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 "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMHazardRecognizer.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMRegisterInfo.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/MC/MCAsmInfo.h"
33 #include "llvm/Support/BranchProbability.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/ADT/STLExtras.h"
39 #define GET_INSTRINFO_CTOR
40 #include "ARMGenInstrInfo.inc"
45 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
46 cl::desc("Enable ARM 2-addr to 3-addr conv"));
48 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
50 unsigned MLxOpc; // MLA / MLS opcode
51 unsigned MulOpc; // Expanded multiplication opcode
52 unsigned AddSubOpc; // Expanded add / sub opcode
53 bool NegAcc; // True if the acc is negated before the add / sub.
54 bool HasLane; // True if instruction has an extra "lane" operand.
57 static const ARM_MLxEntry ARM_MLxTable[] = {
58 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
60 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
61 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
62 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
63 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
64 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
65 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
66 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
67 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
70 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
71 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
72 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
73 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
74 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
75 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
76 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
77 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
80 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
81 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
83 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
84 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
85 assert(false && "Duplicated entries?");
86 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
87 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
91 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
92 // currently defaults to no prepass hazard recognizer.
93 ScheduleHazardRecognizer *ARMBaseInstrInfo::
94 CreateTargetHazardRecognizer(const TargetMachine *TM,
95 const ScheduleDAG *DAG) const {
96 if (usePreRAHazardRecognizer()) {
97 const InstrItineraryData *II = TM->getInstrItineraryData();
98 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
100 return TargetInstrInfoImpl::CreateTargetHazardRecognizer(TM, DAG);
103 ScheduleHazardRecognizer *ARMBaseInstrInfo::
104 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
105 const ScheduleDAG *DAG) const {
106 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
107 return (ScheduleHazardRecognizer *)
108 new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget, DAG);
109 return TargetInstrInfoImpl::CreateTargetPostRAHazardRecognizer(II, DAG);
113 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
114 MachineBasicBlock::iterator &MBBI,
115 LiveVariables *LV) const {
116 // FIXME: Thumb2 support.
121 MachineInstr *MI = MBBI;
122 MachineFunction &MF = *MI->getParent()->getParent();
123 uint64_t TSFlags = MI->getDesc().TSFlags;
125 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
126 default: return NULL;
127 case ARMII::IndexModePre:
130 case ARMII::IndexModePost:
134 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
136 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
140 MachineInstr *UpdateMI = NULL;
141 MachineInstr *MemMI = NULL;
142 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
143 const MCInstrDesc &MCID = MI->getDesc();
144 unsigned NumOps = MCID.getNumOperands();
145 bool isLoad = !MCID.mayStore();
146 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
147 const MachineOperand &Base = MI->getOperand(2);
148 const MachineOperand &Offset = MI->getOperand(NumOps-3);
149 unsigned WBReg = WB.getReg();
150 unsigned BaseReg = Base.getReg();
151 unsigned OffReg = Offset.getReg();
152 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
153 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
156 assert(false && "Unknown indexed op!");
158 case ARMII::AddrMode2: {
159 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
160 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
162 if (ARM_AM::getSOImmVal(Amt) == -1)
163 // Can't encode it in a so_imm operand. This transformation will
164 // add more than 1 instruction. Abandon!
166 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
167 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
168 .addReg(BaseReg).addImm(Amt)
169 .addImm(Pred).addReg(0).addReg(0);
170 } else if (Amt != 0) {
171 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
172 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
173 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
174 get(isSub ? ARM::SUBrs : ARM::ADDrs), WBReg)
175 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
176 .addImm(Pred).addReg(0).addReg(0);
178 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
179 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
180 .addReg(BaseReg).addReg(OffReg)
181 .addImm(Pred).addReg(0).addReg(0);
184 case ARMII::AddrMode3 : {
185 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
186 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
188 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
189 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
190 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
191 .addReg(BaseReg).addImm(Amt)
192 .addImm(Pred).addReg(0).addReg(0);
194 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
195 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
196 .addReg(BaseReg).addReg(OffReg)
197 .addImm(Pred).addReg(0).addReg(0);
202 std::vector<MachineInstr*> NewMIs;
205 MemMI = BuildMI(MF, MI->getDebugLoc(),
206 get(MemOpc), MI->getOperand(0).getReg())
207 .addReg(WBReg).addImm(0).addImm(Pred);
209 MemMI = BuildMI(MF, MI->getDebugLoc(),
210 get(MemOpc)).addReg(MI->getOperand(1).getReg())
211 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
212 NewMIs.push_back(MemMI);
213 NewMIs.push_back(UpdateMI);
216 MemMI = BuildMI(MF, MI->getDebugLoc(),
217 get(MemOpc), MI->getOperand(0).getReg())
218 .addReg(BaseReg).addImm(0).addImm(Pred);
220 MemMI = BuildMI(MF, MI->getDebugLoc(),
221 get(MemOpc)).addReg(MI->getOperand(1).getReg())
222 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
224 UpdateMI->getOperand(0).setIsDead();
225 NewMIs.push_back(UpdateMI);
226 NewMIs.push_back(MemMI);
229 // Transfer LiveVariables states, kill / dead info.
231 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
232 MachineOperand &MO = MI->getOperand(i);
233 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
234 unsigned Reg = MO.getReg();
236 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
238 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
240 LV->addVirtualRegisterDead(Reg, NewMI);
242 if (MO.isUse() && MO.isKill()) {
243 for (unsigned j = 0; j < 2; ++j) {
244 // Look at the two new MI's in reverse order.
