HexagonHardwareLoops.cpp revision 234285
1//===-- HexagonHardwareLoops.cpp - Identify and generate hardware loops ---===//
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 pass identifies loops where we can generate the Hexagon hardware
11// loop instruction.  The hardware loop can perform loop branches with a
12// zero-cycle overhead.
13//
14// The pattern that defines the induction variable can changed depending on
15// prior optimizations.  For example, the IndVarSimplify phase run by 'opt'
16// normalizes induction variables, and the Loop Strength Reduction pass
17// run by 'llc' may also make changes to the induction variable.
18// The pattern detected by this phase is due to running Strength Reduction.
19//
20// Criteria for hardware loops:
21//  - Countable loops (w/ ind. var for a trip count)
22//  - Assumes loops are normalized by IndVarSimplify
23//  - Try inner-most loops first
24//  - No nested hardware loops.
25//  - No function calls in loops.
26//
27//===----------------------------------------------------------------------===//
28
29#define DEBUG_TYPE "hwloops"
30#include "Hexagon.h"
31#include "HexagonTargetMachine.h"
32#include "llvm/Constants.h"
33#include "llvm/PassSupport.h"
34#include "llvm/ADT/DenseMap.h"
35#include "llvm/ADT/Statistic.h"
36#include "llvm/CodeGen/Passes.h"
37#include "llvm/CodeGen/MachineDominators.h"
38#include "llvm/CodeGen/MachineFunction.h"
39#include "llvm/CodeGen/MachineFunctionPass.h"
40#include "llvm/CodeGen/MachineInstrBuilder.h"
41#include "llvm/CodeGen/MachineLoopInfo.h"
42#include "llvm/CodeGen/MachineRegisterInfo.h"
43#include "llvm/CodeGen/RegisterScavenging.h"
44#include "llvm/Support/Debug.h"
45#include "llvm/Support/raw_ostream.h"
46#include "llvm/Target/TargetInstrInfo.h"
47#include <algorithm>
48
49using namespace llvm;
50
51STATISTIC(NumHWLoops, "Number of loops converted to hardware loops");
52
53namespace {
54  class CountValue;
55  struct HexagonHardwareLoops : public MachineFunctionPass {
56    MachineLoopInfo       *MLI;
57    MachineRegisterInfo   *MRI;
58    const TargetInstrInfo *TII;
59
60  public:
61    static char ID;   // Pass identification, replacement for typeid
62
63    HexagonHardwareLoops() : MachineFunctionPass(ID) {}
64
65    virtual bool runOnMachineFunction(MachineFunction &MF);
66
67    const char *getPassName() const { return "Hexagon Hardware Loops"; }
68
69    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
70      AU.setPreservesCFG();
71      AU.addRequired<MachineDominatorTree>();
72      AU.addPreserved<MachineDominatorTree>();
73      AU.addRequired<MachineLoopInfo>();
74      AU.addPreserved<MachineLoopInfo>();
75      MachineFunctionPass::getAnalysisUsage(AU);
76    }
77
78  private:
79    /// getCanonicalInductionVariable - Check to see if the loop has a canonical
80    /// induction variable.
81    /// Should be defined in MachineLoop. Based upon version in class Loop.
82    const MachineInstr *getCanonicalInductionVariable(MachineLoop *L) const;
83
84    /// getTripCount - Return a loop-invariant LLVM register indicating the
85    /// number of times the loop will be executed.  If the trip-count cannot
86    /// be determined, this return null.
87    CountValue *getTripCount(MachineLoop *L) const;
88
89    /// isInductionOperation - Return true if the instruction matches the
90    /// pattern for an opertion that defines an induction variable.
91    bool isInductionOperation(const MachineInstr *MI, unsigned IVReg) const;
92
93    /// isInvalidOperation - Return true if the instruction is not valid within
94    /// a hardware loop.
95    bool isInvalidLoopOperation(const MachineInstr *MI) const;
96
97    /// containsInavlidInstruction - Return true if the loop contains an
98    /// instruction that inhibits using the hardware loop.
99    bool containsInvalidInstruction(MachineLoop *L) const;
100
101    /// converToHardwareLoop - Given a loop, check if we can convert it to a
102    /// hardware loop.  If so, then perform the conversion and return true.
103    bool convertToHardwareLoop(MachineLoop *L);
104
105  };
106
107  char HexagonHardwareLoops::ID = 0;
108
109
110  // CountValue class - Abstraction for a trip count of a loop. A
111  // smaller vesrsion of the MachineOperand class without the concerns
112  // of changing the operand representation.
