1 //===-- PPCCTRLoops.cpp - Identify and generate CTR loops -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass identifies loops where we can generate the PPC branch instructions
11 // that decrement and test the count register (CTR) (bdnz and friends).
12 // This pass is based on the HexagonHardwareLoops pass.
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.
20 // Criteria for CTR 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 CTR loops.
25 // - No function calls in loops.
27 // Note: As with unconverted loops, PPCBranchSelector must be run after this
28 // pass in order to convert long-displacement jumps into jump pairs.
30 //===----------------------------------------------------------------------===//
32 #define DEBUG_TYPE "ctrloops"
34 #include "PPCTargetMachine.h"
35 #include "MCTargetDesc/PPCPredicates.h"
36 #include "llvm/Constants.h"
37 #include "llvm/PassSupport.h"
38 #include "llvm/ADT/DenseMap.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/CodeGen/Passes.h"
41 #include "llvm/CodeGen/MachineDominators.h"
42 #include "llvm/CodeGen/MachineFunction.h"
43 #include "llvm/CodeGen/MachineFunctionPass.h"
44 #include "llvm/CodeGen/MachineInstrBuilder.h"
45 #include "llvm/CodeGen/MachineLoopInfo.h"
46 #include "llvm/CodeGen/MachineRegisterInfo.h"
47 #include "llvm/CodeGen/RegisterScavenging.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Target/TargetInstrInfo.h"
55 STATISTIC(NumCTRLoops, "Number of loops converted to CTR loops");
59 struct PPCCTRLoops : public MachineFunctionPass {
61 MachineRegisterInfo *MRI;
62 const TargetInstrInfo *TII;
65 static char ID; // Pass identification, replacement for typeid
67 PPCCTRLoops() : MachineFunctionPass(ID) {}
69 virtual bool runOnMachineFunction(MachineFunction &MF);
71 const char *getPassName() const { return "PPC CTR Loops"; }
73 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
75 AU.addRequired<MachineDominatorTree>();
76 AU.addPreserved<MachineDominatorTree>();
77 AU.addRequired<MachineLoopInfo>();
78 AU.addPreserved<MachineLoopInfo>();
79 MachineFunctionPass::getAnalysisUsage(AU);
83 /// getCanonicalInductionVariable - Check to see if the loop has a canonical
84 /// induction variable.
85 /// Should be defined in MachineLoop. Based upon version in class Loop.
86 void getCanonicalInductionVariable(MachineLoop *L,
87 SmallVector<MachineInstr *, 4> &IVars,
88 SmallVector<MachineInstr *, 4> &IOps) const;
90 /// getTripCount - Return a loop-invariant LLVM register indicating the
91 /// number of times the loop will be executed. If the trip-count cannot
92 /// be determined, this return null.
93 CountValue *getTripCount(MachineLoop *L,
94 SmallVector<MachineInstr *, 2> &OldInsts) const;
96 /// isInductionOperation - Return true if the instruction matches the
97 /// pattern for an opertion that defines an induction variable.
98 bool isInductionOperation(const MachineInstr *MI, unsigned IVReg) const;
100 /// isInvalidOperation - Return true if the instruction is not valid within
102 bool isInvalidLoopOperation(const MachineInstr *MI) const;
104 /// containsInavlidInstruction - Return true if the loop contains an
105 /// instruction that inhibits using the CTR loop.
106 bool containsInvalidInstruction(MachineLoop *L) const;
108 /// converToCTRLoop - Given a loop, check if we can convert it to a
109 /// CTR loop. If so, then perform the conversion and return true.
110 bool convertToCTRLoop(MachineLoop *L);
112 /// isDead - Return true if the instruction is now dead.
113 bool isDead(const MachineInstr *MI,
114 SmallVector<MachineInstr *, 1> &DeadPhis) const;
116 /// removeIfDead - Remove the instruction if it is now dead.
117 void removeIfDead(MachineInstr *MI);
120 char PPCCTRLoops::ID = 0;
123 // CountValue class - Abstraction for a trip count of a loop. A
124 // smaller vesrsion of the MachineOperand class without the concerns
125 // of changing the operand representation.
