1 //===-- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass ----*- C++ -*-=//
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 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 "ARMBaseInstrInfo.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMRegisterInfo.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.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/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/RegisterScavenging.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/Target/TargetInstrInfo.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
33 #include "llvm/Support/Compiler.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/ADT/DenseMap.h"
36 #include "llvm/ADT/STLExtras.h"
37 #include "llvm/ADT/SmallPtrSet.h"
38 #include "llvm/ADT/SmallSet.h"
39 #include "llvm/ADT/SmallVector.h"
40 #include "llvm/ADT/Statistic.h"
43 STATISTIC(NumLDMGened , "Number of ldm instructions generated");
44 STATISTIC(NumSTMGened , "Number of stm instructions generated");
45 STATISTIC(NumFLDMGened, "Number of fldm instructions generated");
46 STATISTIC(NumFSTMGened, "Number of fstm instructions generated");
47 STATISTIC(NumLdStMoved, "Number of load / store instructions moved");
48 STATISTIC(NumLDRDFormed,"Number of ldrd created before allocation");
49 STATISTIC(NumSTRDFormed,"Number of strd created before allocation");
50 STATISTIC(NumLDRD2LDM, "Number of ldrd instructions turned back into ldm");
51 STATISTIC(NumSTRD2STM, "Number of strd instructions turned back into stm");
52 STATISTIC(NumLDRD2LDR, "Number of ldrd instructions turned back into ldr's");
53 STATISTIC(NumSTRD2STR, "Number of strd instructions turned back into str's");
55 /// ARMAllocLoadStoreOpt - Post- register allocation pass the combine
56 /// load / store instructions to form ldm / stm instructions.
59 struct VISIBILITY_HIDDEN ARMLoadStoreOpt : public MachineFunctionPass {
61 ARMLoadStoreOpt() : MachineFunctionPass(&ID) {}
63 const TargetInstrInfo *TII;
64 const TargetRegisterInfo *TRI;
69 virtual bool runOnMachineFunction(MachineFunction &Fn);
71 virtual const char *getPassName() const {
72 return "ARM load / store optimization pass";
76 struct MemOpQueueEntry {
79 MachineBasicBlock::iterator MBBI;
81 MemOpQueueEntry(int o, int p, MachineBasicBlock::iterator i)
82 : Offset(o), Position(p), MBBI(i), Merged(false) {};
84 typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
85 typedef MemOpQueue::iterator MemOpQueueIter;
87 bool MergeOps(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
88 int Offset, unsigned Base, bool BaseKill, int Opcode,
89 ARMCC::CondCodes Pred, unsigned PredReg, unsigned Scratch,
90 DebugLoc dl, SmallVector<std::pair<unsigned, bool>, 8> &Regs);
91 void MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex, unsigned Base,
92 int Opcode, unsigned Size,
93 ARMCC::CondCodes Pred, unsigned PredReg,
94 unsigned Scratch, MemOpQueue &MemOps,
95 SmallVector<MachineBasicBlock::iterator, 4> &Merges);
97 void AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps);
98 bool FixInvalidRegPairOp(MachineBasicBlock &MBB,
99 MachineBasicBlock::iterator &MBBI);
100 bool MergeBaseUpdateLoadStore(MachineBasicBlock &MBB,
101 MachineBasicBlock::iterator MBBI,
102 const TargetInstrInfo *TII,
104 MachineBasicBlock::iterator &I);
105 bool MergeBaseUpdateLSMultiple(MachineBasicBlock &MBB,
106 MachineBasicBlock::iterator MBBI,
108 MachineBasicBlock::iterator &I);
109 bool LoadStoreMultipleOpti(MachineBasicBlock &MBB);
110 bool MergeReturnIntoLDM(MachineBasicBlock &MBB);
112 char ARMLoadStoreOpt::ID = 0;
115 static int getLoadStoreMultipleOpcode(int Opcode) {
143 default: llvm_unreachable("Unhandled opcode!");
148 static bool isT2i32Load(unsigned Opc) {
149 return Opc == ARM::t2LDRi12 || Opc == ARM::t2LDRi8;
152 static bool isi32Load(unsigned Opc) {
153 return Opc == ARM::LDR || isT2i32Load(Opc);
156 static bool isT2i32Store(unsigned Opc) {
157 return Opc == ARM::t2STRi12 || Opc == ARM::t2STRi8;
160 static bool isi32Store(unsigned Opc) {
161 return Opc == ARM::STR || isT2i32Store(Opc);
164 /// MergeOps - Create and insert a LDM or STM with Base as base register and
165 /// registers in Regs as the register operands that would be loaded / stored.
166 /// It returns true if the transformation is done.
168 ARMLoadStoreOpt::MergeOps(MachineBasicBlock &MBB,
169 MachineBasicBlock::iterator MBBI,
170 int Offset, unsigned Base, bool BaseKill,
171 int Opcode, ARMCC::CondCodes Pred,
172 unsigned PredReg, unsigned Scratch, DebugLoc dl,
173 SmallVector<std::pair<unsigned, bool>, 8> &Regs) {
174 // Only a single register to load / store. Don't bother.
175 unsigned NumRegs = Regs.size();
179 ARM_AM::AMSubMode Mode = ARM_AM::ia;
180 bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
181 if (isAM4 && Offset == 4) {
183 // Thumb2 does not support ldmib / stmib.
186 } else if (isAM4 && Offset == -4 * (int)NumRegs + 4) {
188 // Thumb2 does not support ldmda / stmda.
191 } else if (isAM4 && Offset == -4 * (int)NumRegs) {
193 } else if (Offset != 0) {
194 // If starting offset isn't zero, insert a MI to materialize a new base.
195 // But only do so if it is cost effective, i.e. merging more than two
201 if (isi32Load(Opcode))
202 // If it is a load, then just use one of the destination register to
203 // use as the new base.
