711   if (EndValue->isReg()) {
712     Register R = EndValue->getReg();
713     MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent();
714     if (!MDT->properlyDominates(DefBB, Header)) {
715       int64_t V;
716       if (!checkForImmediate(*EndValue, V))
717         return nullptr;
718     }
719     OldInsts.push_back(MRI->getVRegDef(R));
720   }
721 
722   return computeCount(L, InitialValue, EndValue, IVReg, IVBump, Cmp);
723 }
724 
725 /// Helper function that returns the expression that represents the
726 /// number of times a loop iterates.  The function takes the operands that
727 /// represent the loop start value, loop end value, and induction value.
728 /// Based upon these operands, the function attempts to compute the trip count.
729 CountValue *HexagonHardwareLoops::computeCount(MachineLoop *Loop,
730                                                const MachineOperand *Start,
731                                                const MachineOperand *End,
732                                                Register IVReg,
733                                                int64_t IVBump,
734                                                Comparison::Kind Cmp) const {
735   // Cannot handle comparison EQ, i.e. while (A == B).
736   if (Cmp == Comparison::EQ)
737     return nullptr;
738 
739   // Check if either the start or end values are an assignment of an immediate.
740   // If so, use the immediate value rather than the register.
741   if (Start->isReg()) {
742     const MachineInstr *StartValInstr = MRI->getVRegDef(Start->getReg());
743     if (StartValInstr && (StartValInstr->getOpcode() == Hexagon::A2_tfrsi ||
744                           StartValInstr->getOpcode() == Hexagon::A2_tfrpi))
745       Start = &StartValInstr->getOperand(1);
746   }
747   if (End->isReg()) {
748     const MachineInstr *EndValInstr = MRI->getVRegDef(End->getReg());
749     if (EndValInstr && (EndValInstr->getOpcode() == Hexagon::A2_tfrsi ||
750                         EndValInstr->getOpcode() == Hexagon::A2_tfrpi))
751       End = &EndValInstr->getOperand(1);
752   }
753 
754   if (!Start->isReg() && !Start->isImm())
755     return nullptr;
756   if (!End->isReg() && !End->isImm())
757     return nullptr;
758 
759   bool CmpLess =     Cmp & Comparison::L;
760   bool CmpGreater =  Cmp & Comparison::G;
761   bool CmpHasEqual = Cmp & Comparison::EQ;
762 
763   // Avoid certain wrap-arounds.  This doesn't detect all wrap-arounds.
764   if (CmpLess && IVBump < 0)
765     // Loop going while iv is "less" with the iv value going down.  Must wrap.
766     return nullptr;
767 
768   if (CmpGreater && IVBump > 0)
769     // Loop going while iv is "greater" with the iv value going up.  Must wrap.
770     return nullptr;
771 
772   // Phis that may feed into the loop.
773   LoopFeederMap LoopFeederPhi;
774 
775   // Check if the initial value may be zero and can be decremented in the first
776   // iteration. If the value is zero, the endloop instruction will not decrement
777   // the loop counter, so we shouldn't generate a hardware loop in this case.
778   if (loopCountMayWrapOrUnderFlow(Start, End, Loop->getLoopPreheader(), Loop,
779                                   LoopFeederPhi))
780       return nullptr;
781 
782   if (Start->isImm() && End->isImm()) {
783     // Both, start and end are immediates.
784     int64_t StartV = Start->getImm();
785     int64_t EndV = End->getImm();
786     int64_t Dist = EndV - StartV;
787     if (Dist == 0)
788       return nullptr;
789 
790     bool Exact = (Dist % IVBump) == 0;
791 
792     if (Cmp == Comparison::NE) {
793       if (!Exact)
794         return nullptr;
795       if ((Dist < 0) ^ (IVBump < 0))
796         return nullptr;
797     }
798 
799     // For comparisons that include the final value (i.e. include equality
800     // with the final value), we need to increase the distance by 1.
801     if (CmpHasEqual)
802       Dist = Dist > 0 ? Dist+1 : Dist-1;
803 
804     // For the loop to iterate, CmpLess should imply Dist > 0.  Similarly,
805     // CmpGreater should imply Dist < 0.  These conditions could actually
806     // fail, for example, in unreachable code (which may still appear to be
807     // reachable in the CFG).
808     if ((CmpLess && Dist < 0) || (CmpGreater && Dist > 0))
809       return nullptr;
810 
811     // "Normalized" distance, i.e. with the bump set to +-1.
812     int64_t Dist1 = (IVBump > 0) ? (Dist +  (IVBump - 1)) / IVBump
813                                  : (-Dist + (-IVBump - 1)) / (-IVBump);
814     assert (Dist1 > 0 && "Fishy thing.  Both operands have the same sign.");
815 
816     uint64_t Count = Dist1;
817 
818     if (Count > 0xFFFFFFFFULL)
819       return nullptr;
820 
821     return new CountValue(CountValue::CV_Immediate, Count);
822   }
823 
824   // A general case: Start and End are some values, but the actual
825   // iteration count may not be available.  If it is not, insert
826   // a computation of it into the preheader.
827 
828   // If the induction variable bump is not a power of 2, quit.
829   // Othwerise we'd need a general integer division.
830   if (!isPowerOf2_64(std::abs(IVBump)))
831     return nullptr;
832 
833   MachineBasicBlock *PH = MLI->findLoopPreheader(Loop, SpecPreheader);
834   assert (PH && "Should have a preheader by now");
835   MachineBasicBlock::iterator InsertPos = PH->getFirstTerminator();
836   DebugLoc DL;
837   if (InsertPos != PH->end())
838     DL = InsertPos->getDebugLoc();
839 
840   // If Start is an immediate and End is a register, the trip count
841   // will be "reg - imm".  Hexagon's "subtract immediate" instruction
842   // is actually "reg + -imm".
843 
844   // If the loop IV is going downwards, i.e. if the bump is negative,
845   // then the iteration count (computed as End-Start) will need to be
846   // negated.  To avoid the negation, just swap Start and End.
847   if (IVBump < 0) {
848     std::swap(Start, End);
849     IVBump = -IVBump;
850   }
851   // Cmp may now have a wrong direction, e.g.  LEs may now be GEs.
852   // Signedness, and "including equality" are preserved.
853 
854   bool RegToImm = Start->isReg() && End->isImm(); // for (reg..imm)
855   bool RegToReg = Start->isReg() && End->isReg(); // for (reg..reg)
856 
857   int64_t StartV = 0, EndV = 0;
858   if (Start->isImm())
859     StartV = Start->getImm();
860   if (End->isImm())
861     EndV = End->getImm();
862 
863   int64_t AdjV = 0;
864   // To compute the iteration count, we would need this computation:
865   //   Count = (End - Start + (IVBump-1)) / IVBump
866   // or, when CmpHasEqual:
867   //   Count = (End - Start + (IVBump-1)+1) / IVBump
868   // The "IVBump-1" part is the adjustment (AdjV).  We can avoid
869   // generating an instruction specifically to add it if we can adjust
870   // the immediate values for Start or End.