128 enum CountValueType {
137 Values(unsigned r) : RegNum(r) {}
138 Values(int64_t i) : ImmVal(i) {}
143 CountValue(unsigned r, bool neg) : Kind(CV_Register), Contents(r),
145 explicit CountValue(int64_t i) : Kind(CV_Immediate), Contents(i),
147 CountValueType getType() const { return Kind; }
148 bool isReg() const { return Kind == CV_Register; }
149 bool isImm() const { return Kind == CV_Immediate; }
150 bool isNeg() const { return isNegative; }
152 unsigned getReg() const {
153 assert(isReg() && "Wrong CountValue accessor");
154 return Contents.RegNum;
156 void setReg(unsigned Val) {
157 Contents.RegNum = Val;
159 int64_t getImm() const {
160 assert(isImm() && "Wrong CountValue accessor");
162 return -Contents.ImmVal;
164 return Contents.ImmVal;
166 void setImm(int64_t Val) {
167 Contents.ImmVal = Val;
170 void print(raw_ostream &OS, const TargetMachine *TM = 0) const {
171 if (isReg()) { OS << PrintReg(getReg()); }
172 if (isImm()) { OS << getImm(); }
175 } // end anonymous namespace
178 /// isCompareEquals - Returns true if the instruction is a compare equals
179 /// instruction with an immediate operand.
180 static bool isCompareEqualsImm(const MachineInstr *MI, bool &SignedCmp) {
181 if (MI->getOpcode() == PPC::CMPWI || MI->getOpcode() == PPC::CMPDI) {
184 } else if (MI->getOpcode() == PPC::CMPLWI || MI->getOpcode() == PPC::CMPLDI) {
193 /// createPPCCTRLoops - Factory for creating
194 /// the CTR loop phase.
195 FunctionPass *llvm::createPPCCTRLoops() {
196 return new PPCCTRLoops();
200 bool PPCCTRLoops::runOnMachineFunction(MachineFunction &MF) {
201 DEBUG(dbgs() << "********* PPC CTR Loops *********\n");
203 bool Changed = false;
205 // get the loop information
206 MLI = &getAnalysis<MachineLoopInfo>();
207 // get the register information
208 MRI = &MF.getRegInfo();
209 // the target specific instructio info.
210 TII = MF.getTarget().getInstrInfo();
212 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end();
215 if (!L->getParentLoop()) {
216 Changed |= convertToCTRLoop(L);
223 /// getCanonicalInductionVariable - Check to see if the loop has a canonical
224 /// induction variable. We check for a simple recurrence pattern - an
225 /// integer recurrence that decrements by one each time through the loop and
226 /// ends at zero. If so, return the phi node that corresponds to it.
228 /// Based upon the similar code in LoopInfo except this code is specific to
230 /// This method assumes that the IndVarSimplify pass has been run by 'opt'.
233 PPCCTRLoops::getCanonicalInductionVariable(MachineLoop *L,
234 SmallVector<MachineInstr *, 4> &IVars,
235 SmallVector<MachineInstr *, 4> &IOps) const {
236 MachineBasicBlock *TopMBB = L->getTopBlock();
237 MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin();
238 assert(PI != TopMBB->pred_end() &&
239 "Loop must have more than one incoming edge!");
240 MachineBasicBlock *Backedge = *PI++;
241 if (PI == TopMBB->pred_end()) return; // dead loop
242 MachineBasicBlock *Incoming = *PI++;
243 if (PI != TopMBB->pred_end()) return; // multiple backedges?
245 // make sure there is one incoming and one backedge and determine which
247 if (L->contains(Incoming)) {
248 if (L->contains(Backedge))
250 std::swap(Incoming, Backedge);
251 } else if (!L->contains(Backedge))
254 // Loop over all of the PHI nodes, looking for a canonical induction variable:
255 // - The PHI node is "reg1 = PHI reg2, BB1, reg3, BB2".
256 // - The recurrence comes from the backedge.
257 // - the definition is an induction operatio.n
258 for (MachineBasicBlock::iterator I = TopMBB->begin(), E = TopMBB->end();
259 I != E && I->isPHI(); ++I) {
260 MachineInstr *MPhi = &*I;
261 unsigned DefReg = MPhi->getOperand(0).getReg();
262 for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2) {
263 // Check each operand for the value from the backedge.
264 MachineBasicBlock *MBB = MPhi->getOperand(i+1).getMBB();
265 if (L->contains(MBB)) { // operands comes from the backedge
266 // Check if the definition is an induction operation.