204 NewBase = Regs[NumRegs-1].first;
206 // Use the scratch register to use as a new base.
211 int BaseOpc = !isThumb2
213 : ((Base == ARM::SP) ? ARM::t2ADDrSPi : ARM::t2ADDri);
217 : ((Base == ARM::SP) ? ARM::t2SUBrSPi : ARM::t2SUBri);
220 int ImmedOffset = isThumb2
221 ? ARM_AM::getT2SOImmVal(Offset) : ARM_AM::getSOImmVal(Offset);
222 if (ImmedOffset == -1)
223 // FIXME: Try t2ADDri12 or t2SUBri12?
224 return false; // Probably not worth it then.
226 BuildMI(MBB, MBBI, dl, TII->get(BaseOpc), NewBase)
227 .addReg(Base, getKillRegState(BaseKill)).addImm(Offset)
228 .addImm(Pred).addReg(PredReg).addReg(0);
230 BaseKill = true; // New base is always killed right its use.
233 bool isDPR = Opcode == ARM::FLDD || Opcode == ARM::FSTD;
234 bool isDef = isi32Load(Opcode) || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
235 Opcode = getLoadStoreMultipleOpcode(Opcode);
236 MachineInstrBuilder MIB = (isAM4)
237 ? BuildMI(MBB, MBBI, dl, TII->get(Opcode))
238 .addReg(Base, getKillRegState(BaseKill))
239 .addImm(ARM_AM::getAM4ModeImm(Mode)).addImm(Pred).addReg(PredReg)
240 : BuildMI(MBB, MBBI, dl, TII->get(Opcode))
241 .addReg(Base, getKillRegState(BaseKill))
242 .addImm(ARM_AM::getAM5Opc(Mode, false, isDPR ? NumRegs<<1 : NumRegs))
243 .addImm(Pred).addReg(PredReg);
244 for (unsigned i = 0; i != NumRegs; ++i)
245 MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
246 | getKillRegState(Regs[i].second));
251 /// MergeLDR_STR - Merge a number of load / store instructions into one or more
252 /// load / store multiple instructions.
254 ARMLoadStoreOpt::MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex,
255 unsigned Base, int Opcode, unsigned Size,
256 ARMCC::CondCodes Pred, unsigned PredReg,
257 unsigned Scratch, MemOpQueue &MemOps,
258 SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
259 bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
260 int Offset = MemOps[SIndex].Offset;
261 int SOffset = Offset;
262 unsigned Pos = MemOps[SIndex].Position;
263 MachineBasicBlock::iterator Loc = MemOps[SIndex].MBBI;
264 DebugLoc dl = Loc->getDebugLoc();
265 unsigned PReg = Loc->getOperand(0).getReg();
266 unsigned PRegNum = ARMRegisterInfo::getRegisterNumbering(PReg);
267 bool isKill = Loc->getOperand(0).isKill();
269 SmallVector<std::pair<unsigned,bool>, 8> Regs;
270 Regs.push_back(std::make_pair(PReg, isKill));
271 for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
272 int NewOffset = MemOps[i].Offset;
273 unsigned Reg = MemOps[i].MBBI->getOperand(0).getReg();
274 unsigned RegNum = ARMRegisterInfo::getRegisterNumbering(Reg);
275 isKill = MemOps[i].MBBI->getOperand(0).isKill();
276 // AM4 - register numbers in ascending order.
277 // AM5 - consecutive register numbers in ascending order.
278 if (NewOffset == Offset + (int)Size &&
279 ((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
281 Regs.push_back(std::make_pair(Reg, isKill));
284 // Can't merge this in. Try merge the earlier ones first.
285 if (MergeOps(MBB, ++Loc, SOffset, Base, false, Opcode, Pred, PredReg,
286 Scratch, dl, Regs)) {
287 Merges.push_back(prior(Loc));
288 for (unsigned j = SIndex; j < i; ++j) {
289 MBB.erase(MemOps[j].MBBI);
290 MemOps[j].Merged = true;
293 MergeLDR_STR(MBB, i, Base, Opcode, Size, Pred, PredReg, Scratch,
298 if (MemOps[i].Position > Pos) {
299 Pos = MemOps[i].Position;
300 Loc = MemOps[i].MBBI;
304 bool BaseKill = Loc->findRegisterUseOperandIdx(Base, true) != -1;
305 if (MergeOps(MBB, ++Loc, SOffset, Base, BaseKill, Opcode, Pred, PredReg,
306 Scratch, dl, Regs)) {
307 Merges.push_back(prior(Loc));
308 for (unsigned i = SIndex, e = MemOps.size(); i != e; ++i) {
309 MBB.erase(MemOps[i].MBBI);
310 MemOps[i].Merged = true;
317 static inline bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
318 unsigned Bytes, unsigned Limit,
319 ARMCC::CondCodes Pred, unsigned PredReg){
320 unsigned MyPredReg = 0;
323 if (MI->getOpcode() != ARM::t2SUBri &&
324 MI->getOpcode() != ARM::t2SUBrSPi &&
325 MI->getOpcode() != ARM::t2SUBrSPi12 &&
326 MI->getOpcode() != ARM::tSUBspi &&
327 MI->getOpcode() != ARM::SUBri)
330 // Make sure the offset fits in 8 bits.
331 if (Bytes <= 0 || (Limit && Bytes >= Limit))
334 unsigned Scale = (MI->getOpcode() == ARM::tSUBspi) ? 4 : 1; // FIXME
335 return (MI->getOperand(0).getReg() == Base &&
336 MI->getOperand(1).getReg() == Base &&
337 (MI->getOperand(2).getImm()*Scale) == Bytes &&
338 llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
339 MyPredReg == PredReg);
342 static inline bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
343 unsigned Bytes, unsigned Limit,
344 ARMCC::CondCodes Pred, unsigned PredReg){
345 unsigned MyPredReg = 0;
348 if (MI->getOpcode() != ARM::t2ADDri &&
349 MI->getOpcode() != ARM::t2ADDrSPi &&
350 MI->getOpcode() != ARM::t2ADDrSPi12 &&
351 MI->getOpcode() != ARM::tADDspi &&
352 MI->getOpcode() != ARM::ADDri)
355 if (Bytes <= 0 || (Limit && Bytes >= Limit))
356 // Make sure the offset fits in 8 bits.