871 
872   if (CmpHasEqual) {
873     // Need to add 1 to the total iteration count.
874     if (Start->isImm())
875       StartV--;
876     else if (End->isImm())
877       EndV++;
878     else
879       AdjV += 1;
880   }
881 
882   if (Cmp != Comparison::NE) {
883     if (Start->isImm())
884       StartV -= (IVBump-1);
885     else if (End->isImm())
886       EndV += (IVBump-1);
887     else
888       AdjV += (IVBump-1);
889   }
890 
891   Register R = 0;
892   unsigned SR = 0;
893   if (Start->isReg()) {
894     R = Start->getReg();
895     SR = Start->getSubReg();
896   } else {
897     R = End->getReg();
898     SR = End->getSubReg();
899   }
900   const TargetRegisterClass *RC = MRI->getRegClass(R);
901   // Hardware loops cannot handle 64-bit registers.  If it's a double
902   // register, it has to have a subregister.
903   if (!SR && RC == &Hexagon::DoubleRegsRegClass)
904     return nullptr;
905   const TargetRegisterClass *IntRC = &Hexagon::IntRegsRegClass;
906 
907   // Compute DistR (register with the distance between Start and End).
908   Register DistR;
909   unsigned DistSR;
910 
911   // Avoid special case, where the start value is an imm(0).
912   if (Start->isImm() && StartV == 0) {
913     DistR = End->getReg();
914     DistSR = End->getSubReg();
915   } else {
916     const MCInstrDesc &SubD = RegToReg ? TII->get(Hexagon::A2_sub) :
917                               (RegToImm ? TII->get(Hexagon::A2_subri) :
918                                           TII->get(Hexagon::A2_addi));
919     if (RegToReg || RegToImm) {
920       Register SubR = MRI->createVirtualRegister(IntRC);
921       MachineInstrBuilder SubIB =
922         BuildMI(*PH, InsertPos, DL, SubD, SubR);
923 
924       if (RegToReg)
925         SubIB.addReg(End->getReg(), 0, End->getSubReg())
926           .addReg(Start->getReg(), 0, Start->getSubReg());
927       else
928         SubIB.addImm(EndV)
929           .addReg(Start->getReg(), 0, Start->getSubReg());
930       DistR = SubR;
931     } else {
932       // If the loop has been unrolled, we should use the original loop count
933       // instead of recalculating the value. This will avoid additional
934       // 'Add' instruction.
935       const MachineInstr *EndValInstr = MRI->getVRegDef(End->getReg());
936       if (EndValInstr->getOpcode() == Hexagon::A2_addi &&
937           EndValInstr->getOperand(1).getSubReg() == 0 &&
938           EndValInstr->getOperand(2).getImm() == StartV) {
939         DistR = EndValInstr->getOperand(1).getReg();
940       } else {
941         Register SubR = MRI->createVirtualRegister(IntRC);
942         MachineInstrBuilder SubIB =
943           BuildMI(*PH, InsertPos, DL, SubD, SubR);
944         SubIB.addReg(End->getReg(), 0, End->getSubReg())
945              .addImm(-StartV);
946         DistR = SubR;
947       }
948     }
949     DistSR = 0;
950   }
951 
952   // From DistR, compute AdjR (register with the adjusted distance).
953   Register AdjR;
954   unsigned AdjSR;
955 
956   if (AdjV == 0) {
957     AdjR = DistR;
958     AdjSR = DistSR;
959   } else {
960     // Generate CountR = ADD DistR, AdjVal
961     Register AddR = MRI->createVirtualRegister(IntRC);
962     MCInstrDesc const &AddD = TII->get(Hexagon::A2_addi);
963     BuildMI(*PH, InsertPos, DL, AddD, AddR)
964       .addReg(DistR, 0, DistSR)
965       .addImm(AdjV);
966 
967     AdjR = AddR;
968     AdjSR = 0;
969   }
970 
971   // From AdjR, compute CountR (register with the final count).
972   Register CountR;
973   unsigned CountSR;
974 
975   if (IVBump == 1) {
976     CountR = AdjR;
977     CountSR = AdjSR;
978   } else {
979     // The IV bump is a power of two. Log_2(IV bump) is the shift amount.
980     unsigned Shift = Log2_32(IVBump);
981 
982     // Generate NormR = LSR DistR, Shift.
983     Register LsrR = MRI->createVirtualRegister(IntRC);
984     const MCInstrDesc &LsrD = TII->get(Hexagon::S2_lsr_i_r);
985     BuildMI(*PH, InsertPos, DL, LsrD, LsrR)
986       .addReg(AdjR, 0, AdjSR)
987       .addImm(Shift);
988 
989     CountR = LsrR;
990     CountSR = 0;
991   }
992 
993   return new CountValue(CountValue::CV_Register, CountR, CountSR);
994 }
995 
996 /// Return true if the operation is invalid within hardware loop.
997 bool HexagonHardwareLoops::isInvalidLoopOperation(const MachineInstr *MI,
998                                                   bool IsInnerHWLoop) const {
999   // Call is not allowed because the callee may use a hardware loop except for
1000   // the case when the call never returns.
1001   if (MI->getDesc().isCall())
1002     return !TII->doesNotReturn(*MI);
1003 
1004   // Check if the instruction defines a hardware loop register.
1005   using namespace Hexagon;
1006 
1007   static const Register Regs01[] = { LC0, SA0, LC1, SA1 };
1008   static const Register Regs1[]  = { LC1, SA1 };
1009   auto CheckRegs = IsInnerHWLoop ? ArrayRef(Regs01, std::size(Regs01))
1010                                  : ArrayRef(Regs1, std::size(Regs1));
1011   for (Register R : CheckRegs)
1012     if (MI->modifiesRegister(R, TRI))
1013       return true;
1014 
1015   return false;
1016 }
1017 
1018 /// Return true if the loop contains an instruction that inhibits
1019 /// the use of the hardware loop instruction.
1020 bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L,
1021     bool IsInnerHWLoop) const {
1022   LLVM_DEBUG(dbgs() << "\nhw_loop head, "
1023                     << printMBBReference(**L->block_begin()));
1024   for (MachineBasicBlock *MBB : L->getBlocks()) {
1025     for (const MachineInstr &MI : *MBB) {
1026       if (isInvalidLoopOperation(&MI, IsInnerHWLoop)) {
1027         LLVM_DEBUG(dbgs() << "\nCannot convert to hw_loop due to:";
1028                    MI.dump(););
1029         return true;
1030       }
1031     }
1032   }
1033   return false;
1034 }
1035 
1036 /// Returns true if the instruction is dead.  This was essentially
1037 /// copied from DeadMachineInstructionElim::isDead, but with special cases
1038 /// for inline asm, physical registers and instructions with side effects
1039 /// removed.
1040 bool HexagonHardwareLoops::isDead(const MachineInstr *MI,
1041                               SmallVectorImpl<MachineInstr *> &DeadPhis) const {
1042   // Examine each operand.
