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 "ARMRegisterInfo.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/MachineBasicBlock.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstr.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/Support/Compiler.h"
28 #include "llvm/Target/MRegisterInfo.h"
29 #include "llvm/Target/TargetInstrInfo.h"
30 #include "llvm/Target/TargetMachine.h"
33 STATISTIC(NumLDMGened , "Number of ldm instructions generated");
34 STATISTIC(NumSTMGened , "Number of stm instructions generated");
35 STATISTIC(NumFLDMGened, "Number of fldm instructions generated");
36 STATISTIC(NumFSTMGened, "Number of fstm instructions generated");
39 struct VISIBILITY_HIDDEN ARMLoadStoreOpt : public MachineFunctionPass {
40 const TargetInstrInfo *TII;
41 const MRegisterInfo *MRI;
44 virtual bool runOnMachineFunction(MachineFunction &Fn);
46 virtual const char *getPassName() const {
47 return "ARM load / store optimization pass";
51 struct MemOpQueueEntry {
54 MachineBasicBlock::iterator MBBI;
56 MemOpQueueEntry(int o, int p, MachineBasicBlock::iterator i)
57 : Offset(o), Position(p), MBBI(i), Merged(false) {};
59 typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
60 typedef MemOpQueue::iterator MemOpQueueIter;
62 void AdvanceRS(MachineBasicBlock *MBB, MemOpQueue &MemOps);
64 SmallVector<MachineBasicBlock::iterator, 4>
65 MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex, unsigned Base,
66 int Opcode, unsigned Size, MemOpQueue &MemOps);
68 bool LoadStoreMultipleOpti(MachineBasicBlock &MBB);
69 bool MergeReturnIntoLDM(MachineBasicBlock &MBB);
73 /// createARMLoadStoreOptimizationPass - returns an instance of the load / store
74 /// optimization pass.
75 FunctionPass *llvm::createARMLoadStoreOptimizationPass() {
76 return new ARMLoadStoreOpt();
79 static int getLoadStoreMultipleOpcode(int Opcode) {
104 /// mergeOps - Create and insert a LDM or STM with Base as base register and
105 /// registers in Regs as the register operands that would be loaded / stored.
106 /// It returns true if the transformation is done.
107 static bool mergeOps(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
108 int Offset, unsigned Base, bool BaseKill, int Opcode,
109 SmallVector<std::pair<unsigned, bool>, 8> &Regs,
111 const TargetInstrInfo *TII) {
112 // Only a single register to load / store. Don't bother.
113 unsigned NumRegs = Regs.size();
117 ARM_AM::AMSubMode Mode = ARM_AM::ia;
118 bool isAM4 = Opcode == ARM::LDR || Opcode == ARM::STR;
119 if (isAM4 && Offset == 4)
121 else if (isAM4 && Offset == -4 * (int)NumRegs + 4)
123 else if (isAM4 && Offset == -4 * (int)NumRegs)
125 else if (Offset != 0) {
126 // If starting offset isn't zero, insert a MI to materialize a new base.
127 // But only do so if it is cost effective, i.e. merging more than two
133 if (Opcode == ARM::LDR)
134 // If it is a load, then just use one of the destination register to
135 // use as the new base.
136 NewBase = Regs[NumRegs-1].first;
138 // Try to find a free register to use as a new base.
139 NewBase = RS ? RS->FindUnusedReg(&ARM::GPRRegClass) : (unsigned)ARM::R12;
143 int BaseOpc = ARM::ADDri;
145 BaseOpc = ARM::SUBri;
148 int ImmedOffset = ARM_AM::getSOImmVal(Offset);
149 if (ImmedOffset == -1)
150 return false; // Probably not worth it then.