267 MachineInstr *DI = MRI->getVRegDef(MPhi->getOperand(i).getReg());
268 if (isInductionOperation(DI, DefReg)) {
270 IVars.push_back(MPhi);
278 /// getTripCount - Return a loop-invariant LLVM value indicating the
279 /// number of times the loop will be executed. The trip count can
280 /// be either a register or a constant value. If the trip-count
281 /// cannot be determined, this returns null.
283 /// We find the trip count from the phi instruction that defines the
284 /// induction variable. We follow the links to the CMP instruction
285 /// to get the trip count.
287 /// Based upon getTripCount in LoopInfo.
289 CountValue *PPCCTRLoops::getTripCount(MachineLoop *L,
290 SmallVector<MachineInstr *, 2> &OldInsts) const {
291 MachineBasicBlock *LastMBB = L->getExitingBlock();
292 // Don't generate a CTR loop if the loop has more than one exit.
296 MachineBasicBlock::iterator LastI = LastMBB->getFirstTerminator();
297 if (LastI->getOpcode() != PPC::BCC)
300 // We need to make sure that this compare is defining the condition
301 // register actually used by the terminating branch.
303 unsigned PredReg = LastI->getOperand(1).getReg();
304 DEBUG(dbgs() << "Examining loop with first terminator: " << *LastI);
306 unsigned PredCond = LastI->getOperand(0).getImm();
307 if (PredCond != PPC::PRED_EQ && PredCond != PPC::PRED_NE)
310 // Check that the loop has a induction variable.
311 SmallVector<MachineInstr *, 4> IVars, IOps;
312 getCanonicalInductionVariable(L, IVars, IOps);
313 for (unsigned i = 0; i < IVars.size(); ++i) {
314 MachineInstr *IOp = IOps[i];
315 MachineInstr *IV_Inst = IVars[i];
317 // Canonical loops will end with a 'cmpwi/cmpdi cr, IV, Imm',
318 // if Imm is 0, get the count from the PHI opnd
319 // if Imm is -M, than M is the count
320 // Otherwise, Imm is the count
321 MachineOperand *IV_Opnd;
322 const MachineOperand *InitialValue;
323 if (!L->contains(IV_Inst->getOperand(2).getMBB())) {
324 InitialValue = &IV_Inst->getOperand(1);
325 IV_Opnd = &IV_Inst->getOperand(3);
327 InitialValue = &IV_Inst->getOperand(3);
328 IV_Opnd = &IV_Inst->getOperand(1);
331 DEBUG(dbgs() << "Considering:\n");
332 DEBUG(dbgs() << " induction operation: " << *IOp);
333 DEBUG(dbgs() << " induction variable: " << *IV_Inst);
334 DEBUG(dbgs() << " initial value: " << *InitialValue << "\n");
336 // Look for the cmp instruction to determine if we
337 // can get a useful trip count. The trip count can
338 // be either a register or an immediate. The location
339 // of the value depends upon the type (reg or imm).
340 while ((IV_Opnd = IV_Opnd->getNextOperandForReg())) {
342 MachineInstr *MI = IV_Opnd->getParent();
343 if (L->contains(MI) && isCompareEqualsImm(MI, SignedCmp) &&
344 MI->getOperand(0).getReg() == PredReg) {
346 OldInsts.push_back(MI);
347 OldInsts.push_back(IOp);
349 DEBUG(dbgs() << " compare: " << *MI);
351 const MachineOperand &MO = MI->getOperand(2);
352 assert(MO.isImm() && "IV Cmp Operand should be an immediate");
356 ImmVal = (short) MO.getImm();
358 ImmVal = MO.getImm();
360 const MachineInstr *IV_DefInstr = MRI->getVRegDef(IV_Opnd->getReg());
361 assert(L->contains(IV_DefInstr->getParent()) &&
362 "IV definition should occurs in loop");
363 int64_t iv_value = (short) IV_DefInstr->getOperand(2).getImm();
365 assert(InitialValue->isReg() && "Expecting register for init value");
366 unsigned InitialValueReg = InitialValue->getReg();
368 const MachineInstr *DefInstr = MRI->getVRegDef(InitialValueReg);
370 // Here we need to look for an immediate load (an li or lis/ori pair).
