1 //===-- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass ----*- C++ -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Evan Cheng and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains a pass that performs load / store related peephole
11 // optimizations. This pass should be run after register allocation.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-ldst-opt"
17 #include "ARMAddressingModes.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMRegisterInfo.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineInstr.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/RegisterScavenging.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Target/MRegisterInfo.h"
30 #include "llvm/Target/TargetInstrInfo.h"
31 #include "llvm/Target/TargetMachine.h"
34 STATISTIC(NumLDMGened , "Number of ldm instructions generated");
35 STATISTIC(NumSTMGened , "Number of stm instructions generated");
36 STATISTIC(NumFLDMGened, "Number of fldm instructions generated");
37 STATISTIC(NumFSTMGened, "Number of fstm instructions generated");
40 struct VISIBILITY_HIDDEN ARMLoadStoreOpt : public MachineFunctionPass {
42 ARMLoadStoreOpt() : MachineFunctionPass((intptr_t)&ID) {}
44 const TargetInstrInfo *TII;
45 const MRegisterInfo *MRI;
49 virtual bool runOnMachineFunction(MachineFunction &Fn);
51 virtual const char *getPassName() const {
52 return "ARM load / store optimization pass";
56 struct MemOpQueueEntry {
59 MachineBasicBlock::iterator MBBI;
61 MemOpQueueEntry(int o, int p, MachineBasicBlock::iterator i)
62 : Offset(o), Position(p), MBBI(i), Merged(false) {};
64 typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
65 typedef MemOpQueue::iterator MemOpQueueIter;
67 SmallVector<MachineBasicBlock::iterator, 4>
68 MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex, unsigned Base,
69 int Opcode, unsigned Size, ARMCC::CondCodes Pred,
70 unsigned Scratch, MemOpQueue &MemOps);
72 void AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps);
73 bool LoadStoreMultipleOpti(MachineBasicBlock &MBB);
74 bool MergeReturnIntoLDM(MachineBasicBlock &MBB);
76 char ARMLoadStoreOpt::ID = 0;
79 /// createARMLoadStoreOptimizationPass - returns an instance of the load / store
80 /// optimization pass.
81 FunctionPass *llvm::createARMLoadStoreOptimizationPass() {
82 return new ARMLoadStoreOpt();
85 static int getLoadStoreMultipleOpcode(int Opcode) {
110 /// mergeOps - Create and insert a LDM or STM with Base as base register and
111 /// registers in Regs as the register operands that would be loaded / stored.
112 /// It returns true if the transformation is done.
113 static bool mergeOps(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
114 int Offset, unsigned Base, bool BaseKill, int Opcode,
115 ARMCC::CondCodes Pred, unsigned Scratch,
116 SmallVector<std::pair<unsigned, bool>, 8> &Regs,
117 const TargetInstrInfo *TII) {
118 // Only a single register to load / store. Don't bother.
119 unsigned NumRegs = Regs.size();
123 ARM_AM::AMSubMode Mode = ARM_AM::ia;
124 bool isAM4 = Opcode == ARM::LDR || Opcode == ARM::STR;
125 if (isAM4 && Offset == 4)
127 else if (isAM4 && Offset == -4 * (int)NumRegs + 4)
129 else if (isAM4 && Offset == -4 * (int)NumRegs)
131 else if (Offset != 0) {
132 // If starting offset isn't zero, insert a MI to materialize a new base.
133 // But only do so if it is cost effective, i.e. merging more than two
139 if (Opcode == ARM::LDR)
140 // If it is a load, then just use one of the destination register to
141 // use as the new base.
142 NewBase = Regs[NumRegs-1].first;
144 // Use the scratch register to use as a new base.
149 int BaseOpc = ARM::ADDri;
151 BaseOpc = ARM::SUBri;
154 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
155 if (ImmedOffset == -1)
156 return false; // Probably not worth it then.
158 BuildMI(MBB, MBBI, TII->get(BaseOpc), NewBase)
159 .addReg(Base, false, false, BaseKill).addImm(ImmedOffset).addImm(Pred);
161 BaseKill = true; // New base is always killed right its use.
164 bool isDPR = Opcode == ARM::FLDD || Opcode == ARM::FSTD;
165 bool isDef = Opcode == ARM::LDR || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
166 Opcode = getLoadStoreMultipleOpcode(Opcode);
167 MachineInstrBuilder MIB = (isAM4)
168 ? BuildMI(MBB, MBBI, TII->get(Opcode)).addReg(Base, false, false, BaseKill)
169 .addImm(ARM_AM::getAM4ModeImm(Mode)).addImm(Pred)
170 : BuildMI(MBB, MBBI, TII->get(Opcode)).addReg(Base, false, false, BaseKill)
171 .addImm(ARM_AM::getAM5Opc(Mode, false, isDPR ? NumRegs<<1 : NumRegs))
173 for (unsigned i = 0; i != NumRegs; ++i)
174 MIB = MIB.addReg(Regs[i].first, isDef, false, Regs[i].second);
179 /// MergeLDR_STR - Merge a number of load / store instructions into one or more
180 /// load / store multiple instructions.
