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 "ARMMachineFunctionInfo.h"
19 #include "ARMRegisterInfo.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/CodeGen/MachineBasicBlock.h"
22 #include "llvm/CodeGen/MachineFunctionPass.h"
23 #include "llvm/CodeGen/MachineInstr.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/Target/TargetInstrInfo.h"
29 #include "llvm/Target/TargetMachine.h"
30 #include "llvm/Target/TargetRegisterInfo.h"
31 #include "llvm/Support/Compiler.h"
32 #include "llvm/ADT/DenseMap.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/SmallVector.h"
36 #include "llvm/ADT/Statistic.h"
39 STATISTIC(NumLDMGened , "Number of ldm instructions generated");
40 STATISTIC(NumSTMGened , "Number of stm instructions generated");
41 STATISTIC(NumFLDMGened, "Number of fldm instructions generated");
42 STATISTIC(NumFSTMGened, "Number of fstm instructions generated");
43 STATISTIC(NumLdStMoved, "Number of load / store instructions moved");
44 STATISTIC(NumLDRDFormed,"Number of ldrd created before allocation");
45 STATISTIC(NumSTRDFormed,"Number of strd created before allocation");
46 STATISTIC(NumLDRD2LDM, "Number of ldrd instructions turned back into ldm");
47 STATISTIC(NumSTRD2STM, "Number of strd instructions turned back into stm");
48 STATISTIC(NumLDRD2LDR, "Number of ldrd instructions turned back into ldr's");
49 STATISTIC(NumSTRD2STR, "Number of strd instructions turned back into str's");
51 /// ARMAllocLoadStoreOpt - Post- register allocation pass the combine
52 /// load / store instructions to form ldm / stm instructions.
55 struct VISIBILITY_HIDDEN ARMLoadStoreOpt : public MachineFunctionPass {
57 ARMLoadStoreOpt() : MachineFunctionPass(&ID) {}
59 const TargetInstrInfo *TII;
60 const TargetRegisterInfo *TRI;
64 virtual bool runOnMachineFunction(MachineFunction &Fn);
66 virtual const char *getPassName() const {
67 return "ARM load / store optimization pass";
71 struct MemOpQueueEntry {
74 MachineBasicBlock::iterator MBBI;
76 MemOpQueueEntry(int o, int p, MachineBasicBlock::iterator i)
77 : Offset(o), Position(p), MBBI(i), Merged(false) {};
79 typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
80 typedef MemOpQueue::iterator MemOpQueueIter;
82 bool MergeOps(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
83 int Offset, unsigned Base, bool BaseKill, int Opcode,
84 ARMCC::CondCodes Pred, unsigned PredReg, unsigned Scratch,
85 DebugLoc dl, SmallVector<std::pair<unsigned, bool>, 8> &Regs);
86 void MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex, unsigned Base,
87 int Opcode, unsigned Size,
88 ARMCC::CondCodes Pred, unsigned PredReg,
89 unsigned Scratch, MemOpQueue &MemOps,
90 SmallVector<MachineBasicBlock::iterator, 4> &Merges);
92 void AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps);
93 bool FixInvalidRegPairOp(MachineBasicBlock &MBB,
94 MachineBasicBlock::iterator &MBBI);
95 bool LoadStoreMultipleOpti(MachineBasicBlock &MBB);
96 bool MergeReturnIntoLDM(MachineBasicBlock &MBB);
98 char ARMLoadStoreOpt::ID = 0;
101 static int getLoadStoreMultipleOpcode(int Opcode) {
126 /// MergeOps - Create and insert a LDM or STM with Base as base register and
127 /// registers in Regs as the register operands that would be loaded / stored.
128 /// It returns true if the transformation is done.
130 ARMLoadStoreOpt::MergeOps(MachineBasicBlock &MBB,
131 MachineBasicBlock::iterator MBBI,
132 int Offset, unsigned Base, bool BaseKill,
133 int Opcode, ARMCC::CondCodes Pred,
134 unsigned PredReg, unsigned Scratch, DebugLoc dl,
135 SmallVector<std::pair<unsigned, bool>, 8> &Regs) {
136 // Only a single register to load / store. Don't bother.
137 unsigned NumRegs = Regs.size();
141 ARM_AM::AMSubMode Mode = ARM_AM::ia;
142 bool isAM4 = Opcode == ARM::LDR || Opcode == ARM::STR;
143 if (isAM4 && Offset == 4)
145 else if (isAM4 && Offset == -4 * (int)NumRegs + 4)
147 else if (isAM4 && Offset == -4 * (int)NumRegs)
149 else if (Offset != 0) {
150 // If starting offset isn't zero, insert a MI to materialize a new base.
151 // But only do so if it is cost effective, i.e. merging more than two
157 if (Opcode == ARM::LDR)
158 // If it is a load, then just use one of the destination register to
159 // use as the new base.
160 NewBase = Regs[NumRegs-1].first;
162 // Use the scratch register to use as a new base.
167 int BaseOpc = ARM::ADDri;
169 BaseOpc = ARM::SUBri;
172 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
173 if (ImmedOffset == -1)
174 return false; // Probably not worth it then.