245 MachineInstr *NewMI = NewMIs[j];
246 if (!NewMI->readsRegister(Reg))
248 LV->addVirtualRegisterKilled(Reg, NewMI);
249 if (VI.removeKill(MI))
250 VI.Kills.push_back(NewMI);
258 MFI->insert(MBBI, NewMIs[1]);
259 MFI->insert(MBBI, NewMIs[0]);
265 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
266 MachineBasicBlock *&FBB,
267 SmallVectorImpl<MachineOperand> &Cond,
268 bool AllowModify) const {
269 // If the block has no terminators, it just falls into the block after it.
270 MachineBasicBlock::iterator I = MBB.end();
271 if (I == MBB.begin())
274 while (I->isDebugValue()) {
275 if (I == MBB.begin())
279 if (!isUnpredicatedTerminator(I))
282 // Get the last instruction in the block.
283 MachineInstr *LastInst = I;
285 // If there is only one terminator instruction, process it.
286 unsigned LastOpc = LastInst->getOpcode();
287 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
288 if (isUncondBranchOpcode(LastOpc)) {
289 TBB = LastInst->getOperand(0).getMBB();
292 if (isCondBranchOpcode(LastOpc)) {
293 // Block ends with fall-through condbranch.
294 TBB = LastInst->getOperand(0).getMBB();
295 Cond.push_back(LastInst->getOperand(1));
296 Cond.push_back(LastInst->getOperand(2));
299 return true; // Can't handle indirect branch.
302 // Get the instruction before it if it is a terminator.
303 MachineInstr *SecondLastInst = I;
304 unsigned SecondLastOpc = SecondLastInst->getOpcode();
306 // If AllowModify is true and the block ends with two or more unconditional
307 // branches, delete all but the first unconditional branch.
308 if (AllowModify && isUncondBranchOpcode(LastOpc)) {
309 while (isUncondBranchOpcode(SecondLastOpc)) {
310 LastInst->eraseFromParent();
311 LastInst = SecondLastInst;
312 LastOpc = LastInst->getOpcode();
313 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
314 // Return now the only terminator is an unconditional branch.
315 TBB = LastInst->getOperand(0).getMBB();
319 SecondLastOpc = SecondLastInst->getOpcode();
324 // If there are three terminators, we don't know what sort of block this is.
325 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
328 // If the block ends with a B and a Bcc, handle it.
329 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
330 TBB = SecondLastInst->getOperand(0).getMBB();
331 Cond.push_back(SecondLastInst->getOperand(1));
332 Cond.push_back(SecondLastInst->getOperand(2));
333 FBB = LastInst->getOperand(0).getMBB();
337 // If the block ends with two unconditional branches, handle it. The second
338 // one is not executed, so remove it.
339 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
340 TBB = SecondLastInst->getOperand(0).getMBB();
343 I->eraseFromParent();
347 // ...likewise if it ends with a branch table followed by an unconditional
348 // branch. The branch folder can create these, and we must get rid of them for
349 // correctness of Thumb constant islands.
350 if ((isJumpTableBranchOpcode(SecondLastOpc) ||
351 isIndirectBranchOpcode(SecondLastOpc)) &&
352 isUncondBranchOpcode(LastOpc)) {
355 I->eraseFromParent();
359 // Otherwise, can't handle this.
364 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
365 MachineBasicBlock::iterator I = MBB.end();
366 if (I == MBB.begin()) return 0;
368 while (I->isDebugValue()) {
369 if (I == MBB.begin())
373 if (!isUncondBranchOpcode(I->getOpcode()) &&
374 !isCondBranchOpcode(I->getOpcode()))
377 // Remove the branch.
378 I->eraseFromParent();
382 if (I == MBB.begin()) return 1;
384 if (!isCondBranchOpcode(I->getOpcode()))
387 // Remove the branch.
388 I->eraseFromParent();
393 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
394 MachineBasicBlock *FBB,
395 const SmallVectorImpl<MachineOperand> &Cond,
397 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
398 int BOpc = !AFI->isThumbFunction()
399 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
400 int BccOpc = !AFI->isThumbFunction()
401 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
403 // Shouldn't be a fall through.
404 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
405 assert((Cond.size() == 2 || Cond.size() == 0) &&
406 "ARM branch conditions have two components!");
409 if (Cond.empty()) // Unconditional branch?
410 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
412 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
413 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
417 // Two-way conditional branch.
418 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
419 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
420 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
424 bool ARMBaseInstrInfo::
425 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
426 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
427 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
431 bool ARMBaseInstrInfo::
432 PredicateInstruction(MachineInstr *MI,
433 const SmallVectorImpl<MachineOperand> &Pred) const {
434 unsigned Opc = MI->getOpcode();
435 if (isUncondBranchOpcode(Opc)) {
436 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
437 MI->addOperand(MachineOperand::CreateImm(Pred[0].getImm()));
438 MI->addOperand(MachineOperand::CreateReg(Pred[1].getReg(), false));
442 int PIdx = MI->findFirstPredOperandIdx();
444 MachineOperand &PMO = MI->getOperand(PIdx);
445 PMO.setImm(Pred[0].getImm());
446 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
452 bool ARMBaseInstrInfo::
453 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
454 const SmallVectorImpl<MachineOperand> &Pred2) const {
455 if (Pred1.size() > 2 || Pred2.size() > 2)
458 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
459 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
469 return CC2 == ARMCC::HI;
471 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
473 return CC2 == ARMCC::GT;
475 return CC2 == ARMCC::LT;
479 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
480 std::vector<MachineOperand> &Pred) const {
481 // FIXME: This confuses implicit_def with optional CPSR def.
482 const MCInstrDesc &MCID = MI->getDesc();
483 if (!MCID.getImplicitDefs() && !MCID.hasOptionalDef())
487 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
488 const MachineOperand &MO = MI->getOperand(i);
489 if (MO.isReg() && MO.getReg() == ARM::CPSR) {
498 /// isPredicable - Return true if the specified instruction can be predicated.
