1 //===- ARMBaseInstrInfo.cpp - ARM Instruction Information -------*- 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 the Base ARM implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "ARMBaseInstrInfo.h"
16 #include "ARMConstantPoolValue.h"
17 #include "ARMHazardRecognizer.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMRegisterInfo.h"
20 #include "MCTargetDesc/ARMAddressingModes.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalValue.h"
24 #include "llvm/CodeGen/LiveVariables.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineJumpTableInfo.h"
29 #include "llvm/CodeGen/MachineMemOperand.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/PseudoSourceValue.h"
32 #include "llvm/CodeGen/SelectionDAGNodes.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/Support/BranchProbability.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/ADT/STLExtras.h"
40 #define GET_INSTRINFO_CTOR
41 #include "ARMGenInstrInfo.inc"
46 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
47 cl::desc("Enable ARM 2-addr to 3-addr conv"));
49 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
51 unsigned MLxOpc; // MLA / MLS opcode
52 unsigned MulOpc; // Expanded multiplication opcode
53 unsigned AddSubOpc; // Expanded add / sub opcode
54 bool NegAcc; // True if the acc is negated before the add / sub.
55 bool HasLane; // True if instruction has an extra "lane" operand.
58 static const ARM_MLxEntry ARM_MLxTable[] = {
59 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
61 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
62 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
63 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
64 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
65 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
66 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
67 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
68 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
71 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
72 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
73 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
74 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
75 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
76 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
77 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
78 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
81 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
82 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
84 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
85 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
86 assert(false && "Duplicated entries?");
87 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
88 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
92 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
93 // currently defaults to no prepass hazard recognizer.
94 ScheduleHazardRecognizer *ARMBaseInstrInfo::
95 CreateTargetHazardRecognizer(const TargetMachine *TM,
96 const ScheduleDAG *DAG) const {
97 if (usePreRAHazardRecognizer()) {
98 const InstrItineraryData *II = TM->getInstrItineraryData();
99 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
101 return TargetInstrInfoImpl::CreateTargetHazardRecognizer(TM, DAG);
104 ScheduleHazardRecognizer *ARMBaseInstrInfo::
105 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
106 const ScheduleDAG *DAG) const {
107 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
108 return (ScheduleHazardRecognizer *)
109 new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget, DAG);
110 return TargetInstrInfoImpl::CreateTargetPostRAHazardRecognizer(II, DAG);
114 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
115 MachineBasicBlock::iterator &MBBI,
116 LiveVariables *LV) const {
117 // FIXME: Thumb2 support.
122 MachineInstr *MI = MBBI;
123 MachineFunction &MF = *MI->getParent()->getParent();
124 uint64_t TSFlags = MI->getDesc().TSFlags;
126 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
127 default: return NULL;
128 case ARMII::IndexModePre:
131 case ARMII::IndexModePost:
135 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
137 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
141 MachineInstr *UpdateMI = NULL;
142 MachineInstr *MemMI = NULL;
143 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
144 const MCInstrDesc &MCID = MI->getDesc();
145 unsigned NumOps = MCID.getNumOperands();
146 bool isLoad = !MCID.mayStore();
147 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
148 const MachineOperand &Base = MI->getOperand(2);
149 const MachineOperand &Offset = MI->getOperand(NumOps-3);
150 unsigned WBReg = WB.getReg();
151 unsigned BaseReg = Base.getReg();
152 unsigned OffReg = Offset.getReg();
153 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
154 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
157 assert(false && "Unknown indexed op!");
159 case ARMII::AddrMode2: {
160 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
161 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
163 if (ARM_AM::getSOImmVal(Amt) == -1)
164 // Can't encode it in a so_imm operand. This transformation will
165 // add more than 1 instruction. Abandon!
167 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
168 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
169 .addReg(BaseReg).addImm(Amt)
170 .addImm(Pred).addReg(0).addReg(0);
171 } else if (Amt != 0) {
172 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
173 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
174 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
175 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
176 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
177 .addImm(Pred).addReg(0).addReg(0);
179 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
180 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
181 .addReg(BaseReg).addReg(OffReg)
182 .addImm(Pred).addReg(0).addReg(0);
185 case ARMII::AddrMode3 : {
186 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
187 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
189 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
190 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
191 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
192 .addReg(BaseReg).addImm(Amt)
193 .addImm(Pred).addReg(0).addReg(0);
195 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
196 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
197 .addReg(BaseReg).addReg(OffReg)
198 .addImm(Pred).addReg(0).addReg(0);
203 std::vector<MachineInstr*> NewMIs;
206 MemMI = BuildMI(MF, MI->getDebugLoc(),
207 get(MemOpc), MI->getOperand(0).getReg())
208 .addReg(WBReg).addImm(0).addImm(Pred);
210 MemMI = BuildMI(MF, MI->getDebugLoc(),
211 get(MemOpc)).addReg(MI->getOperand(1).getReg())
212 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
213 NewMIs.push_back(MemMI);
214 NewMIs.push_back(UpdateMI);
217 MemMI = BuildMI(MF, MI->getDebugLoc(),
218 get(MemOpc), MI->getOperand(0).getReg())
219 .addReg(BaseReg).addImm(0).addImm(Pred);
221 MemMI = BuildMI(MF, MI->getDebugLoc(),
222 get(MemOpc)).addReg(MI->getOperand(1).getReg())
223 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
225 UpdateMI->getOperand(0).setIsDead();
226 NewMIs.push_back(UpdateMI);
227 NewMIs.push_back(MemMI);
230 // Transfer LiveVariables states, kill / dead info.
232 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
233 MachineOperand &MO = MI->getOperand(i);
234 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
235 unsigned Reg = MO.getReg();
237 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
239 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
241 LV->addVirtualRegisterDead(Reg, NewMI);
243 if (MO.isUse() && MO.isKill()) {
244 for (unsigned j = 0; j < 2; ++j) {
245 // Look at the two new MI's in reverse order.
246 MachineInstr *NewMI = NewMIs[j];
247 if (!NewMI->readsRegister(Reg))
249 LV->addVirtualRegisterKilled(Reg, NewMI);
250 if (VI.removeKill(MI))
251 VI.Kills.push_back(NewMI);
259 MFI->insert(MBBI, NewMIs[1]);
260 MFI->insert(MBBI, NewMIs[0]);
266 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
267 MachineBasicBlock *&FBB,
268 SmallVectorImpl<MachineOperand> &Cond,
269 bool AllowModify) const {
270 // If the block has no terminators, it just falls into the block after it.
271 MachineBasicBlock::iterator I = MBB.end();
272 if (I == MBB.begin())
275 while (I->isDebugValue()) {
276 if (I == MBB.begin())
280 if (!isUnpredicatedTerminator(I))
283 // Get the last instruction in the block.
284 MachineInstr *LastInst = I;
286 // If there is only one terminator instruction, process it.
287 unsigned LastOpc = LastInst->getOpcode();
288 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
289 if (isUncondBranchOpcode(LastOpc)) {
290 TBB = LastInst->getOperand(0).getMBB();
293 if (isCondBranchOpcode(LastOpc)) {
294 // Block ends with fall-through condbranch.
295 TBB = LastInst->getOperand(0).getMBB();
296 Cond.push_back(LastInst->getOperand(1));
297 Cond.push_back(LastInst->getOperand(2));
300 return true; // Can't handle indirect branch.
303 // Get the instruction before it if it is a terminator.
304 MachineInstr *SecondLastInst = I;
305 unsigned SecondLastOpc = SecondLastInst->getOpcode();
307 // If AllowModify is true and the block ends with two or more unconditional
308 // branches, delete all but the first unconditional branch.
309 if (AllowModify && isUncondBranchOpcode(LastOpc)) {
310 while (isUncondBranchOpcode(SecondLastOpc)) {
311 LastInst->eraseFromParent();
312 LastInst = SecondLastInst;
313 LastOpc = LastInst->getOpcode();
314 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
315 // Return now the only terminator is an unconditional branch.
316 TBB = LastInst->getOperand(0).getMBB();
320 SecondLastOpc = SecondLastInst->getOpcode();
325 // If there are three terminators, we don't know what sort of block this is.
326 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
329 // If the block ends with a B and a Bcc, handle it.