113  class CountValue {
114  public:
115    enum CountValueType {
116      CV_Register,
117      CV_Immediate
118    };
119  private:
120    CountValueType Kind;
121    union Values {
122      unsigned RegNum;
123      int64_t ImmVal;
124      Values(unsigned r) : RegNum(r) {}
125      Values(int64_t i) : ImmVal(i) {}
126    } Contents;
127    bool isNegative;
128
129  public:
130    CountValue(unsigned r, bool neg) : Kind(CV_Register), Contents(r),
131                                       isNegative(neg) {}
132    explicit CountValue(int64_t i) : Kind(CV_Immediate), Contents(i),
133                                     isNegative(i < 0) {}
134    CountValueType getType() const { return Kind; }
135    bool isReg() const { return Kind == CV_Register; }
136    bool isImm() const { return Kind == CV_Immediate; }
137    bool isNeg() const { return isNegative; }
138
139    unsigned getReg() const {
140      assert(isReg() && "Wrong CountValue accessor");
141      return Contents.RegNum;
142    }
143    void setReg(unsigned Val) {
144      Contents.RegNum = Val;
145    }
146    int64_t getImm() const {
147      assert(isImm() && "Wrong CountValue accessor");
148      if (isNegative) {
149        return -Contents.ImmVal;
150      }
151      return Contents.ImmVal;
152    }
153    void setImm(int64_t Val) {
154      Contents.ImmVal = Val;
155    }
156
157    void print(raw_ostream &OS, const TargetMachine *TM = 0) const {
158      if (isReg()) { OS << PrintReg(getReg()); }
159      if (isImm()) { OS << getImm(); }
160    }
161  };
162
163  struct HexagonFixupHwLoops : public MachineFunctionPass {
164  public:
165    static char ID;     // Pass identification, replacement for typeid.
166
167    HexagonFixupHwLoops() : MachineFunctionPass(ID) {}
168
169    virtual bool runOnMachineFunction(MachineFunction &MF);
170
171    const char *getPassName() const { return "Hexagon Hardware Loop Fixup"; }
172
173    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
174      AU.setPreservesCFG();
175      MachineFunctionPass::getAnalysisUsage(AU);
176    }
177
178  private:
179    /// Maximum distance between the loop instr and the basic block.
180    /// Just an estimate.
181    static const unsigned MAX_LOOP_DISTANCE = 200;
182
183    /// fixupLoopInstrs - Check the offset between each loop instruction and
184    /// the loop basic block to determine if we can use the LOOP instruction
185    /// or if we need to set the LC/SA registers explicitly.
186    bool fixupLoopInstrs(MachineFunction &MF);
187
188    /// convertLoopInstr - Add the instruction to set the LC and SA registers
189    /// explicitly.
190    void convertLoopInstr(MachineFunction &MF,
191                          MachineBasicBlock::iterator &MII,
192                          RegScavenger &RS);
193
194  };
195
196  char HexagonFixupHwLoops::ID = 0;
197
198} // end anonymous namespace
199
200
201/// isHardwareLoop - Returns true if the instruction is a hardware loop
202/// instruction.
203static bool isHardwareLoop(const MachineInstr *MI) {
204  return MI->getOpcode() == Hexagon::LOOP0_r ||
205    MI->getOpcode() == Hexagon::LOOP0_i;
206}
207
208/// isCompareEquals - Returns true if the instruction is a compare equals
209/// instruction with an immediate operand.
210static bool isCompareEqualsImm(const MachineInstr *MI) {
211  return MI->getOpcode() == Hexagon::CMPEQri;
212}
213
214
215/// createHexagonHardwareLoops - Factory for creating
216/// the hardware loop phase.
217FunctionPass *llvm::createHexagonHardwareLoops() {
218  return new HexagonHardwareLoops();
219}
220
221
222bool HexagonHardwareLoops::runOnMachineFunction(MachineFunction &MF) {
223  DEBUG(dbgs() << "********* Hexagon Hardware Loops *********\n");
224
225  bool Changed = false;
226
227  // get the loop information
228  MLI = &getAnalysis<MachineLoopInfo>();
229  // get the register information
230  MRI = &MF.getRegInfo();
231  // the target specific instructio info.