499 /// By default, this returns true for every instruction with a
500 /// PredicateOperand.
501 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
502 const MCInstrDesc &MCID = MI->getDesc();
503 if (!MCID.isPredicable())
506 if ((MCID.TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
507 ARMFunctionInfo *AFI =
508 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
509 return AFI->isThumb2Function();
514 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
515 LLVM_ATTRIBUTE_NOINLINE
516 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
518 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
520 assert(JTI < JT.size());
521 return JT[JTI].MBBs.size();
524 /// GetInstSize - Return the size of the specified MachineInstr.
526 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
527 const MachineBasicBlock &MBB = *MI->getParent();
528 const MachineFunction *MF = MBB.getParent();
529 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
531 const MCInstrDesc &MCID = MI->getDesc();
533 return MCID.getSize();
535 // If this machine instr is an inline asm, measure it.
536 if (MI->getOpcode() == ARM::INLINEASM)
537 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
540 unsigned Opc = MI->getOpcode();
542 case TargetOpcode::IMPLICIT_DEF:
543 case TargetOpcode::KILL:
544 case TargetOpcode::PROLOG_LABEL:
545 case TargetOpcode::EH_LABEL:
546 case TargetOpcode::DBG_VALUE:
548 case ARM::MOVi16_ga_pcrel:
549 case ARM::MOVTi16_ga_pcrel:
550 case ARM::t2MOVi16_ga_pcrel:
551 case ARM::t2MOVTi16_ga_pcrel:
554 case ARM::t2MOVi32imm:
556 case ARM::CONSTPOOL_ENTRY:
557 // If this machine instr is a constant pool entry, its size is recorded as
559 return MI->getOperand(2).getImm();
560 case ARM::Int_eh_sjlj_longjmp:
562 case ARM::tInt_eh_sjlj_longjmp:
564 case ARM::Int_eh_sjlj_setjmp:
565 case ARM::Int_eh_sjlj_setjmp_nofp:
567 case ARM::tInt_eh_sjlj_setjmp:
568 case ARM::t2Int_eh_sjlj_setjmp:
569 case ARM::t2Int_eh_sjlj_setjmp_nofp:
577 case ARM::t2TBH_JT: {
578 // These are jumptable branches, i.e. a branch followed by an inlined
579 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
580 // entry is one byte; TBH two byte each.
581 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
582 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
583 unsigned NumOps = MCID.getNumOperands();
584 MachineOperand JTOP =
585 MI->getOperand(NumOps - (MCID.isPredicable() ? 3 : 2));
586 unsigned JTI = JTOP.getIndex();
587 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
589 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
590 assert(JTI < JT.size());
591 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
592 // 4 aligned. The assembler / linker may add 2 byte padding just before
593 // the JT entries. The size does not include this padding; the
594 // constant islands pass does separate bookkeeping for it.
595 // FIXME: If we know the size of the function is less than (1 << 16) *2
596 // bytes, we can use 16-bit entries instead. Then there won't be an
598 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
599 unsigned NumEntries = getNumJTEntries(JT, JTI);
600 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
601 // Make sure the instruction that follows TBB is 2-byte aligned.
602 // FIXME: Constant island pass should insert an "ALIGN" instruction
605 return NumEntries * EntrySize + InstSize;
608 // Otherwise, pseudo-instruction sizes are zero.
611 return 0; // Not reached
614 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
615 MachineBasicBlock::iterator I, DebugLoc DL,
616 unsigned DestReg, unsigned SrcReg,
617 bool KillSrc) const {
618 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
619 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
621 if (GPRDest && GPRSrc) {
622 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
623 .addReg(SrcReg, getKillRegState(KillSrc))));
627 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
628 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
631 if (SPRDest && SPRSrc)
633 else if (GPRDest && SPRSrc)
635 else if (SPRDest && GPRSrc)
637 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
639 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
641 else if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
643 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
646 llvm_unreachable("Impossible reg-to-reg copy");
648 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
649 MIB.addReg(SrcReg, getKillRegState(KillSrc));
650 if (Opc == ARM::VORRq)
651 MIB.addReg(SrcReg, getKillRegState(KillSrc));
652 if (Opc != ARM::VMOVQQ && Opc != ARM::VMOVQQQQ)
657 MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB,
658 unsigned Reg, unsigned SubIdx, unsigned State,
659 const TargetRegisterInfo *TRI) {
661 return MIB.addReg(Reg, State);
663 if (TargetRegisterInfo::isPhysicalRegister(Reg))
664 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
665 return MIB.addReg(Reg, State, SubIdx);
668 void ARMBaseInstrInfo::
669 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
670 unsigned SrcReg, bool isKill, int FI,
671 const TargetRegisterClass *RC,
672 const TargetRegisterInfo *TRI) const {
674 if (I != MBB.end()) DL = I->getDebugLoc();
675 MachineFunction &MF = *MBB.getParent();
676 MachineFrameInfo &MFI = *MF.getFrameInfo();
677 unsigned Align = MFI.getObjectAlignment(FI);
679 MachineMemOperand *MMO =
680 MF.getMachineMemOperand(MachinePointerInfo(
681 PseudoSourceValue::getFixedStack(FI)),
682 MachineMemOperand::MOStore,
683 MFI.getObjectSize(FI),
686 // tGPR is used sometimes in ARM instructions that need to avoid using
687 // certain registers. Just treat it as GPR here. Likewise, rGPR.