152 BuildMI(MBB, MBBI, TII->get(BaseOpc), NewBase)
153 .addReg(Base, false, false, BaseKill).addImm(ImmedOffset);
155 BaseKill = true; // New base is always killed right its use.
158 bool isDPR = Opcode == ARM::FLDD || Opcode == ARM::FSTD;
159 bool isDef = Opcode == ARM::LDR || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
160 Opcode = getLoadStoreMultipleOpcode(Opcode);
161 MachineInstrBuilder MIB = (isAM4)
162 ? BuildMI(MBB, MBBI, TII->get(Opcode)).addReg(Base, false, false, BaseKill)
163 .addImm(ARM_AM::getAM4ModeImm(Mode))
164 : BuildMI(MBB, MBBI, TII->get(Opcode)).addReg(Base, false, false, BaseKill)
165 .addImm(ARM_AM::getAM5Opc(Mode, false, isDPR ? NumRegs<<1 : NumRegs));
166 for (unsigned i = 0; i != NumRegs; ++i)
167 MIB = MIB.addReg(Regs[i].first, isDef, false, Regs[i].second);
172 /// AdvanceRS - Advance register scavenger to just before the earliest memory
173 /// op that is being merged.
174 void ARMLoadStoreOpt::AdvanceRS(MachineBasicBlock *MBB, MemOpQueue &MemOps) {
175 MachineBasicBlock::iterator Loc = MemOps[0].MBBI;
176 unsigned Position = MemOps[0].Position;
177 for (unsigned i = 1, e = MemOps.size(); i != e; ++i) {
178 if (MemOps[i].Position < Position) {
179 Position = MemOps[i].Position;
180 Loc = MemOps[i].MBBI;
184 if (Loc != MBB->begin())
185 RS->forward(prior(Loc));
188 /// MergeLDR_STR - Merge a number of load / store instructions into one or more
189 /// load / store multiple instructions.
190 SmallVector<MachineBasicBlock::iterator, 4>
191 ARMLoadStoreOpt::MergeLDR_STR(MachineBasicBlock &MBB,
192 unsigned SIndex, unsigned Base, int Opcode,
193 unsigned Size, MemOpQueue &MemOps) {
194 if (RS && SIndex == 0)
195 AdvanceRS(&MBB, MemOps);
197 SmallVector<MachineBasicBlock::iterator, 4> Merges;
198 SmallVector<std::pair<unsigned,bool>, 8> Regs;
199 bool isAM4 = Opcode == ARM::LDR || Opcode == ARM::STR;
200 int Offset = MemOps[SIndex].Offset;
201 int SOffset = Offset;
202 unsigned Pos = MemOps[SIndex].Position;
203 MachineBasicBlock::iterator Loc = MemOps[SIndex].MBBI;
204 unsigned PReg = MemOps[SIndex].MBBI->getOperand(0).getReg();
205 unsigned PRegNum = ARMRegisterInfo::getRegisterNumbering(PReg);
206 bool isKill = MemOps[SIndex].MBBI->getOperand(0).isKill();
207 Regs.push_back(std::make_pair(PReg, isKill));
208 for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
209 int NewOffset = MemOps[i].Offset;
210 unsigned Reg = MemOps[i].MBBI->getOperand(0).getReg();
211 unsigned RegNum = ARMRegisterInfo::getRegisterNumbering(Reg);
212 isKill = MemOps[i].MBBI->getOperand(0).isKill();
213 // AM4 - register numbers in ascending order.
214 // AM5 - consecutive register numbers in ascending order.
215 if (NewOffset == Offset + (int)Size &&
216 ((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
218 Regs.push_back(std::make_pair(Reg, isKill));
221 // Can't merge this in. Try merge the earlier ones first.