232  TII = MF.getTarget().getInstrInfo();
233
234  for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end();
235       I != E; ++I) {
236    MachineLoop *L = *I;
237    if (!L->getParentLoop()) {
238      Changed |= convertToHardwareLoop(L);
239    }
240  }
241
242  return Changed;
243}
244
245/// getCanonicalInductionVariable - Check to see if the loop has a canonical
246/// induction variable. We check for a simple recurrence pattern - an
247/// integer recurrence that decrements by one each time through the loop and
248/// ends at zero.  If so, return the phi node that corresponds to it.
249///
250/// Based upon the similar code in LoopInfo except this code is specific to
251/// the machine.
252/// This method assumes that the IndVarSimplify pass has been run by 'opt'.
253///
254const MachineInstr
255*HexagonHardwareLoops::getCanonicalInductionVariable(MachineLoop *L) const {
256  MachineBasicBlock *TopMBB = L->getTopBlock();
257  MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin();
258  assert(PI != TopMBB->pred_end() &&
259         "Loop must have more than one incoming edge!");
260  MachineBasicBlock *Backedge = *PI++;
261  if (PI == TopMBB->pred_end()) return 0;  // dead loop
262  MachineBasicBlock *Incoming = *PI++;
263  if (PI != TopMBB->pred_end()) return 0;  // multiple backedges?
264
265  // make sure there is one incoming and one backedge and determine which
266  // is which.
267  if (L->contains(Incoming)) {
268    if (L->contains(Backedge))
269      return 0;
270    std::swap(Incoming, Backedge);
271  } else if (!L->contains(Backedge))
272    return 0;
273
274  // Loop over all of the PHI nodes, looking for a canonical induction variable:
275  //   - The PHI node is "reg1 = PHI reg2, BB1, reg3, BB2".
276  //   - The recurrence comes from the backedge.
277  //   - the definition is an induction operatio.n
278  for (MachineBasicBlock::iterator I = TopMBB->begin(), E = TopMBB->end();
279       I != E && I->isPHI(); ++I) {
280    const MachineInstr *MPhi = &*I;
281    unsigned DefReg = MPhi->getOperand(0).getReg();
282    for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2) {
283      // Check each operand for the value from the backedge.
284      MachineBasicBlock *MBB = MPhi->getOperand(i+1).getMBB();
285      if (L->contains(MBB)) { // operands comes from the backedge
286        // Check if the definition is an induction operation.
287        const MachineInstr *DI = MRI->getVRegDef(MPhi->getOperand(i).getReg());
288        if (isInductionOperation(DI, DefReg)) {
289          return MPhi;
290        }
291      }
292    }
293  }
294  return 0;
295}
296
297/// getTripCount - Return a loop-invariant LLVM value indicating the
298/// number of times the loop will be executed.  The trip count can
299/// be either a register or a constant value.  If the trip-count
300/// cannot be determined, this returns null.
301///
302/// We find the trip count from the phi instruction that defines the
303/// induction variable.  We follow the links to the CMP instruction
304/// to get the trip count.
305///
306/// Based upon getTripCount in LoopInfo.
307///
308CountValue *HexagonHardwareLoops::getTripCount(MachineLoop *L) const {
309  // Check that the loop has a induction variable.
310  const MachineInstr *IV_Inst = getCanonicalInductionVariable(L);
311  if (IV_Inst == 0) return 0;
312
313  // Canonical loops will end with a 'cmpeq_ri IV, Imm',
314  //  if Imm is 0, get the count from the PHI opnd
315  //  if Imm is -M, than M is the count
316  //  Otherwise, Imm is the count
317  const MachineOperand *IV_Opnd;
318  const MachineOperand *InitialValue;
319  if (!L->contains(IV_Inst->getOperand(2).getMBB())) {
320    InitialValue = &IV_Inst->getOperand(1);
321    IV_Opnd = &IV_Inst->getOperand(3);
322  } else {
323    InitialValue = &IV_Inst->getOperand(3);
324    IV_Opnd = &IV_Inst->getOperand(1);
325  }
326
327  // Look for the cmp instruction to determine if we
328  // can get a useful trip count.  The trip count can
329  // be either a register or an immediate.  The location
330  // of the value depends upon the type (reg or imm).