688 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
689 || RC == ARM::rGPRRegisterClass)
690 RC = ARM::GPRRegisterClass;
692 switch (RC->getID()) {
693 case ARM::GPRRegClassID:
694 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
695 .addReg(SrcReg, getKillRegState(isKill))
696 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
698 case ARM::SPRRegClassID:
699 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
700 .addReg(SrcReg, getKillRegState(isKill))
701 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
703 case ARM::DPRRegClassID:
704 case ARM::DPR_VFP2RegClassID:
705 case ARM::DPR_8RegClassID:
706 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
707 .addReg(SrcReg, getKillRegState(isKill))
708 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
710 case ARM::QPRRegClassID:
711 case ARM::QPR_VFP2RegClassID:
712 case ARM::QPR_8RegClassID:
713 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
714 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64Pseudo))
715 .addFrameIndex(FI).addImm(16)
716 .addReg(SrcReg, getKillRegState(isKill))
717 .addMemOperand(MMO));
719 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
720 .addReg(SrcReg, getKillRegState(isKill))
722 .addMemOperand(MMO));
725 case ARM::QQPRRegClassID:
726 case ARM::QQPR_VFP2RegClassID:
727 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
728 // FIXME: It's possible to only store part of the QQ register if the
729 // spilled def has a sub-register index.
730 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
731 .addFrameIndex(FI).addImm(16)
732 .addReg(SrcReg, getKillRegState(isKill))
733 .addMemOperand(MMO));
735 MachineInstrBuilder MIB =
736 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
739 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
740 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
741 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
742 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
745 case ARM::QQQQPRRegClassID: {
746 MachineInstrBuilder MIB =
747 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
750 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
751 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
752 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
753 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
754 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
755 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
756 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
757 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
761 llvm_unreachable("Unknown regclass!");
766 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
767 int &FrameIndex) const {
768 switch (MI->getOpcode()) {
771 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
772 if (MI->getOperand(1).isFI() &&
773 MI->getOperand(2).isReg() &&
774 MI->getOperand(3).isImm() &&
775 MI->getOperand(2).getReg() == 0 &&
776 MI->getOperand(3).getImm() == 0) {
777 FrameIndex = MI->getOperand(1).getIndex();
778 return MI->getOperand(0).getReg();
786 if (MI->getOperand(1).isFI() &&
787 MI->getOperand(2).isImm() &&
788 MI->getOperand(2).getImm() == 0) {
789 FrameIndex = MI->getOperand(1).getIndex();
790 return MI->getOperand(0).getReg();
793 case ARM::VST1q64Pseudo:
794 if (MI->getOperand(0).isFI() &&
795 MI->getOperand(2).getSubReg() == 0) {
796 FrameIndex = MI->getOperand(0).getIndex();
797 return MI->getOperand(2).getReg();
801 if (MI->getOperand(1).isFI() &&
802 MI->getOperand(0).getSubReg() == 0) {
803 FrameIndex = MI->getOperand(1).getIndex();
804 return MI->getOperand(0).getReg();
812 void ARMBaseInstrInfo::
813 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
814 unsigned DestReg, int FI,
815 const TargetRegisterClass *RC,
816 const TargetRegisterInfo *TRI) const {
818 if (I != MBB.end()) DL = I->getDebugLoc();
819 MachineFunction &MF = *MBB.getParent();
820 MachineFrameInfo &MFI = *MF.getFrameInfo();
821 unsigned Align = MFI.getObjectAlignment(FI);
822 MachineMemOperand *MMO =
823 MF.getMachineMemOperand(
824 MachinePointerInfo(PseudoSourceValue::getFixedStack(FI)),
825 MachineMemOperand::MOLoad,
826 MFI.getObjectSize(FI),
829 // tGPR is used sometimes in ARM instructions that need to avoid using
830 // certain registers. Just treat it as GPR here.
831 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
832 || RC == ARM::rGPRRegisterClass)
833 RC = ARM::GPRRegisterClass;
835 switch (RC->getID()) {
836 case ARM::GPRRegClassID:
837 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
838 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
840 case ARM::SPRRegClassID:
841 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
842 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
844 case ARM::DPRRegClassID:
845 case ARM::DPR_VFP2RegClassID:
846 case ARM::DPR_8RegClassID:
847 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
848 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
850 case ARM::QPRRegClassID:
851 case ARM::QPR_VFP2RegClassID:
852 case ARM::QPR_8RegClassID:
853 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
854 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64Pseudo), DestReg)
855 .addFrameIndex(FI).addImm(16)
856 .addMemOperand(MMO));
858 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
860 .addMemOperand(MMO));
863 case ARM::QQPRRegClassID:
864 case ARM::QQPR_VFP2RegClassID:
865 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
866 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
867 .addFrameIndex(FI).addImm(16)
868 .addMemOperand(MMO));
870 MachineInstrBuilder MIB =
871 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
874 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
875 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
876 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
877 AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
880 case ARM::QQQQPRRegClassID: {
881 MachineInstrBuilder MIB =
882 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
885 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
886 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
887 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
888 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