330 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
331 TBB = SecondLastInst->getOperand(0).getMBB();
332 Cond.push_back(SecondLastInst->getOperand(1));
333 Cond.push_back(SecondLastInst->getOperand(2));
334 FBB = LastInst->getOperand(0).getMBB();
338 // If the block ends with two unconditional branches, handle it. The second
339 // one is not executed, so remove it.
340 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
341 TBB = SecondLastInst->getOperand(0).getMBB();
344 I->eraseFromParent();
348 // ...likewise if it ends with a branch table followed by an unconditional
349 // branch. The branch folder can create these, and we must get rid of them for
350 // correctness of Thumb constant islands.
351 if ((isJumpTableBranchOpcode(SecondLastOpc) ||
352 isIndirectBranchOpcode(SecondLastOpc)) &&
353 isUncondBranchOpcode(LastOpc)) {
356 I->eraseFromParent();
360 // Otherwise, can't handle this.
365 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
366 MachineBasicBlock::iterator I = MBB.end();
367 if (I == MBB.begin()) return 0;
369 while (I->isDebugValue()) {
370 if (I == MBB.begin())
374 if (!isUncondBranchOpcode(I->getOpcode()) &&
375 !isCondBranchOpcode(I->getOpcode()))
378 // Remove the branch.
379 I->eraseFromParent();
383 if (I == MBB.begin()) return 1;
385 if (!isCondBranchOpcode(I->getOpcode()))
388 // Remove the branch.
389 I->eraseFromParent();
394 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
395 MachineBasicBlock *FBB,
396 const SmallVectorImpl<MachineOperand> &Cond,
398 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
399 int BOpc = !AFI->isThumbFunction()
400 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
401 int BccOpc = !AFI->isThumbFunction()
402 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
404 // Shouldn't be a fall through.
405 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
406 assert((Cond.size() == 2 || Cond.size() == 0) &&
407 "ARM branch conditions have two components!");
410 if (Cond.empty()) // Unconditional branch?
411 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
413 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
414 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
418 // Two-way conditional branch.
419 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
420 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
421 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
425 bool ARMBaseInstrInfo::
426 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
427 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
428 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
432 bool ARMBaseInstrInfo::
433 PredicateInstruction(MachineInstr *MI,
434 const SmallVectorImpl<MachineOperand> &Pred) const {
435 unsigned Opc = MI->getOpcode();
436 if (isUncondBranchOpcode(Opc)) {
437 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
438 MI->addOperand(MachineOperand::CreateImm(Pred[0].getImm()));
439 MI->addOperand(MachineOperand::CreateReg(Pred[1].getReg(), false));
443 int PIdx = MI->findFirstPredOperandIdx();
445 MachineOperand &PMO = MI->getOperand(PIdx);
446 PMO.setImm(Pred[0].getImm());
447 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
453 bool ARMBaseInstrInfo::
454 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
455 const SmallVectorImpl<MachineOperand> &Pred2) const {
456 if (Pred1.size() > 2 || Pred2.size() > 2)
459 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
460 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
470 return CC2 == ARMCC::HI;
472 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
474 return CC2 == ARMCC::GT;
476 return CC2 == ARMCC::LT;
480 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
481 std::vector<MachineOperand> &Pred) const {
482 // FIXME: This confuses implicit_def with optional CPSR def.
483 const MCInstrDesc &MCID = MI->getDesc();
484 if (!MCID.getImplicitDefs() && !MCID.hasOptionalDef())
488 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
489 const MachineOperand &MO = MI->getOperand(i);
490 if (MO.isReg() && MO.getReg() == ARM::CPSR) {
499 /// isPredicable - Return true if the specified instruction can be predicated.
500 /// By default, this returns true for every instruction with a
501 /// PredicateOperand.
502 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
503 const MCInstrDesc &MCID = MI->getDesc();
504 if (!MCID.isPredicable())
507 if ((MCID.TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
508 ARMFunctionInfo *AFI =
509 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
510 return AFI->isThumb2Function();
515 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
516 LLVM_ATTRIBUTE_NOINLINE
517 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
519 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
521 assert(JTI < JT.size());
522 return JT[JTI].MBBs.size();
525 /// GetInstSize - Return the size of the specified MachineInstr.
527 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
528 const MachineBasicBlock &MBB = *MI->getParent();
529 const MachineFunction *MF = MBB.getParent();
530 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
532 const MCInstrDesc &MCID = MI->getDesc();
534 return MCID.getSize();
536 // If this machine instr is an inline asm, measure it.
537 if (MI->getOpcode() == ARM::INLINEASM)
538 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
541 unsigned Opc = MI->getOpcode();
543 case TargetOpcode::IMPLICIT_DEF:
544 case TargetOpcode::KILL:
545 case TargetOpcode::PROLOG_LABEL:
546 case TargetOpcode::EH_LABEL:
547 case TargetOpcode::DBG_VALUE:
549 case ARM::MOVi16_ga_pcrel:
550 case ARM::MOVTi16_ga_pcrel:
551 case ARM::t2MOVi16_ga_pcrel:
552 case ARM::t2MOVTi16_ga_pcrel:
555 case ARM::t2MOVi32imm:
557 case ARM::CONSTPOOL_ENTRY:
558 // If this machine instr is a constant pool entry, its size is recorded as
560 return MI->getOperand(2).getImm();
561 case ARM::Int_eh_sjlj_longjmp:
563 case ARM::tInt_eh_sjlj_longjmp:
565 case ARM::Int_eh_sjlj_setjmp:
566 case ARM::Int_eh_sjlj_setjmp_nofp:
568 case ARM::tInt_eh_sjlj_setjmp:
569 case ARM::t2Int_eh_sjlj_setjmp:
570 case ARM::t2Int_eh_sjlj_setjmp_nofp:
578 case ARM::t2TBH_JT: {
579 // These are jumptable branches, i.e. a branch followed by an inlined
580 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
581 // entry is one byte; TBH two byte each.
582 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
583 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
584 unsigned NumOps = MCID.getNumOperands();
585 MachineOperand JTOP =
586 MI->getOperand(NumOps - (MCID.isPredicable() ? 3 : 2));
587 unsigned JTI = JTOP.getIndex();
588 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
590 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
591 assert(JTI < JT.size());
592 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
593 // 4 aligned. The assembler / linker may add 2 byte padding just before
594 // the JT entries. The size does not include this padding; the
595 // constant islands pass does separate bookkeeping for it.
596 // FIXME: If we know the size of the function is less than (1 << 16) *2
597 // bytes, we can use 16-bit entries instead. Then there won't be an
599 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
600 unsigned NumEntries = getNumJTEntries(JT, JTI);
601 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
602 // Make sure the instruction that follows TBB is 2-byte aligned.
603 // FIXME: Constant island pass should insert an "ALIGN" instruction
606 return NumEntries * EntrySize + InstSize;
609 // Otherwise, pseudo-instruction sizes are zero.
612 return 0; // Not reached
615 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
616 MachineBasicBlock::iterator I, DebugLoc DL,
617 unsigned DestReg, unsigned SrcReg,
618 bool KillSrc) const {
619 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
620 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
622 if (GPRDest && GPRSrc) {
623 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
624 .addReg(SrcReg, getKillRegState(KillSrc))));
628 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
629 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
632 if (SPRDest && SPRSrc) {
635 // An even S-S copy may be feeding a NEON v2f32 instruction being used for
636 // f32 operations. In that case, it is better to copy the full D-regs with
637 // a VMOVD since that can be converted to a NEON-domain move by
638 // NEONMoveFix.cpp. Check that MI is the original COPY instruction, and
639 // that it really defines the whole D-register.
640 if ((DestReg - ARM::S0) % 2 == 0 && (SrcReg - ARM::S0) % 2 == 0 &&
641 I != MBB.end() && I->isCopy() &&
642 I->getOperand(0).getReg() == DestReg &&
643 I->getOperand(1).getReg() == SrcReg) {
644 // I is pointing to the ortiginal COPY instruction.
645 // Find the parent D-registers.
646 const TargetRegisterInfo *TRI = &getRegisterInfo();
647 unsigned SrcD = TRI->getMatchingSuperReg(SrcReg, ARM::ssub_0,
649 unsigned DestD = TRI->getMatchingSuperReg(DestReg, ARM::ssub_0,
651 // Be careful to not clobber an INSERT_SUBREG that reads and redefines a
652 // D-register. There must be an <imp-def> of destD, and no <imp-use>.
