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 // tGPR is used sometimes in ARM instructions that need to avoid using
715 // certain registers. Just treat it as GPR here. Likewise, rGPR.
716 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
717 || RC == ARM::rGPRRegisterClass)
718 RC = ARM::GPRRegisterClass;
720 switch (RC->getID()) {
721 case ARM::GPRRegClassID:
722 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
723 .addReg(SrcReg, getKillRegState(isKill))
724 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
726 case ARM::SPRRegClassID:
727 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
728 .addReg(SrcReg, getKillRegState(isKill))
729 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
731 case ARM::DPRRegClassID:
732 case ARM::DPR_VFP2RegClassID:
733 case ARM::DPR_8RegClassID:
734 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
735 .addReg(SrcReg, getKillRegState(isKill))
736 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
738 case ARM::QPRRegClassID:
739 case ARM::QPR_VFP2RegClassID:
740 case ARM::QPR_8RegClassID:
741 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
742 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64Pseudo))
743 .addFrameIndex(FI).addImm(16)
744 .addReg(SrcReg, getKillRegState(isKill))
745 .addMemOperand(MMO));
747 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
748 .addReg(SrcReg, getKillRegState(isKill))
750 .addMemOperand(MMO));
753 case ARM::QQPRRegClassID:
754 case ARM::QQPR_VFP2RegClassID:
755 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
756 // FIXME: It's possible to only store part of the QQ register if the
757 // spilled def has a sub-register index.
758 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
759 .addFrameIndex(FI).addImm(16)
760 .addReg(SrcReg, getKillRegState(isKill))
761 .addMemOperand(MMO));
763 MachineInstrBuilder MIB =
764 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
767 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
768 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
769 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
770 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
773 case ARM::QQQQPRRegClassID: {
774 MachineInstrBuilder MIB =
775 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
778 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
779 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
780 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
781 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
782 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
783 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
784 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
785 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
789 llvm_unreachable("Unknown regclass!");
794 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
795 int &FrameIndex) const {
796 switch (MI->getOpcode()) {
799 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
800 if (MI->getOperand(1).isFI() &&
801 MI->getOperand(2).isReg() &&
802 MI->getOperand(3).isImm() &&
803 MI->getOperand(2).getReg() == 0 &&
804 MI->getOperand(3).getImm() == 0) {
805 FrameIndex = MI->getOperand(1).getIndex();
806 return MI->getOperand(0).getReg();
814 if (MI->getOperand(1).isFI() &&
815 MI->getOperand(2).isImm() &&
816 MI->getOperand(2).getImm() == 0) {
817 FrameIndex = MI->getOperand(1).getIndex();
818 return MI->getOperand(0).getReg();
821 case ARM::VST1q64Pseudo:
822 if (MI->getOperand(0).isFI() &&
823 MI->getOperand(2).getSubReg() == 0) {
824 FrameIndex = MI->getOperand(0).getIndex();
825 return MI->getOperand(2).getReg();
829 if (MI->getOperand(1).isFI() &&
830 MI->getOperand(0).getSubReg() == 0) {
831 FrameIndex = MI->getOperand(1).getIndex();
832 return MI->getOperand(0).getReg();
840 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
841 int &FrameIndex) const {
842 const MachineMemOperand *Dummy;
843 return MI->getDesc().mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
846 void ARMBaseInstrInfo::
847 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
848 unsigned DestReg, int FI,
849 const TargetRegisterClass *RC,
850 const TargetRegisterInfo *TRI) const {
852 if (I != MBB.end()) DL = I->getDebugLoc();
853 MachineFunction &MF = *MBB.getParent();
854 MachineFrameInfo &MFI = *MF.getFrameInfo();
855 unsigned Align = MFI.getObjectAlignment(FI);
856 MachineMemOperand *MMO =
857 MF.getMachineMemOperand(
858 MachinePointerInfo(PseudoSourceValue::getFixedStack(FI)),
859 MachineMemOperand::MOLoad,
860 MFI.getObjectSize(FI),
863 // tGPR is used sometimes in ARM instructions that need to avoid using
864 // certain registers. Just treat it as GPR here.
865 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
866 || RC == ARM::rGPRRegisterClass)
867 RC = ARM::GPRRegisterClass;
869 switch (RC->getID()) {
870 case ARM::GPRRegClassID:
871 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
872 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
874 case ARM::SPRRegClassID:
875 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
876 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
878 case ARM::DPRRegClassID:
879 case ARM::DPR_VFP2RegClassID:
880 case ARM::DPR_8RegClassID:
881 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
882 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
884 case ARM::QPRRegClassID:
885 case ARM::QPR_VFP2RegClassID:
886 case ARM::QPR_8RegClassID:
887 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
888 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64Pseudo), DestReg)
889 .addFrameIndex(FI).addImm(16)
890 .addMemOperand(MMO));
892 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
894 .addMemOperand(MMO));
897 case ARM::QQPRRegClassID:
898 case ARM::QQPR_VFP2RegClassID:
899 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
900 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
901 .addFrameIndex(FI).addImm(16)
902 .addMemOperand(MMO));
904 MachineInstrBuilder MIB =
905 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
908 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
909 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
910 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
911 AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
914 case ARM::QQQQPRRegClassID: {
915 MachineInstrBuilder MIB =
916 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
919 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
920 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
921 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
922 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
923 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI);
924 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI);
925 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI);
926 AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
930 llvm_unreachable("Unknown regclass!");
935 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
936 int &FrameIndex) const {
937 switch (MI->getOpcode()) {
940 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