889 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI);
890 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI);
891 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI);
892 AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
896 llvm_unreachable("Unknown regclass!");
901 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
902 int &FrameIndex) const {
903 switch (MI->getOpcode()) {
906 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
907 if (MI->getOperand(1).isFI() &&
908 MI->getOperand(2).isReg() &&
909 MI->getOperand(3).isImm() &&
910 MI->getOperand(2).getReg() == 0 &&
911 MI->getOperand(3).getImm() == 0) {
912 FrameIndex = MI->getOperand(1).getIndex();
913 return MI->getOperand(0).getReg();
921 if (MI->getOperand(1).isFI() &&
922 MI->getOperand(2).isImm() &&
923 MI->getOperand(2).getImm() == 0) {
924 FrameIndex = MI->getOperand(1).getIndex();
925 return MI->getOperand(0).getReg();
928 case ARM::VLD1q64Pseudo:
929 if (MI->getOperand(1).isFI() &&
930 MI->getOperand(0).getSubReg() == 0) {
931 FrameIndex = MI->getOperand(1).getIndex();
932 return MI->getOperand(0).getReg();
936 if (MI->getOperand(1).isFI() &&
937 MI->getOperand(0).getSubReg() == 0) {
938 FrameIndex = MI->getOperand(1).getIndex();
939 return MI->getOperand(0).getReg();
948 ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
949 int FrameIx, uint64_t Offset,
952 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
953 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
957 /// Create a copy of a const pool value. Update CPI to the new index and return
959 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
960 MachineConstantPool *MCP = MF.getConstantPool();
961 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
963 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
964 assert(MCPE.isMachineConstantPoolEntry() &&
965 "Expecting a machine constantpool entry!");
966 ARMConstantPoolValue *ACPV =
967 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
969 unsigned PCLabelId = AFI->createPICLabelUId();
970 ARMConstantPoolValue *NewCPV = 0;
971 // FIXME: The below assumes PIC relocation model and that the function
972 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
973 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
974 // instructions, so that's probably OK, but is PIC always correct when
976 if (ACPV->isGlobalValue())
977 NewCPV = new ARMConstantPoolValue(ACPV->getGV(), PCLabelId,
979 else if (ACPV->isExtSymbol())
980 NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(),
981 ACPV->getSymbol(), PCLabelId, 4);
982 else if (ACPV->isBlockAddress())
983 NewCPV = new ARMConstantPoolValue(ACPV->getBlockAddress(), PCLabelId,
984 ARMCP::CPBlockAddress, 4);
985 else if (ACPV->isLSDA())
986 NewCPV = new ARMConstantPoolValue(MF.getFunction(), PCLabelId,
989 llvm_unreachable("Unexpected ARM constantpool value type!!");
990 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
994 void ARMBaseInstrInfo::
995 reMaterialize(MachineBasicBlock &MBB,
996 MachineBasicBlock::iterator I,
997 unsigned DestReg, unsigned SubIdx,
998 const MachineInstr *Orig,
999 const TargetRegisterInfo &TRI) const {
1000 unsigned Opcode = Orig->getOpcode();
1003 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1004 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1008 case ARM::tLDRpci_pic:
1009 case ARM::t2LDRpci_pic: {
1010 MachineFunction &MF = *MBB.getParent();
1011 unsigned CPI = Orig->getOperand(1).getIndex();
1012 unsigned PCLabelId = duplicateCPV(MF, CPI);
1013 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1015 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1016 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1023 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1024 MachineInstr *MI = TargetInstrInfoImpl::duplicate(Orig, MF);
1025 switch(Orig->getOpcode()) {
1026 case ARM::tLDRpci_pic:
1027 case ARM::t2LDRpci_pic: {
1028 unsigned CPI = Orig->getOperand(1).getIndex();
1029 unsigned PCLabelId = duplicateCPV(MF, CPI);
1030 Orig->getOperand(1).setIndex(CPI);
1031 Orig->getOperand(2).setImm(PCLabelId);
1038 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1039 const MachineInstr *MI1,
1040 const MachineRegisterInfo *MRI) const {
1041 int Opcode = MI0->getOpcode();
1042 if (Opcode == ARM::t2LDRpci ||
1043 Opcode == ARM::t2LDRpci_pic ||
1044 Opcode == ARM::tLDRpci ||
1045 Opcode == ARM::tLDRpci_pic ||
1046 Opcode == ARM::MOV_ga_dyn ||
1047 Opcode == ARM::MOV_ga_pcrel ||
1048 Opcode == ARM::MOV_ga_pcrel_ldr ||
1049 Opcode == ARM::t2MOV_ga_dyn ||
1050 Opcode == ARM::t2MOV_ga_pcrel) {
1051 if (MI1->getOpcode() != Opcode)
1053 if (MI0->getNumOperands() != MI1->getNumOperands())
1056 const MachineOperand &MO0 = MI0->getOperand(1);
1057 const MachineOperand &MO1 = MI1->getOperand(1);
1058 if (MO0.getOffset() != MO1.getOffset())
1061 if (Opcode == ARM::MOV_ga_dyn ||
1062 Opcode == ARM::MOV_ga_pcrel ||
1063 Opcode == ARM::MOV_ga_pcrel_ldr ||
1064 Opcode == ARM::t2MOV_ga_dyn ||
1065 Opcode == ARM::t2MOV_ga_pcrel)
1066 // Ignore the PC labels.
1067 return MO0.getGlobal() == MO1.getGlobal();
1069 const MachineFunction *MF = MI0->getParent()->getParent();
1070 const MachineConstantPool *MCP = MF->getConstantPool();
1071 int CPI0 = MO0.getIndex();
1072 int CPI1 = MO1.getIndex();
1073 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1074 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1075 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1076 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1077 if (isARMCP0 && isARMCP1) {
1078 ARMConstantPoolValue *ACPV0 =
1079 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1080 ARMConstantPoolValue *ACPV1 =
1081 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1082 return ACPV0->hasSameValue(ACPV1);
1083 } else if (!isARMCP0 && !isARMCP1) {
1084 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1087 } else if (Opcode == ARM::PICLDR) {
1088 if (MI1->getOpcode() != Opcode)
1090 if (MI0->getNumOperands() != MI1->getNumOperands())
1093 unsigned Addr0 = MI0->getOperand(1).getReg();
1094 unsigned Addr1 = MI1->getOperand(1).getReg();
1095 if (Addr0 != Addr1) {
1097 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1098 !TargetRegisterInfo::isVirtualRegister(Addr1))
1101 // This assumes SSA form.