176 BuildMI(MBB, MBBI, dl, TII->get(BaseOpc), NewBase)
177 .addReg(Base, getKillRegState(BaseKill)).addImm(ImmedOffset)
178 .addImm(Pred).addReg(PredReg).addReg(0);
180 BaseKill = true; // New base is always killed right its use.
183 bool isDPR = Opcode == ARM::FLDD || Opcode == ARM::FSTD;
184 bool isDef = Opcode == ARM::LDR || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
185 Opcode = getLoadStoreMultipleOpcode(Opcode);
186 MachineInstrBuilder MIB = (isAM4)
187 ? BuildMI(MBB, MBBI, dl, TII->get(Opcode))
188 .addReg(Base, getKillRegState(BaseKill))
189 .addImm(ARM_AM::getAM4ModeImm(Mode)).addImm(Pred).addReg(PredReg)
190 : BuildMI(MBB, MBBI, dl, TII->get(Opcode))
191 .addReg(Base, getKillRegState(BaseKill))
192 .addImm(ARM_AM::getAM5Opc(Mode, false, isDPR ? NumRegs<<1 : NumRegs))
193 .addImm(Pred).addReg(PredReg);
194 for (unsigned i = 0; i != NumRegs; ++i)
195 MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
196 | getKillRegState(Regs[i].second));
201 /// MergeLDR_STR - Merge a number of load / store instructions into one or more
202 /// load / store multiple instructions.
204 ARMLoadStoreOpt::MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex,
205 unsigned Base, int Opcode, unsigned Size,
206 ARMCC::CondCodes Pred, unsigned PredReg,
207 unsigned Scratch, MemOpQueue &MemOps,
208 SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
209 bool isAM4 = Opcode == ARM::LDR || Opcode == ARM::STR;
210 int Offset = MemOps[SIndex].Offset;
211 int SOffset = Offset;
212 unsigned Pos = MemOps[SIndex].Position;
213 MachineBasicBlock::iterator Loc = MemOps[SIndex].MBBI;
214 DebugLoc dl = Loc->getDebugLoc();
215 unsigned PReg = Loc->getOperand(0).getReg();
216 unsigned PRegNum = ARMRegisterInfo::getRegisterNumbering(PReg);
217 bool isKill = Loc->getOperand(0).isKill();
219 SmallVector<std::pair<unsigned,bool>, 8> Regs;
220 Regs.push_back(std::make_pair(PReg, isKill));
221 for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
222 int NewOffset = MemOps[i].Offset;
223 unsigned Reg = MemOps[i].MBBI->getOperand(0).getReg();
224 unsigned RegNum = ARMRegisterInfo::getRegisterNumbering(Reg);
225 isKill = MemOps[i].MBBI->getOperand(0).isKill();
226 // AM4 - register numbers in ascending order.
227 // AM5 - consecutive register numbers in ascending order.
228 if (NewOffset == Offset + (int)Size &&
229 ((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
231 Regs.push_back(std::make_pair(Reg, isKill));
234 // Can't merge this in. Try merge the earlier ones first.
235 if (MergeOps(MBB, ++Loc, SOffset, Base, false, Opcode, Pred, PredReg,
236 Scratch, dl, Regs)) {
237 Merges.push_back(prior(Loc));
238 for (unsigned j = SIndex; j < i; ++j) {
239 MBB.erase(MemOps[j].MBBI);
240 MemOps[j].Merged = true;
243 MergeLDR_STR(MBB, i, Base, Opcode, Size, Pred, PredReg, Scratch,
248 if (MemOps[i].Position > Pos) {
249 Pos = MemOps[i].Position;
250 Loc = MemOps[i].MBBI;
254 bool BaseKill = Loc->findRegisterUseOperandIdx(Base, true) != -1;
255 if (MergeOps(MBB, ++Loc, SOffset, Base, BaseKill, Opcode, Pred, PredReg,
256 Scratch, dl, Regs)) {
257 Merges.push_back(prior(Loc));
258 for (unsigned i = SIndex, e = MemOps.size(); i != e; ++i) {
259 MBB.erase(MemOps[i].MBBI);
260 MemOps[i].Merged = true;
267 /// getInstrPredicate - If instruction is predicated, returns its predicate
268 /// condition, otherwise returns AL. It also returns the condition code
269 /// register by reference.