1043   for (const MachineOperand &MO : MI->operands()) {
1044     if (!MO.isReg() || !MO.isDef())
1045       continue;
1046 
1047     Register Reg = MO.getReg();
1048     if (MRI->use_nodbg_empty(Reg))
1049       continue;
1050 
1051     using use_nodbg_iterator = MachineRegisterInfo::use_nodbg_iterator;
1052 
1053     // This instruction has users, but if the only user is the phi node for the
1054     // parent block, and the only use of that phi node is this instruction, then
1055     // this instruction is dead: both it (and the phi node) can be removed.
1056     use_nodbg_iterator I = MRI->use_nodbg_begin(Reg);
1057     use_nodbg_iterator End = MRI->use_nodbg_end();
1058     if (std::next(I) != End || !I->getParent()->isPHI())
1059       return false;
1060 
1061     MachineInstr *OnePhi = I->getParent();
1062     for (const MachineOperand &OPO : OnePhi->operands()) {
1063       if (!OPO.isReg() || !OPO.isDef())
1064         continue;
1065 
1066       Register OPReg = OPO.getReg();
1067       use_nodbg_iterator nextJ;
1068       for (use_nodbg_iterator J = MRI->use_nodbg_begin(OPReg);
1069            J != End; J = nextJ) {
1070         nextJ = std::next(J);
1071         MachineOperand &Use = *J;
1072         MachineInstr *UseMI = Use.getParent();
1073 
1074         // If the phi node has a user that is not MI, bail.
1075         if (MI != UseMI)
1076           return false;
1077       }
1078     }
1079     DeadPhis.push_back(OnePhi);
1080   }
1081 
1082   // If there are no defs with uses, the instruction is dead.
1083   return true;
1084 }
1085 
1086 void HexagonHardwareLoops::removeIfDead(MachineInstr *MI) {
1087   // This procedure was essentially copied from DeadMachineInstructionElim.
1088 
1089   SmallVector<MachineInstr*, 1> DeadPhis;
1090   if (isDead(MI, DeadPhis)) {
1091     LLVM_DEBUG(dbgs() << "HW looping will remove: " << *MI);
1092 
1093     // It is possible that some DBG_VALUE instructions refer to this
1094     // instruction.  Examine each def operand for such references;
1095     // if found, mark the DBG_VALUE as undef (but don't delete it).
1096     for (const MachineOperand &MO : MI->operands()) {
1097       if (!MO.isReg() || !MO.isDef())
1098         continue;
1099       Register Reg = MO.getReg();
1100       // We use make_early_inc_range here because setReg below invalidates the
1101       // iterator.
1102       for (MachineOperand &MO :
1103            llvm::make_early_inc_range(MRI->use_operands(Reg))) {
1104         MachineInstr *UseMI = MO.getParent();
1105         if (UseMI == MI)
1106           continue;
1107         if (MO.isDebug())
1108           MO.setReg(0U);
1109       }
1110     }
1111 
1112     MI->eraseFromParent();
1113     for (unsigned i = 0; i < DeadPhis.size(); ++i)
1114       DeadPhis[i]->eraseFromParent();
1115   }
1116 }
1117 
1118 /// Check if the loop is a candidate for converting to a hardware
1119 /// loop.  If so, then perform the transformation.
1120 ///
1121 /// This function works on innermost loops first.  A loop can be converted
1122 /// if it is a counting loop; either a register value or an immediate.
1123 ///
1124 /// The code makes several assumptions about the representation of the loop
1125 /// in llvm.
1126 bool HexagonHardwareLoops::convertToHardwareLoop(MachineLoop *L,
1127                                                  bool &RecL0used,
1128                                                  bool &RecL1used) {
1129   // This is just to confirm basic correctness.
1130   assert(L->getHeader() && "Loop without a header?");
1131 
1132   bool Changed = false;
1133   bool L0Used = false;
1134   bool L1Used = false;
1135 
1136   // Process nested loops first.
1137   for (MachineLoop *I : *L) {
1138     Changed |= convertToHardwareLoop(I, RecL0used, RecL1used);
1139     L0Used |= RecL0used;
1140     L1Used |= RecL1used;
1141   }
1142 
1143   // If a nested loop has been converted, then we can't convert this loop.
1144   if (Changed && L0Used && L1Used)
1145     return Changed;
1146 
1147   unsigned LOOP_i;
1148   unsigned LOOP_r;
1149   unsigned ENDLOOP;
1150 
1151   // Flag used to track loopN instruction:
1152   // 1 - Hardware loop is being generated for the inner most loop.
1153   // 0 - Hardware loop is being generated for the outer loop.
1154   unsigned IsInnerHWLoop = 1;
1155 
1156   if (L0Used) {
1157     LOOP_i = Hexagon::J2_loop1i;
1158     LOOP_r = Hexagon::J2_loop1r;
1159     ENDLOOP = Hexagon::ENDLOOP1;
1160     IsInnerHWLoop = 0;
1161   } else {
1162     LOOP_i = Hexagon::J2_loop0i;
1163     LOOP_r = Hexagon::J2_loop0r;
1164     ENDLOOP = Hexagon::ENDLOOP0;
1165   }
1166 
1167 #ifndef NDEBUG
1168   // Stop trying after reaching the limit (if any).
1169   int Limit = HWLoopLimit;
1170   if (Limit >= 0) {
1171     if (Counter >= HWLoopLimit)
1172       return false;
1173     Counter++;
1174   }
1175 #endif
1176 
1177   // Does the loop contain any invalid instructions?
1178   if (containsInvalidInstruction(L, IsInnerHWLoop))
1179     return false;
1180 
1181   MachineBasicBlock *LastMBB = L->findLoopControlBlock();
1182   // Don't generate hw loop if the loop has more than one exit.
1183   if (!LastMBB)
1184     return false;
1185 
1186   MachineBasicBlock::iterator LastI = LastMBB->getFirstTerminator();
1187   if (LastI == LastMBB->end())
1188     return false;
1189 
1190   // Is the induction variable bump feeding the latch condition?
1191   if (!fixupInductionVariable(L))
1192     return false;
1193 
1194   // Ensure the loop has a preheader: the loop instruction will be
1195   // placed there.
1196   MachineBasicBlock *Preheader = MLI->findLoopPreheader(L, SpecPreheader);
1197   if (!Preheader) {
1198     Preheader = createPreheaderForLoop(L);
1199     if (!Preheader)
1200       return false;
1201   }
1202 
1203   MachineBasicBlock::iterator InsertPos = Preheader->getFirstTerminator();
1204 
1205   SmallVector<MachineInstr*, 2> OldInsts;
1206   // Are we able to determine the trip count for the loop?
1207   CountValue *TripCount = getLoopTripCount(L, OldInsts);
1208   if (!TripCount)
1209     return false;
1210 
1211   // Is the trip count available in the preheader?
1212   if (TripCount->isReg()) {
1213     // There will be a use of the register inserted into the preheader,
1214     // so make sure that the register is actually defined at that point.