222 if (mergeOps(MBB, ++Loc, SOffset, Base, false, Opcode, Regs, RS, TII)) {
223 Merges.push_back(prior(Loc));
224 for (unsigned j = SIndex; j < i; ++j) {
225 MBB.erase(MemOps[j].MBBI);
226 MemOps[j].Merged = true;
229 SmallVector<MachineBasicBlock::iterator, 4> Merges2 =
230 MergeLDR_STR(MBB, i, Base, Opcode, Size, MemOps);
231 Merges.append(Merges2.begin(), Merges2.end());
235 if (MemOps[i].Position > Pos) {
236 Pos = MemOps[i].Position;
237 Loc = MemOps[i].MBBI;
241 bool BaseKill = Loc->findRegisterUseOperand(Base, true) != NULL;
242 if (mergeOps(MBB, ++Loc, SOffset, Base, BaseKill, Opcode, Regs, RS, TII)) {
243 Merges.push_back(prior(Loc));
244 for (unsigned i = SIndex, e = MemOps.size(); i != e; ++i) {
245 MBB.erase(MemOps[i].MBBI);
246 MemOps[i].Merged = true;
253 static inline bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
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);
261 static inline bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
263 return (MI && MI->getOpcode() == ARM::ADDri &&
264 MI->getOperand(0).getReg() == Base &&
265 MI->getOperand(1).getReg() == Base &&
266 ARM_AM::getAM2Offset(MI->getOperand(2).getImm()) == Bytes);
269 static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
270 switch (MI->getOpcode()) {
282 return (MI->getNumOperands() - 2) * 4;
287 return ARM_AM::getAM5Offset(MI->getOperand(1).getImm()) * 4;
291 /// mergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
292 /// register into the LDM/STM/FLDM{D|S}/FSTM{D|S} op when possible:
294 /// stmia rn, <ra, rb, rc>
295 /// rn := rn + 4 * 3;
297 /// stmia rn!, <ra, rb, rc>
299 /// rn := rn - 4 * 3;
300 /// ldmia rn, <ra, rb, rc>
302 /// ldmdb rn!, <ra, rb, rc>
303 static bool mergeBaseUpdateLSMultiple(MachineBasicBlock &MBB,
304 MachineBasicBlock::iterator MBBI) {
305 MachineInstr *MI = MBBI;
306 unsigned Base = MI->getOperand(0).getReg();
307 unsigned Bytes = getLSMultipleTransferSize(MI);
308 int Opcode = MI->getOpcode();
309 bool isAM4 = Opcode == ARM::LDM || Opcode == ARM::STM;
312 if (ARM_AM::getAM4WBFlag(MI->getOperand(1).getImm()))
315 // Can't use the updating AM4 sub-mode if the base register is also a dest
316 // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
317 for (unsigned i = 2, e = MI->getNumOperands(); i != e; ++i) {
318 if (MI->getOperand(i).getReg() == Base)
322 ARM_AM::AMSubMode Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
323 if (MBBI != MBB.begin()) {
324 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
325 if (Mode == ARM_AM::ia &&
326 isMatchingDecrement(PrevMBBI, Base, Bytes)) {
327 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::db, true));
330 } else if (Mode == ARM_AM::ib &&
331 isMatchingDecrement(PrevMBBI, Base, Bytes)) {
332 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::da, true));
338 if (MBBI != MBB.end()) {
339 MachineBasicBlock::iterator NextMBBI = next(MBBI);
340 if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
341 isMatchingIncrement(NextMBBI, Base, Bytes)) {
342 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
345 } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
346 isMatchingDecrement(NextMBBI, Base, Bytes)) {
347 MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
353 // FLDM{D|S}, FSTM{D|S} addressing mode 5 ops.
354 if (ARM_AM::getAM5WBFlag(MI->getOperand(1).getImm()))
357 ARM_AM::AMSubMode Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
358 unsigned Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
359 if (MBBI != MBB.begin()) {
360 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
361 if (Mode == ARM_AM::ia &&
362 isMatchingDecrement(PrevMBBI, Base, Bytes)) {
363 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::db, true, Offset));
369 if (MBBI != MBB.end()) {
370 MachineBasicBlock::iterator NextMBBI = next(MBBI);
371 if (Mode == ARM_AM::ia &&
372 isMatchingIncrement(NextMBBI, Base, Bytes)) {
373 MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::ia, true, Offset));
383 static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
385 case ARM::LDR: return ARM::LDR_PRE;
386 case ARM::STR: return ARM::STR_PRE;
387 case ARM::FLDS: return ARM::FLDMS;
388 case ARM::FLDD: return ARM::FLDMD;
389 case ARM::FSTS: return ARM::FSTMS;
390 case ARM::FSTD: return ARM::FSTMD;
396 static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc) {
398 case ARM::LDR: return ARM::LDR_POST;
399 case ARM::STR: return ARM::STR_POST;
400 case ARM::FLDS: return ARM::FLDMS;
401 case ARM::FLDD: return ARM::FLDMD;
402 case ARM::FSTS: return ARM::FSTMS;
403 case ARM::FSTD: return ARM::FSTMD;
409 /// mergeBaseUpdateLoadStore - Fold proceeding/trailing inc/dec of base
410 /// register into the LDR/STR/FLD{D|S}/FST{D|S} op when possible:
411 static bool mergeBaseUpdateLoadStore(MachineBasicBlock &MBB,
412 MachineBasicBlock::iterator MBBI,
413 const TargetInstrInfo *TII) {
414 MachineInstr *MI = MBBI;
415 unsigned Base = MI->getOperand(1).getReg();
416 bool BaseKill = MI->getOperand(1).isKill();
417 unsigned Bytes = getLSMultipleTransferSize(MI);
418 int Opcode = MI->getOpcode();
419 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
420 if ((isAM2 && ARM_AM::getAM2Offset(MI->getOperand(3).getImm()) != 0) ||
421 (!isAM2 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0))
424 bool isLd = Opcode == ARM::LDR || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
425 // Can't do the merge if the destination register is the same as the would-be
426 // writeback register.