371 if (DefInstr && (DefInstr->getOpcode() == PPC::ORI8 ||
372 DefInstr->getOpcode() == PPC::ORI)) {
373 int64_t start = (short) DefInstr->getOperand(2).getImm();
374 const MachineInstr *DefInstr2 =
375 MRI->getVRegDef(DefInstr->getOperand(0).getReg());
376 if (DefInstr2 && (DefInstr2->getOpcode() == PPC::LIS8 ||
377 DefInstr2->getOpcode() == PPC::LIS)) {
378 DEBUG(dbgs() << " initial constant: " << *DefInstr);
379 DEBUG(dbgs() << " initial constant: " << *DefInstr2);
381 start |= int64_t(short(DefInstr2->getOperand(1).getImm())) << 16;
383 int64_t count = ImmVal - start;
384 if ((count % iv_value) != 0) {
387 return new CountValue(count/iv_value);
389 } else if (DefInstr && (DefInstr->getOpcode() == PPC::LI8 ||
390 DefInstr->getOpcode() == PPC::LI)) {
391 DEBUG(dbgs() << " initial constant: " << *DefInstr);
393 int64_t count = ImmVal - int64_t(short(DefInstr->getOperand(1).getImm()));
394 if ((count % iv_value) != 0) {
397 return new CountValue(count/iv_value);
398 } else if (iv_value == 1 || iv_value == -1) {
399 // We can't determine a constant starting value.
401 return new CountValue(InitialValueReg, iv_value > 0);
403 // FIXME: handle non-zero end value.
405 // FIXME: handle non-unit increments (we might not want to introduce division
406 // but we can handle some 2^n cases with shifts).
414 /// isInductionOperation - return true if the operation is matches the
415 /// pattern that defines an induction variable:
419 PPCCTRLoops::isInductionOperation(const MachineInstr *MI,
420 unsigned IVReg) const {
421 return ((MI->getOpcode() == PPC::ADDI || MI->getOpcode() == PPC::ADDI8) &&
422 MI->getOperand(1).isReg() && // could be a frame index instead
423 MI->getOperand(1).getReg() == IVReg);
426 /// isInvalidOperation - Return true if the operation is invalid within
429 PPCCTRLoops::isInvalidLoopOperation(const MachineInstr *MI) const {
431 // call is not allowed because the callee may use a CTR loop
432 if (MI->getDesc().isCall()) {
435 // check if the instruction defines a CTR loop register
436 // (this will also catch nested CTR loops)
437 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
438 const MachineOperand &MO = MI->getOperand(i);
439 if (MO.isReg() && MO.isDef() &&
440 (MO.getReg() == PPC::CTR || MO.getReg() == PPC::CTR8)) {
447 /// containsInvalidInstruction - Return true if the loop contains
448 /// an instruction that inhibits the use of the CTR loop function.
450 bool PPCCTRLoops::containsInvalidInstruction(MachineLoop *L) const {
451 const std::vector<MachineBasicBlock*> Blocks = L->getBlocks();
452 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
453 MachineBasicBlock *MBB = Blocks[i];
454 for (MachineBasicBlock::iterator
455 MII = MBB->begin(), E = MBB->end(); MII != E; ++MII) {
456 const MachineInstr *MI = &*MII;
457 if (isInvalidLoopOperation(MI)) {
465 /// isDead returns true if the instruction is dead
466 /// (this was essentially copied from DeadMachineInstructionElim::isDead, but
467 /// with special cases for inline asm, physical registers and instructions with
468 /// side effects removed)
469 bool PPCCTRLoops::isDead(const MachineInstr *MI,
470 SmallVector<MachineInstr *, 1> &DeadPhis) const {
471 // Examine each operand.
472 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
473 const MachineOperand &MO = MI->getOperand(i);
474 if (MO.isReg() && MO.isDef()) {
475 unsigned Reg = MO.getReg();
476 if (!MRI->use_nodbg_empty(Reg)) {
477 // This instruction has users, but if the only user is the phi node for the
478 // parent block, and the only use of that phi node is this instruction, then
479 // this instruction is dead: both it (and the phi node) can be removed.
480 MachineRegisterInfo::use_iterator I = MRI->use_begin(Reg);
481 if (llvm::next(I) == MRI->use_end() &&
482 I.getOperand().getParent()->isPHI()) {
483 MachineInstr *OnePhi = I.getOperand().getParent();
485 for (unsigned j = 0, f = OnePhi->getNumOperands(); j != f; ++j) {
486 const MachineOperand &OPO = OnePhi->getOperand(j);
487 if (OPO.isReg() && OPO.isDef()) {
488 unsigned OPReg = OPO.getReg();
490 MachineRegisterInfo::use_iterator nextJ;
491 for (MachineRegisterInfo::use_iterator J = MRI->use_begin(OPReg),
492 E = MRI->use_end(); J!=E; J=nextJ) {
493 nextJ = llvm::next(J);
494 MachineOperand& Use = J.getOperand();
495 MachineInstr *UseMI = Use.getParent();
498 // The phi node has a user that is not MI, bail...