359 unsigned Scale = (MI->getOpcode() == ARM::tADDspi) ? 4 : 1; // FIXME
360 return (MI->getOperand(0).getReg() == Base &&
361 MI->getOperand(1).getReg() == Base &&
362 (MI->getOperand(2).getImm()*Scale) == Bytes &&
363 llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
364 MyPredReg == PredReg);
367 static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
368 switch (MI->getOpcode()) {
386 return (MI->getNumOperands() - 4) * 4;
391 return ARM_AM::getAM5Offset(MI->getOperand(1).getImm()) * 4;
395 /// MergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
396 /// register into the LDM/STM/FLDM{D|S}/FSTM{D|S} op when possible:
398 /// stmia rn, <ra, rb, rc>
399 /// rn := rn + 4 * 3;
401 /// stmia rn!, <ra, rb, rc>
403 /// rn := rn - 4 * 3;
404 /// ldmia rn, <ra, rb, rc>
406 /// ldmdb rn!, <ra, rb, rc>
407 bool ARMLoadStoreOpt::MergeBaseUpdateLSMultiple(MachineBasicBlock &MBB,
408 MachineBasicBlock::iterator MBBI,
410 MachineBasicBlock::iterator &I) {
411 MachineInstr *MI = MBBI;
412 unsigned Base = MI->getOperand(0).getReg();
413 unsigned Bytes = getLSMultipleTransferSize(MI);
414 unsigned PredReg = 0;
415 ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
416 int Opcode = MI->getOpcode();
417 bool isAM4 = Opcode == ARM::LDM || Opcode == ARM::t2LDM ||
418 Opcode == ARM::STM || Opcode == ARM::t2STM;
421 if (ARM_AM::getAM4WBFlag(MI->getOperand(1).getImm()))
424 // Can't use the updating AM4 sub-mode if the base register is also a dest
425 // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
426 for (unsigned i = 3, e = MI->getNumOperands(); i != e; ++i) {
427 if (MI->getOperand(i).getReg() == Base)
431 ARM_AM::AMSubMode Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
432 if (MBBI != MBB.begin()) {
433 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
434 if (Mode == ARM_AM::ia &&
435 isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
436 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::db, true));
439 } else if (Mode == ARM_AM::ib &&
440 isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
441 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::da, true));
447 if (MBBI != MBB.end()) {
448 MachineBasicBlock::iterator NextMBBI = next(MBBI);
449 if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
450 isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
451 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
458 } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
459 isMatchingDecrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
460 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
470 // FLDM{D|S}, FSTM{D|S} addressing mode 5 ops.
471 if (ARM_AM::getAM5WBFlag(MI->getOperand(1).getImm()))
474 ARM_AM::AMSubMode Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
475 unsigned Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
476 if (MBBI != MBB.begin()) {
477 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
478 if (Mode == ARM_AM::ia &&
479 isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
480 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::db, true, Offset));
486 if (MBBI != MBB.end()) {
487 MachineBasicBlock::iterator NextMBBI = next(MBBI);
488 if (Mode == ARM_AM::ia &&
489 isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
490 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::ia, true, Offset));
504 static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
506 case ARM::LDR: return ARM::LDR_PRE;
507 case ARM::STR: return ARM::STR_PRE;
508 case ARM::FLDS: return ARM::FLDMS;
509 case ARM::FLDD: return ARM::FLDMD;
510 case ARM::FSTS: return ARM::FSTMS;
511 case ARM::FSTD: return ARM::FSTMD;
514 return ARM::t2LDR_PRE;
517 return ARM::t2STR_PRE;
518 default: llvm_unreachable("Unhandled opcode!");
523 static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc) {
525 case ARM::LDR: return ARM::LDR_POST;
526 case ARM::STR: return ARM::STR_POST;
527 case ARM::FLDS: return ARM::FLDMS;
528 case ARM::FLDD: return ARM::FLDMD;
529 case ARM::FSTS: return ARM::FSTMS;
530 case ARM::FSTD: return ARM::FSTMD;
533 return ARM::t2LDR_POST;
536 return ARM::t2STR_POST;
537 default: llvm_unreachable("Unhandled opcode!");
542 /// MergeBaseUpdateLoadStore - Fold proceeding/trailing inc/dec of base
543 /// register into the LDR/STR/FLD{D|S}/FST{D|S} op when possible:
544 bool ARMLoadStoreOpt::MergeBaseUpdateLoadStore(MachineBasicBlock &MBB,
545 MachineBasicBlock::iterator MBBI,
546 const TargetInstrInfo *TII,
548 MachineBasicBlock::iterator &I) {
549 MachineInstr *MI = MBBI;
550 unsigned Base = MI->getOperand(1).getReg();
551 bool BaseKill = MI->getOperand(1).isKill();
552 unsigned Bytes = getLSMultipleTransferSize(MI);
553 int Opcode = MI->getOpcode();
554 DebugLoc dl = MI->getDebugLoc();
555 bool isAM5 = Opcode == ARM::FLDD || Opcode == ARM::FLDS ||
556 Opcode == ARM::FSTD || Opcode == ARM::FSTS;
557 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
558 if (isAM2 && ARM_AM::getAM2Offset(MI->getOperand(3).getImm()) != 0)
560 else if (isAM5 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0)
562 else if (isT2i32Load(Opcode) || isT2i32Store(Opcode))
563 if (MI->getOperand(2).getImm() != 0)
566 bool isLd = isi32Load(Opcode) || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
567 // Can't do the merge if the destination register is the same as the would-be
568 // writeback register.