941 if (MI->getOperand(1).isFI() &&
942 MI->getOperand(2).isReg() &&
943 MI->getOperand(3).isImm() &&
944 MI->getOperand(2).getReg() == 0 &&
945 MI->getOperand(3).getImm() == 0) {
946 FrameIndex = MI->getOperand(1).getIndex();
947 return MI->getOperand(0).getReg();
955 if (MI->getOperand(1).isFI() &&
956 MI->getOperand(2).isImm() &&
957 MI->getOperand(2).getImm() == 0) {
958 FrameIndex = MI->getOperand(1).getIndex();
959 return MI->getOperand(0).getReg();
962 case ARM::VLD1q64Pseudo:
963 if (MI->getOperand(1).isFI() &&
964 MI->getOperand(0).getSubReg() == 0) {
965 FrameIndex = MI->getOperand(1).getIndex();
966 return MI->getOperand(0).getReg();
970 if (MI->getOperand(1).isFI() &&
971 MI->getOperand(0).getSubReg() == 0) {
972 FrameIndex = MI->getOperand(1).getIndex();
973 return MI->getOperand(0).getReg();
981 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
982 int &FrameIndex) const {
983 const MachineMemOperand *Dummy;
984 return MI->getDesc().mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
988 ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
989 int FrameIx, uint64_t Offset,
992 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
993 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
997 /// Create a copy of a const pool value. Update CPI to the new index and return
999 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1000 MachineConstantPool *MCP = MF.getConstantPool();
1001 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1003 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1004 assert(MCPE.isMachineConstantPoolEntry() &&
1005 "Expecting a machine constantpool entry!");
1006 ARMConstantPoolValue *ACPV =
1007 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1009 unsigned PCLabelId = AFI->createPICLabelUId();
1010 ARMConstantPoolValue *NewCPV = 0;
1011 // FIXME: The below assumes PIC relocation model and that the function
1012 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1013 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1014 // instructions, so that's probably OK, but is PIC always correct when
1016 if (ACPV->isGlobalValue())
1017 NewCPV = new ARMConstantPoolValue(ACPV->getGV(), PCLabelId,
1019 else if (ACPV->isExtSymbol())
1020 NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(),
1021 ACPV->getSymbol(), PCLabelId, 4);
1022 else if (ACPV->isBlockAddress())
1023 NewCPV = new ARMConstantPoolValue(ACPV->getBlockAddress(), PCLabelId,
1024 ARMCP::CPBlockAddress, 4);
1025 else if (ACPV->isLSDA())
1026 NewCPV = new ARMConstantPoolValue(MF.getFunction(), PCLabelId,
1029 llvm_unreachable("Unexpected ARM constantpool value type!!");
1030 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1034 void ARMBaseInstrInfo::
1035 reMaterialize(MachineBasicBlock &MBB,
1036 MachineBasicBlock::iterator I,
1037 unsigned DestReg, unsigned SubIdx,
1038 const MachineInstr *Orig,
1039 const TargetRegisterInfo &TRI) const {
1040 unsigned Opcode = Orig->getOpcode();
1043 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1044 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1048 case ARM::tLDRpci_pic:
1049 case ARM::t2LDRpci_pic: {
1050 MachineFunction &MF = *MBB.getParent();
1051 unsigned CPI = Orig->getOperand(1).getIndex();
1052 unsigned PCLabelId = duplicateCPV(MF, CPI);
1053 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1055 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1056 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1063 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1064 MachineInstr *MI = TargetInstrInfoImpl::duplicate(Orig, MF);
1065 switch(Orig->getOpcode()) {
1066 case ARM::tLDRpci_pic:
1067 case ARM::t2LDRpci_pic: {
1068 unsigned CPI = Orig->getOperand(1).getIndex();
1069 unsigned PCLabelId = duplicateCPV(MF, CPI);
1070 Orig->getOperand(1).setIndex(CPI);
1071 Orig->getOperand(2).setImm(PCLabelId);
1078 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1079 const MachineInstr *MI1,
1080 const MachineRegisterInfo *MRI) const {
1081 int Opcode = MI0->getOpcode();
1082 if (Opcode == ARM::t2LDRpci ||
1083 Opcode == ARM::t2LDRpci_pic ||
1084 Opcode == ARM::tLDRpci ||
1085 Opcode == ARM::tLDRpci_pic ||
1086 Opcode == ARM::MOV_ga_dyn ||
1087 Opcode == ARM::MOV_ga_pcrel ||
1088 Opcode == ARM::MOV_ga_pcrel_ldr ||
1089 Opcode == ARM::t2MOV_ga_dyn ||
1090 Opcode == ARM::t2MOV_ga_pcrel) {
1091 if (MI1->getOpcode() != Opcode)
1093 if (MI0->getNumOperands() != MI1->getNumOperands())
1096 const MachineOperand &MO0 = MI0->getOperand(1);
1097 const MachineOperand &MO1 = MI1->getOperand(1);
1098 if (MO0.getOffset() != MO1.getOffset())
1101 if (Opcode == ARM::MOV_ga_dyn ||
1102 Opcode == ARM::MOV_ga_pcrel ||
1103 Opcode == ARM::MOV_ga_pcrel_ldr ||
1104 Opcode == ARM::t2MOV_ga_dyn ||
1105 Opcode == ARM::t2MOV_ga_pcrel)
1106 // Ignore the PC labels.
1107 return MO0.getGlobal() == MO1.getGlobal();
1109 const MachineFunction *MF = MI0->getParent()->getParent();
1110 const MachineConstantPool *MCP = MF->getConstantPool();
1111 int CPI0 = MO0.getIndex();
1112 int CPI1 = MO1.getIndex();
1113 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1114 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1115 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1116 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1117 if (isARMCP0 && isARMCP1) {
1118 ARMConstantPoolValue *ACPV0 =
1119 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1120 ARMConstantPoolValue *ACPV1 =
1121 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1122 return ACPV0->hasSameValue(ACPV1);
1123 } else if (!isARMCP0 && !isARMCP1) {
1124 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1127 } else if (Opcode == ARM::PICLDR) {
1128 if (MI1->getOpcode() != Opcode)
1130 if (MI0->getNumOperands() != MI1->getNumOperands())
1133 unsigned Addr0 = MI0->getOperand(1).getReg();
1134 unsigned Addr1 = MI1->getOperand(1).getReg();
1135 if (Addr0 != Addr1) {
1137 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1138 !TargetRegisterInfo::isVirtualRegister(Addr1))
1141 // This assumes SSA form.
1142 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1143 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1144 // Check if the loaded value, e.g. a constantpool of a global address, are
1146 if (!produceSameValue(Def0, Def1, MRI))
1150 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1151 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1152 const MachineOperand &MO0 = MI0->getOperand(i);
1153 const MachineOperand &MO1 = MI1->getOperand(i);
1154 if (!MO0.isIdenticalTo(MO1))
1160 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1163 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1164 /// determine if two loads are loading from the same base address. It should
1165 /// only return true if the base pointers are the same and the only differences
1166 /// between the two addresses is the offset. It also returns the offsets by
1168 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1170 int64_t &Offset2) const {
1171 // Don't worry about Thumb: just ARM and Thumb2.