1215     MachineInstr *TCDef = MRI->getVRegDef(TripCount->getReg());
1216     MachineBasicBlock *BBDef = TCDef->getParent();
1217     if (!MDT->dominates(BBDef, Preheader))
1218       return false;
1219   }
1220 
1221   // Determine the loop start.
1222   MachineBasicBlock *TopBlock = L->getTopBlock();
1223   MachineBasicBlock *ExitingBlock = L->findLoopControlBlock();
1224   MachineBasicBlock *LoopStart = nullptr;
1225   if (ExitingBlock !=  L->getLoopLatch()) {
1226     MachineBasicBlock *TB = nullptr, *FB = nullptr;
1227     SmallVector<MachineOperand, 2> Cond;
1228 
1229     if (TII->analyzeBranch(*ExitingBlock, TB, FB, Cond, false))
1230       return false;
1231 
1232     if (L->contains(TB))
1233       LoopStart = TB;
1234     else if (L->contains(FB))
1235       LoopStart = FB;
1236     else
1237       return false;
1238   }
1239   else
1240     LoopStart = TopBlock;
1241 
1242   // Convert the loop to a hardware loop.
1243   LLVM_DEBUG(dbgs() << "Change to hardware loop at "; L->dump());
1244   DebugLoc DL;
1245   if (InsertPos != Preheader->end())
1246     DL = InsertPos->getDebugLoc();
1247 
1248   if (TripCount->isReg()) {
1249     // Create a copy of the loop count register.
1250     Register CountReg = MRI->createVirtualRegister(&Hexagon::IntRegsRegClass);
1251     BuildMI(*Preheader, InsertPos, DL, TII->get(TargetOpcode::COPY), CountReg)
1252       .addReg(TripCount->getReg(), 0, TripCount->getSubReg());
1253     // Add the Loop instruction to the beginning of the loop.
1254     BuildMI(*Preheader, InsertPos, DL, TII->get(LOOP_r)).addMBB(LoopStart)
1255       .addReg(CountReg);
1256   } else {
1257     assert(TripCount->isImm() && "Expecting immediate value for trip count");
1258     // Add the Loop immediate instruction to the beginning of the loop,
1259     // if the immediate fits in the instructions.  Otherwise, we need to
1260     // create a new virtual register.
1261     int64_t CountImm = TripCount->getImm();
1262     if (!TII->isValidOffset(LOOP_i, CountImm, TRI)) {
1263       Register CountReg = MRI->createVirtualRegister(&Hexagon::IntRegsRegClass);
1264       BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::A2_tfrsi), CountReg)
1265         .addImm(CountImm);
1266       BuildMI(*Preheader, InsertPos, DL, TII->get(LOOP_r))
1267         .addMBB(LoopStart).addReg(CountReg);
1268     } else
1269       BuildMI(*Preheader, InsertPos, DL, TII->get(LOOP_i))
1270         .addMBB(LoopStart).addImm(CountImm);
1271   }
1272 
1273   // Make sure the loop start always has a reference in the CFG.
1274   LoopStart->setMachineBlockAddressTaken();
1275 
1276   // Replace the loop branch with an endloop instruction.
1277   DebugLoc LastIDL = LastI->getDebugLoc();
1278   BuildMI(*LastMBB, LastI, LastIDL, TII->get(ENDLOOP)).addMBB(LoopStart);
1279 
1280   // The loop ends with either:
1281   //  - a conditional branch followed by an unconditional branch, or
1282   //  - a conditional branch to the loop start.
1283   if (LastI->getOpcode() == Hexagon::J2_jumpt ||
1284       LastI->getOpcode() == Hexagon::J2_jumpf) {
1285     // Delete one and change/add an uncond. branch to out of the loop.
1286     MachineBasicBlock *BranchTarget = LastI->getOperand(1).getMBB();
1287     LastI = LastMBB->erase(LastI);
1288     if (!L->contains(BranchTarget)) {
1289       if (LastI != LastMBB->end())
1290         LastI = LastMBB->erase(LastI);
1291       SmallVector<MachineOperand, 0> Cond;
1292       TII->insertBranch(*LastMBB, BranchTarget, nullptr, Cond, LastIDL);
1293     }
1294   } else {
1295     // Conditional branch to loop start; just delete it.
1296     LastMBB->erase(LastI);
1297   }
1298   delete TripCount;
1299 
1300   // The induction operation and the comparison may now be
1301   // unneeded. If these are unneeded, then remove them.
1302   for (unsigned i = 0; i < OldInsts.size(); ++i)
1303     removeIfDead(OldInsts[i]);
1304 
1305   ++NumHWLoops;
1306 
1307   // Set RecL1used and RecL0used only after hardware loop has been
1308   // successfully generated. Doing it earlier can cause wrong loop instruction
1309   // to be used.
1310   if (L0Used) // Loop0 was already used. So, the correct loop must be loop1.
1311     RecL1used = true;
1312   else
1313     RecL0used = true;
1314 
1315   return true;
1316 }
1317 
1318 bool HexagonHardwareLoops::orderBumpCompare(MachineInstr *BumpI,
1319                                             MachineInstr *CmpI) {
1320   assert (BumpI != CmpI && "Bump and compare in the same instruction?");
1321 
1322   MachineBasicBlock *BB = BumpI->getParent();
1323   if (CmpI->getParent() != BB)
1324     return false;
1325 
1326   using instr_iterator = MachineBasicBlock::instr_iterator;
1327 
1328   // Check if things are in order to begin with.
1329   for (instr_iterator I(BumpI), E = BB->instr_end(); I != E; ++I)
1330     if (&*I == CmpI)
1331       return true;
1332 
1333   // Out of order.
1334   Register PredR = CmpI->getOperand(0).getReg();
1335   bool FoundBump = false;
1336   instr_iterator CmpIt = CmpI->getIterator(), NextIt = std::next(CmpIt);
1337   for (instr_iterator I = NextIt, E = BB->instr_end(); I != E; ++I) {
1338     MachineInstr *In = &*I;
1339     for (unsigned i = 0, n = In->getNumOperands(); i < n; ++i) {
1340       MachineOperand &MO = In->getOperand(i);
1341       if (MO.isReg() && MO.isUse()) {
1342         if (MO.getReg() == PredR)  // Found an intervening use of PredR.
1343           return false;
1344       }
1345     }
1346 
1347     if (In == BumpI) {
1348       BB->splice(++BumpI->getIterator(), BB, CmpI->getIterator());
1349       FoundBump = true;
1350       break;
1351     }
1352   }
1353   assert (FoundBump && "Cannot determine instruction order");
1354   return FoundBump;
1355 }
1356 
1357 /// This function is required to break recursion. Visiting phis in a loop may
1358 /// result in recursion during compilation. We break the recursion by making
1359 /// sure that we visit a MachineOperand and its definition in a
1360 /// MachineInstruction only once. If we attempt to visit more than once, then
1361 /// there is recursion, and will return false.