270 static ARMCC::CondCodes getInstrPredicate(MachineInstr *MI, unsigned &PredReg) {
271 int PIdx = MI->findFirstPredOperandIdx();
277 PredReg = MI->getOperand(PIdx+1).getReg();
278 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
281 static inline bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
282 unsigned Bytes, ARMCC::CondCodes Pred,
284 unsigned MyPredReg = 0;
285 return (MI && MI->getOpcode() == ARM::SUBri &&
286 MI->getOperand(0).getReg() == Base &&
287 MI->getOperand(1).getReg() == Base &&
288 ARM_AM::getAM2Offset(MI->getOperand(2).getImm()) == Bytes &&
289 getInstrPredicate(MI, MyPredReg) == Pred &&
290 MyPredReg == PredReg);
293 static inline bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
294 unsigned Bytes, ARMCC::CondCodes Pred,
296 unsigned MyPredReg = 0;
297 return (MI && MI->getOpcode() == ARM::ADDri &&
298 MI->getOperand(0).getReg() == Base &&
299 MI->getOperand(1).getReg() == Base &&
300 ARM_AM::getAM2Offset(MI->getOperand(2).getImm()) == Bytes &&
301 getInstrPredicate(MI, MyPredReg) == Pred &&
302 MyPredReg == PredReg);
305 static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
306 switch (MI->getOpcode()) {
318 return (MI->getNumOperands() - 4) * 4;
323 return ARM_AM::getAM5Offset(MI->getOperand(1).getImm()) * 4;
327 /// mergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
328 /// register into the LDM/STM/FLDM{D|S}/FSTM{D|S} op when possible:
330 /// stmia rn, <ra, rb, rc>
331 /// rn := rn + 4 * 3;
333 /// stmia rn!, <ra, rb, rc>
335 /// rn := rn - 4 * 3;
336 /// ldmia rn, <ra, rb, rc>
338 /// ldmdb rn!, <ra, rb, rc>
339 static bool mergeBaseUpdateLSMultiple(MachineBasicBlock &MBB,
340 MachineBasicBlock::iterator MBBI,
342 MachineBasicBlock::iterator &I) {
343 MachineInstr *MI = MBBI;
344 unsigned Base = MI->getOperand(0).getReg();
345 unsigned Bytes = getLSMultipleTransferSize(MI);
346 unsigned PredReg = 0;
347 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
348 int Opcode = MI->getOpcode();
349 bool isAM4 = Opcode == ARM::LDM || Opcode == ARM::STM;
352 if (ARM_AM::getAM4WBFlag(MI->getOperand(1).getImm()))
355 // Can't use the updating AM4 sub-mode if the base register is also a dest
356 // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
357 for (unsigned i = 3, e = MI->getNumOperands(); i != e; ++i) {
358 if (MI->getOperand(i).getReg() == Base)
362 ARM_AM::AMSubMode Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
363 if (MBBI != MBB.begin()) {
364 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
365 if (Mode == ARM_AM::ia &&
366 isMatchingDecrement(PrevMBBI, Base, Bytes, Pred, PredReg)) {
367 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::db, true));
370 } else if (Mode == ARM_AM::ib &&
371 isMatchingDecrement(PrevMBBI, Base, Bytes, Pred, PredReg)) {
372 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::da, true));
378 if (MBBI != MBB.end()) {
379 MachineBasicBlock::iterator NextMBBI = next(MBBI);
380 if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
381 isMatchingIncrement(NextMBBI, Base, Bytes, Pred, PredReg)) {
382 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
389 } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
390 isMatchingDecrement(NextMBBI, Base, Bytes, Pred, PredReg)) {
391 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
401 // FLDM{D|S}, FSTM{D|S} addressing mode 5 ops.
402 if (ARM_AM::getAM5WBFlag(MI->getOperand(1).getImm()))
405 ARM_AM::AMSubMode Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
406 unsigned Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
407 if (MBBI != MBB.begin()) {
408 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
409 if (Mode == ARM_AM::ia &&
410 isMatchingDecrement(PrevMBBI, Base, Bytes, Pred, PredReg)) {
411 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::db, true, Offset));
417 if (MBBI != MBB.end()) {
418 MachineBasicBlock::iterator NextMBBI = next(MBBI);
419 if (Mode == ARM_AM::ia &&
420 isMatchingIncrement(NextMBBI, Base, Bytes, Pred, PredReg)) {
421 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::ia, true, Offset));
435 static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
437 case ARM::LDR: return ARM::LDR_PRE;
438 case ARM::STR: return ARM::STR_PRE;
439 case ARM::FLDS: return ARM::FLDMS;
440 case ARM::FLDD: return ARM::FLDMD;
441 case ARM::FSTS: return ARM::FSTMS;
442 case ARM::FSTD: return ARM::FSTMD;
448 static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc) {
450 case ARM::LDR: return ARM::LDR_POST;
451 case ARM::STR: return ARM::STR_POST;
452 case ARM::FLDS: return ARM::FLDMS;
453 case ARM::FLDD: return ARM::FLDMD;
454 case ARM::FSTS: return ARM::FSTMS;
455 case ARM::FSTD: return ARM::FSTMD;
461 /// mergeBaseUpdateLoadStore - Fold proceeding/trailing inc/dec of base
462 /// register into the LDR/STR/FLD{D|S}/FST{D|S} op when possible:
463 static bool mergeBaseUpdateLoadStore(MachineBasicBlock &MBB,
464 MachineBasicBlock::iterator MBBI,
465 const TargetInstrInfo *TII,
467 MachineBasicBlock::iterator &I) {
468 MachineInstr *MI = MBBI;
469 unsigned Base = MI->getOperand(1).getReg();
470 bool BaseKill = MI->getOperand(1).isKill();
471 unsigned Bytes = getLSMultipleTransferSize(MI);
472 int Opcode = MI->getOpcode();
473 DebugLoc dl = MI->getDebugLoc();
474 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
475 if ((isAM2 && ARM_AM::getAM2Offset(MI->getOperand(3).getImm()) != 0) ||
476 (!isAM2 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0))
479 bool isLd = Opcode == ARM::LDR || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
480 // Can't do the merge if the destination register is the same as the would-be
481 // writeback register.