1172 if (Subtarget.isThumb1Only()) return false;
1174 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1177 switch (Load1->getMachineOpcode()) {
1190 case ARM::t2LDRSHi8:
1192 case ARM::t2LDRSHi12:
1196 switch (Load2->getMachineOpcode()) {
1209 case ARM::t2LDRSHi8:
1211 case ARM::t2LDRSHi12:
1215 // Check if base addresses and chain operands match.
1216 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1217 Load1->getOperand(4) != Load2->getOperand(4))
1220 // Index should be Reg0.
1221 if (Load1->getOperand(3) != Load2->getOperand(3))
1224 // Determine the offsets.
1225 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1226 isa<ConstantSDNode>(Load2->getOperand(1))) {
1227 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1228 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1235 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1236 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1237 /// be scheduled togther. On some targets if two loads are loading from
1238 /// addresses in the same cache line, it's better if they are scheduled
1239 /// together. This function takes two integers that represent the load offsets
1240 /// from the common base address. It returns true if it decides it's desirable
1241 /// to schedule the two loads together. "NumLoads" is the number of loads that
1242 /// have already been scheduled after Load1.
1243 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1244 int64_t Offset1, int64_t Offset2,
1245 unsigned NumLoads) const {
1246 // Don't worry about Thumb: just ARM and Thumb2.
1247 if (Subtarget.isThumb1Only()) return false;
1249 assert(Offset2 > Offset1);
1251 if ((Offset2 - Offset1) / 8 > 64)
1254 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1255 return false; // FIXME: overly conservative?
1257 // Four loads in a row should be sufficient.
1264 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1265 const MachineBasicBlock *MBB,
1266 const MachineFunction &MF) const {
1267 // Debug info is never a scheduling boundary. It's necessary to be explicit
1268 // due to the special treatment of IT instructions below, otherwise a
1269 // dbg_value followed by an IT will result in the IT instruction being
1270 // considered a scheduling hazard, which is wrong. It should be the actual
1271 // instruction preceding the dbg_value instruction(s), just like it is
1272 // when debug info is not present.
1273 if (MI->isDebugValue())
1276 // Terminators and labels can't be scheduled around.
1277 if (MI->getDesc().isTerminator() || MI->isLabel())
1280 // Treat the start of the IT block as a scheduling boundary, but schedule
1281 // t2IT along with all instructions following it.
1282 // FIXME: This is a big hammer. But the alternative is to add all potential
1283 // true and anti dependencies to IT block instructions as implicit operands
1284 // to the t2IT instruction. The added compile time and complexity does not
1286 MachineBasicBlock::const_iterator I = MI;
1287 // Make sure to skip any dbg_value instructions
1288 while (++I != MBB->end() && I->isDebugValue())
1290 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1293 // Don't attempt to schedule around any instruction that defines
1294 // a stack-oriented pointer, as it's unlikely to be profitable. This
1295 // saves compile time, because it doesn't require every single
1296 // stack slot reference to depend on the instruction that does the
1298 if (MI->definesRegister(ARM::SP))
1304 bool ARMBaseInstrInfo::
1305 isProfitableToIfCvt(MachineBasicBlock &MBB,
1306 unsigned NumCycles, unsigned ExtraPredCycles,
1307 const BranchProbability &Probability) const {
1311 // Attempt to estimate the relative costs of predication versus branching.
1312 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1313 UnpredCost /= Probability.getDenominator();
1314 UnpredCost += 1; // The branch itself
1315 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1317 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1320 bool ARMBaseInstrInfo::
1321 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1322 unsigned TCycles, unsigned TExtra,
1323 MachineBasicBlock &FMBB,
1324 unsigned FCycles, unsigned FExtra,
1325 const BranchProbability &Probability) const {
1326 if (!TCycles || !FCycles)
1329 // Attempt to estimate the relative costs of predication versus branching.
1330 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1331 TUnpredCost /= Probability.getDenominator();
1333 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1334 unsigned FUnpredCost = Comp * FCycles;
1335 FUnpredCost /= Probability.getDenominator();
1337 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1338 UnpredCost += 1; // The branch itself
1339 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1341 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1344 /// getInstrPredicate - If instruction is predicated, returns its predicate
1345 /// condition, otherwise returns AL. It also returns the condition code
1346 /// register by reference.
1348 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1349 int PIdx = MI->findFirstPredOperandIdx();
1355 PredReg = MI->getOperand(PIdx+1).getReg();
1356 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1360 int llvm::getMatchingCondBranchOpcode(int Opc) {
1363 else if (Opc == ARM::tB)
1365 else if (Opc == ARM::t2B)
1368 llvm_unreachable("Unknown unconditional branch opcode!");
1373 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1374 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1375 unsigned DestReg, unsigned BaseReg, int NumBytes,
1376 ARMCC::CondCodes Pred, unsigned PredReg,
1377 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1378 bool isSub = NumBytes < 0;
1379 if (isSub) NumBytes = -NumBytes;
1382 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1383 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1384 assert(ThisVal && "Didn't extract field correctly");
1386 // We will handle these bits from offset, clear them.
1387 NumBytes &= ~ThisVal;
1389 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1391 // Build the new ADD / SUB.
1392 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1393 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1394 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1395 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1396 .setMIFlags(MIFlags);
1401 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1402 unsigned FrameReg, int &Offset,
1403 const ARMBaseInstrInfo &TII) {
1404 unsigned Opcode = MI.getOpcode();
1405 const MCInstrDesc &Desc = MI.getDesc();
1406 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1409 // Memory operands in inline assembly always use AddrMode2.
1410 if (Opcode == ARM::INLINEASM)
1411 AddrMode = ARMII::AddrMode2;
1413 if (Opcode == ARM::ADDri) {
1414 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1416 // Turn it into a move.
1417 MI.setDesc(TII.get(ARM::MOVr));
1418 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1419 MI.RemoveOperand(FrameRegIdx+1);
1422 } else if (Offset < 0) {
1425 MI.setDesc(TII.get(ARM::SUBri));
1428 // Common case: small offset, fits into instruction.