653 if (I->definesRegister(DestD, TRI) && !I->readsRegister(DestD, TRI)) {
658 KillSrc = I->killsRegister(SrcReg, TRI);
661 } else if (GPRDest && SPRSrc)
663 else if (SPRDest && GPRSrc)
665 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
667 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
669 else if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
671 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
674 llvm_unreachable("Impossible reg-to-reg copy");
676 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
677 MIB.addReg(SrcReg, getKillRegState(KillSrc));
678 if (Opc == ARM::VORRq)
679 MIB.addReg(SrcReg, getKillRegState(KillSrc));
680 if (Opc != ARM::VMOVQQ && Opc != ARM::VMOVQQQQ)
685 MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB,
686 unsigned Reg, unsigned SubIdx, unsigned State,
687 const TargetRegisterInfo *TRI) {
689 return MIB.addReg(Reg, State);
691 if (TargetRegisterInfo::isPhysicalRegister(Reg))
692 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
693 return MIB.addReg(Reg, State, SubIdx);
696 void ARMBaseInstrInfo::
697 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
698 unsigned SrcReg, bool isKill, int FI,
699 const TargetRegisterClass *RC,
700 const TargetRegisterInfo *TRI) const {
702 if (I != MBB.end()) DL = I->getDebugLoc();
703 MachineFunction &MF = *MBB.getParent();
704 MachineFrameInfo &MFI = *MF.getFrameInfo();
705 unsigned Align = MFI.getObjectAlignment(FI);
707 MachineMemOperand *MMO =
708 MF.getMachineMemOperand(MachinePointerInfo(
709 PseudoSourceValue::getFixedStack(FI)),
710 MachineMemOperand::MOStore,
711 MFI.getObjectSize(FI),
714 switch (RC->getSize()) {
716 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
717 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
718 .addReg(SrcReg, getKillRegState(isKill))
719 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
720 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
721 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
722 .addReg(SrcReg, getKillRegState(isKill))
723 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
725 llvm_unreachable("Unknown reg class!");
728 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
729 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
730 .addReg(SrcReg, getKillRegState(isKill))
731 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
733 llvm_unreachable("Unknown reg class!");
736 if (ARM::QPRRegClass.hasSubClassEq(RC)) {
737 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
738 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64Pseudo))
739 .addFrameIndex(FI).addImm(16)
740 .addReg(SrcReg, getKillRegState(isKill))
741 .addMemOperand(MMO));
743 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
744 .addReg(SrcReg, getKillRegState(isKill))
746 .addMemOperand(MMO));
749 llvm_unreachable("Unknown reg class!");
752 if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
753 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
754 // FIXME: It's possible to only store part of the QQ register if the
755 // spilled def has a sub-register index.
756 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
757 .addFrameIndex(FI).addImm(16)
758 .addReg(SrcReg, getKillRegState(isKill))
759 .addMemOperand(MMO));
761 MachineInstrBuilder MIB =
762 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
765 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
766 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
767 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
768 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
771 llvm_unreachable("Unknown reg class!");
774 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
775 MachineInstrBuilder MIB =
776 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
779 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
780 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
781 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
782 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
783 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
784 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
785 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
786 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
788 llvm_unreachable("Unknown reg class!");
791 llvm_unreachable("Unknown reg class!");
796 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
797 int &FrameIndex) const {
798 switch (MI->getOpcode()) {
801 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
802 if (MI->getOperand(1).isFI() &&
803 MI->getOperand(2).isReg() &&
804 MI->getOperand(3).isImm() &&
805 MI->getOperand(2).getReg() == 0 &&
806 MI->getOperand(3).getImm() == 0) {
807 FrameIndex = MI->getOperand(1).getIndex();
808 return MI->getOperand(0).getReg();
816 if (MI->getOperand(1).isFI() &&
817 MI->getOperand(2).isImm() &&
818 MI->getOperand(2).getImm() == 0) {
819 FrameIndex = MI->getOperand(1).getIndex();
820 return MI->getOperand(0).getReg();
823 case ARM::VST1q64Pseudo:
824 if (MI->getOperand(0).isFI() &&
825 MI->getOperand(2).getSubReg() == 0) {
826 FrameIndex = MI->getOperand(0).getIndex();
827 return MI->getOperand(2).getReg();
831 if (MI->getOperand(1).isFI() &&
832 MI->getOperand(0).getSubReg() == 0) {
833 FrameIndex = MI->getOperand(1).getIndex();
834 return MI->getOperand(0).getReg();
842 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
843 int &FrameIndex) const {
844 const MachineMemOperand *Dummy;
845 return MI->getDesc().mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
848 void ARMBaseInstrInfo::
849 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
850 unsigned DestReg, int FI,
851 const TargetRegisterClass *RC,
852 const TargetRegisterInfo *TRI) const {
854 if (I != MBB.end()) DL = I->getDebugLoc();
855 MachineFunction &MF = *MBB.getParent();
856 MachineFrameInfo &MFI = *MF.getFrameInfo();
857 unsigned Align = MFI.getObjectAlignment(FI);
858 MachineMemOperand *MMO =
859 MF.getMachineMemOperand(
860 MachinePointerInfo(PseudoSourceValue::getFixedStack(FI)),
861 MachineMemOperand::MOLoad,
862 MFI.getObjectSize(FI),
865 switch (RC->getSize()) {
867 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
868 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
869 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
871 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
872 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
873 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
875 llvm_unreachable("Unknown reg class!");
878 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
879 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
880 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
882 llvm_unreachable("Unknown reg class!");
885 if (ARM::QPRRegClass.hasSubClassEq(RC)) {
886 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
887 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64Pseudo), DestReg)
888 .addFrameIndex(FI).addImm(16)
889 .addMemOperand(MMO));
891 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
893 .addMemOperand(MMO));
896 llvm_unreachable("Unknown reg class!");
899 if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
900 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
901 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
902 .addFrameIndex(FI).addImm(16)
903 .addMemOperand(MMO));
905 MachineInstrBuilder MIB =
906 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
909 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
910 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
911 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
912 AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
915 llvm_unreachable("Unknown reg class!");
918 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
919 MachineInstrBuilder MIB =
920 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
923 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
924 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
925 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
926 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
927 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI);
928 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI);
929 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI);
930 AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
932 llvm_unreachable("Unknown reg class!");
935 llvm_unreachable("Unknown regclass!");
940 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
941 int &FrameIndex) const {
942 switch (MI->getOpcode()) {
945 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
946 if (MI->getOperand(1).isFI() &&
947 MI->getOperand(2).isReg() &&
948 MI->getOperand(3).isImm() &&
949 MI->getOperand(2).getReg() == 0 &&
950 MI->getOperand(3).getImm() == 0) {
951 FrameIndex = MI->getOperand(1).getIndex();
952 return MI->getOperand(0).getReg();
960 if (MI->getOperand(1).isFI() &&
961 MI->getOperand(2).isImm() &&
962 MI->getOperand(2).getImm() == 0) {
963 FrameIndex = MI->getOperand(1).getIndex();
964 return MI->getOperand(0).getReg();
967 case ARM::VLD1q64Pseudo:
968 if (MI->getOperand(1).isFI() &&
969 MI->getOperand(0).getSubReg() == 0) {
970 FrameIndex = MI->getOperand(1).getIndex();
971 return MI->getOperand(0).getReg();
975 if (MI->getOperand(1).isFI() &&
976 MI->getOperand(0).getSubReg() == 0) {
977 FrameIndex = MI->getOperand(1).getIndex();
978 return MI->getOperand(0).getReg();
986 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
987 int &FrameIndex) const {
988 const MachineMemOperand *Dummy;
989 return MI->getDesc().mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
993 ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
994 int FrameIx, uint64_t Offset,