482 if (isLd && MI->getOperand(0).getReg() == Base)
485 unsigned PredReg = 0;
486 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
487 bool DoMerge = false;
488 ARM_AM::AddrOpc AddSub = ARM_AM::add;
490 if (MBBI != MBB.begin()) {
491 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
492 if (isMatchingDecrement(PrevMBBI, Base, Bytes, Pred, PredReg)) {
494 AddSub = ARM_AM::sub;
495 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
496 } else if (isAM2 && isMatchingIncrement(PrevMBBI, Base, Bytes,
499 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
505 if (!DoMerge && MBBI != MBB.end()) {
506 MachineBasicBlock::iterator NextMBBI = next(MBBI);
507 if (isAM2 && isMatchingDecrement(NextMBBI, Base, Bytes, Pred, PredReg)) {
509 AddSub = ARM_AM::sub;
510 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
511 } else if (isMatchingIncrement(NextMBBI, Base, Bytes, Pred, PredReg)) {
513 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
527 bool isDPR = NewOpc == ARM::FLDMD || NewOpc == ARM::FSTMD;
528 unsigned Offset = isAM2 ? ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift)
529 : ARM_AM::getAM5Opc((AddSub == ARM_AM::sub) ? ARM_AM::db : ARM_AM::ia,
530 true, isDPR ? 2 : 1);
533 // LDR_PRE, LDR_POST;
534 BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
535 .addReg(Base, RegState::Define)
536 .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
539 BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
540 .addReg(Base, getKillRegState(BaseKill))
541 .addImm(Offset).addImm(Pred).addReg(PredReg)
542 .addReg(MI->getOperand(0).getReg(), RegState::Define);
544 MachineOperand &MO = MI->getOperand(0);
546 // STR_PRE, STR_POST;
547 BuildMI(MBB, MBBI, dl, TII->get(NewOpc), Base)
548 .addReg(MO.getReg(), getKillRegState(MO.isKill()))
549 .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
552 BuildMI(MBB, MBBI, dl, TII->get(NewOpc)).addReg(Base).addImm(Offset)
553 .addImm(Pred).addReg(PredReg)
554 .addReg(MO.getReg(), getKillRegState(MO.isKill()));
561 /// isMemoryOp - Returns true if instruction is a memory operations (that this
562 /// pass is capable of operating on).
563 static bool isMemoryOp(MachineInstr *MI) {
564 int Opcode = MI->getOpcode();
569 return MI->getOperand(1).isReg() && MI->getOperand(2).getReg() == 0;
572 return MI->getOperand(1).isReg();
575 return MI->getOperand(1).isReg();
580 /// AdvanceRS - Advance register scavenger to just before the earliest memory
581 /// op that is being merged.
582 void ARMLoadStoreOpt::AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps) {
583 MachineBasicBlock::iterator Loc = MemOps[0].MBBI;
584 unsigned Position = MemOps[0].Position;
585 for (unsigned i = 1, e = MemOps.size(); i != e; ++i) {
586 if (MemOps[i].Position < Position) {
587 Position = MemOps[i].Position;
588 Loc = MemOps[i].MBBI;
592 if (Loc != MBB.begin())
593 RS->forward(prior(Loc));
596 static int getMemoryOpOffset(const MachineInstr *MI) {
597 int Opcode = MI->getOpcode();
598 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
599 bool isAM3 = Opcode == ARM::LDRD || Opcode == ARM::STRD;
600 unsigned NumOperands = MI->getDesc().getNumOperands();
601 unsigned OffField = MI->getOperand(NumOperands-3).getImm();
603 ? ARM_AM::getAM2Offset(OffField)
604 : (isAM3 ? ARM_AM::getAM3Offset(OffField)
605 : ARM_AM::getAM5Offset(OffField) * 4);
607 if (ARM_AM::getAM2Op(OffField) == ARM_AM::sub)
610 if (ARM_AM::getAM3Op(OffField) == ARM_AM::sub)
613 if (ARM_AM::getAM5Op(OffField) == ARM_AM::sub)
619 static void InsertLDR_STR(MachineBasicBlock &MBB,
620 MachineBasicBlock::iterator &MBBI,
621 int OffImm, bool isDef,
622 DebugLoc dl, unsigned NewOpc,
623 unsigned Reg, bool RegKill,
624 unsigned BaseReg, bool BaseKill,
625 unsigned OffReg, bool OffKill,
626 ARMCC::CondCodes Pred, unsigned PredReg,
627 const TargetInstrInfo *TII) {
630 Offset = ARM_AM::getAM2Opc(ARM_AM::sub, -OffImm, ARM_AM::no_shift);
632 Offset = ARM_AM::getAM2Opc(ARM_AM::add, OffImm, ARM_AM::no_shift);
634 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc), Reg)
635 .addReg(BaseReg, getKillRegState(BaseKill))
636 .addReg(OffReg, getKillRegState(OffKill))
638 .addImm(Pred).addReg(PredReg);
640 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
641 .addReg(Reg, getKillRegState(RegKill))
642 .addReg(BaseReg, getKillRegState(BaseKill))
643 .addReg(OffReg, getKillRegState(OffKill))
645 .addImm(Pred).addReg(PredReg);
648 bool ARMLoadStoreOpt::FixInvalidRegPairOp(MachineBasicBlock &MBB,
649 MachineBasicBlock::iterator &MBBI) {
650 MachineInstr *MI = &*MBBI;
651 unsigned Opcode = MI->getOpcode();
652 if (Opcode == ARM::LDRD || Opcode == ARM::STRD) {
653 unsigned EvenReg = MI->getOperand(0).getReg();
654 unsigned OddReg = MI->getOperand(1).getReg();
655 unsigned EvenRegNum = TRI->getDwarfRegNum(EvenReg, false);