569 if (isLd && MI->getOperand(0).getReg() == Base)
572 unsigned PredReg = 0;
573 ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
574 bool DoMerge = false;
575 ARM_AM::AddrOpc AddSub = ARM_AM::add;
577 // AM2 - 12 bits, thumb2 - 8 bits.
578 unsigned Limit = isAM5 ? 0 : (isAM2 ? 0x1000 : 0x100);
579 if (MBBI != MBB.begin()) {
580 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
581 if (isMatchingDecrement(PrevMBBI, Base, Bytes, Limit, Pred, PredReg)) {
583 AddSub = ARM_AM::sub;
584 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
586 isMatchingIncrement(PrevMBBI, Base, Bytes, Limit,Pred,PredReg)) {
588 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
594 if (!DoMerge && MBBI != MBB.end()) {
595 MachineBasicBlock::iterator NextMBBI = next(MBBI);
597 isMatchingDecrement(NextMBBI, Base, Bytes, Limit, Pred, PredReg)) {
599 AddSub = ARM_AM::sub;
600 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
601 } else if (isMatchingIncrement(NextMBBI, Base, Bytes, Limit,Pred,PredReg)) {
603 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
617 bool isDPR = NewOpc == ARM::FLDMD || NewOpc == ARM::FSTMD;
620 Offset = ARM_AM::getAM5Opc((AddSub == ARM_AM::sub)
622 : ARM_AM::ia, true, (isDPR ? 2 : 1));
624 Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
626 Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
630 BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
631 .addReg(Base, getKillRegState(BaseKill))
632 .addImm(Offset).addImm(Pred).addReg(PredReg)
633 .addReg(MI->getOperand(0).getReg(), RegState::Define);
635 // LDR_PRE, LDR_POST,
636 BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
637 .addReg(Base, RegState::Define)
638 .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
640 // t2LDR_PRE, t2LDR_POST
641 BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
642 .addReg(Base, RegState::Define)
643 .addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
645 MachineOperand &MO = MI->getOperand(0);
648 BuildMI(MBB, MBBI, dl, TII->get(NewOpc)).addReg(Base).addImm(Offset)
649 .addImm(Pred).addReg(PredReg)
650 .addReg(MO.getReg(), getKillRegState(MO.isKill()));
653 BuildMI(MBB, MBBI, dl, TII->get(NewOpc), Base)
654 .addReg(MO.getReg(), getKillRegState(MO.isKill()))
655 .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
657 // t2STR_PRE, t2STR_POST
658 BuildMI(MBB, MBBI, dl, TII->get(NewOpc), Base)
659 .addReg(MO.getReg(), getKillRegState(MO.isKill()))
660 .addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
667 /// isMemoryOp - Returns true if instruction is a memory operations (that this
668 /// pass is capable of operating on).
669 static bool isMemoryOp(const MachineInstr *MI) {
670 int Opcode = MI->getOpcode();
675 return MI->getOperand(1).isReg() && MI->getOperand(2).getReg() == 0;
678 return MI->getOperand(1).isReg();
681 return MI->getOperand(1).isReg();
691 /// AdvanceRS - Advance register scavenger to just before the earliest memory
692 /// op that is being merged.
693 void ARMLoadStoreOpt::AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps) {
694 MachineBasicBlock::iterator Loc = MemOps[0].MBBI;
695 unsigned Position = MemOps[0].Position;
696 for (unsigned i = 1, e = MemOps.size(); i != e; ++i) {
697 if (MemOps[i].Position < Position) {
698 Position = MemOps[i].Position;
699 Loc = MemOps[i].MBBI;
703 if (Loc != MBB.begin())
704 RS->forward(prior(Loc));
707 static int getMemoryOpOffset(const MachineInstr *MI) {
708 int Opcode = MI->getOpcode();
709 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
710 bool isAM3 = Opcode == ARM::LDRD || Opcode == ARM::STRD;
711 unsigned NumOperands = MI->getDesc().getNumOperands();
712 unsigned OffField = MI->getOperand(NumOperands-3).getImm();
714 if (Opcode == ARM::t2LDRi12 || Opcode == ARM::t2LDRi8 ||
715 Opcode == ARM::t2STRi12 || Opcode == ARM::t2STRi8 ||
716 Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8)
720 ? ARM_AM::getAM2Offset(OffField)
721 : (isAM3 ? ARM_AM::getAM3Offset(OffField)
722 : ARM_AM::getAM5Offset(OffField) * 4);
724 if (ARM_AM::getAM2Op(OffField) == ARM_AM::sub)
727 if (ARM_AM::getAM3Op(OffField) == ARM_AM::sub)
730 if (ARM_AM::getAM5Op(OffField) == ARM_AM::sub)
736 static void InsertLDR_STR(MachineBasicBlock &MBB,
737 MachineBasicBlock::iterator &MBBI,
738 int OffImm, bool isDef,
739 DebugLoc dl, unsigned NewOpc,
740 unsigned Reg, bool RegDeadKill,
741 unsigned BaseReg, bool BaseKill,
742 unsigned OffReg, bool OffKill,
743 ARMCC::CondCodes Pred, unsigned PredReg,
744 const TargetInstrInfo *TII) {
747 Offset = ARM_AM::getAM2Opc(ARM_AM::sub, -OffImm, ARM_AM::no_shift);
749 Offset = ARM_AM::getAM2Opc(ARM_AM::add, OffImm, ARM_AM::no_shift);
751 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
752 .addReg(Reg, getDefRegState(true) | getDeadRegState(RegDeadKill))
753 .addReg(BaseReg, getKillRegState(BaseKill))
754 .addReg(OffReg, getKillRegState(OffKill))
756 .addImm(Pred).addReg(PredReg);
758 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
759 .addReg(Reg, getKillRegState(RegDeadKill))
760 .addReg(BaseReg, getKillRegState(BaseKill))
761 .addReg(OffReg, getKillRegState(OffKill))
763 .addImm(Pred).addReg(PredReg);
766 bool ARMLoadStoreOpt::FixInvalidRegPairOp(MachineBasicBlock &MBB,
767 MachineBasicBlock::iterator &MBBI) {
768 MachineInstr *MI = &*MBBI;
769 unsigned Opcode = MI->getOpcode();
770 if (Opcode == ARM::LDRD || Opcode == ARM::STRD) {
771 unsigned EvenReg = MI->getOperand(0).getReg();
772 unsigned OddReg = MI->getOperand(1).getReg();
773 unsigned EvenRegNum = TRI->getDwarfRegNum(EvenReg, false);
774 unsigned OddRegNum = TRI->getDwarfRegNum(OddReg, false);
775 if ((EvenRegNum & 1) == 0 && (EvenRegNum + 1) == OddRegNum)
778 bool isLd = Opcode == ARM::LDRD;
779 bool EvenDeadKill = isLd ?