181 SmallVector<MachineBasicBlock::iterator, 4>
182 ARMLoadStoreOpt::MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex,
183 unsigned Base, int Opcode, unsigned Size,
184 ARMCC::CondCodes Pred, unsigned Scratch,
185 MemOpQueue &MemOps) {
186 SmallVector<MachineBasicBlock::iterator, 4> Merges;
187 bool isAM4 = Opcode == ARM::LDR || Opcode == ARM::STR;
188 int Offset = MemOps[SIndex].Offset;
189 int SOffset = Offset;
190 unsigned Pos = MemOps[SIndex].Position;
191 MachineBasicBlock::iterator Loc = MemOps[SIndex].MBBI;
192 unsigned PReg = MemOps[SIndex].MBBI->getOperand(0).getReg();
193 unsigned PRegNum = ARMRegisterInfo::getRegisterNumbering(PReg);
194 bool isKill = MemOps[SIndex].MBBI->getOperand(0).isKill();
196 SmallVector<std::pair<unsigned,bool>, 8> Regs;
197 Regs.push_back(std::make_pair(PReg, isKill));
198 for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
199 int NewOffset = MemOps[i].Offset;
200 unsigned Reg = MemOps[i].MBBI->getOperand(0).getReg();
201 unsigned RegNum = ARMRegisterInfo::getRegisterNumbering(Reg);
202 isKill = MemOps[i].MBBI->getOperand(0).isKill();
203 // AM4 - register numbers in ascending order.
204 // AM5 - consecutive register numbers in ascending order.
205 if (NewOffset == Offset + (int)Size &&
206 ((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
208 Regs.push_back(std::make_pair(Reg, isKill));
211 // Can't merge this in. Try merge the earlier ones first.
212 if (mergeOps(MBB, ++Loc, SOffset, Base, false, Opcode, Pred, Scratch,
214 Merges.push_back(prior(Loc));
215 for (unsigned j = SIndex; j < i; ++j) {
216 MBB.erase(MemOps[j].MBBI);
217 MemOps[j].Merged = true;
220 SmallVector<MachineBasicBlock::iterator, 4> Merges2 =
221 MergeLDR_STR(MBB, i, Base, Opcode, Size, Pred, Scratch, MemOps);
222 Merges.append(Merges2.begin(), Merges2.end());
226 if (MemOps[i].Position > Pos) {
227 Pos = MemOps[i].Position;
228 Loc = MemOps[i].MBBI;
232 bool BaseKill = Loc->findRegisterUseOperandIdx(Base, true) != -1;
233 if (mergeOps(MBB, ++Loc, SOffset, Base, BaseKill, Opcode, Pred, Scratch,
235 Merges.push_back(prior(Loc));
236 for (unsigned i = SIndex, e = MemOps.size(); i != e; ++i) {
237 MBB.erase(MemOps[i].MBBI);
238 MemOps[i].Merged = true;
245 /// getInstrPredicate - If instruction is predicated, returns its predicate
246 /// condition, otherwise returns AL.