656 unsigned OddRegNum = TRI->getDwarfRegNum(OddReg, false);
657 if ((EvenRegNum & 1) == 0 && (EvenRegNum + 1) == OddRegNum)
660 bool isLd = Opcode == ARM::LDRD;
661 bool EvenKill = isLd ? false : MI->getOperand(0).isKill();
662 bool OddKill = isLd ? false : MI->getOperand(1).isKill();
663 const MachineOperand &BaseOp = MI->getOperand(2);
664 unsigned BaseReg = BaseOp.getReg();
665 bool BaseKill = BaseOp.isKill();
666 const MachineOperand &OffOp = MI->getOperand(3);
667 unsigned OffReg = OffOp.getReg();
668 bool OffKill = OffOp.isKill();
669 int OffImm = getMemoryOpOffset(MI);
670 unsigned PredReg = 0;
671 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
673 if (OddRegNum > EvenRegNum && OffReg == 0 && OffImm == 0) {
674 // Ascending register numbers and no offset. It's safe to change it to a
676 unsigned NewOpc = (Opcode == ARM::LDRD) ? ARM::LDM : ARM::STM;
678 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
679 .addReg(BaseReg, getKillRegState(BaseKill))
680 .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
681 .addImm(Pred).addReg(PredReg)
682 .addReg(EvenReg, getDefRegState(isLd))
683 .addReg(OddReg, getDefRegState(isLd));
686 BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
687 .addReg(BaseReg, getKillRegState(BaseKill))
688 .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
689 .addImm(Pred).addReg(PredReg)
690 .addReg(EvenReg, getKillRegState(EvenKill))
691 .addReg(OddReg, getKillRegState(OddKill));
695 // Split into two instructions.
696 unsigned NewOpc = (Opcode == ARM::LDRD) ? ARM::LDR : ARM::STR;
697 DebugLoc dl = MBBI->getDebugLoc();
698 // If this is a load and base register is killed, it may have been
699 // re-defed by the load, make sure the first load does not clobber it.
701 (BaseKill || OffKill) &&
702 (TRI->regsOverlap(EvenReg, BaseReg) ||
703 (OffReg && TRI->regsOverlap(EvenReg, OffReg)))) {
704 assert(!TRI->regsOverlap(OddReg, BaseReg) &&
705 (!OffReg || !TRI->regsOverlap(OddReg, OffReg)));
706 InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc, OddReg, OddKill,
707 BaseReg, false, OffReg, false, Pred, PredReg, TII);
708 InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc, EvenReg, EvenKill,
709 BaseReg, BaseKill, OffReg, OffKill, Pred, PredReg, TII);
711 InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc, EvenReg, EvenKill,
712 BaseReg, false, OffReg, false, Pred, PredReg, TII);
713 InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc, OddReg, OddKill,
714 BaseReg, BaseKill, OffReg, OffKill, Pred, PredReg, TII);
728 /// LoadStoreMultipleOpti - An optimization pass to turn multiple LDR / STR
729 /// ops of the same base and incrementing offset into LDM / STM ops.
730 bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
731 unsigned NumMerges = 0;
732 unsigned NumMemOps = 0;
734 unsigned CurrBase = 0;
736 unsigned CurrSize = 0;
737 ARMCC::CondCodes CurrPred = ARMCC::AL;
738 unsigned CurrPredReg = 0;
739 unsigned Position = 0;
740 SmallVector<MachineBasicBlock::iterator,4> Merges;
742 RS->enterBasicBlock(&MBB);
743 MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
745 if (FixInvalidRegPairOp(MBB, MBBI))
748 bool Advance = false;
749 bool TryMerge = false;
750 bool Clobber = false;
752 bool isMemOp = isMemoryOp(MBBI);
754 int Opcode = MBBI->getOpcode();
755 unsigned Size = getLSMultipleTransferSize(MBBI);
756 unsigned Base = MBBI->getOperand(1).getReg();
757 unsigned PredReg = 0;
758 ARMCC::CondCodes Pred = getInstrPredicate(MBBI, PredReg);
759 int Offset = getMemoryOpOffset(MBBI);
762 // r5 := ldr [r5, #4]
763 // r6 := ldr [r5, #8]
765 // The second ldr has effectively broken the chain even though it
766 // looks like the later ldr(s) use the same base register. Try to
767 // merge the ldr's so far, including this one. But don't try to
768 // combine the following ldr(s).
769 Clobber = (Opcode == ARM::LDR && Base == MBBI->getOperand(0).getReg());
770 if (CurrBase == 0 && !Clobber) {
771 // Start of a new chain.
776 CurrPredReg = PredReg;
777 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
786 if (CurrOpc == Opcode && CurrBase == Base && CurrPred == Pred) {
787 // No need to match PredReg.
788 // Continue adding to the queue.
789 if (Offset > MemOps.back().Offset) {
790 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
794 for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
796 if (Offset < I->Offset) {
797 MemOps.insert(I, MemOpQueueEntry(Offset, Position, MBBI));
801 } else if (Offset == I->Offset) {
802 // Collision! This can't be merged!
819 // Try to find a free register to use as a new base in case it's needed.
820 // First advance to the instruction just before the start of the chain.
821 AdvanceRS(MBB, MemOps);
822 // Find a scratch register. Make sure it's a call clobbered register or
823 // a spilled callee-saved register.