1362 bool HexagonHardwareLoops::isLoopFeeder(MachineLoop *L, MachineBasicBlock *A,
1363                                         MachineInstr *MI,
1364                                         const MachineOperand *MO,
1365                                         LoopFeederMap &LoopFeederPhi) const {
1366   if (LoopFeederPhi.find(MO->getReg()) == LoopFeederPhi.end()) {
1367     LLVM_DEBUG(dbgs() << "\nhw_loop head, "
1368                       << printMBBReference(**L->block_begin()));
1369     // Ignore all BBs that form Loop.
1370     if (llvm::is_contained(L->getBlocks(), A))
1371       return false;
1372     MachineInstr *Def = MRI->getVRegDef(MO->getReg());
1373     LoopFeederPhi.insert(std::make_pair(MO->getReg(), Def));
1374     return true;
1375   } else
1376     // Already visited node.
1377     return false;
1378 }
1379 
1380 /// Return true if a Phi may generate a value that can underflow.
1381 /// This function calls loopCountMayWrapOrUnderFlow for each Phi operand.
1382 bool HexagonHardwareLoops::phiMayWrapOrUnderflow(
1383     MachineInstr *Phi, const MachineOperand *EndVal, MachineBasicBlock *MBB,
1384     MachineLoop *L, LoopFeederMap &LoopFeederPhi) const {
1385   assert(Phi->isPHI() && "Expecting a Phi.");
1386   // Walk through each Phi, and its used operands. Make sure that
1387   // if there is recursion in Phi, we won't generate hardware loops.
1388   for (int i = 1, n = Phi->getNumOperands(); i < n; i += 2)
1389     if (isLoopFeeder(L, MBB, Phi, &(Phi->getOperand(i)), LoopFeederPhi))
1390       if (loopCountMayWrapOrUnderFlow(&(Phi->getOperand(i)), EndVal,
1391                                       Phi->getParent(), L, LoopFeederPhi))
1392         return true;
1393   return false;
1394 }
1395 
1396 /// Return true if the induction variable can underflow in the first iteration.
1397 /// An example, is an initial unsigned value that is 0 and is decrement in the
1398 /// first itertion of a do-while loop.  In this case, we cannot generate a
1399 /// hardware loop because the endloop instruction does not decrement the loop
1400 /// counter if it is <= 1. We only need to perform this analysis if the
1401 /// initial value is a register.
1402 ///
1403 /// This function assumes the initial value may underfow unless proven
1404 /// otherwise. If the type is signed, then we don't care because signed
1405 /// underflow is undefined. We attempt to prove the initial value is not
1406 /// zero by perfoming a crude analysis of the loop counter. This function
1407 /// checks if the initial value is used in any comparison prior to the loop
1408 /// and, if so, assumes the comparison is a range check. This is inexact,
1409 /// but will catch the simple cases.
1410 bool HexagonHardwareLoops::loopCountMayWrapOrUnderFlow(
1411     const MachineOperand *InitVal, const MachineOperand *EndVal,
1412     MachineBasicBlock *MBB, MachineLoop *L,
1413     LoopFeederMap &LoopFeederPhi) const {
1414   // Only check register values since they are unknown.
1415   if (!InitVal->isReg())
1416     return false;
1417 
1418   if (!EndVal->isImm())
1419     return false;
1420 
1421   // A register value that is assigned an immediate is a known value, and it
1422   // won't underflow in the first iteration.
1423   int64_t Imm;
1424   if (checkForImmediate(*InitVal, Imm))
1425     return (EndVal->getImm() == Imm);
1426 
1427   Register Reg = InitVal->getReg();
1428 
1429   // We don't know the value of a physical register.
1430   if (!Reg.isVirtual())
1431     return true;
1432 
1433   MachineInstr *Def = MRI->getVRegDef(Reg);
1434   if (!Def)
1435     return true;
1436 
1437   // If the initial value is a Phi or copy and the operands may not underflow,
1438   // then the definition cannot be underflow either.
1439   if (Def->isPHI() && !phiMayWrapOrUnderflow(Def, EndVal, Def->getParent(),
1440                                              L, LoopFeederPhi))
1441     return false;
1442   if (Def->isCopy() && !loopCountMayWrapOrUnderFlow(&(Def->getOperand(1)),
1443                                                     EndVal, Def->getParent(),
1444                                                     L, LoopFeederPhi))
1445     return false;
1446 
1447   // Iterate over the uses of the initial value. If the initial value is used
1448   // in a compare, then we assume this is a range check that ensures the loop
1449   // doesn't underflow. This is not an exact test and should be improved.
1450   for (MachineRegisterInfo::use_instr_nodbg_iterator I = MRI->use_instr_nodbg_begin(Reg),
1451          E = MRI->use_instr_nodbg_end(); I != E; ++I) {
1452     MachineInstr *MI = &*I;
1453     Register CmpReg1, CmpReg2;
1454     int64_t CmpMask = 0, CmpValue = 0;
1455 
1456     if (!TII->analyzeCompare(*MI, CmpReg1, CmpReg2, CmpMask, CmpValue))
1457       continue;
1458 
1459     MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1460     SmallVector<MachineOperand, 2> Cond;
1461     if (TII->analyzeBranch(*MI->getParent(), TBB, FBB, Cond, false))
1462       continue;
1463 
1464     Comparison::Kind Cmp =
1465         getComparisonKind(MI->getOpcode(), nullptr, nullptr, 0);
1466     if (Cmp == 0)
1467       continue;
1468     if (TII->predOpcodeHasNot(Cond) ^ (TBB != MBB))
1469       Cmp = Comparison::getNegatedComparison(Cmp);
1470     if (CmpReg2 != 0 && CmpReg2 == Reg)
1471       Cmp = Comparison::getSwappedComparison(Cmp);
1472 
1473     // Signed underflow is undefined.
1474     if (Comparison::isSigned(Cmp))
1475       return false;
1476 
1477     // Check if there is a comparison of the initial value. If the initial value
1478     // is greater than or not equal to another value, then assume this is a
1479     // range check.
1480     if ((Cmp & Comparison::G) || Cmp == Comparison::NE)
1481       return false;
1482   }
1483 
1484   // OK - this is a hack that needs to be improved. We really need to analyze
1485   // the instructions performed on the initial value. This works on the simplest
1486   // cases only.
1487   if (!Def->isCopy() && !Def->isPHI())
1488     return false;
1489 
1490   return true;
1491 }
1492 
1493 bool HexagonHardwareLoops::checkForImmediate(const MachineOperand &MO,
1494                                              int64_t &Val) const {
1495   if (MO.isImm()) {
1496     Val = MO.getImm();
1497     return true;
1498   }
1499   if (!MO.isReg())
1500     return false;
1501 
1502   // MO is a register. Check whether it is defined as an immediate value,
1503   // and if so, get the value of it in TV. That value will then need to be
1504   // processed to handle potential subregisters in MO.