247 static ARMCC::CondCodes getInstrPredicate(MachineInstr *MI) {
248 int PIdx = MI->findFirstPredOperandIdx();
249 return PIdx == -1 ? ARMCC::AL
250 : (ARMCC::CondCodes)MI->getOperand(PIdx).getImmedValue();
253 static inline bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
254 unsigned Bytes, ARMCC::CondCodes Pred) {
255 return (MI && MI->getOpcode() == ARM::SUBri &&
256 MI->getOperand(0).getReg() == Base &&
257 MI->getOperand(1).getReg() == Base &&
258 ARM_AM::getAM2Offset(MI->getOperand(2).getImm()) == Bytes &&
259 getInstrPredicate(MI) == Pred);
262 static inline bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
263 unsigned Bytes, ARMCC::CondCodes Pred) {
264 return (MI && MI->getOpcode() == ARM::ADDri &&
265 MI->getOperand(0).getReg() == Base &&
266 MI->getOperand(1).getReg() == Base &&
267 ARM_AM::getAM2Offset(MI->getOperand(2).getImm()) == Bytes &&
268 getInstrPredicate(MI) == Pred);
271 static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
272 switch (MI->getOpcode()) {
284 return (MI->getNumOperands() - 3) * 4;
289 return ARM_AM::getAM5Offset(MI->getOperand(1).getImm()) * 4;
293 /// mergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
294 /// register into the LDM/STM/FLDM{D|S}/FSTM{D|S} op when possible:
296 /// stmia rn, <ra, rb, rc>
297 /// rn := rn + 4 * 3;
299 /// stmia rn!, <ra, rb, rc>
301 /// rn := rn - 4 * 3;
302 /// ldmia rn, <ra, rb, rc>
304 /// ldmdb rn!, <ra, rb, rc>
305 static bool mergeBaseUpdateLSMultiple(MachineBasicBlock &MBB,
306 MachineBasicBlock::iterator MBBI) {
307 MachineInstr *MI = MBBI;
308 unsigned Base = MI->getOperand(0).getReg();
309 unsigned Bytes = getLSMultipleTransferSize(MI);
310 ARMCC::CondCodes Pred = getInstrPredicate(MI);
311 int Opcode = MI->getOpcode();
312 bool isAM4 = Opcode == ARM::LDM || Opcode == ARM::STM;
315 if (ARM_AM::getAM4WBFlag(MI->getOperand(1).getImm()))
318 // Can't use the updating AM4 sub-mode if the base register is also a dest
319 // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
320 for (unsigned i = 3, e = MI->getNumOperands(); i != e; ++i) {
321 if (MI->getOperand(i).getReg() == Base)
325 ARM_AM::AMSubMode Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
326 if (MBBI != MBB.begin()) {
327 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
328 if (Mode == ARM_AM::ia &&
329 isMatchingDecrement(PrevMBBI, Base, Bytes, Pred)) {
330 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::db, true));
333 } else if (Mode == ARM_AM::ib &&
334 isMatchingDecrement(PrevMBBI, Base, Bytes, Pred)) {
335 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::da, true));
341 if (MBBI != MBB.end()) {
342 MachineBasicBlock::iterator NextMBBI = next(MBBI);
343 if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
344 isMatchingIncrement(NextMBBI, Base, Bytes, Pred)) {
345 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
348 } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
349 isMatchingDecrement(NextMBBI, Base, Bytes, Pred)) {
350 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
356 // FLDM{D|S}, FSTM{D|S} addressing mode 5 ops.
357 if (ARM_AM::getAM5WBFlag(MI->getOperand(1).getImm()))
360 ARM_AM::AMSubMode Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
361 unsigned Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
362 if (MBBI != MBB.begin()) {
363 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
364 if (Mode == ARM_AM::ia &&
365 isMatchingDecrement(PrevMBBI, Base, Bytes, Pred)) {
366 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::db, true, Offset));
372 if (MBBI != MBB.end()) {
373 MachineBasicBlock::iterator NextMBBI = next(MBBI);
374 if (Mode == ARM_AM::ia &&
375 isMatchingIncrement(NextMBBI, Base, Bytes, Pred)) {
376 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::ia, true, Offset));
386 static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
388 case ARM::LDR: return ARM::LDR_PRE;
389 case ARM::STR: return ARM::STR_PRE;
390 case ARM::FLDS: return ARM::FLDMS;
391 case ARM::FLDD: return ARM::FLDMD;
392 case ARM::FSTS: return ARM::FSTMS;
393 case ARM::FSTD: return ARM::FSTMD;
399 static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc) {
401 case ARM::LDR: return ARM::LDR_POST;
402 case ARM::STR: return ARM::STR_POST;
403 case ARM::FLDS: return ARM::FLDMS;
404 case ARM::FLDD: return ARM::FLDMD;
405 case ARM::FSTS: return ARM::FSTMS;
406 case ARM::FSTD: return ARM::FSTMD;
412 /// mergeBaseUpdateLoadStore - Fold proceeding/trailing inc/dec of base
413 /// register into the LDR/STR/FLD{D|S}/FST{D|S} op when possible:
414 static bool mergeBaseUpdateLoadStore(MachineBasicBlock &MBB,
415 MachineBasicBlock::iterator MBBI,
416 const TargetInstrInfo *TII) {
417 MachineInstr *MI = MBBI;
418 unsigned Base = MI->getOperand(1).getReg();
419 bool BaseKill = MI->getOperand(1).isKill();
420 unsigned Bytes = getLSMultipleTransferSize(MI);
421 int Opcode = MI->getOpcode();
422 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
423 if ((isAM2 && ARM_AM::getAM2Offset(MI->getOperand(3).getImm()) != 0) ||
424 (!isAM2 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0))
427 bool isLd = Opcode == ARM::LDR || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
428 // Can't do the merge if the destination register is the same as the would-be
429 // writeback register.