427 if (isLd && MI->getOperand(0).getReg() == Base)
430 bool DoMerge = false;
431 ARM_AM::AddrOpc AddSub = ARM_AM::add;
433 if (MBBI != MBB.begin()) {
434 MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
435 if (isMatchingDecrement(PrevMBBI, Base, Bytes)) {
437 AddSub = ARM_AM::sub;
438 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
439 } else if (isAM2 && isMatchingIncrement(PrevMBBI, Base, Bytes)) {
441 NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
447 if (!DoMerge && MBBI != MBB.end()) {
448 MachineBasicBlock::iterator NextMBBI = next(MBBI);
449 if (isAM2 && isMatchingDecrement(NextMBBI, Base, Bytes)) {
451 AddSub = ARM_AM::sub;
452 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
453 } else if (isMatchingIncrement(NextMBBI, Base, Bytes)) {
455 NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
464 bool isDPR = NewOpc == ARM::FLDMD || NewOpc == ARM::FSTMD;
465 unsigned Offset = isAM2 ? ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift)
466 : ARM_AM::getAM5Opc((AddSub == ARM_AM::sub) ? ARM_AM::db : ARM_AM::ia,
467 true, isDPR ? 2 : 1);
470 // LDR_PRE, LDR_POST;
471 BuildMI(MBB, MBBI, TII->get(NewOpc), MI->getOperand(0).getReg())
473 .addReg(Base).addReg(0).addImm(Offset);
475 BuildMI(MBB, MBBI, TII->get(NewOpc)).addReg(Base, false, false, BaseKill)
476 .addImm(Offset).addReg(MI->getOperand(0).getReg(), true);
478 MachineOperand &MO = MI->getOperand(0);
480 // STR_PRE, STR_POST;
481 BuildMI(MBB, MBBI, TII->get(NewOpc), Base)
482 .addReg(MO.getReg(), false, false, MO.isKill())
483 .addReg(Base).addReg(0).addImm(Offset);
485 BuildMI(MBB, MBBI, TII->get(NewOpc)).addReg(Base)
486 .addImm(Offset).addReg(MO.getReg(), false, false, MO.isKill());
493 /// isMemoryOp - Returns true if instruction is a memory operations (that this
494 /// pass is capable of operating on).
495 static bool isMemoryOp(MachineInstr *MI) {
496 int Opcode = MI->getOpcode();
501 return MI->getOperand(1).isRegister() && MI->getOperand(2).getReg() == 0;
504 return MI->getOperand(1).isRegister();
507 return MI->getOperand(1).isRegister();
512 /// LoadStoreMultipleOpti - An optimization pass to turn multiple LDR / STR
513 /// ops of the same base and incrementing offset into LDM / STM ops.