1505   int64_t TV;
1506 
1507   Register R = MO.getReg();
1508   if (!R.isVirtual())
1509     return false;
1510   MachineInstr *DI = MRI->getVRegDef(R);
1511   unsigned DOpc = DI->getOpcode();
1512   switch (DOpc) {
1513     case TargetOpcode::COPY:
1514     case Hexagon::A2_tfrsi:
1515     case Hexagon::A2_tfrpi:
1516     case Hexagon::CONST32:
1517     case Hexagon::CONST64:
1518       // Call recursively to avoid an extra check whether operand(1) is
1519       // indeed an immediate (it could be a global address, for example),
1520       // plus we can handle COPY at the same time.
1521       if (!checkForImmediate(DI->getOperand(1), TV))
1522         return false;
1523       break;
1524     case Hexagon::A2_combineii:
1525     case Hexagon::A4_combineir:
1526     case Hexagon::A4_combineii:
1527     case Hexagon::A4_combineri:
1528     case Hexagon::A2_combinew: {
1529       const MachineOperand &S1 = DI->getOperand(1);
1530       const MachineOperand &S2 = DI->getOperand(2);
1531       int64_t V1, V2;
1532       if (!checkForImmediate(S1, V1) || !checkForImmediate(S2, V2))
1533         return false;
1534       TV = V2 | (static_cast<uint64_t>(V1) << 32);
1535       break;
1536     }
1537     case TargetOpcode::REG_SEQUENCE: {
1538       const MachineOperand &S1 = DI->getOperand(1);
1539       const MachineOperand &S3 = DI->getOperand(3);
1540       int64_t V1, V3;
1541       if (!checkForImmediate(S1, V1) || !checkForImmediate(S3, V3))
1542         return false;
1543       unsigned Sub2 = DI->getOperand(2).getImm();
1544       unsigned Sub4 = DI->getOperand(4).getImm();
1545       if (Sub2 == Hexagon::isub_lo && Sub4 == Hexagon::isub_hi)
1546         TV = V1 | (V3 << 32);
1547       else if (Sub2 == Hexagon::isub_hi && Sub4 == Hexagon::isub_lo)
1548         TV = V3 | (V1 << 32);
1549       else
1550         llvm_unreachable("Unexpected form of REG_SEQUENCE");
1551       break;
1552     }
1553 
1554     default:
1555       return false;
1556   }
1557 
1558   // By now, we should have successfully obtained the immediate value defining
1559   // the register referenced in MO. Handle a potential use of a subregister.
1560   switch (MO.getSubReg()) {
1561     case Hexagon::isub_lo:
1562       Val = TV & 0xFFFFFFFFULL;
1563       break;
1564     case Hexagon::isub_hi:
1565       Val = (TV >> 32) & 0xFFFFFFFFULL;
1566       break;
1567     default:
1568       Val = TV;
1569       break;
1570   }
1571   return true;
1572 }
1573 
1574 void HexagonHardwareLoops::setImmediate(MachineOperand &MO, int64_t Val) {
1575   if (MO.isImm()) {
1576     MO.setImm(Val);
1577     return;
1578   }
1579 
1580   assert(MO.isReg());
1581   Register R = MO.getReg();
1582   MachineInstr *DI = MRI->getVRegDef(R);
1583 
1584   const TargetRegisterClass *RC = MRI->getRegClass(R);
1585   Register NewR = MRI->createVirtualRegister(RC);
1586   MachineBasicBlock &B = *DI->getParent();
1587   DebugLoc DL = DI->getDebugLoc();
1588   BuildMI(B, DI, DL, TII->get(DI->getOpcode()), NewR).addImm(Val);
1589   MO.setReg(NewR);
1590 }
1591 
1592 bool HexagonHardwareLoops::fixupInductionVariable(MachineLoop *L) {
1593   MachineBasicBlock *Header = L->getHeader();
1594   MachineBasicBlock *Latch = L->getLoopLatch();
1595   MachineBasicBlock *ExitingBlock = L->findLoopControlBlock();
1596 
1597   if (!(Header && Latch && ExitingBlock))
1598     return false;
1599 
1600   // These data structures follow the same concept as the corresponding
1601   // ones in findInductionRegister (where some comments are).
1602   using RegisterBump = std::pair<Register, int64_t>;
1603   using RegisterInduction = std::pair<Register, RegisterBump>;
1604   using RegisterInductionSet = std::set<RegisterInduction>;
1605 
1606   // Register candidates for induction variables, with their associated bumps.
1607   RegisterInductionSet IndRegs;
1608 
1609   // Look for induction patterns:
1610   //   %1 = PHI ..., [ latch, %2 ]
1611   //   %2 = ADD %1, imm
1612   using instr_iterator = MachineBasicBlock::instr_iterator;
1613 
1614   for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
1615        I != E && I->isPHI(); ++I) {
1616     MachineInstr *Phi = &*I;
1617 
1618     // Have a PHI instruction.
1619     for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) {
1620       if (Phi->getOperand(i+1).getMBB() != Latch)
1621         continue;
1622 
1623       Register PhiReg = Phi->getOperand(i).getReg();
1624       MachineInstr *DI = MRI->getVRegDef(PhiReg);
1625 
1626       if (DI->getDesc().isAdd()) {
1627         // If the register operand to the add/sub is the PHI we are looking
1628         // at, this meets the induction pattern.
1629         Register IndReg = DI->getOperand(1).getReg();
1630         MachineOperand &Opnd2 = DI->getOperand(2);
1631         int64_t V;
1632         if (MRI->getVRegDef(IndReg) == Phi && checkForImmediate(Opnd2, V)) {
1633           Register UpdReg = DI->getOperand(0).getReg();
1634           IndRegs.insert(std::make_pair(UpdReg, std::make_pair(IndReg, V)));
1635         }
1636       }
1637     }  // for (i)
1638   }  // for (instr)
1639 
1640   if (IndRegs.empty())
1641     return false;
1642 
1643   MachineBasicBlock *TB = nullptr, *FB = nullptr;
1644   SmallVector<MachineOperand,2> Cond;
1645   // analyzeBranch returns true if it fails to analyze branch.
1646   bool NotAnalyzed = TII->analyzeBranch(*ExitingBlock, TB, FB, Cond, false);
1647   if (NotAnalyzed || Cond.empty())
1648     return false;
1649 
1650   if (ExitingBlock != Latch && (TB == Latch || FB == Latch)) {
1651     MachineBasicBlock *LTB = nullptr, *LFB = nullptr;
1652     SmallVector<MachineOperand,2> LCond;
1653     bool NotAnalyzed = TII->analyzeBranch(*Latch, LTB, LFB, LCond, false);
1654     if (NotAnalyzed)
1655       return false;
1656 
1657     // Since latch is not the exiting block, the latch branch should be an
1658     // unconditional branch to the loop header.
1659     if (TB == Latch)
1660       TB = (LTB == Header) ? LTB : LFB;
1661     else
1662       FB = (LTB == Header) ? LTB : LFB;
1663   }
1664   if (TB != Header) {
1665     if (FB != Header) {
1666       // The latch/exit block does not go back to the header.
1667       return false;
1668     }
1669     // FB is the header (i.e., uncond. jump to branch header)
1670     // In this case, the LoopBody -> TB should not be a back edge otherwise
1671     // it could result in an infinite loop after conversion to hw_loop.