331  while ((IV_Opnd = IV_Opnd->getNextOperandForReg())) {
332    const MachineInstr *MI = IV_Opnd->getParent();
333    if (L->contains(MI) && isCompareEqualsImm(MI)) {
334      const MachineOperand &MO = MI->getOperand(2);
335      assert(MO.isImm() && "IV Cmp Operand should be 0");
336      int64_t ImmVal = MO.getImm();
337
338      const MachineInstr *IV_DefInstr = MRI->getVRegDef(IV_Opnd->getReg());
339      assert(L->contains(IV_DefInstr->getParent()) &&
340             "IV definition should occurs in loop");
341      int64_t iv_value = IV_DefInstr->getOperand(2).getImm();
342
343      if (ImmVal == 0) {
344        // Make sure the induction variable changes by one on each iteration.
345        if (iv_value != 1 && iv_value != -1) {
346          return 0;
347        }
348        return new CountValue(InitialValue->getReg(), iv_value > 0);
349      } else {
350        assert(InitialValue->isReg() && "Expecting register for init value");
351        const MachineInstr *DefInstr = MRI->getVRegDef(InitialValue->getReg());
352        if (DefInstr && DefInstr->getOpcode() == Hexagon::TFRI) {
353          int64_t count = ImmVal - DefInstr->getOperand(1).getImm();
354          if ((count % iv_value) != 0) {
355            return 0;
356          }
357          return new CountValue(count/iv_value);
358        }
359      }
360    }
361  }
362  return 0;
363}
364
365/// isInductionOperation - return true if the operation is matches the
366/// pattern that defines an induction variable:
367///    add iv, c
368///
369bool
370HexagonHardwareLoops::isInductionOperation(const MachineInstr *MI,
371                                           unsigned IVReg) const {
372  return (MI->getOpcode() ==
373          Hexagon::ADD_ri && MI->getOperand(1).getReg() == IVReg);
374}
375
376/// isInvalidOperation - Return true if the operation is invalid within
377/// hardware loop.
378bool
379HexagonHardwareLoops::isInvalidLoopOperation(const MachineInstr *MI) const {
380
381  // call is not allowed because the callee may use a hardware loop
382  if (MI->getDesc().isCall()) {
383    return true;
384  }
385  // do not allow nested hardware loops
386  if (isHardwareLoop(MI)) {
387    return true;
388  }
389  // check if the instruction defines a hardware loop register
390  for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
391    const MachineOperand &MO = MI->getOperand(i);
392    if (MO.isReg() && MO.isDef() &&
393        (MO.getReg() == Hexagon::LC0 || MO.getReg() == Hexagon::LC1 ||
394         MO.getReg() == Hexagon::SA0 || MO.getReg() == Hexagon::SA0)) {
395      return true;
396    }
397  }
398  return false;
399}
400
401/// containsInvalidInstruction - Return true if the loop contains
402/// an instruction that inhibits the use of the hardware loop function.
403///
404bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L) const {
405  const std::vector<MachineBasicBlock*> Blocks = L->getBlocks();
406  for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
407    MachineBasicBlock *MBB = Blocks[i];
408    for (MachineBasicBlock::iterator
409           MII = MBB->begin(), E = MBB->end(); MII != E; ++MII) {
410      const MachineInstr *MI = &*MII;
411      if (isInvalidLoopOperation(MI)) {
412        return true;
413      }
414    }
415  }
416  return false;
417}
418
419/// converToHardwareLoop - check if the loop is a candidate for
420/// converting to a hardware loop.  If so, then perform the
421/// transformation.
422///
423/// This function works on innermost loops first.  A loop can
424/// be converted if it is a counting loop; either a register
425/// value or an immediate.
426///
427/// The code makes several assumptions about the representation
428/// of the loop in llvm.
429bool HexagonHardwareLoops::convertToHardwareLoop(MachineLoop *L) {
430  bool Changed = false;
431  // Process nested loops first.
432  for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
433    Changed |= convertToHardwareLoop(*I);
434  }
435  // If a nested loop has been converted, then we can't convert this loop.
436  if (Changed) {
437    return Changed;
438  }
439  // Are we able to determine the trip count for the loop?
440  CountValue *TripCount = getTripCount(L);
441  if (TripCount == 0) {
442    return false;
443  }
444  // Does the loop contain any invalid instructions?
445  if (containsInvalidInstruction(L)) {
446    return false;
447  }
448  MachineBasicBlock *Preheader = L->getLoopPreheader();
449  // No preheader means there's not place for the loop instr.
450  if (Preheader == 0) {
451    return false;
452  }
453  MachineBasicBlock::iterator InsertPos = Preheader->getFirstTerminator();
454
455  MachineBasicBlock *LastMBB = L->getExitingBlock();
456  // Don't generate hw loop if the loop has more than one exit.