430 if (isLd && MI->getOperand(0).getReg() == Base)
433 ARMCC::CondCodes Pred = getInstrPredicate(MI);
434 bool DoMerge = false;
435 ARM_AM::AddrOpc AddSub = ARM_AM::add;
437 if (MBBI != MBB.begin()) {
438 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
439 if (isMatchingDecrement(PrevMBBI, Base, Bytes, Pred)) {
441 AddSub = ARM_AM::sub;
442 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
443 } else if (isAM2 && isMatchingIncrement(PrevMBBI, Base, Bytes, Pred)) {
445 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
451 if (!DoMerge && MBBI != MBB.end()) {
452 MachineBasicBlock::iterator NextMBBI = next(MBBI);
453 if (isAM2 && isMatchingDecrement(NextMBBI, Base, Bytes, Pred)) {
455 AddSub = ARM_AM::sub;
456 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
457 } else if (isMatchingIncrement(NextMBBI, Base, Bytes, Pred)) {
459 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
468 bool isDPR = NewOpc == ARM::FLDMD || NewOpc == ARM::FSTMD;
469 unsigned Offset = isAM2 ? ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift)
470 : ARM_AM::getAM5Opc((AddSub == ARM_AM::sub) ? ARM_AM::db : ARM_AM::ia,
471 true, isDPR ? 2 : 1);
474 // LDR_PRE, LDR_POST;
475 BuildMI(MBB, MBBI, TII->get(NewOpc), MI->getOperand(0).getReg())
477 .addReg(Base).addReg(0).addImm(Offset).addImm(Pred);
480 BuildMI(MBB, MBBI, TII->get(NewOpc)).addReg(Base, false, false, BaseKill)
481 .addImm(Offset).addImm(Pred).addReg(MI->getOperand(0).getReg(), true);
483 MachineOperand &MO = MI->getOperand(0);
485 // STR_PRE, STR_POST;
486 BuildMI(MBB, MBBI, TII->get(NewOpc), Base)
487 .addReg(MO.getReg(), false, false, MO.isKill())
488 .addReg(Base).addReg(0).addImm(Offset).addImm(Pred);
491 BuildMI(MBB, MBBI, TII->get(NewOpc)).addReg(Base).addImm(Offset)
492 .addImm(Pred).addReg(MO.getReg(), false, false, MO.isKill());
499 /// isMemoryOp - Returns true if instruction is a memory operations (that this
500 /// pass is capable of operating on).
501 static bool isMemoryOp(MachineInstr *MI) {
502 int Opcode = MI->getOpcode();
507 return MI->getOperand(1).isRegister() && MI->getOperand(2).getReg() == 0;
510 return MI->getOperand(1).isRegister();
513 return MI->getOperand(1).isRegister();
518 /// AdvanceRS - Advance register scavenger to just before the earliest memory
519 /// op that is being merged.
520 void ARMLoadStoreOpt::AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps) {
521 MachineBasicBlock::iterator Loc = MemOps[0].MBBI;
522 unsigned Position = MemOps[0].Position;
523 for (unsigned i = 1, e = MemOps.size(); i != e; ++i) {
524 if (MemOps[i].Position < Position) {
525 Position = MemOps[i].Position;
526 Loc = MemOps[i].MBBI;
530 if (Loc != MBB.begin())
531 RS->forward(prior(Loc));
534 /// LoadStoreMultipleOpti - An optimization pass to turn multiple LDR / STR
535 /// ops of the same base and incrementing offset into LDM / STM ops.