1102 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1103 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1104 // Check if the loaded value, e.g. a constantpool of a global address, are
1106 if (!produceSameValue(Def0, Def1, MRI))
1110 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1111 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1112 const MachineOperand &MO0 = MI0->getOperand(i);
1113 const MachineOperand &MO1 = MI1->getOperand(i);
1114 if (!MO0.isIdenticalTo(MO1))
1120 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1123 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1124 /// determine if two loads are loading from the same base address. It should
1125 /// only return true if the base pointers are the same and the only differences
1126 /// between the two addresses is the offset. It also returns the offsets by
1128 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1130 int64_t &Offset2) const {
1131 // Don't worry about Thumb: just ARM and Thumb2.
1132 if (Subtarget.isThumb1Only()) return false;
1134 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1137 switch (Load1->getMachineOpcode()) {
1150 case ARM::t2LDRSHi8:
1152 case ARM::t2LDRSHi12:
1156 switch (Load2->getMachineOpcode()) {
1169 case ARM::t2LDRSHi8:
1171 case ARM::t2LDRSHi12:
1175 // Check if base addresses and chain operands match.
1176 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1177 Load1->getOperand(4) != Load2->getOperand(4))
1180 // Index should be Reg0.
1181 if (Load1->getOperand(3) != Load2->getOperand(3))
1184 // Determine the offsets.
1185 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1186 isa<ConstantSDNode>(Load2->getOperand(1))) {
1187 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1188 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1195 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1196 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1197 /// be scheduled togther. On some targets if two loads are loading from
1198 /// addresses in the same cache line, it's better if they are scheduled
1199 /// together. This function takes two integers that represent the load offsets
1200 /// from the common base address. It returns true if it decides it's desirable
1201 /// to schedule the two loads together. "NumLoads" is the number of loads that
1202 /// have already been scheduled after Load1.
1203 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1204 int64_t Offset1, int64_t Offset2,
1205 unsigned NumLoads) const {
1206 // Don't worry about Thumb: just ARM and Thumb2.
1207 if (Subtarget.isThumb1Only()) return false;
1209 assert(Offset2 > Offset1);
1211 if ((Offset2 - Offset1) / 8 > 64)
1214 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1215 return false; // FIXME: overly conservative?
1217 // Four loads in a row should be sufficient.
1224 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1225 const MachineBasicBlock *MBB,
1226 const MachineFunction &MF) const {
1227 // Debug info is never a scheduling boundary. It's necessary to be explicit
1228 // due to the special treatment of IT instructions below, otherwise a
1229 // dbg_value followed by an IT will result in the IT instruction being
1230 // considered a scheduling hazard, which is wrong. It should be the actual
1231 // instruction preceding the dbg_value instruction(s), just like it is
1232 // when debug info is not present.
1233 if (MI->isDebugValue())
1236 // Terminators and labels can't be scheduled around.
1237 if (MI->getDesc().isTerminator() || MI->isLabel())
1240 // Treat the start of the IT block as a scheduling boundary, but schedule
1241 // t2IT along with all instructions following it.
1242 // FIXME: This is a big hammer. But the alternative is to add all potential
1243 // true and anti dependencies to IT block instructions as implicit operands
1244 // to the t2IT instruction. The added compile time and complexity does not
1246 MachineBasicBlock::const_iterator I = MI;
1247 // Make sure to skip any dbg_value instructions
1248 while (++I != MBB->end() && I->isDebugValue())
1250 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1253 // Don't attempt to schedule around any instruction that defines
1254 // a stack-oriented pointer, as it's unlikely to be profitable. This
1255 // saves compile time, because it doesn't require every single
1256 // stack slot reference to depend on the instruction that does the
1258 if (MI->definesRegister(ARM::SP))
1264 bool ARMBaseInstrInfo::
1265 isProfitableToIfCvt(MachineBasicBlock &MBB,
1266 unsigned NumCycles, unsigned ExtraPredCycles,
1267 const BranchProbability &Probability) const {
1271 // Attempt to estimate the relative costs of predication versus branching.
1272 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1273 UnpredCost /= Probability.getDenominator();
1274 UnpredCost += 1; // The branch itself
1275 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1277 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1280 bool ARMBaseInstrInfo::
1281 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1282 unsigned TCycles, unsigned TExtra,
1283 MachineBasicBlock &FMBB,
1284 unsigned FCycles, unsigned FExtra,
1285 const BranchProbability &Probability) const {
1286 if (!TCycles || !FCycles)
1289 // Attempt to estimate the relative costs of predication versus branching.
1290 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1291 TUnpredCost /= Probability.getDenominator();
1293 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1294 unsigned FUnpredCost = Comp * FCycles;
1295 FUnpredCost /= Probability.getDenominator();
1297 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1298 UnpredCost += 1; // The branch itself
1299 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1301 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1304 /// getInstrPredicate - If instruction is predicated, returns its predicate
1305 /// condition, otherwise returns AL. It also returns the condition code
1306 /// register by reference.
1308 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1309 int PIdx = MI->findFirstPredOperandIdx();
1315 PredReg = MI->getOperand(PIdx+1).getReg();
1316 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1320 int llvm::getMatchingCondBranchOpcode(int Opc) {
1323 else if (Opc == ARM::tB)
1325 else if (Opc == ARM::t2B)
1328 llvm_unreachable("Unknown unconditional branch opcode!");
1333 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1334 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1335 unsigned DestReg, unsigned BaseReg, int NumBytes,
1336 ARMCC::CondCodes Pred, unsigned PredReg,
1337 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1338 bool isSub = NumBytes < 0;
1339 if (isSub) NumBytes = -NumBytes;
1342 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1343 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1344 assert(ThisVal && "Didn't extract field correctly");
1346 // We will handle these bits from offset, clear them.
1347 NumBytes &= ~ThisVal;
1349 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1351 // Build the new ADD / SUB.