780 MI->getOperand(0).isDead() : MI->getOperand(0).isKill();
781 bool OddDeadKill = isLd ?
782 MI->getOperand(1).isDead() : MI->getOperand(1).isKill();
783 const MachineOperand &BaseOp = MI->getOperand(2);
784 unsigned BaseReg = BaseOp.getReg();
785 bool BaseKill = BaseOp.isKill();
786 const MachineOperand &OffOp = MI->getOperand(3);
787 unsigned OffReg = OffOp.getReg();
788 bool OffKill = OffOp.isKill();
789 int OffImm = getMemoryOpOffset(MI);
790 unsigned PredReg = 0;
791 ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
793 if (OddRegNum > EvenRegNum && OffReg == 0 && OffImm == 0) {
794 // Ascending register numbers and no offset. It's safe to change it to a
796 unsigned NewOpc = (Opcode == ARM::LDRD) ? ARM::LDM : ARM::STM;
798 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
799 .addReg(BaseReg, getKillRegState(BaseKill))
800 .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
801 .addImm(Pred).addReg(PredReg)
802 .addReg(EvenReg, getDefRegState(isLd) | getDeadRegState(EvenDeadKill))
803 .addReg(OddReg, getDefRegState(isLd) | getDeadRegState(OddDeadKill));
806 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
807 .addReg(BaseReg, getKillRegState(BaseKill))
808 .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
809 .addImm(Pred).addReg(PredReg)
810 .addReg(EvenReg, getKillRegState(EvenDeadKill))
811 .addReg(OddReg, getKillRegState(OddDeadKill));
815 // Split into two instructions.
816 unsigned NewOpc = (Opcode == ARM::LDRD) ? ARM::LDR : ARM::STR;
817 DebugLoc dl = MBBI->getDebugLoc();
818 // If this is a load and base register is killed, it may have been
819 // re-defed by the load, make sure the first load does not clobber it.
821 (BaseKill || OffKill) &&
822 (TRI->regsOverlap(EvenReg, BaseReg) ||
823 (OffReg && TRI->regsOverlap(EvenReg, OffReg)))) {
824 assert(!TRI->regsOverlap(OddReg, BaseReg) &&
825 (!OffReg || !TRI->regsOverlap(OddReg, OffReg)));
826 InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc, OddReg, OddDeadKill,
827 BaseReg, false, OffReg, false, Pred, PredReg, TII);
828 InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc, EvenReg, EvenDeadKill,
829 BaseReg, BaseKill, OffReg, OffKill, Pred, PredReg, TII);
831 InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
832 EvenReg, EvenDeadKill, BaseReg, false, OffReg, false,
834 InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
835 OddReg, OddDeadKill, BaseReg, BaseKill, OffReg, OffKill,
850 /// LoadStoreMultipleOpti - An optimization pass to turn multiple LDR / STR
851 /// ops of the same base and incrementing offset into LDM / STM ops.
852 bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
853 unsigned NumMerges = 0;
854 unsigned NumMemOps = 0;
856 unsigned CurrBase = 0;
858 unsigned CurrSize = 0;
859 ARMCC::CondCodes CurrPred = ARMCC::AL;
860 unsigned CurrPredReg = 0;
861 unsigned Position = 0;
862 SmallVector<MachineBasicBlock::iterator,4> Merges;
864 RS->enterBasicBlock(&MBB);
865 MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
867 if (FixInvalidRegPairOp(MBB, MBBI))
870 bool Advance = false;
871 bool TryMerge = false;
872 bool Clobber = false;
874 bool isMemOp = isMemoryOp(MBBI);
876 int Opcode = MBBI->getOpcode();
877 unsigned Size = getLSMultipleTransferSize(MBBI);
878 unsigned Base = MBBI->getOperand(1).getReg();
879 unsigned PredReg = 0;
880 ARMCC::CondCodes Pred = llvm::getInstrPredicate(MBBI, PredReg);
881 int Offset = getMemoryOpOffset(MBBI);
884 // r5 := ldr [r5, #4]
885 // r6 := ldr [r5, #8]
887 // The second ldr has effectively broken the chain even though it
888 // looks like the later ldr(s) use the same base register. Try to
889 // merge the ldr's so far, including this one. But don't try to
890 // combine the following ldr(s).
891 Clobber = (isi32Load(Opcode) && Base == MBBI->getOperand(0).getReg());
892 if (CurrBase == 0 && !Clobber) {
893 // Start of a new chain.
898 CurrPredReg = PredReg;
899 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
908 if (CurrOpc == Opcode && CurrBase == Base && CurrPred == Pred) {
909 // No need to match PredReg.
910 // Continue adding to the queue.
911 if (Offset > MemOps.back().Offset) {
912 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
916 for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
918 if (Offset < I->Offset) {
919 MemOps.insert(I, MemOpQueueEntry(Offset, Position, MBBI));
923 } else if (Offset == I->Offset) {
924 // Collision! This can't be merged!