536 bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
537 unsigned NumMerges = 0;
538 unsigned NumMemOps = 0;
540 unsigned CurrBase = 0;
542 unsigned CurrSize = 0;
543 ARMCC::CondCodes CurrPred = ARMCC::AL;
544 unsigned Position = 0;
546 RS->enterBasicBlock(&MBB);
547 MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
549 bool Advance = false;
550 bool TryMerge = false;
551 bool Clobber = false;
553 bool isMemOp = isMemoryOp(MBBI);
555 int Opcode = MBBI->getOpcode();
556 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
557 unsigned Size = getLSMultipleTransferSize(MBBI);
558 unsigned Base = MBBI->getOperand(1).getReg();
559 ARMCC::CondCodes Pred = getInstrPredicate(MBBI);
560 const TargetInstrDescriptor *TID = MBBI->getInstrDescriptor();
561 unsigned OffField = MBBI->getOperand(TID->numOperands-2).getImm();
563 ? ARM_AM::getAM2Offset(OffField) : ARM_AM::getAM5Offset(OffField) * 4;
565 if (ARM_AM::getAM2Op(OffField) == ARM_AM::sub)
568 if (ARM_AM::getAM5Op(OffField) == ARM_AM::sub)
573 // r5 := ldr [r5, #4]
574 // r6 := ldr [r5, #8]
576 // The second ldr has effectively broken the chain even though it
577 // looks like the later ldr(s) use the same base register. Try to
578 // merge the ldr's so far, including this one. But don't try to
579 // combine the following ldr(s).
580 Clobber = (Opcode == ARM::LDR && Base == MBBI->getOperand(0).getReg());
581 if (CurrBase == 0 && !Clobber) {
582 // Start of a new chain.
587 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
596 if (CurrOpc == Opcode && CurrBase == Base && CurrPred == Pred) {
597 // Continue adding to the queue.
598 if (Offset > MemOps.back().Offset) {
599 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
603 for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
605 if (Offset < I->Offset) {
606 MemOps.insert(I, MemOpQueueEntry(Offset, Position, MBBI));
610 } else if (Offset == I->Offset) {
611 // Collision! This can't be merged!
628 // Try to find a free register to use as a new base in case it's needed.
629 // First advance to the instruction just before the start of the chain.
630 AdvanceRS(MBB, MemOps);
631 // Find a scratch register. Make sure it's a call clobbered register or
632 // a spilled callee-saved register.
633 unsigned Scratch = RS->FindUnusedReg(&ARM::GPRRegClass, true);
635 Scratch = RS->FindUnusedReg(&ARM::GPRRegClass,
636 AFI->getSpilledCSRegisters());
637 // Process the load / store instructions.
638 RS->forward(prior(MBBI));
641 SmallVector<MachineBasicBlock::iterator,4> MBBII =
642 MergeLDR_STR(MBB, 0, CurrBase, CurrOpc, CurrSize, CurrPred,
645 // Try folding preceeding/trailing base inc/dec into the generated
647 for (unsigned i = 0, e = MBBII.size(); i < e; ++i)
648 if (mergeBaseUpdateLSMultiple(MBB, MBBII[i]))
650 NumMerges += MBBII.size();
652 // Try folding preceeding/trailing base inc/dec into those load/store
653 // that were not merged to form LDM/STM ops.
654 for (unsigned i = 0; i != NumMemOps; ++i)
655 if (!MemOps[i].Merged)
656 if (mergeBaseUpdateLoadStore(MBB, MemOps[i].MBBI, TII))
659 // RS may be pointing to an instruction that's deleted.
660 RS->skipTo(prior(MBBI));
666 CurrPred = ARMCC::AL;
672 // If iterator hasn't been advanced and this is not a memory op, skip it.
673 // It can't start a new chain anyway.
674 if (!Advance && !isMemOp && MBBI != E) {
680 return NumMerges > 0;
683 /// MergeReturnIntoLDM - If this is a exit BB, try merging the return op
684 /// (bx lr) into the preceeding stack restore so it directly restore the value
686 /// ldmfd sp!, {r7, lr}
689 /// ldmfd sp!, {r7, pc}
690 bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
691 if (MBB.empty()) return false;
693 MachineBasicBlock::iterator MBBI = prior(MBB.end());
694 if (MBBI->getOpcode() == ARM::BX_RET && MBBI != MBB.begin()) {
695 MachineInstr *PrevMI = prior(MBBI);
696 if (PrevMI->getOpcode() == ARM::LDM) {
697 MachineOperand &MO = PrevMI->getOperand(PrevMI->getNumOperands()-1);
698 if (MO.getReg() == ARM::LR) {
699 PrevMI->setInstrDescriptor(TII->get(ARM::LDM_RET));
709 bool ARMLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
710 const TargetMachine &TM = Fn.getTarget();
711 AFI = Fn.getInfo<ARMFunctionInfo>();
712 TII = TM.getInstrInfo();
713 MRI = TM.getRegisterInfo();
714 RS = new RegScavenger();
716 bool Modified = false;
717 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
719 MachineBasicBlock &MBB = *MFI;
720 Modified |= LoadStoreMultipleOpti(MBB);
721 Modified |= MergeReturnIntoLDM(MBB);