514 bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
515 unsigned NumMerges = 0;
516 unsigned NumMemOps = 0;
518 unsigned CurrBase = 0;
520 unsigned CurrSize = 0;
521 unsigned Position = 0;
523 if (RS) RS->enterBasicBlock(&MBB);
524 MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
526 bool Advance = false;
527 bool TryMerge = false;
528 bool Clobber = false;
530 bool isMemOp = isMemoryOp(MBBI);
532 int Opcode = MBBI->getOpcode();
533 bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
534 unsigned Size = getLSMultipleTransferSize(MBBI);
535 unsigned Base = MBBI->getOperand(1).getReg();
536 unsigned OffIdx = MBBI->getNumOperands()-1;
537 unsigned OffField = MBBI->getOperand(OffIdx).getImm();
539 ? ARM_AM::getAM2Offset(OffField) : ARM_AM::getAM5Offset(OffField) * 4;
541 if (ARM_AM::getAM2Op(OffField) == ARM_AM::sub)
544 if (ARM_AM::getAM5Op(OffField) == ARM_AM::sub)
549 // r5 := ldr [r5, #4]
550 // r6 := ldr [r5, #8]
552 // The second ldr has effectively broken the chain even though it
553 // looks like the later ldr(s) use the same base register. Try to
554 // merge the ldr's so far, including this one. But don't try to
555 // combine the following ldr(s).
556 Clobber = (Opcode == ARM::LDR && Base == MBBI->getOperand(0).getReg());
557 if (CurrBase == 0 && !Clobber) {
558 // Start of a new chain.
562 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
571 if (CurrOpc == Opcode && CurrBase == Base) {
572 // Continue adding to the queue.
573 if (Offset > MemOps.back().Offset) {
574 MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
578 for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
580 if (Offset < I->Offset) {
581 MemOps.insert(I, MemOpQueueEntry(Offset, Position, MBBI));
585 } else if (Offset == I->Offset) {
586 // Collision! This can't be merged!
603 SmallVector<MachineBasicBlock::iterator,4> MBBII =
604 MergeLDR_STR(MBB, 0, CurrBase, CurrOpc, CurrSize, MemOps);
605 // Try folding preceeding/trailing base inc/dec into the generated
607 for (unsigned i = 0, e = MBBII.size(); i < e; ++i)
608 if (mergeBaseUpdateLSMultiple(MBB, MBBII[i]))
610 NumMerges += MBBII.size();
613 // Try folding preceeding/trailing base inc/dec into those load/store
614 // that were not merged to form LDM/STM ops.
615 for (unsigned i = 0; i != NumMemOps; ++i)
616 if (!MemOps[i].Merged)
617 if (mergeBaseUpdateLoadStore(MBB, MemOps[i].MBBI, TII))
627 // If iterator hasn't been advanced and this is not a memory op, skip it.
628 // It can't start a new chain anyway.
629 if (!Advance && !isMemOp && MBBI != E) {
635 return NumMerges > 0;
638 /// MergeReturnIntoLDM - If this is a exit BB, try merging the return op
639 /// (bx lr) into the preceeding stack restore so it directly restore the value
641 /// ldmfd sp!, {r7, lr}
644 /// ldmfd sp!, {r7, pc}
645 bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
646 if (MBB.empty()) return false;
648 MachineBasicBlock::iterator MBBI = prior(MBB.end());
649 if (MBBI->getOpcode() == ARM::BX_RET && MBBI != MBB.begin()) {
650 MachineInstr *PrevMI = prior(MBBI);
651 if (PrevMI->getOpcode() == ARM::LDM) {
652 MachineOperand &MO = PrevMI->getOperand(PrevMI->getNumOperands()-1);
653 if (MO.getReg() == ARM::LR) {
654 PrevMI->setInstrDescriptor(TII->get(ARM::LDM_RET));
664 bool ARMLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
665 const TargetMachine &TM = Fn.getTarget();
666 TII = TM.getInstrInfo();
667 MRI = TM.getRegisterInfo();
668 RS = MRI->requiresRegisterScavenging(Fn) ? new RegScavenger() : NULL;
670 bool Modified = false;
671 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
673 MachineBasicBlock &MBB = *MFI;
674 Modified |= LoadStoreMultipleOpti(MBB);
675 Modified |= MergeReturnIntoLDM(MBB);