1352 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1353 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1354 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1355 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1356 .setMIFlags(MIFlags);
1361 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1362 unsigned FrameReg, int &Offset,
1363 const ARMBaseInstrInfo &TII) {
1364 unsigned Opcode = MI.getOpcode();
1365 const MCInstrDesc &Desc = MI.getDesc();
1366 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1369 // Memory operands in inline assembly always use AddrMode2.
1370 if (Opcode == ARM::INLINEASM)
1371 AddrMode = ARMII::AddrMode2;
1373 if (Opcode == ARM::ADDri) {
1374 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1376 // Turn it into a move.
1377 MI.setDesc(TII.get(ARM::MOVr));
1378 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1379 MI.RemoveOperand(FrameRegIdx+1);
1382 } else if (Offset < 0) {
1385 MI.setDesc(TII.get(ARM::SUBri));
1388 // Common case: small offset, fits into instruction.
1389 if (ARM_AM::getSOImmVal(Offset) != -1) {
1390 // Replace the FrameIndex with sp / fp
1391 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1392 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1397 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1399 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1400 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1402 // We will handle these bits from offset, clear them.
1403 Offset &= ~ThisImmVal;
1405 // Get the properly encoded SOImmVal field.
1406 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1407 "Bit extraction didn't work?");
1408 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1410 unsigned ImmIdx = 0;
1412 unsigned NumBits = 0;
1415 case ARMII::AddrMode_i12: {
1416 ImmIdx = FrameRegIdx + 1;
1417 InstrOffs = MI.getOperand(ImmIdx).getImm();
1421 case ARMII::AddrMode2: {
1422 ImmIdx = FrameRegIdx+2;
1423 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1424 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1429 case ARMII::AddrMode3: {
1430 ImmIdx = FrameRegIdx+2;
1431 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1432 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1437 case ARMII::AddrMode4:
1438 case ARMII::AddrMode6:
1439 // Can't fold any offset even if it's zero.
1441 case ARMII::AddrMode5: {
1442 ImmIdx = FrameRegIdx+1;
1443 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1444 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1451 llvm_unreachable("Unsupported addressing mode!");
1455 Offset += InstrOffs * Scale;
1456 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1462 // Attempt to fold address comp. if opcode has offset bits
1464 // Common case: small offset, fits into instruction.
1465 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1466 int ImmedOffset = Offset / Scale;
1467 unsigned Mask = (1 << NumBits) - 1;
1468 if ((unsigned)Offset <= Mask * Scale) {
1469 // Replace the FrameIndex with sp
1470 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1471 // FIXME: When addrmode2 goes away, this will simplify (like the
1472 // T2 version), as the LDR.i12 versions don't need the encoding
1473 // tricks for the offset value.
1475 if (AddrMode == ARMII::AddrMode_i12)
1476 ImmedOffset = -ImmedOffset;
1478 ImmedOffset |= 1 << NumBits;
1480 ImmOp.ChangeToImmediate(ImmedOffset);
1485 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1486 ImmedOffset = ImmedOffset & Mask;
1488 if (AddrMode == ARMII::AddrMode_i12)
1489 ImmedOffset = -ImmedOffset;
1491 ImmedOffset |= 1 << NumBits;
1493 ImmOp.ChangeToImmediate(ImmedOffset);
1494 Offset &= ~(Mask*Scale);
1498 Offset = (isSub) ? -Offset : Offset;
1502 bool ARMBaseInstrInfo::
1503 AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpMask,
1504 int &CmpValue) const {
1505 switch (MI->getOpcode()) {
1509 SrcReg = MI->getOperand(0).getReg();
1511 CmpValue = MI->getOperand(1).getImm();
1515 SrcReg = MI->getOperand(0).getReg();
1516 CmpMask = MI->getOperand(1).getImm();
1524 /// isSuitableForMask - Identify a suitable 'and' instruction that
1525 /// operates on the given source register and applies the same mask
1526 /// as a 'tst' instruction. Provide a limited look-through for copies.
1527 /// When successful, MI will hold the found instruction.
1528 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
1529 int CmpMask, bool CommonUse) {
1530 switch (MI->getOpcode()) {
1533 if (CmpMask != MI->getOperand(2).getImm())
1535 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
1539 // Walk down one instruction which is potentially an 'and'.
1540 const MachineInstr &Copy = *MI;
1541 MachineBasicBlock::iterator AND(
1542 llvm::next(MachineBasicBlock::iterator(MI)));
1543 if (AND == MI->getParent()->end()) return false;
1545 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
1553 /// OptimizeCompareInstr - Convert the instruction supplying the argument to the
1554 /// comparison into one that sets the zero bit in the flags register.
1555 bool ARMBaseInstrInfo::
1556 OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
1557 int CmpValue, const MachineRegisterInfo *MRI) const {
1561 MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
1562 if (llvm::next(DI) != MRI->def_end())
1563 // Only support one definition.
1566 MachineInstr *MI = &*DI;
1568 // Masked compares sometimes use the same register as the corresponding 'and'.
1569 if (CmpMask != ~0) {
1570 if (!isSuitableForMask(MI, SrcReg, CmpMask, false)) {
1572 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
1573 UE = MRI->use_end(); UI != UE; ++UI) {
1574 if (UI->getParent() != CmpInstr->getParent()) continue;
1575 MachineInstr *PotentialAND = &*UI;
1576 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true))
1581 if (!MI) return false;
1585 // Conservatively refuse to convert an instruction which isn't in the same BB
1586 // as the comparison.
1587 if (MI->getParent() != CmpInstr->getParent())
1590 // Check that CPSR isn't set between the comparison instruction and the one we
1592 MachineBasicBlock::const_iterator I = CmpInstr, E = MI,
1593 B = MI->getParent()->begin();
1595 // Early exit if CmpInstr is at the beginning of the BB.