457  if (LastMBB == 0) {
458    return false;
459  }
460  MachineBasicBlock::iterator LastI = LastMBB->getFirstTerminator();
461
462  // Determine the loop start.
463  MachineBasicBlock *LoopStart = L->getTopBlock();
464  if (L->getLoopLatch() != LastMBB) {
465    // When the exit and latch are not the same, use the latch block as the
466    // start.
467    // The loop start address is used only after the 1st iteration, and the loop
468    // latch may contains instrs. that need to be executed after the 1st iter.
469    LoopStart = L->getLoopLatch();
470    // Make sure the latch is a successor of the exit, otherwise it won't work.
471    if (!LastMBB->isSuccessor(LoopStart)) {
472      return false;
473    }
474  }
475
476  // Convert the loop to a hardware loop
477  DEBUG(dbgs() << "Change to hardware loop at "; L->dump());
478
479  if (TripCount->isReg()) {
480    // Create a copy of the loop count register.
481    MachineFunction *MF = LastMBB->getParent();
482    const TargetRegisterClass *RC =
483      MF->getRegInfo().getRegClass(TripCount->getReg());
484    unsigned CountReg = MF->getRegInfo().createVirtualRegister(RC);
485    BuildMI(*Preheader, InsertPos, InsertPos->getDebugLoc(),
486            TII->get(TargetOpcode::COPY), CountReg).addReg(TripCount->getReg());
487    if (TripCount->isNeg()) {
488      unsigned CountReg1 = CountReg;
489      CountReg = MF->getRegInfo().createVirtualRegister(RC);
490      BuildMI(*Preheader, InsertPos, InsertPos->getDebugLoc(),
491              TII->get(Hexagon::NEG), CountReg).addReg(CountReg1);
492    }
493
494    // Add the Loop instruction to the begining of the loop.
495    BuildMI(*Preheader, InsertPos, InsertPos->getDebugLoc(),
496            TII->get(Hexagon::LOOP0_r)).addMBB(LoopStart).addReg(CountReg);
497  } else {
498    assert(TripCount->isImm() && "Expecting immedate vaule for trip count");
499    // Add the Loop immediate instruction to the beginning of the loop.
500    int64_t CountImm = TripCount->getImm();
501    BuildMI(*Preheader, InsertPos, InsertPos->getDebugLoc(),
502            TII->get(Hexagon::LOOP0_i)).addMBB(LoopStart).addImm(CountImm);
503  }
504
505  // Make sure the loop start always has a reference in the CFG.  We need to
506  // create a BlockAddress operand to get this mechanism to work both the
507  // MachineBasicBlock and BasicBlock objects need the flag set.
508  LoopStart->setHasAddressTaken();
509  // This line is needed to set the hasAddressTaken flag on the BasicBlock
510  // object
511  BlockAddress::get(const_cast<BasicBlock *>(LoopStart->getBasicBlock()));
512
513  // Replace the loop branch with an endloop instruction.
514  DebugLoc dl = LastI->getDebugLoc();
515  BuildMI(*LastMBB, LastI, dl, TII->get(Hexagon::ENDLOOP0)).addMBB(LoopStart);
516
517  // The loop ends with either:
518  //  - a conditional branch followed by an unconditional branch, or
519  //  - a conditional branch to the loop start.
520  if (LastI->getOpcode() == Hexagon::JMP_c ||
521      LastI->getOpcode() == Hexagon::JMP_cNot) {
522    // delete one and change/add an uncond. branch to out of the loop
523    MachineBasicBlock *BranchTarget = LastI->getOperand(1).getMBB();
524    LastI = LastMBB->erase(LastI);
525    if (!L->contains(BranchTarget)) {
526      if (LastI != LastMBB->end()) {
527        TII->RemoveBranch(*LastMBB);
528      }
529      SmallVector<MachineOperand, 0> Cond;
530      TII->InsertBranch(*LastMBB, BranchTarget, 0, Cond, dl);
531    }
532  } else {
533    // Conditional branch to loop start; just delete it.
534    LastMBB->erase(LastI);
535  }
536  delete TripCount;
537
538  ++NumHWLoops;
539  return true;
540}
541
542/// createHexagonFixupHwLoops - Factory for creating the hardware loop
543/// phase.