997 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
998 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
1002 /// Create a copy of a const pool value. Update CPI to the new index and return
1004 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1005 MachineConstantPool *MCP = MF.getConstantPool();
1006 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1008 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1009 assert(MCPE.isMachineConstantPoolEntry() &&
1010 "Expecting a machine constantpool entry!");
1011 ARMConstantPoolValue *ACPV =
1012 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1014 unsigned PCLabelId = AFI->createPICLabelUId();
1015 ARMConstantPoolValue *NewCPV = 0;
1016 // FIXME: The below assumes PIC relocation model and that the function
1017 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1018 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1019 // instructions, so that's probably OK, but is PIC always correct when
1021 if (ACPV->isGlobalValue())
1022 NewCPV = new ARMConstantPoolValue(ACPV->getGV(), PCLabelId,
1024 else if (ACPV->isExtSymbol())
1025 NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(),
1026 ACPV->getSymbol(), PCLabelId, 4);
1027 else if (ACPV->isBlockAddress())
1028 NewCPV = new ARMConstantPoolValue(ACPV->getBlockAddress(), PCLabelId,
1029 ARMCP::CPBlockAddress, 4);
1030 else if (ACPV->isLSDA())
1031 NewCPV = new ARMConstantPoolValue(MF.getFunction(), PCLabelId,
1034 llvm_unreachable("Unexpected ARM constantpool value type!!");
1035 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1039 void ARMBaseInstrInfo::
1040 reMaterialize(MachineBasicBlock &MBB,
1041 MachineBasicBlock::iterator I,
1042 unsigned DestReg, unsigned SubIdx,
1043 const MachineInstr *Orig,
1044 const TargetRegisterInfo &TRI) const {
1045 unsigned Opcode = Orig->getOpcode();
1048 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1049 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1053 case ARM::tLDRpci_pic:
1054 case ARM::t2LDRpci_pic: {
1055 MachineFunction &MF = *MBB.getParent();
1056 unsigned CPI = Orig->getOperand(1).getIndex();
1057 unsigned PCLabelId = duplicateCPV(MF, CPI);
1058 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1060 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1061 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1068 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1069 MachineInstr *MI = TargetInstrInfoImpl::duplicate(Orig, MF);
1070 switch(Orig->getOpcode()) {
1071 case ARM::tLDRpci_pic:
1072 case ARM::t2LDRpci_pic: {
1073 unsigned CPI = Orig->getOperand(1).getIndex();
1074 unsigned PCLabelId = duplicateCPV(MF, CPI);
1075 Orig->getOperand(1).setIndex(CPI);
1076 Orig->getOperand(2).setImm(PCLabelId);
1083 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1084 const MachineInstr *MI1,
1085 const MachineRegisterInfo *MRI) const {
1086 int Opcode = MI0->getOpcode();
1087 if (Opcode == ARM::t2LDRpci ||
1088 Opcode == ARM::t2LDRpci_pic ||
1089 Opcode == ARM::tLDRpci ||
1090 Opcode == ARM::tLDRpci_pic ||
1091 Opcode == ARM::MOV_ga_dyn ||
1092 Opcode == ARM::MOV_ga_pcrel ||
1093 Opcode == ARM::MOV_ga_pcrel_ldr ||
1094 Opcode == ARM::t2MOV_ga_dyn ||
1095 Opcode == ARM::t2MOV_ga_pcrel) {
1096 if (MI1->getOpcode() != Opcode)
1098 if (MI0->getNumOperands() != MI1->getNumOperands())
1101 const MachineOperand &MO0 = MI0->getOperand(1);
1102 const MachineOperand &MO1 = MI1->getOperand(1);
1103 if (MO0.getOffset() != MO1.getOffset())
1106 if (Opcode == ARM::MOV_ga_dyn ||
1107 Opcode == ARM::MOV_ga_pcrel ||
1108 Opcode == ARM::MOV_ga_pcrel_ldr ||
1109 Opcode == ARM::t2MOV_ga_dyn ||
1110 Opcode == ARM::t2MOV_ga_pcrel)
1111 // Ignore the PC labels.
1112 return MO0.getGlobal() == MO1.getGlobal();
1114 const MachineFunction *MF = MI0->getParent()->getParent();
1115 const MachineConstantPool *MCP = MF->getConstantPool();
1116 int CPI0 = MO0.getIndex();
1117 int CPI1 = MO1.getIndex();
1118 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1119 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1120 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1121 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1122 if (isARMCP0 && isARMCP1) {
1123 ARMConstantPoolValue *ACPV0 =
1124 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1125 ARMConstantPoolValue *ACPV1 =
1126 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1127 return ACPV0->hasSameValue(ACPV1);
1128 } else if (!isARMCP0 && !isARMCP1) {
1129 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1132 } else if (Opcode == ARM::PICLDR) {
1133 if (MI1->getOpcode() != Opcode)
1135 if (MI0->getNumOperands() != MI1->getNumOperands())
1138 unsigned Addr0 = MI0->getOperand(1).getReg();
1139 unsigned Addr1 = MI1->getOperand(1).getReg();
1140 if (Addr0 != Addr1) {
1142 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1143 !TargetRegisterInfo::isVirtualRegister(Addr1))
1146 // This assumes SSA form.
1147 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1148 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1149 // Check if the loaded value, e.g. a constantpool of a global address, are
1151 if (!produceSameValue(Def0, Def1, MRI))
1155 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1156 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1157 const MachineOperand &MO0 = MI0->getOperand(i);
1158 const MachineOperand &MO1 = MI1->getOperand(i);
1159 if (!MO0.isIdenticalTo(MO1))
1165 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1168 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1169 /// determine if two loads are loading from the same base address. It should
1170 /// only return true if the base pointers are the same and the only differences
1171 /// between the two addresses is the offset. It also returns the offsets by
1173 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1175 int64_t &Offset2) const {
1176 // Don't worry about Thumb: just ARM and Thumb2.
1177 if (Subtarget.isThumb1Only()) return false;
1179 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1182 switch (Load1->getMachineOpcode()) {
1195 case ARM::t2LDRSHi8:
1197 case ARM::t2LDRSHi12:
1201 switch (Load2->getMachineOpcode()) {
1214 case ARM::t2LDRSHi8:
1216 case ARM::t2LDRSHi12:
1220 // Check if base addresses and chain operands match.
1221 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1222 Load1->getOperand(4) != Load2->getOperand(4))
1225 // Index should be Reg0.
1226 if (Load1->getOperand(3) != Load2->getOperand(3))
1229 // Determine the offsets.
1230 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1231 isa<ConstantSDNode>(Load2->getOperand(1))) {
1232 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1233 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1240 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1241 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1242 /// be scheduled togther. On some targets if two loads are loading from
1243 /// addresses in the same cache line, it's better if they are scheduled
1244 /// together. This function takes two integers that represent the load offsets
1245 /// from the common base address. It returns true if it decides it's desirable
1246 /// to schedule the two loads together. "NumLoads" is the number of loads that
1247 /// have already been scheduled after Load1.
1248 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1249 int64_t Offset1, int64_t Offset2,
1250 unsigned NumLoads) const {
1251 // Don't worry about Thumb: just ARM and Thumb2.
1252 if (Subtarget.isThumb1Only()) return false;
1254 assert(Offset2 > Offset1);
1256 if ((Offset2 - Offset1) / 8 > 64)
1259 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1260 return false; // FIXME: overly conservative?
1262 // Four loads in a row should be sufficient.
1269 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1270 const MachineBasicBlock *MBB,
1271 const MachineFunction &MF) const {
1272 // Debug info is never a scheduling boundary. It's necessary to be explicit
1273 // due to the special treatment of IT instructions below, otherwise a
1274 // dbg_value followed by an IT will result in the IT instruction being
1275 // considered a scheduling hazard, which is wrong. It should be the actual
1276 // instruction preceding the dbg_value instruction(s), just like it is
1277 // when debug info is not present.
1278 if (MI->isDebugValue())
1281 // Terminators and labels can't be scheduled around.
1282 if (MI->getDesc().isTerminator() || MI->isLabel())
1285 // Treat the start of the IT block as a scheduling boundary, but schedule
1286 // t2IT along with all instructions following it.
1287 // FIXME: This is a big hammer. But the alternative is to add all potential
1288 // true and anti dependencies to IT block instructions as implicit operands
1289 // to the t2IT instruction. The added compile time and complexity does not
1291 MachineBasicBlock::const_iterator I = MI;
1292 // Make sure to skip any dbg_value instructions
1293 while (++I != MBB->end() && I->isDebugValue())
1295 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1298 // Don't attempt to schedule around any instruction that defines
1299 // a stack-oriented pointer, as it's unlikely to be profitable. This
1300 // saves compile time, because it doesn't require every single
1301 // stack slot reference to depend on the instruction that does the
1303 if (MI->definesRegister(ARM::SP))
1309 bool ARMBaseInstrInfo::
1310 isProfitableToIfCvt(MachineBasicBlock &MBB,
1311 unsigned NumCycles, unsigned ExtraPredCycles,
1312 const BranchProbability &Probability) const {
1316 // Attempt to estimate the relative costs of predication versus branching.