1429 if (ARM_AM::getSOImmVal(Offset) != -1) {
1430 // Replace the FrameIndex with sp / fp
1431 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1432 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1437 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1439 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1440 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1442 // We will handle these bits from offset, clear them.
1443 Offset &= ~ThisImmVal;
1445 // Get the properly encoded SOImmVal field.
1446 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1447 "Bit extraction didn't work?");
1448 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1450 unsigned ImmIdx = 0;
1452 unsigned NumBits = 0;
1455 case ARMII::AddrMode_i12: {
1456 ImmIdx = FrameRegIdx + 1;
1457 InstrOffs = MI.getOperand(ImmIdx).getImm();
1461 case ARMII::AddrMode2: {
1462 ImmIdx = FrameRegIdx+2;
1463 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1464 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1469 case ARMII::AddrMode3: {
1470 ImmIdx = FrameRegIdx+2;
1471 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1472 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1477 case ARMII::AddrMode4:
1478 case ARMII::AddrMode6:
1479 // Can't fold any offset even if it's zero.
1481 case ARMII::AddrMode5: {
1482 ImmIdx = FrameRegIdx+1;
1483 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1484 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1491 llvm_unreachable("Unsupported addressing mode!");
1495 Offset += InstrOffs * Scale;
1496 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1502 // Attempt to fold address comp. if opcode has offset bits
1504 // Common case: small offset, fits into instruction.
1505 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1506 int ImmedOffset = Offset / Scale;
1507 unsigned Mask = (1 << NumBits) - 1;
1508 if ((unsigned)Offset <= Mask * Scale) {
1509 // Replace the FrameIndex with sp
1510 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1511 // FIXME: When addrmode2 goes away, this will simplify (like the
1512 // T2 version), as the LDR.i12 versions don't need the encoding
1513 // tricks for the offset value.
1515 if (AddrMode == ARMII::AddrMode_i12)
1516 ImmedOffset = -ImmedOffset;
1518 ImmedOffset |= 1 << NumBits;
1520 ImmOp.ChangeToImmediate(ImmedOffset);
1525 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1526 ImmedOffset = ImmedOffset & Mask;
1528 if (AddrMode == ARMII::AddrMode_i12)
1529 ImmedOffset = -ImmedOffset;
1531 ImmedOffset |= 1 << NumBits;
1533 ImmOp.ChangeToImmediate(ImmedOffset);
1534 Offset &= ~(Mask*Scale);
1538 Offset = (isSub) ? -Offset : Offset;
1542 bool ARMBaseInstrInfo::
1543 AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpMask,
1544 int &CmpValue) const {
1545 switch (MI->getOpcode()) {
1549 SrcReg = MI->getOperand(0).getReg();
1551 CmpValue = MI->getOperand(1).getImm();
1555 SrcReg = MI->getOperand(0).getReg();
1556 CmpMask = MI->getOperand(1).getImm();
1564 /// isSuitableForMask - Identify a suitable 'and' instruction that
1565 /// operates on the given source register and applies the same mask
1566 /// as a 'tst' instruction. Provide a limited look-through for copies.
1567 /// When successful, MI will hold the found instruction.
1568 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
1569 int CmpMask, bool CommonUse) {
1570 switch (MI->getOpcode()) {
1573 if (CmpMask != MI->getOperand(2).getImm())
1575 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
1579 // Walk down one instruction which is potentially an 'and'.
1580 const MachineInstr &Copy = *MI;
1581 MachineBasicBlock::iterator AND(
1582 llvm::next(MachineBasicBlock::iterator(MI)));
1583 if (AND == MI->getParent()->end()) return false;
1585 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
1593 /// OptimizeCompareInstr - Convert the instruction supplying the argument to the
1594 /// comparison into one that sets the zero bit in the flags register.
1595 bool ARMBaseInstrInfo::
1596 OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
1597 int CmpValue, const MachineRegisterInfo *MRI) const {
1601 MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
1602 if (llvm::next(DI) != MRI->def_end())
1603 // Only support one definition.
1606 MachineInstr *MI = &*DI;
1608 // Masked compares sometimes use the same register as the corresponding 'and'.
1609 if (CmpMask != ~0) {
1610 if (!isSuitableForMask(MI, SrcReg, CmpMask, false)) {
1612 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
1613 UE = MRI->use_end(); UI != UE; ++UI) {
1614 if (UI->getParent() != CmpInstr->getParent()) continue;
1615 MachineInstr *PotentialAND = &*UI;
1616 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true))
1621 if (!MI) return false;
1625 // Conservatively refuse to convert an instruction which isn't in the same BB
1626 // as the comparison.
1627 if (MI->getParent() != CmpInstr->getParent())
1630 // Check that CPSR isn't set between the comparison instruction and the one we
1632 MachineBasicBlock::const_iterator I = CmpInstr, E = MI,
1633 B = MI->getParent()->begin();
1635 // Early exit if CmpInstr is at the beginning of the BB.
1636 if (I == B) return false;
1639 for (; I != E; --I) {
1640 const MachineInstr &Instr = *I;
1642 for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) {
1643 const MachineOperand &MO = Instr.getOperand(IO);
1644 if (!MO.isReg()) continue;
1646 // This instruction modifies or uses CPSR after the one we want to
1647 // change. We can't do this transformation.
1648 if (MO.getReg() == ARM::CPSR)
1653 // The 'and' is below the comparison instruction.
1657 // Set the "zero" bit in CPSR.
1658 switch (MI->getOpcode()) {
1692 case ARM::t2EORri: {
1693 // Scan forward for the use of CPSR, if it's a conditional code requires
1694 // checking of V bit, then this is not safe to do. If we can't find the
1695 // CPSR use (i.e. used in another block), then it's not safe to perform
1696 // the optimization.
1697 bool isSafe = false;
1699 E = MI->getParent()->end();
1700 while (!isSafe && ++I != E) {
1701 const MachineInstr &Instr = *I;
1702 for (unsigned IO = 0, EO = Instr.getNumOperands();
1703 !isSafe && IO != EO; ++IO) {
1704 const MachineOperand &MO = Instr.getOperand(IO);
1705 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
1711 // Condition code is after the operand before CPSR.