941 // Try to find a free register to use as a new base in case it's needed.
942 // First advance to the instruction just before the start of the chain.
943 AdvanceRS(MBB, MemOps);
944 // Find a scratch register.
945 unsigned Scratch = RS->FindUnusedReg(ARM::GPRRegisterClass);
946 // Process the load / store instructions.
947 RS->forward(prior(MBBI));
951 MergeLDR_STR(MBB, 0, CurrBase, CurrOpc, CurrSize,
952 CurrPred, CurrPredReg, Scratch, MemOps, Merges);
954 // Try folding preceeding/trailing base inc/dec into the generated
956 for (unsigned i = 0, e = Merges.size(); i < e; ++i)
957 if (MergeBaseUpdateLSMultiple(MBB, Merges[i], Advance, MBBI))
959 NumMerges += Merges.size();
961 // Try folding preceeding/trailing base inc/dec into those load/store
962 // that were not merged to form LDM/STM ops.
963 for (unsigned i = 0; i != NumMemOps; ++i)
964 if (!MemOps[i].Merged)
965 if (MergeBaseUpdateLoadStore(MBB, MemOps[i].MBBI, TII,Advance,MBBI))
968 // RS may be pointing to an instruction that's deleted.
969 RS->skipTo(prior(MBBI));
970 } else if (NumMemOps == 1) {
971 // Try folding preceeding/trailing base inc/dec into the single
973 if (MergeBaseUpdateLoadStore(MBB, MemOps[0].MBBI, TII, Advance, MBBI)) {
975 RS->forward(prior(MBBI));
982 CurrPred = ARMCC::AL;
989 // If iterator hasn't been advanced and this is not a memory op, skip it.
990 // It can't start a new chain anyway.
991 if (!Advance && !isMemOp && MBBI != E) {
997 return NumMerges > 0;
1001 struct OffsetCompare {
1002 bool operator()(const MachineInstr *LHS, const MachineInstr *RHS) const {
1003 int LOffset = getMemoryOpOffset(LHS);
1004 int ROffset = getMemoryOpOffset(RHS);
1005 assert(LHS == RHS || LOffset != ROffset);
1006 return LOffset > ROffset;
1011 /// MergeReturnIntoLDM - If this is a exit BB, try merging the return op
1012 /// (bx lr) into the preceeding stack restore so it directly restore the value
1014 /// ldmfd sp!, {r7, lr}
1017 /// ldmfd sp!, {r7, pc}
1018 bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
1019 if (MBB.empty()) return false;
1021 MachineBasicBlock::iterator MBBI = prior(MBB.end());
1022 if (MBBI != MBB.begin() &&
1023 (MBBI->getOpcode() == ARM::BX_RET || MBBI->getOpcode() == ARM::tBX_RET)) {
1024 MachineInstr *PrevMI = prior(MBBI);
1025 if (PrevMI->getOpcode() == ARM::LDM || PrevMI->getOpcode() == ARM::t2LDM) {
1026 MachineOperand &MO = PrevMI->getOperand(PrevMI->getNumOperands()-1);
1027 if (MO.getReg() != ARM::LR)
1029 unsigned NewOpc = isThumb2 ? ARM::t2LDM_RET : ARM::LDM_RET;
1030 PrevMI->setDesc(TII->get(NewOpc));
1039 bool ARMLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
1040 const TargetMachine &TM = Fn.getTarget();
1041 AFI = Fn.getInfo<ARMFunctionInfo>();
1042 TII = TM.getInstrInfo();
1043 TRI = TM.getRegisterInfo();
1044 RS = new RegScavenger();
1045 isThumb2 = AFI->isThumb2Function();
1047 bool Modified = false;
1048 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
1050 MachineBasicBlock &MBB = *MFI;
1051 Modified |= LoadStoreMultipleOpti(MBB);
1052 Modified |= MergeReturnIntoLDM(MBB);
1060 /// ARMPreAllocLoadStoreOpt - Pre- register allocation pass that move
1061 /// load / stores from consecutive locations close to make it more
1062 /// likely they will be combined later.
1065 struct VISIBILITY_HIDDEN ARMPreAllocLoadStoreOpt : public MachineFunctionPass{
1067 ARMPreAllocLoadStoreOpt() : MachineFunctionPass(&ID) {}
1069 const TargetData *TD;
1070 const TargetInstrInfo *TII;
1071 const TargetRegisterInfo *TRI;
1072 const ARMSubtarget *STI;
1073 MachineRegisterInfo *MRI;
1075 virtual bool runOnMachineFunction(MachineFunction &Fn);
1077 virtual const char *getPassName() const {
1078 return "ARM pre- register allocation load / store optimization pass";
1082 bool CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1, DebugLoc &dl,
1083 unsigned &NewOpc, unsigned &EvenReg,
1084 unsigned &OddReg, unsigned &BaseReg,
1085 unsigned &OffReg, unsigned &Offset,
1086 unsigned &PredReg, ARMCC::CondCodes &Pred);
1087 bool RescheduleOps(MachineBasicBlock *MBB,
1088 SmallVector<MachineInstr*, 4> &Ops,
1089 unsigned Base, bool isLd,
1090 DenseMap<MachineInstr*, unsigned> &MI2LocMap);
1091 bool RescheduleLoadStoreInstrs(MachineBasicBlock *MBB);
1093 char ARMPreAllocLoadStoreOpt::ID = 0;
1096 bool ARMPreAllocLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
1097 TD = Fn.getTarget().getTargetData();
1098 TII = Fn.getTarget().getInstrInfo();
1099 TRI = Fn.getTarget().getRegisterInfo();
1100 STI = &Fn.getTarget().getSubtarget<ARMSubtarget>();
1101 MRI = &Fn.getRegInfo();
1103 bool Modified = false;
1104 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
1106 Modified |= RescheduleLoadStoreInstrs(MFI);
1111 static bool IsSafeAndProfitableToMove(bool isLd, unsigned Base,
1112 MachineBasicBlock::iterator I,
1113 MachineBasicBlock::iterator E,
1114 SmallPtrSet<MachineInstr*, 4> &MemOps,
1115 SmallSet<unsigned, 4> &MemRegs,
1116 const TargetRegisterInfo *TRI) {
1117 // Are there stores / loads / calls between them?