824 unsigned Scratch = RS->FindUnusedReg(&ARM::GPRRegClass, true);
826 Scratch = RS->FindUnusedReg(&ARM::GPRRegClass,
827 AFI->getSpilledCSRegisters());
828 // Process the load / store instructions.
829 RS->forward(prior(MBBI));
833 MergeLDR_STR(MBB, 0, CurrBase, CurrOpc, CurrSize,
834 CurrPred, CurrPredReg, Scratch, MemOps, Merges);
836 // Try folding preceeding/trailing base inc/dec into the generated
838 for (unsigned i = 0, e = Merges.size(); i < e; ++i)
839 if (mergeBaseUpdateLSMultiple(MBB, Merges[i], Advance, MBBI))
841 NumMerges += Merges.size();
843 // Try folding preceeding/trailing base inc/dec into those load/store
844 // that were not merged to form LDM/STM ops.
845 for (unsigned i = 0; i != NumMemOps; ++i)
846 if (!MemOps[i].Merged)
847 if (mergeBaseUpdateLoadStore(MBB, MemOps[i].MBBI, TII,Advance,MBBI))
850 // RS may be pointing to an instruction that's deleted.
851 RS->skipTo(prior(MBBI));
852 } else if (NumMemOps == 1) {
853 // Try folding preceeding/trailing base inc/dec into the single
855 if (mergeBaseUpdateLoadStore(MBB, MemOps[0].MBBI, TII, Advance, MBBI)) {
857 RS->forward(prior(MBBI));
864 CurrPred = ARMCC::AL;
871 // If iterator hasn't been advanced and this is not a memory op, skip it.
872 // It can't start a new chain anyway.
873 if (!Advance && !isMemOp && MBBI != E) {
879 return NumMerges > 0;
883 struct OffsetCompare {
884 bool operator()(const MachineInstr *LHS, const MachineInstr *RHS) const {
885 int LOffset = getMemoryOpOffset(LHS);
886 int ROffset = getMemoryOpOffset(RHS);
887 assert(LHS == RHS || LOffset != ROffset);
888 return LOffset > ROffset;
893 /// MergeReturnIntoLDM - If this is a exit BB, try merging the return op
894 /// (bx lr) into the preceeding stack restore so it directly restore the value
896 /// ldmfd sp!, {r7, lr}
899 /// ldmfd sp!, {r7, pc}
900 bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
901 if (MBB.empty()) return false;
903 MachineBasicBlock::iterator MBBI = prior(MBB.end());
904 if (MBBI->getOpcode() == ARM::BX_RET && MBBI != MBB.begin()) {
905 MachineInstr *PrevMI = prior(MBBI);
906 if (PrevMI->getOpcode() == ARM::LDM) {
907 MachineOperand &MO = PrevMI->getOperand(PrevMI->getNumOperands()-1);
908 if (MO.getReg() == ARM::LR) {
909 PrevMI->setDesc(TII->get(ARM::LDM_RET));
919 bool ARMLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
920 const TargetMachine &TM = Fn.getTarget();
921 AFI = Fn.getInfo<ARMFunctionInfo>();
922 TII = TM.getInstrInfo();
923 TRI = TM.getRegisterInfo();
924 RS = new RegScavenger();
926 bool Modified = false;
927 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
929 MachineBasicBlock &MBB = *MFI;
930 Modified |= LoadStoreMultipleOpti(MBB);
931 Modified |= MergeReturnIntoLDM(MBB);
939 /// ARMPreAllocLoadStoreOpt - Pre- register allocation pass that move
940 /// load / stores from consecutive locations close to make it more
941 /// likely they will be combined later.
944 struct VISIBILITY_HIDDEN ARMPreAllocLoadStoreOpt : public MachineFunctionPass{
946 ARMPreAllocLoadStoreOpt() : MachineFunctionPass(&ID) {}
948 const TargetData *TD;
949 const TargetInstrInfo *TII;
950 const TargetRegisterInfo *TRI;
951 const ARMSubtarget *STI;
952 MachineRegisterInfo *MRI;
954 virtual bool runOnMachineFunction(MachineFunction &Fn);
956 virtual const char *getPassName() const {
957 return "ARM pre- register allocation load / store optimization pass";
961 bool CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1, DebugLoc &dl,
962 unsigned &NewOpc, unsigned &EvenReg,
963 unsigned &OddReg, unsigned &BaseReg,
964 unsigned &OffReg, unsigned &Offset,
965 unsigned &PredReg, ARMCC::CondCodes &Pred);
966 bool RescheduleOps(MachineBasicBlock *MBB,
967 SmallVector<MachineInstr*, 4> &Ops,
968 unsigned Base, bool isLd,
969 DenseMap<MachineInstr*, unsigned> &MI2LocMap);
970 bool RescheduleLoadStoreInstrs(MachineBasicBlock *MBB);
972 char ARMPreAllocLoadStoreOpt::ID = 0;
975 bool ARMPreAllocLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
976 TD = Fn.getTarget().getTargetData();
977 TII = Fn.getTarget().getInstrInfo();
978 TRI = Fn.getTarget().getRegisterInfo();
979 STI = &Fn.getTarget().getSubtarget<ARMSubtarget>();
980 MRI = &Fn.getRegInfo();
982 bool Modified = false;
983 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
985 Modified |= RescheduleLoadStoreInstrs(MFI);
990 static bool IsSafeToMove(bool isLd, unsigned Base,
991 MachineBasicBlock::iterator I,
992 MachineBasicBlock::iterator E,
993 SmallPtrSet<MachineInstr*, 4> MoveOps,
994 const TargetRegisterInfo *TRI) {
995 // Are there stores / loads / calls between them?