1317 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1318 UnpredCost /= Probability.getDenominator();
1319 UnpredCost += 1; // The branch itself
1320 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1322 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1325 bool ARMBaseInstrInfo::
1326 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1327 unsigned TCycles, unsigned TExtra,
1328 MachineBasicBlock &FMBB,
1329 unsigned FCycles, unsigned FExtra,
1330 const BranchProbability &Probability) const {
1331 if (!TCycles || !FCycles)
1334 // Attempt to estimate the relative costs of predication versus branching.
1335 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1336 TUnpredCost /= Probability.getDenominator();
1338 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1339 unsigned FUnpredCost = Comp * FCycles;
1340 FUnpredCost /= Probability.getDenominator();
1342 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1343 UnpredCost += 1; // The branch itself
1344 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1346 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1349 /// getInstrPredicate - If instruction is predicated, returns its predicate
1350 /// condition, otherwise returns AL. It also returns the condition code
1351 /// register by reference.
1353 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1354 int PIdx = MI->findFirstPredOperandIdx();
1360 PredReg = MI->getOperand(PIdx+1).getReg();
1361 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1365 int llvm::getMatchingCondBranchOpcode(int Opc) {
1368 else if (Opc == ARM::tB)
1370 else if (Opc == ARM::t2B)
1373 llvm_unreachable("Unknown unconditional branch opcode!");
1378 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1379 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1380 unsigned DestReg, unsigned BaseReg, int NumBytes,
1381 ARMCC::CondCodes Pred, unsigned PredReg,
1382 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1383 bool isSub = NumBytes < 0;
1384 if (isSub) NumBytes = -NumBytes;
1387 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1388 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1389 assert(ThisVal && "Didn't extract field correctly");
1391 // We will handle these bits from offset, clear them.
1392 NumBytes &= ~ThisVal;
1394 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1396 // Build the new ADD / SUB.
1397 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1398 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1399 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1400 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1401 .setMIFlags(MIFlags);
1406 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1407 unsigned FrameReg, int &Offset,
1408 const ARMBaseInstrInfo &TII) {
1409 unsigned Opcode = MI.getOpcode();
1410 const MCInstrDesc &Desc = MI.getDesc();
1411 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1414 // Memory operands in inline assembly always use AddrMode2.
1415 if (Opcode == ARM::INLINEASM)
1416 AddrMode = ARMII::AddrMode2;
1418 if (Opcode == ARM::ADDri) {
1419 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1421 // Turn it into a move.
1422 MI.setDesc(TII.get(ARM::MOVr));
1423 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1424 MI.RemoveOperand(FrameRegIdx+1);
1427 } else if (Offset < 0) {
1430 MI.setDesc(TII.get(ARM::SUBri));
1433 // Common case: small offset, fits into instruction.
1434 if (ARM_AM::getSOImmVal(Offset) != -1) {
1435 // Replace the FrameIndex with sp / fp
1436 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1437 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1442 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1444 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1445 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1447 // We will handle these bits from offset, clear them.
1448 Offset &= ~ThisImmVal;
1450 // Get the properly encoded SOImmVal field.
1451 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1452 "Bit extraction didn't work?");
1453 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1455 unsigned ImmIdx = 0;
1457 unsigned NumBits = 0;
1460 case ARMII::AddrMode_i12: {
1461 ImmIdx = FrameRegIdx + 1;
1462 InstrOffs = MI.getOperand(ImmIdx).getImm();
1466 case ARMII::AddrMode2: {
1467 ImmIdx = FrameRegIdx+2;
1468 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1469 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1474 case ARMII::AddrMode3: {
1475 ImmIdx = FrameRegIdx+2;
1476 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1477 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1482 case ARMII::AddrMode4:
1483 case ARMII::AddrMode6:
1484 // Can't fold any offset even if it's zero.
1486 case ARMII::AddrMode5: {
1487 ImmIdx = FrameRegIdx+1;
1488 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1489 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1496 llvm_unreachable("Unsupported addressing mode!");
1500 Offset += InstrOffs * Scale;
1501 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1507 // Attempt to fold address comp. if opcode has offset bits
1509 // Common case: small offset, fits into instruction.
1510 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1511 int ImmedOffset = Offset / Scale;
1512 unsigned Mask = (1 << NumBits) - 1;
1513 if ((unsigned)Offset <= Mask * Scale) {
1514 // Replace the FrameIndex with sp
1515 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1516 // FIXME: When addrmode2 goes away, this will simplify (like the
1517 // T2 version), as the LDR.i12 versions don't need the encoding
1518 // tricks for the offset value.
1520 if (AddrMode == ARMII::AddrMode_i12)
1521 ImmedOffset = -ImmedOffset;
1523 ImmedOffset |= 1 << NumBits;
1525 ImmOp.ChangeToImmediate(ImmedOffset);
1530 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1531 ImmedOffset = ImmedOffset & Mask;
1533 if (AddrMode == ARMII::AddrMode_i12)
1534 ImmedOffset = -ImmedOffset;
1536 ImmedOffset |= 1 << NumBits;
1538 ImmOp.ChangeToImmediate(ImmedOffset);
1539 Offset &= ~(Mask*Scale);
1543 Offset = (isSub) ? -Offset : Offset;
1547 bool ARMBaseInstrInfo::
1548 AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpMask,
1549 int &CmpValue) const {
1550 switch (MI->getOpcode()) {
1554 SrcReg = MI->getOperand(0).getReg();
1556 CmpValue = MI->getOperand(1).getImm();
1560 SrcReg = MI->getOperand(0).getReg();
1561 CmpMask = MI->getOperand(1).getImm();
1569 /// isSuitableForMask - Identify a suitable 'and' instruction that
1570 /// operates on the given source register and applies the same mask
1571 /// as a 'tst' instruction. Provide a limited look-through for copies.
1572 /// When successful, MI will hold the found instruction.
1573 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
1574 int CmpMask, bool CommonUse) {
1575 switch (MI->getOpcode()) {
1578 if (CmpMask != MI->getOperand(2).getImm())
1580 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
1584 // Walk down one instruction which is potentially an 'and'.
1585 const MachineInstr &Copy = *MI;
1586 MachineBasicBlock::iterator AND(
1587 llvm::next(MachineBasicBlock::iterator(MI)));
1588 if (AND == MI->getParent()->end()) return false;
1590 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
1598 /// OptimizeCompareInstr - Convert the instruction supplying the argument to the
1599 /// comparison into one that sets the zero bit in the flags register.
1600 bool ARMBaseInstrInfo::
1601 OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
1602 int CmpValue, const MachineRegisterInfo *MRI) const {
1606 MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
1607 if (llvm::next(DI) != MRI->def_end())
1608 // Only support one definition.
1611 MachineInstr *MI = &*DI;
1613 // Masked compares sometimes use the same register as the corresponding 'and'.
1614 if (CmpMask != ~0) {
1615 if (!isSuitableForMask(MI, SrcReg, CmpMask, false)) {
1617 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
1618 UE = MRI->use_end(); UI != UE; ++UI) {
1619 if (UI->getParent() != CmpInstr->getParent()) continue;
1620 MachineInstr *PotentialAND = &*UI;
1621 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true))
1626 if (!MI) return false;
1630 // Conservatively refuse to convert an instruction which isn't in the same BB
1631 // as the comparison.
1632 if (MI->getParent() != CmpInstr->getParent())
1635 // Check that CPSR isn't set between the comparison instruction and the one we
1637 MachineBasicBlock::const_iterator I = CmpInstr, E = MI,
1638 B = MI->getParent()->begin();
1640 // Early exit if CmpInstr is at the beginning of the BB.
1641 if (I == B) return false;
1644 for (; I != E; --I) {
1645 const MachineInstr &Instr = *I;
1647 for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) {
1648 const MachineOperand &MO = Instr.getOperand(IO);
1649 if (!MO.isReg()) continue;
1651 // This instruction modifies or uses CPSR after the one we want to
1652 // change. We can't do this transformation.
1653 if (MO.getReg() == ARM::CPSR)
1658 // The 'and' is below the comparison instruction.
1662 // Set the "zero" bit in CPSR.
1663 switch (MI->getOpcode()) {
1697 case ARM::t2EORri: {
1698 // Scan forward for the use of CPSR, if it's a conditional code requires
1699 // checking of V bit, then this is not safe to do. If we can't find the
1700 // CPSR use (i.e. used in another block), then it's not safe to perform
1701 // the optimization.