1118 // FIXME: This is overly conservative. We should make use of alias information
1120 SmallSet<unsigned, 4> AddedRegPressure;
1122 if (MemOps.count(&*I))
1124 const TargetInstrDesc &TID = I->getDesc();
1125 if (TID.isCall() || TID.isTerminator() || TID.hasUnmodeledSideEffects())
1127 if (isLd && TID.mayStore())
1132 // It's not safe to move the first 'str' down.
1135 // str r4, [r0, #+4]
1139 for (unsigned j = 0, NumOps = I->getNumOperands(); j != NumOps; ++j) {
1140 MachineOperand &MO = I->getOperand(j);
1143 unsigned Reg = MO.getReg();
1144 if (MO.isDef() && TRI->regsOverlap(Reg, Base))
1146 if (Reg != Base && !MemRegs.count(Reg))
1147 AddedRegPressure.insert(Reg);
1151 // Estimate register pressure increase due to the transformation.
1152 if (MemRegs.size() <= 4)
1153 // Ok if we are moving small number of instructions.
1155 return AddedRegPressure.size() <= MemRegs.size() * 2;
1159 ARMPreAllocLoadStoreOpt::CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1,
1161 unsigned &NewOpc, unsigned &EvenReg,
1162 unsigned &OddReg, unsigned &BaseReg,
1163 unsigned &OffReg, unsigned &Offset,
1165 ARMCC::CondCodes &Pred) {
1166 // FIXME: FLDS / FSTS -> FLDD / FSTD
1167 unsigned Opcode = Op0->getOpcode();
1168 if (Opcode == ARM::LDR)
1170 else if (Opcode == ARM::STR)
1175 // Must sure the base address satisfies i64 ld / st alignment requirement.
1176 if (!Op0->hasOneMemOperand() ||
1177 !Op0->memoperands_begin()->getValue() ||
1178 Op0->memoperands_begin()->isVolatile())
1181 unsigned Align = Op0->memoperands_begin()->getAlignment();
1182 unsigned ReqAlign = STI->hasV6Ops()
1183 ? TD->getPrefTypeAlignment(
1184 Type::getInt64Ty(Op0->getParent()->getParent()->getFunction()->getContext()))
1185 : 8; // Pre-v6 need 8-byte align
1186 if (Align < ReqAlign)
1189 // Then make sure the immediate offset fits.
1190 int OffImm = getMemoryOpOffset(Op0);
1191 ARM_AM::AddrOpc AddSub = ARM_AM::add;
1193 AddSub = ARM_AM::sub;
1196 if (OffImm >= 256) // 8 bits
1198 Offset = ARM_AM::getAM3Opc(AddSub, OffImm);
1200 EvenReg = Op0->getOperand(0).getReg();
1201 OddReg = Op1->getOperand(0).getReg();
1202 if (EvenReg == OddReg)
1204 BaseReg = Op0->getOperand(1).getReg();
1205 OffReg = Op0->getOperand(2).getReg();
1206 Pred = llvm::getInstrPredicate(Op0, PredReg);
1207 dl = Op0->getDebugLoc();
1211 bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
1212 SmallVector<MachineInstr*, 4> &Ops,
1213 unsigned Base, bool isLd,
1214 DenseMap<MachineInstr*, unsigned> &MI2LocMap) {
1215 bool RetVal = false;
1217 // Sort by offset (in reverse order).
1218 std::sort(Ops.begin(), Ops.end(), OffsetCompare());
1220 // The loads / stores of the same base are in order. Scan them from first to
1221 // last and check for the followins:
1222 // 1. Any def of base.
1224 while (Ops.size() > 1) {
1225 unsigned FirstLoc = ~0U;
1226 unsigned LastLoc = 0;
1227 MachineInstr *FirstOp = 0;
1228 MachineInstr *LastOp = 0;
1230 unsigned LastOpcode = 0;
1231 unsigned LastBytes = 0;
1232 unsigned NumMove = 0;
1233 for (int i = Ops.size() - 1; i >= 0; --i) {
1234 MachineInstr *Op = Ops[i];
1235 unsigned Loc = MI2LocMap[Op];
1236 if (Loc <= FirstLoc) {
1240 if (Loc >= LastLoc) {
1245 unsigned Opcode = Op->getOpcode();
1246 if (LastOpcode && Opcode != LastOpcode)
1249 int Offset = getMemoryOpOffset(Op);
1250 unsigned Bytes = getLSMultipleTransferSize(Op);
1252 if (Bytes != LastBytes || Offset != (LastOffset + (int)Bytes))
1255 LastOffset = Offset;
1257 LastOpcode = Opcode;
1258 if (++NumMove == 8) // FIXME: Tune
1265 SmallPtrSet<MachineInstr*, 4> MemOps;
1266 SmallSet<unsigned, 4> MemRegs;
1267 for (int i = NumMove-1; i >= 0; --i) {
1268 MemOps.insert(Ops[i]);
1269 MemRegs.insert(Ops[i]->getOperand(0).getReg());
1272 // Be conservative, if the instructions are too far apart, don't
1273 // move them. We want to limit the increase of register pressure.
1274 bool DoMove = (LastLoc - FirstLoc) <= NumMove*4; // FIXME: Tune this.
1276 DoMove = IsSafeAndProfitableToMove(isLd, Base, FirstOp, LastOp,
1277 MemOps, MemRegs, TRI);
1279 for (unsigned i = 0; i != NumMove; ++i)
1282 // This is the new location for the loads / stores.