1712 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
1731 // Toggle the optional operand to CPSR.
1732 MI->getOperand(5).setReg(ARM::CPSR);
1733 MI->getOperand(5).setIsDef(true);
1734 CmpInstr->eraseFromParent();
1742 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
1743 MachineInstr *DefMI, unsigned Reg,
1744 MachineRegisterInfo *MRI) const {
1745 // Fold large immediates into add, sub, or, xor.
1746 unsigned DefOpc = DefMI->getOpcode();
1747 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
1749 if (!DefMI->getOperand(1).isImm())
1750 // Could be t2MOVi32imm <ga:xx>
1753 if (!MRI->hasOneNonDBGUse(Reg))
1756 unsigned UseOpc = UseMI->getOpcode();
1757 unsigned NewUseOpc = 0;
1758 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
1759 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
1760 bool Commute = false;
1762 default: return false;
1770 case ARM::t2EORrr: {
1771 Commute = UseMI->getOperand(2).getReg() != Reg;
1778 NewUseOpc = ARM::SUBri;
1784 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
1786 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
1787 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
1790 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
1791 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
1792 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
1796 case ARM::t2SUBrr: {
1800 NewUseOpc = ARM::t2SUBri;
1805 case ARM::t2EORrr: {
1806 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
1808 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
1809 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
1812 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
1813 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
1814 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
1822 unsigned OpIdx = Commute ? 2 : 1;
1823 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
1824 bool isKill = UseMI->getOperand(OpIdx).isKill();
1825 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
1826 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
1827 *UseMI, UseMI->getDebugLoc(),
1828 get(NewUseOpc), NewReg)
1829 .addReg(Reg1, getKillRegState(isKill))
1830 .addImm(SOImmValV1)));
1831 UseMI->setDesc(get(NewUseOpc));
1832 UseMI->getOperand(1).setReg(NewReg);
1833 UseMI->getOperand(1).setIsKill();
1834 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
1835 DefMI->eraseFromParent();
1840 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1841 const MachineInstr *MI) const {
1842 if (!ItinData || ItinData->isEmpty())
1845 const MCInstrDesc &Desc = MI->getDesc();
1846 unsigned Class = Desc.getSchedClass();
1847 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
1851 unsigned Opc = MI->getOpcode();
1854 llvm_unreachable("Unexpected multi-uops instruction!");
1860 // The number of uOps for load / store multiple are determined by the number
1863 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
1864 // same cycle. The scheduling for the first load / store must be done
1865 // separately by assuming the the address is not 64-bit aligned.
1867 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
1868 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
1869 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
1871 case ARM::VLDMDIA_UPD:
1872 case ARM::VLDMDDB_UPD:
1874 case ARM::VLDMSIA_UPD:
1875 case ARM::VLDMSDB_UPD:
1877 case ARM::VSTMDIA_UPD:
1878 case ARM::VSTMDDB_UPD:
1880 case ARM::VSTMSIA_UPD:
1881 case ARM::VSTMSDB_UPD: {
1882 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
1883 return (NumRegs / 2) + (NumRegs % 2) + 1;
1886 case ARM::LDMIA_RET:
1891 case ARM::LDMIA_UPD:
1892 case ARM::LDMDA_UPD:
1893 case ARM::LDMDB_UPD:
1894 case ARM::LDMIB_UPD:
1899 case ARM::STMIA_UPD:
1900 case ARM::STMDA_UPD:
1901 case ARM::STMDB_UPD:
1902 case ARM::STMIB_UPD:
1904 case ARM::tLDMIA_UPD:
1906 case ARM::tSTMIA_UPD:
1910 case ARM::t2LDMIA_RET:
1913 case ARM::t2LDMIA_UPD:
1914 case ARM::t2LDMDB_UPD:
1917 case ARM::t2STMIA_UPD:
1918 case ARM::t2STMDB_UPD: {
1919 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
1920 if (Subtarget.isCortexA8()) {
1923 // 4 registers would be issued: 2, 2.
1924 // 5 registers would be issued: 2, 2, 1.
1925 UOps = (NumRegs / 2);
1929 } else if (Subtarget.isCortexA9()) {
1930 UOps = (NumRegs / 2);
1931 // If there are odd number of registers or if it's not 64-bit aligned,
1932 // then it takes an extra AGU (Address Generation Unit) cycle.
1933 if ((NumRegs % 2) ||
1934 !MI->hasOneMemOperand() ||
1935 (*MI->memoperands_begin())->getAlignment() < 8)
1939 // Assume the worst.
1947 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
1948 const MCInstrDesc &DefMCID,
1950 unsigned DefIdx, unsigned DefAlign) const {
1951 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1953 // Def is the address writeback.
1954 return ItinData->getOperandCycle(DefClass, DefIdx);
1957 if (Subtarget.isCortexA8()) {
1958 // (regno / 2) + (regno % 2) + 1
1959 DefCycle = RegNo / 2 + 1;
1962 } else if (Subtarget.isCortexA9()) {
1964 bool isSLoad = false;
1966 switch (DefMCID.getOpcode()) {
1969 case ARM::VLDMSIA_UPD:
1970 case ARM::VLDMSDB_UPD:
1975 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
1976 // then it takes an extra cycle.
1977 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
1980 // Assume the worst.
1981 DefCycle = RegNo + 2;
1988 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
1989 const MCInstrDesc &DefMCID,
1991 unsigned DefIdx, unsigned DefAlign) const {
1992 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1994 // Def is the address writeback.
1995 return ItinData->getOperandCycle(DefClass, DefIdx);
1998 if (Subtarget.isCortexA8()) {
1999 // 4 registers would be issued: 1, 2, 1.
2000 // 5 registers would be issued: 1, 2, 2.
2001 DefCycle = RegNo / 2;
2004 // Result latency is issue cycle + 2: E2.
2006 } else if (Subtarget.isCortexA9()) {
2007 DefCycle = (RegNo / 2);
2008 // If there are odd number of registers or if it's not 64-bit aligned,
2009 // then it takes an extra AGU (Address Generation Unit) cycle.
2010 if ((RegNo % 2) || DefAlign < 8)
2012 // Result latency is AGU cycles + 2.