1672     // This case can happen when the Latch has two jumps like this:
1673     // Jmp_c OuterLoopHeader <-- TB
1674     // Jmp   InnerLoopHeader <-- FB
1675     if (MDT->dominates(TB, FB))
1676       return false;
1677   }
1678 
1679   // Expecting a predicate register as a condition.  It won't be a hardware
1680   // predicate register at this point yet, just a vreg.
1681   // HexagonInstrInfo::analyzeBranch for negated branches inserts imm(0)
1682   // into Cond, followed by the predicate register.  For non-negated branches
1683   // it's just the register.
1684   unsigned CSz = Cond.size();
1685   if (CSz != 1 && CSz != 2)
1686     return false;
1687 
1688   if (!Cond[CSz-1].isReg())
1689     return false;
1690 
1691   Register P = Cond[CSz - 1].getReg();
1692   MachineInstr *PredDef = MRI->getVRegDef(P);
1693 
1694   if (!PredDef->isCompare())
1695     return false;
1696 
1697   SmallSet<Register,2> CmpRegs;
1698   MachineOperand *CmpImmOp = nullptr;
1699 
1700   // Go over all operands to the compare and look for immediate and register
1701   // operands.  Assume that if the compare has a single register use and a
1702   // single immediate operand, then the register is being compared with the
1703   // immediate value.
1704   for (MachineOperand &MO : PredDef->operands()) {
1705     if (MO.isReg()) {
1706       // Skip all implicit references.  In one case there was:
1707       //   %140 = FCMPUGT32_rr %138, %139, implicit %usr
1708       if (MO.isImplicit())
1709         continue;
1710       if (MO.isUse()) {
1711         if (!isImmediate(MO)) {
1712           CmpRegs.insert(MO.getReg());
1713           continue;
1714         }
1715         // Consider the register to be the "immediate" operand.
1716         if (CmpImmOp)
1717           return false;
1718         CmpImmOp = &MO;
1719       }
1720     } else if (MO.isImm()) {
1721       if (CmpImmOp)    // A second immediate argument?  Confusing.  Bail out.
1722         return false;
1723       CmpImmOp = &MO;
1724     }
1725   }
1726 
1727   if (CmpRegs.empty())
1728     return false;
1729 
1730   // Check if the compared register follows the order we want.  Fix if needed.
1731   for (RegisterInductionSet::iterator I = IndRegs.begin(), E = IndRegs.end();
1732        I != E; ++I) {
1733     // This is a success.  If the register used in the comparison is one that
1734     // we have identified as a bumped (updated) induction register, there is
1735     // nothing to do.
1736     if (CmpRegs.count(I->first))
1737       return true;
1738 
1739     // Otherwise, if the register being compared comes out of a PHI node,
1740     // and has been recognized as following the induction pattern, and is
1741     // compared against an immediate, we can fix it.
1742     const RegisterBump &RB = I->second;
1743     if (CmpRegs.count(RB.first)) {
1744       if (!CmpImmOp) {
1745         // If both operands to the compare instruction are registers, see if
1746         // it can be changed to use induction register as one of the operands.
1747         MachineInstr *IndI = nullptr;
1748         MachineInstr *nonIndI = nullptr;
1749         MachineOperand *IndMO = nullptr;
1750         MachineOperand *nonIndMO = nullptr;
1751 
1752         for (unsigned i = 1, n = PredDef->getNumOperands(); i < n; ++i) {
1753           MachineOperand &MO = PredDef->getOperand(i);
1754           if (MO.isReg() && MO.getReg() == RB.first) {
1755             LLVM_DEBUG(dbgs() << "\n DefMI(" << i
1756                               << ") = " << *(MRI->getVRegDef(I->first)));
1757             if (IndI)
1758               return false;
1759 
1760             IndI = MRI->getVRegDef(I->first);
1761             IndMO = &MO;
1762           } else if (MO.isReg()) {
1763             LLVM_DEBUG(dbgs() << "\n DefMI(" << i
1764                               << ") = " << *(MRI->getVRegDef(MO.getReg())));
1765             if (nonIndI)
1766               return false;
1767 
1768             nonIndI = MRI->getVRegDef(MO.getReg());
1769             nonIndMO = &MO;
1770           }
1771         }
1772         if (IndI && nonIndI &&
1773             nonIndI->getOpcode() == Hexagon::A2_addi &&
1774             nonIndI->getOperand(2).isImm() &&
1775             nonIndI->getOperand(2).getImm() == - RB.second) {
1776           bool Order = orderBumpCompare(IndI, PredDef);
1777           if (Order) {
1778             IndMO->setReg(I->first);
1779             nonIndMO->setReg(nonIndI->getOperand(1).getReg());
1780             return true;
1781           }
1782         }
1783         return false;
1784       }
1785 
1786       // It is not valid to do this transformation on an unsigned comparison
1787       // because it may underflow.
1788       Comparison::Kind Cmp =
1789           getComparisonKind(PredDef->getOpcode(), nullptr, nullptr, 0);
1790       if (!Cmp || Comparison::isUnsigned(Cmp))
1791         return false;
1792 
1793       // If the register is being compared against an immediate, try changing
1794       // the compare instruction to use induction register and adjust the
1795       // immediate operand.
1796       int64_t CmpImm = getImmediate(*CmpImmOp);
1797       int64_t V = RB.second;
1798       // Handle Overflow (64-bit).
1799       if (((V > 0) && (CmpImm > INT64_MAX - V)) ||
1800           ((V < 0) && (CmpImm < INT64_MIN - V)))
1801         return false;
1802       CmpImm += V;
1803       // Most comparisons of register against an immediate value allow
1804       // the immediate to be constant-extended. There are some exceptions
1805       // though. Make sure the new combination will work.
1806       if (CmpImmOp->isImm() && !TII->isExtendable(*PredDef) &&
1807           !TII->isValidOffset(PredDef->getOpcode(), CmpImm, TRI, false))
1808         return false;
1809 
1810       // Make sure that the compare happens after the bump.  Otherwise,
1811       // after the fixup, the compare would use a yet-undefined register.
1812       MachineInstr *BumpI = MRI->getVRegDef(I->first);
1813       bool Order = orderBumpCompare(BumpI, PredDef);
1814       if (!Order)
1815         return false;
1816 
1817       // Finally, fix the compare instruction.
1818       setImmediate(*CmpImmOp, CmpImm);
1819       for (MachineOperand &MO : PredDef->operands()) {
1820         if (MO.isReg() && MO.getReg() == RB.first) {
1821           MO.setReg(I->first);
1822           return true;
1823         }
1824       }
1825     }
1826   }
1827 
1828   return false;
1829 }
1830 
1831 /// createPreheaderForLoop - Create a preheader for a given loop.