996 // FIXME: This is overly conservative. We should make use of alias information
999 const TargetInstrDesc &TID = I->getDesc();
1000 if (TID.isCall() || TID.isTerminator() || TID.hasUnmodeledSideEffects())
1002 if (isLd && TID.mayStore())
1007 // It's not safe to move the first 'str' down.
1010 // str r4, [r0, #+4]
1011 if (TID.mayStore() && !MoveOps.count(&*I))
1014 for (unsigned j = 0, NumOps = I->getNumOperands(); j != NumOps; ++j) {
1015 MachineOperand &MO = I->getOperand(j);
1016 if (MO.isReg() && MO.isDef() && TRI->regsOverlap(MO.getReg(), Base))
1024 ARMPreAllocLoadStoreOpt::CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1,
1026 unsigned &NewOpc, unsigned &EvenReg,
1027 unsigned &OddReg, unsigned &BaseReg,
1028 unsigned &OffReg, unsigned &Offset,
1030 ARMCC::CondCodes &Pred) {
1031 // FIXME: FLDS / FSTS -> FLDD / FSTD
1032 unsigned Opcode = Op0->getOpcode();
1033 if (Opcode == ARM::LDR)
1035 else if (Opcode == ARM::STR)
1040 // Must sure the base address satisfies i64 ld / st alignment requirement.
1041 if (!Op0->hasOneMemOperand() ||
1042 !Op0->memoperands_begin()->getValue() ||
1043 Op0->memoperands_begin()->isVolatile())
1046 unsigned Align = Op0->memoperands_begin()->getAlignment();
1047 unsigned ReqAlign = STI->hasV6Ops()
1048 ? TD->getPrefTypeAlignment(Type::Int64Ty) : 8; // Pre-v6 need 8-byte align
1049 if (Align < ReqAlign)
1052 // Then make sure the immediate offset fits.
1053 int OffImm = getMemoryOpOffset(Op0);
1054 ARM_AM::AddrOpc AddSub = ARM_AM::add;
1056 AddSub = ARM_AM::sub;
1059 if (OffImm >= 256) // 8 bits
1061 Offset = ARM_AM::getAM3Opc(AddSub, OffImm);
1063 EvenReg = Op0->getOperand(0).getReg();
1064 OddReg = Op1->getOperand(0).getReg();
1065 if (EvenReg == OddReg)
1067 BaseReg = Op0->getOperand(1).getReg();
1068 OffReg = Op0->getOperand(2).getReg();
1069 Pred = getInstrPredicate(Op0, PredReg);
1070 dl = Op0->getDebugLoc();
1074 bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
1075 SmallVector<MachineInstr*, 4> &Ops,
1076 unsigned Base, bool isLd,
1077 DenseMap<MachineInstr*, unsigned> &MI2LocMap) {
1078 bool RetVal = false;
1080 // Sort by offset (in reverse order).
1081 std::sort(Ops.begin(), Ops.end(), OffsetCompare());
1083 // The loads / stores of the same base are in order. Scan them from first to
1084 // last and check for the followins:
1085 // 1. Any def of base.
1087 while (Ops.size() > 1) {
1088 unsigned FirstLoc = ~0U;
1089 unsigned LastLoc = 0;
1090 MachineInstr *FirstOp = 0;
1091 MachineInstr *LastOp = 0;
1093 unsigned LastOpcode = 0;
1094 unsigned LastBytes = 0;
1095 unsigned NumMove = 0;
1096 for (int i = Ops.size() - 1; i >= 0; --i) {
1097 MachineInstr *Op = Ops[i];
1098 unsigned Loc = MI2LocMap[Op];
1099 if (Loc <= FirstLoc) {
1103 if (Loc >= LastLoc) {
1108 unsigned Opcode = Op->getOpcode();
1109 if (LastOpcode && Opcode != LastOpcode)
1112 int Offset = getMemoryOpOffset(Op);
1113 unsigned Bytes = getLSMultipleTransferSize(Op);
1115 if (Bytes != LastBytes || Offset != (LastOffset + (int)Bytes))
1118 LastOffset = Offset;
1120 LastOpcode = Opcode;
1128 SmallPtrSet<MachineInstr*, 4> MoveOps;
1129 for (int i = NumMove-1; i >= 0; --i)
1130 MoveOps.insert(Ops[i]);
1132 // Be conservative, if the instructions are too far apart, don't
1133 // move them. We want to limit the increase of register pressure.
1134 bool DoMove = (LastLoc - FirstLoc) < NumMove*4;
1136 DoMove = IsSafeToMove(isLd, Base, FirstOp, LastOp, MoveOps, TRI);
1138 for (unsigned i = 0; i != NumMove; ++i)
1141 // This is the new location for the loads / stores.
1142 MachineBasicBlock::iterator InsertPos = isLd ? FirstOp : LastOp;
1143 while (InsertPos != MBB->end() && MoveOps.count(InsertPos))
1146 // If we are moving a pair of loads / stores, see if it makes sense
1147 // to try to allocate a pair of registers that can form register pairs.