2015 // Assume the worst.
2016 DefCycle = RegNo + 2;
2023 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
2024 const MCInstrDesc &UseMCID,
2026 unsigned UseIdx, unsigned UseAlign) const {
2027 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2029 return ItinData->getOperandCycle(UseClass, UseIdx);
2032 if (Subtarget.isCortexA8()) {
2033 // (regno / 2) + (regno % 2) + 1
2034 UseCycle = RegNo / 2 + 1;
2037 } else if (Subtarget.isCortexA9()) {
2039 bool isSStore = false;
2041 switch (UseMCID.getOpcode()) {
2044 case ARM::VSTMSIA_UPD:
2045 case ARM::VSTMSDB_UPD:
2050 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2051 // then it takes an extra cycle.
2052 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2055 // Assume the worst.
2056 UseCycle = RegNo + 2;
2063 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2064 const MCInstrDesc &UseMCID,
2066 unsigned UseIdx, unsigned UseAlign) const {
2067 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2069 return ItinData->getOperandCycle(UseClass, UseIdx);
2072 if (Subtarget.isCortexA8()) {
2073 UseCycle = RegNo / 2;
2078 } else if (Subtarget.isCortexA9()) {
2079 UseCycle = (RegNo / 2);
2080 // If there are odd number of registers or if it's not 64-bit aligned,
2081 // then it takes an extra AGU (Address Generation Unit) cycle.
2082 if ((RegNo % 2) || UseAlign < 8)
2085 // Assume the worst.
2092 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2093 const MCInstrDesc &DefMCID,
2094 unsigned DefIdx, unsigned DefAlign,
2095 const MCInstrDesc &UseMCID,
2096 unsigned UseIdx, unsigned UseAlign) const {
2097 unsigned DefClass = DefMCID.getSchedClass();
2098 unsigned UseClass = UseMCID.getSchedClass();
2100 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2101 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2103 // This may be a def / use of a variable_ops instruction, the operand
2104 // latency might be determinable dynamically. Let the target try to
2107 bool LdmBypass = false;
2108 switch (DefMCID.getOpcode()) {
2110 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2114 case ARM::VLDMDIA_UPD:
2115 case ARM::VLDMDDB_UPD:
2117 case ARM::VLDMSIA_UPD:
2118 case ARM::VLDMSDB_UPD:
2119 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2122 case ARM::LDMIA_RET:
2127 case ARM::LDMIA_UPD:
2128 case ARM::LDMDA_UPD:
2129 case ARM::LDMDB_UPD:
2130 case ARM::LDMIB_UPD:
2132 case ARM::tLDMIA_UPD:
2134 case ARM::t2LDMIA_RET:
2137 case ARM::t2LDMIA_UPD:
2138 case ARM::t2LDMDB_UPD:
2140 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2145 // We can't seem to determine the result latency of the def, assume it's 2.
2149 switch (UseMCID.getOpcode()) {
2151 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
2155 case ARM::VSTMDIA_UPD:
2156 case ARM::VSTMDDB_UPD:
2158 case ARM::VSTMSIA_UPD:
2159 case ARM::VSTMSDB_UPD:
2160 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2167 case ARM::STMIA_UPD:
2168 case ARM::STMDA_UPD:
2169 case ARM::STMDB_UPD:
2170 case ARM::STMIB_UPD:
2172 case ARM::tSTMIA_UPD:
2177 case ARM::t2STMIA_UPD:
2178 case ARM::t2STMDB_UPD:
2179 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2184 // Assume it's read in the first stage.
2187 UseCycle = DefCycle - UseCycle + 1;
2190 // It's a variable_ops instruction so we can't use DefIdx here. Just use
2191 // first def operand.
2192 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
2195 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
2196 UseClass, UseIdx)) {
2205 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2206 const MachineInstr *DefMI, unsigned DefIdx,
2207 const MachineInstr *UseMI, unsigned UseIdx) const {
2208 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
2209 DefMI->isRegSequence() || DefMI->isImplicitDef())
2212 const MCInstrDesc &DefMCID = DefMI->getDesc();
2213 if (!ItinData || ItinData->isEmpty())
2214 return DefMCID.mayLoad() ? 3 : 1;
2216 const MCInstrDesc &UseMCID = UseMI->getDesc();
2217 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
2218 if (DefMO.getReg() == ARM::CPSR) {
2219 if (DefMI->getOpcode() == ARM::FMSTAT) {
2220 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
2221 return Subtarget.isCortexA9() ? 1 : 20;
2224 // CPSR set and branch can be paired in the same cycle.
2225 if (UseMCID.isBranch())
2229 unsigned DefAlign = DefMI->hasOneMemOperand()
2230 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
2231 unsigned UseAlign = UseMI->hasOneMemOperand()
2232 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
2233 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2234 UseMCID, UseIdx, UseAlign);
2237 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2238 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2239 // variants are one cycle cheaper.
2240 switch (DefMCID.getOpcode()) {
2244 unsigned ShOpVal = DefMI->getOperand(3).getImm();
2245 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2247 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2254 case ARM::t2LDRSHs: {
2255 // Thumb2 mode: lsl only.