544FunctionPass *llvm::createHexagonFixupHwLoops() {
545  return new HexagonFixupHwLoops();
546}
547
548bool HexagonFixupHwLoops::runOnMachineFunction(MachineFunction &MF) {
549  DEBUG(dbgs() << "****** Hexagon Hardware Loop Fixup ******\n");
550
551  bool Changed = fixupLoopInstrs(MF);
552  return Changed;
553}
554
555/// fixupLoopInsts - For Hexagon, if the loop label is to far from the
556/// loop instruction then we need to set the LC0 and SA0 registers
557/// explicitly instead of using LOOP(start,count).  This function
558/// checks the distance, and generates register assignments if needed.
559///
560/// This function makes two passes over the basic blocks.  The first
561/// pass computes the offset of the basic block from the start.
562/// The second pass checks all the loop instructions.
563bool HexagonFixupHwLoops::fixupLoopInstrs(MachineFunction &MF) {
564
565  // Offset of the current instruction from the start.
566  unsigned InstOffset = 0;
567  // Map for each basic block to it's first instruction.
568  DenseMap<MachineBasicBlock*, unsigned> BlockToInstOffset;
569
570  // First pass - compute the offset of each basic block.
571  for (MachineFunction::iterator MBB = MF.begin(), MBBe = MF.end();
572       MBB != MBBe; ++MBB) {
573    BlockToInstOffset[MBB] = InstOffset;
574    InstOffset += (MBB->size() * 4);
575  }
576
577  // Second pass - check each loop instruction to see if it needs to
578  // be converted.
579  InstOffset = 0;
580  bool Changed = false;
581  RegScavenger RS;
582
583  // Loop over all the basic blocks.
584  for (MachineFunction::iterator MBB = MF.begin(), MBBe = MF.end();
585       MBB != MBBe; ++MBB) {
586    InstOffset = BlockToInstOffset[MBB];
587    RS.enterBasicBlock(MBB);
588
589    // Loop over all the instructions.
590    MachineBasicBlock::iterator MIE = MBB->end();
591    MachineBasicBlock::iterator MII = MBB->begin();
592    while (MII != MIE) {
593      if (isHardwareLoop(MII)) {
594        RS.forward(MII);
595        assert(MII->getOperand(0).isMBB() &&
596               "Expect a basic block as loop operand");
597        int diff = InstOffset - BlockToInstOffset[MII->getOperand(0).getMBB()];
598        diff = (diff > 0 ? diff : -diff);
599        if ((unsigned)diff > MAX_LOOP_DISTANCE) {
600          // Convert to explicity setting LC0 and SA0.
601          convertLoopInstr(MF, MII, RS);
602          MII = MBB->erase(MII);
603          Changed = true;
604        } else {
605          ++MII;
606        }
607      } else {
608        ++MII;
609      }
610      InstOffset += 4;
611    }
612  }
613
614  return Changed;
615
616}
617
618/// convertLoopInstr - convert a loop instruction to a sequence of instructions
619/// that set the lc and sa register explicitly.
620void HexagonFixupHwLoops::convertLoopInstr(MachineFunction &MF,
621                                           MachineBasicBlock::iterator &MII,
622                                           RegScavenger &RS) {
623  const TargetInstrInfo *TII = MF.getTarget().getInstrInfo();
624  MachineBasicBlock *MBB = MII->getParent();
625  DebugLoc DL = MII->getDebugLoc();
626  unsigned Scratch = RS.scavengeRegister(Hexagon::IntRegsRegisterClass, MII, 0);
627
628  // First, set the LC0 with the trip count.
629  if (MII->getOperand(1).isReg()) {
630    // Trip count is a register
631    BuildMI(*MBB, MII, DL, TII->get(Hexagon::TFCR), Hexagon::LC0)
632      .addReg(MII->getOperand(1).getReg());
633  } else {
634    // Trip count is an immediate.
635    BuildMI(*MBB, MII, DL, TII->get(Hexagon::TFRI), Scratch)
636      .addImm(MII->getOperand(1).getImm());
637    BuildMI(*MBB, MII, DL, TII->get(Hexagon::TFCR), Hexagon::LC0)
638      .addReg(Scratch);
639  }
640  // Then, set the SA0 with the loop start address.
641  BuildMI(*MBB, MII, DL, TII->get(Hexagon::CONST32_Label), Scratch)
642    .addMBB(MII->getOperand(0).getMBB());
643  BuildMI(*MBB, MII, DL, TII->get(Hexagon::TFCR), Hexagon::SA0).addReg(Scratch);
644}
645