1148 MachineInstr *Op0 = Ops.back();
1149 MachineInstr *Op1 = Ops[Ops.size()-2];
1150 unsigned EvenReg = 0, OddReg = 0;
1151 unsigned BaseReg = 0, OffReg = 0, PredReg = 0;
1152 ARMCC::CondCodes Pred = ARMCC::AL;
1153 unsigned NewOpc = 0;
1154 unsigned Offset = 0;
1156 if (NumMove == 2 && CanFormLdStDWord(Op0, Op1, dl, NewOpc,
1157 EvenReg, OddReg, BaseReg, OffReg,
1158 Offset, PredReg, Pred)) {
1162 // Form the pair instruction.
1164 BuildMI(*MBB, InsertPos, dl, TII->get(NewOpc))
1165 .addReg(EvenReg, RegState::Define)
1166 .addReg(OddReg, RegState::Define)
1167 .addReg(BaseReg).addReg(0).addImm(Offset)
1168 .addImm(Pred).addReg(PredReg);
1171 BuildMI(*MBB, InsertPos, dl, TII->get(NewOpc))
1174 .addReg(BaseReg).addReg(0).addImm(Offset)
1175 .addImm(Pred).addReg(PredReg);
1181 // Add register allocation hints to form register pairs.
1182 MRI->setRegAllocationHint(EvenReg, ARMRI::RegPairEven, OddReg);
1183 MRI->setRegAllocationHint(OddReg, ARMRI::RegPairOdd, EvenReg);
1185 for (unsigned i = 0; i != NumMove; ++i) {
1186 MachineInstr *Op = Ops.back();
1188 MBB->splice(InsertPos, MBB, Op);
1192 NumLdStMoved += NumMove;
1202 ARMPreAllocLoadStoreOpt::RescheduleLoadStoreInstrs(MachineBasicBlock *MBB) {
1203 bool RetVal = false;
1205 DenseMap<MachineInstr*, unsigned> MI2LocMap;
1206 DenseMap<unsigned, SmallVector<MachineInstr*, 4> > Base2LdsMap;
1207 DenseMap<unsigned, SmallVector<MachineInstr*, 4> > Base2StsMap;
1208 SmallVector<unsigned, 4> LdBases;
1209 SmallVector<unsigned, 4> StBases;
1212 MachineBasicBlock::iterator MBBI = MBB->begin();
1213 MachineBasicBlock::iterator E = MBB->end();
1215 for (; MBBI != E; ++MBBI) {
1216 MachineInstr *MI = MBBI;
1217 const TargetInstrDesc &TID = MI->getDesc();
1218 if (TID.isCall() || TID.isTerminator()) {
1219 // Stop at barriers.
1224 MI2LocMap[MI] = Loc++;
1225 if (!isMemoryOp(MI))
1227 unsigned PredReg = 0;
1228 if (getInstrPredicate(MI, PredReg) != ARMCC::AL)
1231 int Opcode = MI->getOpcode();
1232 bool isLd = Opcode == ARM::LDR ||
1233 Opcode == ARM::FLDS || Opcode == ARM::FLDD;
1234 unsigned Base = MI->getOperand(1).getReg();
1235 int Offset = getMemoryOpOffset(MI);
1237 bool StopHere = false;
1239 DenseMap<unsigned, SmallVector<MachineInstr*, 4> >::iterator BI =
1240 Base2LdsMap.find(Base);
1241 if (BI != Base2LdsMap.end()) {
1242 for (unsigned i = 0, e = BI->second.size(); i != e; ++i) {
1243 if (Offset == getMemoryOpOffset(BI->second[i])) {
1249 BI->second.push_back(MI);
1251 SmallVector<MachineInstr*, 4> MIs;
1253 Base2LdsMap[Base] = MIs;
1254 LdBases.push_back(Base);
1257 DenseMap<unsigned, SmallVector<MachineInstr*, 4> >::iterator BI =
1258 Base2StsMap.find(Base);
1259 if (BI != Base2StsMap.end()) {
1260 for (unsigned i = 0, e = BI->second.size(); i != e; ++i) {
1261 if (Offset == getMemoryOpOffset(BI->second[i])) {
1267 BI->second.push_back(MI);
1269 SmallVector<MachineInstr*, 4> MIs;
1271 Base2StsMap[Base] = MIs;
1272 StBases.push_back(Base);
1277 // Found a duplicate (a base+offset combination that's seen earlier). Backtrack.
1283 // Re-schedule loads.
1284 for (unsigned i = 0, e = LdBases.size(); i != e; ++i) {
1285 unsigned Base = LdBases[i];
1286 SmallVector<MachineInstr*, 4> &Lds = Base2LdsMap[Base];
1288 RetVal |= RescheduleOps(MBB, Lds, Base, true, MI2LocMap);
1291 // Re-schedule stores.
1292 for (unsigned i = 0, e = StBases.size(); i != e; ++i) {
1293 unsigned Base = StBases[i];
1294 SmallVector<MachineInstr*, 4> &Sts = Base2StsMap[Base];
1296 RetVal |= RescheduleOps(MBB, Sts, Base, false, MI2LocMap);
1300 Base2LdsMap.clear();
1301 Base2StsMap.clear();
1311 /// createARMLoadStoreOptimizationPass - returns an instance of the load / store
1312 /// optimization pass.
1313 FunctionPass *llvm::createARMLoadStoreOptimizationPass(bool PreAlloc) {
1315 return new ARMPreAllocLoadStoreOpt();
1316 return new ARMLoadStoreOpt();