1702 bool isSafe = false;
1704 E = MI->getParent()->end();
1705 while (!isSafe && ++I != E) {
1706 const MachineInstr &Instr = *I;
1707 for (unsigned IO = 0, EO = Instr.getNumOperands();
1708 !isSafe && IO != EO; ++IO) {
1709 const MachineOperand &MO = Instr.getOperand(IO);
1710 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
1716 // Condition code is after the operand before CPSR.
1717 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
1736 // Toggle the optional operand to CPSR.
1737 MI->getOperand(5).setReg(ARM::CPSR);
1738 MI->getOperand(5).setIsDef(true);
1739 CmpInstr->eraseFromParent();
1747 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
1748 MachineInstr *DefMI, unsigned Reg,
1749 MachineRegisterInfo *MRI) const {
1750 // Fold large immediates into add, sub, or, xor.
1751 unsigned DefOpc = DefMI->getOpcode();
1752 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
1754 if (!DefMI->getOperand(1).isImm())
1755 // Could be t2MOVi32imm <ga:xx>
1758 if (!MRI->hasOneNonDBGUse(Reg))
1761 unsigned UseOpc = UseMI->getOpcode();
1762 unsigned NewUseOpc = 0;
1763 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
1764 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
1765 bool Commute = false;
1767 default: return false;
1775 case ARM::t2EORrr: {
1776 Commute = UseMI->getOperand(2).getReg() != Reg;
1783 NewUseOpc = ARM::SUBri;
1789 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
1791 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
1792 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
1795 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
1796 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
1797 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
1801 case ARM::t2SUBrr: {
1805 NewUseOpc = ARM::t2SUBri;
1810 case ARM::t2EORrr: {
1811 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
1813 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
1814 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
1817 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
1818 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
1819 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
1827 unsigned OpIdx = Commute ? 2 : 1;
1828 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
1829 bool isKill = UseMI->getOperand(OpIdx).isKill();
1830 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
1831 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
1832 *UseMI, UseMI->getDebugLoc(),
1833 get(NewUseOpc), NewReg)
1834 .addReg(Reg1, getKillRegState(isKill))
1835 .addImm(SOImmValV1)));
1836 UseMI->setDesc(get(NewUseOpc));
1837 UseMI->getOperand(1).setReg(NewReg);
1838 UseMI->getOperand(1).setIsKill();
1839 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
1840 DefMI->eraseFromParent();
1845 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1846 const MachineInstr *MI) const {
1847 if (!ItinData || ItinData->isEmpty())
1850 const MCInstrDesc &Desc = MI->getDesc();
1851 unsigned Class = Desc.getSchedClass();
1852 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
1856 unsigned Opc = MI->getOpcode();
1859 llvm_unreachable("Unexpected multi-uops instruction!");
1865 // The number of uOps for load / store multiple are determined by the number
1868 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
1869 // same cycle. The scheduling for the first load / store must be done
1870 // separately by assuming the the address is not 64-bit aligned.
1872 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
1873 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
1874 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
1876 case ARM::VLDMDIA_UPD:
1877 case ARM::VLDMDDB_UPD:
1879 case ARM::VLDMSIA_UPD:
1880 case ARM::VLDMSDB_UPD:
1882 case ARM::VSTMDIA_UPD:
1883 case ARM::VSTMDDB_UPD:
1885 case ARM::VSTMSIA_UPD:
1886 case ARM::VSTMSDB_UPD: {
1887 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
1888 return (NumRegs / 2) + (NumRegs % 2) + 1;
1891 case ARM::LDMIA_RET:
1896 case ARM::LDMIA_UPD:
1897 case ARM::LDMDA_UPD:
1898 case ARM::LDMDB_UPD:
1899 case ARM::LDMIB_UPD:
1904 case ARM::STMIA_UPD:
1905 case ARM::STMDA_UPD:
1906 case ARM::STMDB_UPD:
1907 case ARM::STMIB_UPD:
1909 case ARM::tLDMIA_UPD:
1911 case ARM::tSTMIA_UPD:
1915 case ARM::t2LDMIA_RET:
1918 case ARM::t2LDMIA_UPD:
1919 case ARM::t2LDMDB_UPD:
1922 case ARM::t2STMIA_UPD:
1923 case ARM::t2STMDB_UPD: {
1924 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
1925 if (Subtarget.isCortexA8()) {
1928 // 4 registers would be issued: 2, 2.
1929 // 5 registers would be issued: 2, 2, 1.
1930 UOps = (NumRegs / 2);
1934 } else if (Subtarget.isCortexA9()) {
1935 UOps = (NumRegs / 2);
1936 // If there are odd number of registers or if it's not 64-bit aligned,
1937 // then it takes an extra AGU (Address Generation Unit) cycle.
1938 if ((NumRegs % 2) ||
1939 !MI->hasOneMemOperand() ||
1940 (*MI->memoperands_begin())->getAlignment() < 8)
1944 // Assume the worst.
1952 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
1953 const MCInstrDesc &DefMCID,
1955 unsigned DefIdx, unsigned DefAlign) const {
1956 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1958 // Def is the address writeback.
1959 return ItinData->getOperandCycle(DefClass, DefIdx);
1962 if (Subtarget.isCortexA8()) {
1963 // (regno / 2) + (regno % 2) + 1
1964 DefCycle = RegNo / 2 + 1;
1967 } else if (Subtarget.isCortexA9()) {
1969 bool isSLoad = false;
1971 switch (DefMCID.getOpcode()) {
1974 case ARM::VLDMSIA_UPD:
1975 case ARM::VLDMSDB_UPD:
1980 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
1981 // then it takes an extra cycle.
1982 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
1985 // Assume the worst.
1986 DefCycle = RegNo + 2;
1993 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
1994 const MCInstrDesc &DefMCID,
1996 unsigned DefIdx, unsigned DefAlign) const {
1997 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1999 // Def is the address writeback.
2000 return ItinData->getOperandCycle(DefClass, DefIdx);
2003 if (Subtarget.isCortexA8()) {
2004 // 4 registers would be issued: 1, 2, 1.
2005 // 5 registers would be issued: 1, 2, 2.
2006 DefCycle = RegNo / 2;
2009 // Result latency is issue cycle + 2: E2.
2011 } else if (Subtarget.isCortexA9()) {
2012 DefCycle = (RegNo / 2);
2013 // If there are odd number of registers or if it's not 64-bit aligned,
2014 // then it takes an extra AGU (Address Generation Unit) cycle.
2015 if ((RegNo % 2) || DefAlign < 8)
2017 // Result latency is AGU cycles + 2.
2020 // Assume the worst.
2021 DefCycle = RegNo + 2;
2028 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
2029 const MCInstrDesc &UseMCID,
2031 unsigned UseIdx, unsigned UseAlign) const {
2032 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2034 return ItinData->getOperandCycle(UseClass, UseIdx);
2037 if (Subtarget.isCortexA8()) {
2038 // (regno / 2) + (regno % 2) + 1
2039 UseCycle = RegNo / 2 + 1;
2042 } else if (Subtarget.isCortexA9()) {
2044 bool isSStore = false;
2046 switch (UseMCID.getOpcode()) {
2049 case ARM::VSTMSIA_UPD:
2050 case ARM::VSTMSDB_UPD:
2055 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2056 // then it takes an extra cycle.
2057 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2060 // Assume the worst.
2061 UseCycle = RegNo + 2;
2068 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2069 const MCInstrDesc &UseMCID,
2071 unsigned UseIdx, unsigned UseAlign) const {
2072 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2074 return ItinData->getOperandCycle(UseClass, UseIdx);
2077 if (Subtarget.isCortexA8()) {
2078 UseCycle = RegNo / 2;
2083 } else if (Subtarget.isCortexA9()) {
2084 UseCycle = (RegNo / 2);
2085 // If there are odd number of registers or if it's not 64-bit aligned,
2086 // then it takes an extra AGU (Address Generation Unit) cycle.
2087 if ((RegNo % 2) || UseAlign < 8)
2090 // Assume the worst.