2256 unsigned ShAmt = DefMI->getOperand(3).getImm();
2257 if (ShAmt == 0 || ShAmt == 2)
2264 if (DefAlign < 8 && Subtarget.isCortexA9())
2265 switch (DefMCID.getOpcode()) {
2271 case ARM::VLD1q8_UPD:
2272 case ARM::VLD1q16_UPD:
2273 case ARM::VLD1q32_UPD:
2274 case ARM::VLD1q64_UPD:
2281 case ARM::VLD2d8_UPD:
2282 case ARM::VLD2d16_UPD:
2283 case ARM::VLD2d32_UPD:
2284 case ARM::VLD2q8_UPD:
2285 case ARM::VLD2q16_UPD:
2286 case ARM::VLD2q32_UPD:
2291 case ARM::VLD3d8_UPD:
2292 case ARM::VLD3d16_UPD:
2293 case ARM::VLD3d32_UPD:
2294 case ARM::VLD1d64T_UPD:
2295 case ARM::VLD3q8_UPD:
2296 case ARM::VLD3q16_UPD:
2297 case ARM::VLD3q32_UPD:
2302 case ARM::VLD4d8_UPD:
2303 case ARM::VLD4d16_UPD:
2304 case ARM::VLD4d32_UPD:
2305 case ARM::VLD1d64Q_UPD:
2306 case ARM::VLD4q8_UPD:
2307 case ARM::VLD4q16_UPD:
2308 case ARM::VLD4q32_UPD:
2309 case ARM::VLD1DUPq8:
2310 case ARM::VLD1DUPq16:
2311 case ARM::VLD1DUPq32:
2312 case ARM::VLD1DUPq8_UPD:
2313 case ARM::VLD1DUPq16_UPD:
2314 case ARM::VLD1DUPq32_UPD:
2315 case ARM::VLD2DUPd8:
2316 case ARM::VLD2DUPd16:
2317 case ARM::VLD2DUPd32:
2318 case ARM::VLD2DUPd8_UPD:
2319 case ARM::VLD2DUPd16_UPD:
2320 case ARM::VLD2DUPd32_UPD:
2321 case ARM::VLD4DUPd8:
2322 case ARM::VLD4DUPd16:
2323 case ARM::VLD4DUPd32:
2324 case ARM::VLD4DUPd8_UPD:
2325 case ARM::VLD4DUPd16_UPD:
2326 case ARM::VLD4DUPd32_UPD:
2328 case ARM::VLD1LNd16:
2329 case ARM::VLD1LNd32:
2330 case ARM::VLD1LNd8_UPD:
2331 case ARM::VLD1LNd16_UPD:
2332 case ARM::VLD1LNd32_UPD:
2334 case ARM::VLD2LNd16:
2335 case ARM::VLD2LNd32:
2336 case ARM::VLD2LNq16:
2337 case ARM::VLD2LNq32:
2338 case ARM::VLD2LNd8_UPD:
2339 case ARM::VLD2LNd16_UPD:
2340 case ARM::VLD2LNd32_UPD:
2341 case ARM::VLD2LNq16_UPD:
2342 case ARM::VLD2LNq32_UPD:
2344 case ARM::VLD4LNd16:
2345 case ARM::VLD4LNd32:
2346 case ARM::VLD4LNq16:
2347 case ARM::VLD4LNq32:
2348 case ARM::VLD4LNd8_UPD:
2349 case ARM::VLD4LNd16_UPD:
2350 case ARM::VLD4LNd32_UPD:
2351 case ARM::VLD4LNq16_UPD:
2352 case ARM::VLD4LNq32_UPD:
2353 // If the address is not 64-bit aligned, the latencies of these
2354 // instructions increases by one.
2363 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2364 SDNode *DefNode, unsigned DefIdx,
2365 SDNode *UseNode, unsigned UseIdx) const {
2366 if (!DefNode->isMachineOpcode())
2369 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
2371 if (isZeroCost(DefMCID.Opcode))
2374 if (!ItinData || ItinData->isEmpty())
2375 return DefMCID.mayLoad() ? 3 : 1;
2377 if (!UseNode->isMachineOpcode()) {
2378 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
2379 if (Subtarget.isCortexA9())
2380 return Latency <= 2 ? 1 : Latency - 1;
2382 return Latency <= 3 ? 1 : Latency - 2;
2385 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
2386 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
2387 unsigned DefAlign = !DefMN->memoperands_empty()
2388 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
2389 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
2390 unsigned UseAlign = !UseMN->memoperands_empty()
2391 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
2392 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2393 UseMCID, UseIdx, UseAlign);
2396 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2397 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2398 // variants are one cycle cheaper.
2399 switch (DefMCID.getOpcode()) {
2404 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2405 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2407 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2414 case ARM::t2LDRSHs: {
2415 // Thumb2 mode: lsl only.
2417 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2418 if (ShAmt == 0 || ShAmt == 2)
2425 if (DefAlign < 8 && Subtarget.isCortexA9())
2426 switch (DefMCID.getOpcode()) {
2428 case ARM::VLD1q8Pseudo:
2429 case ARM::VLD1q16Pseudo:
2430 case ARM::VLD1q32Pseudo:
2431 case ARM::VLD1q64Pseudo:
2432 case ARM::VLD1q8Pseudo_UPD:
2433 case ARM::VLD1q16Pseudo_UPD:
2434 case ARM::VLD1q32Pseudo_UPD:
2435 case ARM::VLD1q64Pseudo_UPD:
2436 case ARM::VLD2d8Pseudo:
2437 case ARM::VLD2d16Pseudo:
2438 case ARM::VLD2d32Pseudo:
2439 case ARM::VLD2q8Pseudo:
2440 case ARM::VLD2q16Pseudo:
2441 case ARM::VLD2q32Pseudo:
2442 case ARM::VLD2d8Pseudo_UPD:
2443 case ARM::VLD2d16Pseudo_UPD:
2444 case ARM::VLD2d32Pseudo_UPD:
2445 case ARM::VLD2q8Pseudo_UPD:
2446 case ARM::VLD2q16Pseudo_UPD:
2447 case ARM::VLD2q32Pseudo_UPD:
2448 case ARM::VLD3d8Pseudo:
2449 case ARM::VLD3d16Pseudo:
2450 case ARM::VLD3d32Pseudo:
2451 case ARM::VLD1d64TPseudo:
2452 case ARM::VLD3d8Pseudo_UPD:
2453 case ARM::VLD3d16Pseudo_UPD:
2454 case ARM::VLD3d32Pseudo_UPD:
2455 case ARM::VLD1d64TPseudo_UPD:
2456 case ARM::VLD3q8Pseudo_UPD:
2457 case ARM::VLD3q16Pseudo_UPD:
2458 case ARM::VLD3q32Pseudo_UPD:
2459 case ARM::VLD3q8oddPseudo:
2460 case ARM::VLD3q16oddPseudo:
2461 case ARM::VLD3q32oddPseudo:
2462 case ARM::VLD3q8oddPseudo_UPD:
2463 case ARM::VLD3q16oddPseudo_UPD:
2464 case ARM::VLD3q32oddPseudo_UPD:
2465 case ARM::VLD4d8Pseudo:
2466 case ARM::VLD4d16Pseudo:
2467 case ARM::VLD4d32Pseudo:
2468 case ARM::VLD1d64QPseudo:
2469 case ARM::VLD4d8Pseudo_UPD:
2470 case ARM::VLD4d16Pseudo_UPD:
2471 case ARM::VLD4d32Pseudo_UPD:
2472 case ARM::VLD1d64QPseudo_UPD:
2473 case ARM::VLD4q8Pseudo_UPD:
2474 case ARM::VLD4q16Pseudo_UPD:
2475 case ARM::VLD4q32Pseudo_UPD:
2476 case ARM::VLD4q8oddPseudo:
2477 case ARM::VLD4q16oddPseudo:
2478 case ARM::VLD4q32oddPseudo:
2479 case ARM::VLD4q8oddPseudo_UPD:
2480 case ARM::VLD4q16oddPseudo_UPD:
2481 case ARM::VLD4q32oddPseudo_UPD:
2482 case ARM::VLD1DUPq8Pseudo:
2483 case ARM::VLD1DUPq16Pseudo:
2484 case ARM::VLD1DUPq32Pseudo:
2485 case ARM::VLD1DUPq8Pseudo_UPD:
2486 case ARM::VLD1DUPq16Pseudo_UPD:
2487 case ARM::VLD1DUPq32Pseudo_UPD:
2488 case ARM::VLD2DUPd8Pseudo:
2489 case ARM::VLD2DUPd16Pseudo:
2490 case ARM::VLD2DUPd32Pseudo:
2491 case ARM::VLD2DUPd8Pseudo_UPD:
2492 case ARM::VLD2DUPd16Pseudo_UPD:
2493 case ARM::VLD2DUPd32Pseudo_UPD:
2494 case ARM::VLD4DUPd8Pseudo:
2495 case ARM::VLD4DUPd16Pseudo:
2496 case ARM::VLD4DUPd32Pseudo:
2497 case ARM::VLD4DUPd8Pseudo_UPD:
2498 case ARM::VLD4DUPd16Pseudo_UPD:
2499 case ARM::VLD4DUPd32Pseudo_UPD:
2500 case ARM::VLD1LNq8Pseudo:
2501 case ARM::VLD1LNq16Pseudo:
2502 case ARM::VLD1LNq32Pseudo:
2503 case ARM::VLD1LNq8Pseudo_UPD:
2504 case ARM::VLD1LNq16Pseudo_UPD:
2505 case ARM::VLD1LNq32Pseudo_UPD:
2506 case ARM::VLD2LNd8Pseudo:
2507 case ARM::VLD2LNd16Pseudo:
2508 case ARM::VLD2LNd32Pseudo:
2509 case ARM::VLD2LNq16Pseudo:
2510 case ARM::VLD2LNq32Pseudo:
2511 case ARM::VLD2LNd8Pseudo_UPD:
2512 case ARM::VLD2LNd16Pseudo_UPD:
2513 case ARM::VLD2LNd32Pseudo_UPD:
2514 case ARM::VLD2LNq16Pseudo_UPD:
2515 case ARM::VLD2LNq32Pseudo_UPD:
2516 case ARM::VLD4LNd8Pseudo:
2517 case ARM::VLD4LNd16Pseudo:
2518 case ARM::VLD4LNd32Pseudo:
2519 case ARM::VLD4LNq16Pseudo:
2520 case ARM::VLD4LNq32Pseudo:
2521 case ARM::VLD4LNd8Pseudo_UPD:
2522 case ARM::VLD4LNd16Pseudo_UPD:
2523 case ARM::VLD4LNd32Pseudo_UPD:
2524 case ARM::VLD4LNq16Pseudo_UPD:
2525 case ARM::VLD4LNq32Pseudo_UPD:
2526 // If the address is not 64-bit aligned, the latencies of these
2527 // instructions increases by one.
2535 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2536 const MachineInstr *MI,
2537 unsigned *PredCost) const {
2538 if (MI->isCopyLike() || MI->isInsertSubreg() ||
2539 MI->isRegSequence() || MI->isImplicitDef())
2542 if (!ItinData || ItinData->isEmpty())
2545 const MCInstrDesc &MCID = MI->getDesc();
2546 unsigned Class = MCID.getSchedClass();
2547 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
2548 if (PredCost && MCID.hasImplicitDefOfPhysReg(ARM::CPSR))
2549 // When predicated, CPSR is an additional source operand for CPSR updating
2550 // instructions, this apparently increases their latencies.
2553 return ItinData->getStageLatency(Class);
2554 return getNumMicroOps(ItinData, MI);
2557 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2558 SDNode *Node) const {
2559 if (!Node->isMachineOpcode())
2562 if (!ItinData || ItinData->isEmpty())
2565 unsigned Opcode = Node->getMachineOpcode();
2568 return ItinData->getStageLatency(get(Opcode).getSchedClass());
2575 bool ARMBaseInstrInfo::
2576 hasHighOperandLatency(const InstrItineraryData *ItinData,
2577 const MachineRegisterInfo *MRI,
2578 const MachineInstr *DefMI, unsigned DefIdx,
2579 const MachineInstr *UseMI, unsigned UseIdx) const {
2580 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2581 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
2582 if (Subtarget.isCortexA8() &&
2583 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
2584 // CortexA8 VFP instructions are not pipelined.
2587 // Hoist VFP / NEON instructions with 4 or higher latency.
2588 int Latency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
2591 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
2592 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
2595 bool ARMBaseInstrInfo::
2596 hasLowDefLatency(const InstrItineraryData *ItinData,
2597 const MachineInstr *DefMI, unsigned DefIdx) const {
2598 if (!ItinData || ItinData->isEmpty())
2601 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2602 if (DDomain == ARMII::DomainGeneral) {
2603 unsigned DefClass = DefMI->getDesc().getSchedClass();
2604 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2605 return (DefCycle != -1 && DefCycle <= 2);
2611 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
2612 unsigned &AddSubOpc,
2613 bool &NegAcc, bool &HasLane) const {
2614 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
2615 if (I == MLxEntryMap.end())
2618 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
2619 MulOpc = Entry.MulOpc;
2620 AddSubOpc = Entry.AddSubOpc;
2621 NegAcc = Entry.NegAcc;
2622 HasLane = Entry.HasLane;