505 DeadPhis.push_back(OnePhi);
507 // This def has a non-debug use. Don't delete the instruction!
514 // If there are no defs with uses, the instruction is dead.
518 void PPCCTRLoops::removeIfDead(MachineInstr *MI) {
519 // This procedure was essentially copied from DeadMachineInstructionElim
521 SmallVector<MachineInstr *, 1> DeadPhis;
522 if (isDead(MI, DeadPhis)) {
523 DEBUG(dbgs() << "CTR looping will remove: " << *MI);
525 // It is possible that some DBG_VALUE instructions refer to this
526 // instruction. Examine each def operand for such references;
527 // if found, mark the DBG_VALUE as undef (but don't delete it).
528 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
529 const MachineOperand &MO = MI->getOperand(i);
530 if (!MO.isReg() || !MO.isDef())
532 unsigned Reg = MO.getReg();
533 MachineRegisterInfo::use_iterator nextI;
534 for (MachineRegisterInfo::use_iterator I = MRI->use_begin(Reg),
535 E = MRI->use_end(); I!=E; I=nextI) {
536 nextI = llvm::next(I); // I is invalidated by the setReg
537 MachineOperand& Use = I.getOperand();
538 MachineInstr *UseMI = Use.getParent();
541 if (Use.isDebug()) // this might also be a instr -> phi -> instr case
542 // which can also be removed.
543 UseMI->getOperand(0).setReg(0U);
547 MI->eraseFromParent();
548 for (unsigned i = 0; i < DeadPhis.size(); ++i) {
549 DeadPhis[i]->eraseFromParent();
554 /// converToCTRLoop - check if the loop is a candidate for
555 /// converting to a CTR loop. If so, then perform the
558 /// This function works on innermost loops first. A loop can
559 /// be converted if it is a counting loop; either a register
560 /// value or an immediate.
562 /// The code makes several assumptions about the representation
563 /// of the loop in llvm.
564 bool PPCCTRLoops::convertToCTRLoop(MachineLoop *L) {
565 bool Changed = false;
566 // Process nested loops first.
567 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
568 Changed |= convertToCTRLoop(*I);
570 // If a nested loop has been converted, then we can't convert this loop.
575 SmallVector<MachineInstr *, 2> OldInsts;
576 // Are we able to determine the trip count for the loop?
577 CountValue *TripCount = getTripCount(L, OldInsts);
578 if (TripCount == 0) {
579 DEBUG(dbgs() << "failed to get trip count!\n");
582 // Does the loop contain any invalid instructions?
583 if (containsInvalidInstruction(L)) {
586 MachineBasicBlock *Preheader = L->getLoopPreheader();
587 // No preheader means there's not place for the loop instr.
588 if (Preheader == 0) {
591 MachineBasicBlock::iterator InsertPos = Preheader->getFirstTerminator();
594 if (InsertPos != Preheader->end())
595 dl = InsertPos->getDebugLoc();
597 MachineBasicBlock *LastMBB = L->getExitingBlock();
598 // Don't generate CTR loop if the loop has more than one exit.
602 MachineBasicBlock::iterator LastI = LastMBB->getFirstTerminator();
604 // Determine the loop start.
605 MachineBasicBlock *LoopStart = L->getTopBlock();
606 if (L->getLoopLatch() != LastMBB) {
607 // When the exit and latch are not the same, use the latch block as the
609 // The loop start address is used only after the 1st iteration, and the loop
610 // latch may contains instrs. that need to be executed after the 1st iter.
611 LoopStart = L->getLoopLatch();
612 // Make sure the latch is a successor of the exit, otherwise it won't work.
613 if (!LastMBB->isSuccessor(LoopStart)) {
618 // Convert the loop to a CTR loop
619 DEBUG(dbgs() << "Change to CTR loop at "; L->dump());
621 MachineFunction *MF = LastMBB->getParent();
622 const PPCSubtarget &Subtarget = MF->getTarget().getSubtarget<PPCSubtarget>();
623 bool isPPC64 = Subtarget.isPPC64();
625 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
626 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass;
627 const TargetRegisterClass *RC = isPPC64 ? G8RC : GPRC;
630 if (TripCount->isReg()) {
631 // Create a copy of the loop count register.