2097 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2098 const MCInstrDesc &DefMCID,
2099 unsigned DefIdx, unsigned DefAlign,
2100 const MCInstrDesc &UseMCID,
2101 unsigned UseIdx, unsigned UseAlign) const {
2102 unsigned DefClass = DefMCID.getSchedClass();
2103 unsigned UseClass = UseMCID.getSchedClass();
2105 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2106 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2108 // This may be a def / use of a variable_ops instruction, the operand
2109 // latency might be determinable dynamically. Let the target try to
2112 bool LdmBypass = false;
2113 switch (DefMCID.getOpcode()) {
2115 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2119 case ARM::VLDMDIA_UPD:
2120 case ARM::VLDMDDB_UPD:
2122 case ARM::VLDMSIA_UPD:
2123 case ARM::VLDMSDB_UPD:
2124 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2127 case ARM::LDMIA_RET:
2132 case ARM::LDMIA_UPD:
2133 case ARM::LDMDA_UPD:
2134 case ARM::LDMDB_UPD:
2135 case ARM::LDMIB_UPD:
2137 case ARM::tLDMIA_UPD:
2139 case ARM::t2LDMIA_RET:
2142 case ARM::t2LDMIA_UPD:
2143 case ARM::t2LDMDB_UPD:
2145 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2150 // We can't seem to determine the result latency of the def, assume it's 2.
2154 switch (UseMCID.getOpcode()) {
2156 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
2160 case ARM::VSTMDIA_UPD:
2161 case ARM::VSTMDDB_UPD:
2163 case ARM::VSTMSIA_UPD:
2164 case ARM::VSTMSDB_UPD:
2165 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2172 case ARM::STMIA_UPD:
2173 case ARM::STMDA_UPD:
2174 case ARM::STMDB_UPD:
2175 case ARM::STMIB_UPD:
2177 case ARM::tSTMIA_UPD:
2182 case ARM::t2STMIA_UPD:
2183 case ARM::t2STMDB_UPD:
2184 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2189 // Assume it's read in the first stage.
2192 UseCycle = DefCycle - UseCycle + 1;
2195 // It's a variable_ops instruction so we can't use DefIdx here. Just use
2196 // first def operand.
2197 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
2200 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
2201 UseClass, UseIdx)) {
2210 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2211 const MachineInstr *DefMI, unsigned DefIdx,
2212 const MachineInstr *UseMI, unsigned UseIdx) const {
2213 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
2214 DefMI->isRegSequence() || DefMI->isImplicitDef())
2217 const MCInstrDesc &DefMCID = DefMI->getDesc();
2218 if (!ItinData || ItinData->isEmpty())
2219 return DefMCID.mayLoad() ? 3 : 1;
2221 const MCInstrDesc &UseMCID = UseMI->getDesc();
2222 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
2223 if (DefMO.getReg() == ARM::CPSR) {
2224 if (DefMI->getOpcode() == ARM::FMSTAT) {
2225 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
2226 return Subtarget.isCortexA9() ? 1 : 20;
2229 // CPSR set and branch can be paired in the same cycle.
2230 if (UseMCID.isBranch())
2234 unsigned DefAlign = DefMI->hasOneMemOperand()
2235 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
2236 unsigned UseAlign = UseMI->hasOneMemOperand()
2237 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
2238 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2239 UseMCID, UseIdx, UseAlign);
2242 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2243 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2244 // variants are one cycle cheaper.
2245 switch (DefMCID.getOpcode()) {
2249 unsigned ShOpVal = DefMI->getOperand(3).getImm();
2250 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2252 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2259 case ARM::t2LDRSHs: {
2260 // Thumb2 mode: lsl only.
2261 unsigned ShAmt = DefMI->getOperand(3).getImm();
2262 if (ShAmt == 0 || ShAmt == 2)
2269 if (DefAlign < 8 && Subtarget.isCortexA9())
2270 switch (DefMCID.getOpcode()) {
2276 case ARM::VLD1q8_UPD:
2277 case ARM::VLD1q16_UPD:
2278 case ARM::VLD1q32_UPD:
2279 case ARM::VLD1q64_UPD:
2286 case ARM::VLD2d8_UPD:
2287 case ARM::VLD2d16_UPD:
2288 case ARM::VLD2d32_UPD:
2289 case ARM::VLD2q8_UPD:
2290 case ARM::VLD2q16_UPD:
2291 case ARM::VLD2q32_UPD:
2296 case ARM::VLD3d8_UPD:
2297 case ARM::VLD3d16_UPD:
2298 case ARM::VLD3d32_UPD:
2299 case ARM::VLD1d64T_UPD:
2300 case ARM::VLD3q8_UPD:
2301 case ARM::VLD3q16_UPD:
2302 case ARM::VLD3q32_UPD:
2307 case ARM::VLD4d8_UPD:
2308 case ARM::VLD4d16_UPD:
2309 case ARM::VLD4d32_UPD:
2310 case ARM::VLD1d64Q_UPD:
2311 case ARM::VLD4q8_UPD:
2312 case ARM::VLD4q16_UPD:
2313 case ARM::VLD4q32_UPD:
2314 case ARM::VLD1DUPq8:
2315 case ARM::VLD1DUPq16:
2316 case ARM::VLD1DUPq32:
2317 case ARM::VLD1DUPq8_UPD:
2318 case ARM::VLD1DUPq16_UPD:
2319 case ARM::VLD1DUPq32_UPD:
2320 case ARM::VLD2DUPd8:
2321 case ARM::VLD2DUPd16:
2322 case ARM::VLD2DUPd32:
2323 case ARM::VLD2DUPd8_UPD:
2324 case ARM::VLD2DUPd16_UPD:
2325 case ARM::VLD2DUPd32_UPD:
2326 case ARM::VLD4DUPd8:
2327 case ARM::VLD4DUPd16:
2328 case ARM::VLD4DUPd32:
2329 case ARM::VLD4DUPd8_UPD:
2330 case ARM::VLD4DUPd16_UPD:
2331 case ARM::VLD4DUPd32_UPD:
2333 case ARM::VLD1LNd16:
2334 case ARM::VLD1LNd32:
2335 case ARM::VLD1LNd8_UPD:
2336 case ARM::VLD1LNd16_UPD:
2337 case ARM::VLD1LNd32_UPD:
2339 case ARM::VLD2LNd16:
2340 case ARM::VLD2LNd32:
2341 case ARM::VLD2LNq16:
2342 case ARM::VLD2LNq32:
2343 case ARM::VLD2LNd8_UPD:
2344 case ARM::VLD2LNd16_UPD:
2345 case ARM::VLD2LNd32_UPD:
2346 case ARM::VLD2LNq16_UPD:
2347 case ARM::VLD2LNq32_UPD:
2349 case ARM::VLD4LNd16:
2350 case ARM::VLD4LNd32:
2351 case ARM::VLD4LNq16:
2352 case ARM::VLD4LNq32:
2353 case ARM::VLD4LNd8_UPD:
2354 case ARM::VLD4LNd16_UPD:
2355 case ARM::VLD4LNd32_UPD:
2356 case ARM::VLD4LNq16_UPD:
2357 case ARM::VLD4LNq32_UPD:
2358 // If the address is not 64-bit aligned, the latencies of these
2359 // instructions increases by one.
2368 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2369 SDNode *DefNode, unsigned DefIdx,
2370 SDNode *UseNode, unsigned UseIdx) const {
2371 if (!DefNode->isMachineOpcode())
2374 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
2376 if (isZeroCost(DefMCID.Opcode))
2379 if (!ItinData || ItinData->isEmpty())
2380 return DefMCID.mayLoad() ? 3 : 1;
2382 if (!UseNode->isMachineOpcode()) {
2383 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
2384 if (Subtarget.isCortexA9())
2385 return Latency <= 2 ? 1 : Latency - 1;
2387 return Latency <= 3 ? 1 : Latency - 2;
2390 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
2391 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
2392 unsigned DefAlign = !DefMN->memoperands_empty()
2393 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
2394 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
2395 unsigned UseAlign = !UseMN->memoperands_empty()
2396 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
2397 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2398 UseMCID, UseIdx, UseAlign);
2401 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2402 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2403 // variants are one cycle cheaper.
2404 switch (DefMCID.getOpcode()) {
2409 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2410 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2412 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2419 case ARM::t2LDRSHs: {
2420 // Thumb2 mode: lsl only.