1283 MachineBasicBlock::iterator InsertPos = isLd ? FirstOp : LastOp;
1284 while (InsertPos != MBB->end() && MemOps.count(InsertPos))
1287 // If we are moving a pair of loads / stores, see if it makes sense
1288 // to try to allocate a pair of registers that can form register pairs.
1289 MachineInstr *Op0 = Ops.back();
1290 MachineInstr *Op1 = Ops[Ops.size()-2];
1291 unsigned EvenReg = 0, OddReg = 0;
1292 unsigned BaseReg = 0, OffReg = 0, PredReg = 0;
1293 ARMCC::CondCodes Pred = ARMCC::AL;
1294 unsigned NewOpc = 0;
1295 unsigned Offset = 0;
1297 if (NumMove == 2 && CanFormLdStDWord(Op0, Op1, dl, NewOpc,
1298 EvenReg, OddReg, BaseReg, OffReg,
1299 Offset, PredReg, Pred)) {
1303 // Form the pair instruction.
1305 BuildMI(*MBB, InsertPos, dl, TII->get(NewOpc))
1306 .addReg(EvenReg, RegState::Define)
1307 .addReg(OddReg, RegState::Define)
1308 .addReg(BaseReg).addReg(0).addImm(Offset)
1309 .addImm(Pred).addReg(PredReg);
1312 BuildMI(*MBB, InsertPos, dl, TII->get(NewOpc))
1315 .addReg(BaseReg).addReg(0).addImm(Offset)
1316 .addImm(Pred).addReg(PredReg);
1322 // Add register allocation hints to form register pairs.
1323 MRI->setRegAllocationHint(EvenReg, ARMRI::RegPairEven, OddReg);
1324 MRI->setRegAllocationHint(OddReg, ARMRI::RegPairOdd, EvenReg);
1326 for (unsigned i = 0; i != NumMove; ++i) {
1327 MachineInstr *Op = Ops.back();
1329 MBB->splice(InsertPos, MBB, Op);
1333 NumLdStMoved += NumMove;
1343 ARMPreAllocLoadStoreOpt::RescheduleLoadStoreInstrs(MachineBasicBlock *MBB) {
1344 bool RetVal = false;
1346 DenseMap<MachineInstr*, unsigned> MI2LocMap;
1347 DenseMap<unsigned, SmallVector<MachineInstr*, 4> > Base2LdsMap;
1348 DenseMap<unsigned, SmallVector<MachineInstr*, 4> > Base2StsMap;
1349 SmallVector<unsigned, 4> LdBases;
1350 SmallVector<unsigned, 4> StBases;
1353 MachineBasicBlock::iterator MBBI = MBB->begin();
1354 MachineBasicBlock::iterator E = MBB->end();
1356 for (; MBBI != E; ++MBBI) {
1357 MachineInstr *MI = MBBI;
1358 const TargetInstrDesc &TID = MI->getDesc();
1359 if (TID.isCall() || TID.isTerminator()) {
1360 // Stop at barriers.
1365 MI2LocMap[MI] = Loc++;
1366 if (!isMemoryOp(MI))
1368 unsigned PredReg = 0;
1369 if (llvm::getInstrPredicate(MI, PredReg) != ARMCC::AL)
1372 int Opcode = MI->getOpcode();
1373 bool isLd = Opcode == ARM::LDR ||
1374 Opcode == ARM::FLDS || Opcode == ARM::FLDD;
1375 unsigned Base = MI->getOperand(1).getReg();
1376 int Offset = getMemoryOpOffset(MI);
1378 bool StopHere = false;
1380 DenseMap<unsigned, SmallVector<MachineInstr*, 4> >::iterator BI =
1381 Base2LdsMap.find(Base);
1382 if (BI != Base2LdsMap.end()) {
1383 for (unsigned i = 0, e = BI->second.size(); i != e; ++i) {
1384 if (Offset == getMemoryOpOffset(BI->second[i])) {
1390 BI->second.push_back(MI);
1392 SmallVector<MachineInstr*, 4> MIs;
1394 Base2LdsMap[Base] = MIs;
1395 LdBases.push_back(Base);
1398 DenseMap<unsigned, SmallVector<MachineInstr*, 4> >::iterator BI =
1399 Base2StsMap.find(Base);
1400 if (BI != Base2StsMap.end()) {
1401 for (unsigned i = 0, e = BI->second.size(); i != e; ++i) {
1402 if (Offset == getMemoryOpOffset(BI->second[i])) {
1408 BI->second.push_back(MI);
1410 SmallVector<MachineInstr*, 4> MIs;
1412 Base2StsMap[Base] = MIs;
1413 StBases.push_back(Base);
1418 // Found a duplicate (a base+offset combination that's seen earlier).
1425 // Re-schedule loads.
1426 for (unsigned i = 0, e = LdBases.size(); i != e; ++i) {
1427 unsigned Base = LdBases[i];
1428 SmallVector<MachineInstr*, 4> &Lds = Base2LdsMap[Base];
1430 RetVal |= RescheduleOps(MBB, Lds, Base, true, MI2LocMap);
1433 // Re-schedule stores.
1434 for (unsigned i = 0, e = StBases.size(); i != e; ++i) {
1435 unsigned Base = StBases[i];
1436 SmallVector<MachineInstr*, 4> &Sts = Base2StsMap[Base];
1438 RetVal |= RescheduleOps(MBB, Sts, Base, false, MI2LocMap);
1442 Base2LdsMap.clear();
1443 Base2StsMap.clear();
1453 /// createARMLoadStoreOptimizationPass - returns an instance of the load / store
1454 /// optimization pass.
1455 FunctionPass *llvm::createARMLoadStoreOptimizationPass(bool PreAlloc) {
1457 return new ARMPreAllocLoadStoreOpt();
1458 return new ARMLoadStoreOpt();