1832 MachineBasicBlock *HexagonHardwareLoops::createPreheaderForLoop(
1833       MachineLoop *L) {
1834   if (MachineBasicBlock *TmpPH = MLI->findLoopPreheader(L, SpecPreheader))
1835     return TmpPH;
1836   if (!HWCreatePreheader)
1837     return nullptr;
1838 
1839   MachineBasicBlock *Header = L->getHeader();
1840   MachineBasicBlock *Latch = L->getLoopLatch();
1841   MachineBasicBlock *ExitingBlock = L->findLoopControlBlock();
1842   MachineFunction *MF = Header->getParent();
1843   DebugLoc DL;
1844 
1845 #ifndef NDEBUG
1846   if ((!PHFn.empty()) && (PHFn != MF->getName()))
1847     return nullptr;
1848 #endif
1849 
1850   if (!Latch || !ExitingBlock || Header->hasAddressTaken())
1851     return nullptr;
1852 
1853   using instr_iterator = MachineBasicBlock::instr_iterator;
1854 
1855   // Verify that all existing predecessors have analyzable branches
1856   // (or no branches at all).
1857   using MBBVector = std::vector<MachineBasicBlock *>;
1858 
1859   MBBVector Preds(Header->pred_begin(), Header->pred_end());
1860   SmallVector<MachineOperand,2> Tmp1;
1861   MachineBasicBlock *TB = nullptr, *FB = nullptr;
1862 
1863   if (TII->analyzeBranch(*ExitingBlock, TB, FB, Tmp1, false))
1864     return nullptr;
1865 
1866   for (MachineBasicBlock *PB : Preds) {
1867     bool NotAnalyzed = TII->analyzeBranch(*PB, TB, FB, Tmp1, false);
1868     if (NotAnalyzed)
1869       return nullptr;
1870   }
1871 
1872   MachineBasicBlock *NewPH = MF->CreateMachineBasicBlock();
1873   MF->insert(Header->getIterator(), NewPH);
1874 
1875   if (Header->pred_size() > 2) {
1876     // Ensure that the header has only two predecessors: the preheader and
1877     // the loop latch.  Any additional predecessors of the header should
1878     // join at the newly created preheader. Inspect all PHI nodes from the
1879     // header and create appropriate corresponding PHI nodes in the preheader.
1880 
1881     for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
1882          I != E && I->isPHI(); ++I) {
1883       MachineInstr *PN = &*I;
1884 
1885       const MCInstrDesc &PD = TII->get(TargetOpcode::PHI);
1886       MachineInstr *NewPN = MF->CreateMachineInstr(PD, DL);
1887       NewPH->insert(NewPH->end(), NewPN);
1888 
1889       Register PR = PN->getOperand(0).getReg();
1890       const TargetRegisterClass *RC = MRI->getRegClass(PR);
1891       Register NewPR = MRI->createVirtualRegister(RC);
1892       NewPN->addOperand(MachineOperand::CreateReg(NewPR, true));
1893 
1894       // Copy all non-latch operands of a header's PHI node to the newly
1895       // created PHI node in the preheader.
1896       for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) {
1897         Register PredR = PN->getOperand(i).getReg();
1898         unsigned PredRSub = PN->getOperand(i).getSubReg();
1899         MachineBasicBlock *PredB = PN->getOperand(i+1).getMBB();
1900         if (PredB == Latch)
1901           continue;
1902 
1903         MachineOperand MO = MachineOperand::CreateReg(PredR, false);
1904         MO.setSubReg(PredRSub);
1905         NewPN->addOperand(MO);
1906         NewPN->addOperand(MachineOperand::CreateMBB(PredB));
1907       }
1908 
1909       // Remove copied operands from the old PHI node and add the value
1910       // coming from the preheader's PHI.
1911       for (int i = PN->getNumOperands()-2; i > 0; i -= 2) {
1912         MachineBasicBlock *PredB = PN->getOperand(i+1).getMBB();
1913         if (PredB != Latch) {
1914           PN->removeOperand(i+1);
1915           PN->removeOperand(i);
1916         }
1917       }
1918       PN->addOperand(MachineOperand::CreateReg(NewPR, false));
1919       PN->addOperand(MachineOperand::CreateMBB(NewPH));
1920     }
1921   } else {
1922     assert(Header->pred_size() == 2);
1923 
1924     // The header has only two predecessors, but the non-latch predecessor
1925     // is not a preheader (e.g. it has other successors, etc.)
1926     // In such a case we don't need any extra PHI nodes in the new preheader,
1927     // all we need is to adjust existing PHIs in the header to now refer to
1928     // the new preheader.
1929     for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
1930          I != E && I->isPHI(); ++I) {
1931       MachineInstr *PN = &*I;
1932       for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) {
1933         MachineOperand &MO = PN->getOperand(i+1);
1934         if (MO.getMBB() != Latch)
1935           MO.setMBB(NewPH);
1936       }
1937     }
1938   }
1939 
1940   // "Reroute" the CFG edges to link in the new preheader.
1941   // If any of the predecessors falls through to the header, insert a branch
1942   // to the new preheader in that place.
1943   SmallVector<MachineOperand,1> Tmp2;
1944   SmallVector<MachineOperand,1> EmptyCond;
1945 
1946   TB = FB = nullptr;
1947 
1948   for (MachineBasicBlock *PB : Preds) {
1949     if (PB != Latch) {
1950       Tmp2.clear();
1951       bool NotAnalyzed = TII->analyzeBranch(*PB, TB, FB, Tmp2, false);
1952       (void)NotAnalyzed; // suppress compiler warning
1953       assert (!NotAnalyzed && "Should be analyzable!");
1954       if (TB != Header && (Tmp2.empty() || FB != Header))
1955         TII->insertBranch(*PB, NewPH, nullptr, EmptyCond, DL);
1956       PB->ReplaceUsesOfBlockWith(Header, NewPH);
1957     }
1958   }
1959 
1960   // It can happen that the latch block will fall through into the header.
1961   // Insert an unconditional branch to the header.
1962   TB = FB = nullptr;
1963   bool LatchNotAnalyzed = TII->analyzeBranch(*Latch, TB, FB, Tmp2, false);
1964   (void)LatchNotAnalyzed; // suppress compiler warning
1965   assert (!LatchNotAnalyzed && "Should be analyzable!");
1966   if (!TB && !FB)
1967     TII->insertBranch(*Latch, Header, nullptr, EmptyCond, DL);
1968 
1969   // Finally, the branch from the preheader to the header.
1970   TII->insertBranch(*NewPH, Header, nullptr, EmptyCond, DL);
1971   NewPH->addSuccessor(Header);
1972 
1973   MachineLoop *ParentLoop = L->getParentLoop();
1974   if (ParentLoop)
1975     ParentLoop->addBasicBlockToLoop(NewPH, MLI->getBase());
1976 
1977   // Update the dominator information with the new preheader.
1978   if (MDT) {
1979     if (MachineDomTreeNode *HN = MDT->getNode(Header)) {
1980       if (MachineDomTreeNode *DHN = HN->getIDom()) {
1981         MDT->addNewBlock(NewPH, DHN->getBlock());
1982         MDT->changeImmediateDominator(Header, NewPH);
1983       }
1984     }
1985   }
1986 
1987   return NewPH;
1988 }
1989