2422 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2423 if (ShAmt == 0 || ShAmt == 2)
2430 if (DefAlign < 8 && Subtarget.isCortexA9())
2431 switch (DefMCID.getOpcode()) {
2433 case ARM::VLD1q8Pseudo:
2434 case ARM::VLD1q16Pseudo:
2435 case ARM::VLD1q32Pseudo:
2436 case ARM::VLD1q64Pseudo:
2437 case ARM::VLD1q8Pseudo_UPD:
2438 case ARM::VLD1q16Pseudo_UPD:
2439 case ARM::VLD1q32Pseudo_UPD:
2440 case ARM::VLD1q64Pseudo_UPD:
2441 case ARM::VLD2d8Pseudo:
2442 case ARM::VLD2d16Pseudo:
2443 case ARM::VLD2d32Pseudo:
2444 case ARM::VLD2q8Pseudo:
2445 case ARM::VLD2q16Pseudo:
2446 case ARM::VLD2q32Pseudo:
2447 case ARM::VLD2d8Pseudo_UPD:
2448 case ARM::VLD2d16Pseudo_UPD:
2449 case ARM::VLD2d32Pseudo_UPD:
2450 case ARM::VLD2q8Pseudo_UPD:
2451 case ARM::VLD2q16Pseudo_UPD:
2452 case ARM::VLD2q32Pseudo_UPD:
2453 case ARM::VLD3d8Pseudo:
2454 case ARM::VLD3d16Pseudo:
2455 case ARM::VLD3d32Pseudo:
2456 case ARM::VLD1d64TPseudo:
2457 case ARM::VLD3d8Pseudo_UPD:
2458 case ARM::VLD3d16Pseudo_UPD:
2459 case ARM::VLD3d32Pseudo_UPD:
2460 case ARM::VLD1d64TPseudo_UPD:
2461 case ARM::VLD3q8Pseudo_UPD:
2462 case ARM::VLD3q16Pseudo_UPD:
2463 case ARM::VLD3q32Pseudo_UPD:
2464 case ARM::VLD3q8oddPseudo:
2465 case ARM::VLD3q16oddPseudo:
2466 case ARM::VLD3q32oddPseudo:
2467 case ARM::VLD3q8oddPseudo_UPD:
2468 case ARM::VLD3q16oddPseudo_UPD:
2469 case ARM::VLD3q32oddPseudo_UPD:
2470 case ARM::VLD4d8Pseudo:
2471 case ARM::VLD4d16Pseudo:
2472 case ARM::VLD4d32Pseudo:
2473 case ARM::VLD1d64QPseudo:
2474 case ARM::VLD4d8Pseudo_UPD:
2475 case ARM::VLD4d16Pseudo_UPD:
2476 case ARM::VLD4d32Pseudo_UPD:
2477 case ARM::VLD1d64QPseudo_UPD:
2478 case ARM::VLD4q8Pseudo_UPD:
2479 case ARM::VLD4q16Pseudo_UPD:
2480 case ARM::VLD4q32Pseudo_UPD:
2481 case ARM::VLD4q8oddPseudo:
2482 case ARM::VLD4q16oddPseudo:
2483 case ARM::VLD4q32oddPseudo:
2484 case ARM::VLD4q8oddPseudo_UPD:
2485 case ARM::VLD4q16oddPseudo_UPD:
2486 case ARM::VLD4q32oddPseudo_UPD:
2487 case ARM::VLD1DUPq8Pseudo:
2488 case ARM::VLD1DUPq16Pseudo:
2489 case ARM::VLD1DUPq32Pseudo:
2490 case ARM::VLD1DUPq8Pseudo_UPD:
2491 case ARM::VLD1DUPq16Pseudo_UPD:
2492 case ARM::VLD1DUPq32Pseudo_UPD:
2493 case ARM::VLD2DUPd8Pseudo:
2494 case ARM::VLD2DUPd16Pseudo:
2495 case ARM::VLD2DUPd32Pseudo:
2496 case ARM::VLD2DUPd8Pseudo_UPD:
2497 case ARM::VLD2DUPd16Pseudo_UPD:
2498 case ARM::VLD2DUPd32Pseudo_UPD:
2499 case ARM::VLD4DUPd8Pseudo:
2500 case ARM::VLD4DUPd16Pseudo:
2501 case ARM::VLD4DUPd32Pseudo:
2502 case ARM::VLD4DUPd8Pseudo_UPD:
2503 case ARM::VLD4DUPd16Pseudo_UPD:
2504 case ARM::VLD4DUPd32Pseudo_UPD:
2505 case ARM::VLD1LNq8Pseudo:
2506 case ARM::VLD1LNq16Pseudo:
2507 case ARM::VLD1LNq32Pseudo:
2508 case ARM::VLD1LNq8Pseudo_UPD:
2509 case ARM::VLD1LNq16Pseudo_UPD:
2510 case ARM::VLD1LNq32Pseudo_UPD:
2511 case ARM::VLD2LNd8Pseudo:
2512 case ARM::VLD2LNd16Pseudo:
2513 case ARM::VLD2LNd32Pseudo:
2514 case ARM::VLD2LNq16Pseudo:
2515 case ARM::VLD2LNq32Pseudo:
2516 case ARM::VLD2LNd8Pseudo_UPD:
2517 case ARM::VLD2LNd16Pseudo_UPD:
2518 case ARM::VLD2LNd32Pseudo_UPD:
2519 case ARM::VLD2LNq16Pseudo_UPD:
2520 case ARM::VLD2LNq32Pseudo_UPD:
2521 case ARM::VLD4LNd8Pseudo:
2522 case ARM::VLD4LNd16Pseudo:
2523 case ARM::VLD4LNd32Pseudo:
2524 case ARM::VLD4LNq16Pseudo:
2525 case ARM::VLD4LNq32Pseudo:
2526 case ARM::VLD4LNd8Pseudo_UPD:
2527 case ARM::VLD4LNd16Pseudo_UPD:
2528 case ARM::VLD4LNd32Pseudo_UPD:
2529 case ARM::VLD4LNq16Pseudo_UPD:
2530 case ARM::VLD4LNq32Pseudo_UPD:
2531 // If the address is not 64-bit aligned, the latencies of these
2532 // instructions increases by one.
2540 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2541 const MachineInstr *MI,
2542 unsigned *PredCost) const {
2543 if (MI->isCopyLike() || MI->isInsertSubreg() ||
2544 MI->isRegSequence() || MI->isImplicitDef())
2547 if (!ItinData || ItinData->isEmpty())
2550 const MCInstrDesc &MCID = MI->getDesc();
2551 unsigned Class = MCID.getSchedClass();
2552 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
2553 if (PredCost && MCID.hasImplicitDefOfPhysReg(ARM::CPSR))
2554 // When predicated, CPSR is an additional source operand for CPSR updating
2555 // instructions, this apparently increases their latencies.
2558 return ItinData->getStageLatency(Class);
2559 return getNumMicroOps(ItinData, MI);
2562 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2563 SDNode *Node) const {
2564 if (!Node->isMachineOpcode())
2567 if (!ItinData || ItinData->isEmpty())
2570 unsigned Opcode = Node->getMachineOpcode();
2573 return ItinData->getStageLatency(get(Opcode).getSchedClass());
2580 bool ARMBaseInstrInfo::
2581 hasHighOperandLatency(const InstrItineraryData *ItinData,
2582 const MachineRegisterInfo *MRI,
2583 const MachineInstr *DefMI, unsigned DefIdx,
2584 const MachineInstr *UseMI, unsigned UseIdx) const {
2585 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2586 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
2587 if (Subtarget.isCortexA8() &&
2588 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
2589 // CortexA8 VFP instructions are not pipelined.
2592 // Hoist VFP / NEON instructions with 4 or higher latency.
2593 int Latency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
2596 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
2597 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
2600 bool ARMBaseInstrInfo::
2601 hasLowDefLatency(const InstrItineraryData *ItinData,
2602 const MachineInstr *DefMI, unsigned DefIdx) const {
2603 if (!ItinData || ItinData->isEmpty())
2606 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2607 if (DDomain == ARMII::DomainGeneral) {
2608 unsigned DefClass = DefMI->getDesc().getSchedClass();
2609 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2610 return (DefCycle != -1 && DefCycle <= 2);
2616 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
2617 unsigned &AddSubOpc,
2618 bool &NegAcc, bool &HasLane) const {
2619 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
2620 if (I == MLxEntryMap.end())
2623 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
2624 MulOpc = Entry.MulOpc;
2625 AddSubOpc = Entry.AddSubOpc;
2626 NegAcc = Entry.NegAcc;
2627 HasLane = Entry.HasLane;