632 const TargetRegisterClass *SrcRC =
633 MF->getRegInfo().getRegClass(TripCount->getReg());
634 CountReg = MF->getRegInfo().createVirtualRegister(RC);
635 unsigned CopyOp = (isPPC64 && SrcRC == GPRC) ?
636 (unsigned) PPC::EXTSW_32_64 :
637 (unsigned) TargetOpcode::COPY;
638 BuildMI(*Preheader, InsertPos, dl,
639 TII->get(CopyOp), CountReg).addReg(TripCount->getReg());
640 if (TripCount->isNeg()) {
641 unsigned CountReg1 = CountReg;
642 CountReg = MF->getRegInfo().createVirtualRegister(RC);
643 BuildMI(*Preheader, InsertPos, dl,
644 TII->get(isPPC64 ? PPC::NEG8 : PPC::NEG),
645 CountReg).addReg(CountReg1);
648 assert(TripCount->isImm() && "Expecting immedate vaule for trip count");
649 // Put the trip count in a register for transfer into the count register.
651 int64_t CountImm = TripCount->getImm();
652 assert(!TripCount->isNeg() && "Constant trip count must be positive");
654 CountReg = MF->getRegInfo().createVirtualRegister(RC);
655 if (CountImm > 0xFFFF) {
656 BuildMI(*Preheader, InsertPos, dl,
657 TII->get(isPPC64 ? PPC::LIS8 : PPC::LIS),
658 CountReg).addImm(CountImm >> 16);
659 unsigned CountReg1 = CountReg;
660 CountReg = MF->getRegInfo().createVirtualRegister(RC);
661 BuildMI(*Preheader, InsertPos, dl,
662 TII->get(isPPC64 ? PPC::ORI8 : PPC::ORI),
663 CountReg).addReg(CountReg1).addImm(CountImm & 0xFFFF);
665 BuildMI(*Preheader, InsertPos, dl,
666 TII->get(isPPC64 ? PPC::LI8 : PPC::LI),
667 CountReg).addImm(CountImm);
671 // Add the mtctr instruction to the beginning of the loop.
672 BuildMI(*Preheader, InsertPos, dl,
673 TII->get(isPPC64 ? PPC::MTCTR8 : PPC::MTCTR)).addReg(CountReg,
674 TripCount->isImm() ? RegState::Kill : 0);
676 // Make sure the loop start always has a reference in the CFG. We need to
677 // create a BlockAddress operand to get this mechanism to work both the
678 // MachineBasicBlock and BasicBlock objects need the flag set.
679 LoopStart->setHasAddressTaken();
680 // This line is needed to set the hasAddressTaken flag on the BasicBlock
682 BlockAddress::get(const_cast<BasicBlock *>(LoopStart->getBasicBlock()));
684 // Replace the loop branch with a bdnz instruction.
685 dl = LastI->getDebugLoc();
686 const std::vector<MachineBasicBlock*> Blocks = L->getBlocks();
687 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
688 MachineBasicBlock *MBB = Blocks[i];
689 if (MBB != Preheader)
690 MBB->addLiveIn(isPPC64 ? PPC::CTR8 : PPC::CTR);
693 // The loop ends with either:
694 // - a conditional branch followed by an unconditional branch, or
695 // - a conditional branch to the loop start.
696 assert(LastI->getOpcode() == PPC::BCC &&
697 "loop end must start with a BCC instruction");
698 // Either the BCC branches to the beginning of the loop, or it
699 // branches out of the loop and there is an unconditional branch
700 // to the start of the loop.
701 MachineBasicBlock *BranchTarget = LastI->getOperand(2).getMBB();
702 BuildMI(*LastMBB, LastI, dl,
703 TII->get((BranchTarget == LoopStart) ?
704 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
705 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(BranchTarget);
707 // Conditional branch; just delete it.
708 DEBUG(dbgs() << "Removing old branch: " << *LastI);
709 LastMBB->erase(LastI);
713 // The induction operation (add) and the comparison (cmpwi) may now be
714 // unneeded. If these are unneeded, then remove them.
715 for (unsigned i = 0; i < OldInsts.size(); ++i)
716 removeIfDead(OldInsts[i]);