1596 if (I == B) return false;
1599 for (; I != E; --I) {
1600 const MachineInstr &Instr = *I;
1602 for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) {
1603 const MachineOperand &MO = Instr.getOperand(IO);
1604 if (!MO.isReg()) continue;
1606 // This instruction modifies or uses CPSR after the one we want to
1607 // change. We can't do this transformation.
1608 if (MO.getReg() == ARM::CPSR)
1613 // The 'and' is below the comparison instruction.
1617 // Set the "zero" bit in CPSR.
1618 switch (MI->getOpcode()) {
1652 case ARM::t2EORri: {
1653 // Scan forward for the use of CPSR, if it's a conditional code requires
1654 // checking of V bit, then this is not safe to do. If we can't find the
1655 // CPSR use (i.e. used in another block), then it's not safe to perform
1656 // the optimization.
1657 bool isSafe = false;
1659 E = MI->getParent()->end();
1660 while (!isSafe && ++I != E) {
1661 const MachineInstr &Instr = *I;
1662 for (unsigned IO = 0, EO = Instr.getNumOperands();
1663 !isSafe && IO != EO; ++IO) {
1664 const MachineOperand &MO = Instr.getOperand(IO);
1665 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
1671 // Condition code is after the operand before CPSR.
1672 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
1691 // Toggle the optional operand to CPSR.
1692 MI->getOperand(5).setReg(ARM::CPSR);
1693 MI->getOperand(5).setIsDef(true);
1694 CmpInstr->eraseFromParent();
1702 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
1703 MachineInstr *DefMI, unsigned Reg,
1704 MachineRegisterInfo *MRI) const {
1705 // Fold large immediates into add, sub, or, xor.
1706 unsigned DefOpc = DefMI->getOpcode();
1707 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
1709 if (!DefMI->getOperand(1).isImm())
1710 // Could be t2MOVi32imm <ga:xx>
1713 if (!MRI->hasOneNonDBGUse(Reg))
1716 unsigned UseOpc = UseMI->getOpcode();
1717 unsigned NewUseOpc = 0;
1718 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
1719 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
1720 bool Commute = false;
1722 default: return false;
1730 case ARM::t2EORrr: {
1731 Commute = UseMI->getOperand(2).getReg() != Reg;
1738 NewUseOpc = ARM::SUBri;
1744 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
1746 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
1747 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
1750 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
1751 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
1752 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
1756 case ARM::t2SUBrr: {
1760 NewUseOpc = ARM::t2SUBri;
1765 case ARM::t2EORrr: {
1766 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
1768 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
1769 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
1772 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
1773 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
1774 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
1782 unsigned OpIdx = Commute ? 2 : 1;
1783 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
1784 bool isKill = UseMI->getOperand(OpIdx).isKill();
1785 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
1786 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
1787 *UseMI, UseMI->getDebugLoc(),
1788 get(NewUseOpc), NewReg)
1789 .addReg(Reg1, getKillRegState(isKill))
1790 .addImm(SOImmValV1)));
1791 UseMI->setDesc(get(NewUseOpc));
1792 UseMI->getOperand(1).setReg(NewReg);
1793 UseMI->getOperand(1).setIsKill();
1794 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
1795 DefMI->eraseFromParent();
1800 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1801 const MachineInstr *MI) const {
1802 if (!ItinData || ItinData->isEmpty())
1805 const MCInstrDesc &Desc = MI->getDesc();
1806 unsigned Class = Desc.getSchedClass();
1807 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
1811 unsigned Opc = MI->getOpcode();
1814 llvm_unreachable("Unexpected multi-uops instruction!");
1820 // The number of uOps for load / store multiple are determined by the number
1823 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
1824 // same cycle. The scheduling for the first load / store must be done
1825 // separately by assuming the the address is not 64-bit aligned.
1827 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
1828 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
1829 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
1831 case ARM::VLDMDIA_UPD:
1832 case ARM::VLDMDDB_UPD:
1834 case ARM::VLDMSIA_UPD:
1835 case ARM::VLDMSDB_UPD:
1837 case ARM::VSTMDIA_UPD:
1838 case ARM::VSTMDDB_UPD:
1840 case ARM::VSTMSIA_UPD:
1841 case ARM::VSTMSDB_UPD: {
1842 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
1843 return (NumRegs / 2) + (NumRegs % 2) + 1;
1846 case ARM::LDMIA_RET:
1851 case ARM::LDMIA_UPD:
1852 case ARM::LDMDA_UPD:
1853 case ARM::LDMDB_UPD:
1854 case ARM::LDMIB_UPD:
1859 case ARM::STMIA_UPD:
1860 case ARM::STMDA_UPD:
1861 case ARM::STMDB_UPD:
1862 case ARM::STMIB_UPD:
1864 case ARM::tLDMIA_UPD:
1866 case ARM::tSTMIA_UPD:
1870 case ARM::t2LDMIA_RET:
1873 case ARM::t2LDMIA_UPD:
1874 case ARM::t2LDMDB_UPD:
1877 case ARM::t2STMIA_UPD:
1878 case ARM::t2STMDB_UPD: {
1879 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
1880 if (Subtarget.isCortexA8()) {
1883 // 4 registers would be issued: 2, 2.
1884 // 5 registers would be issued: 2, 2, 1.
1885 UOps = (NumRegs / 2);
1889 } else if (Subtarget.isCortexA9()) {
1890 UOps = (NumRegs / 2);
1891 // If there are odd number of registers or if it's not 64-bit aligned,
1892 // then it takes an extra AGU (Address Generation Unit) cycle.
1893 if ((NumRegs % 2) ||
1894 !MI->hasOneMemOperand() ||
1895 (*MI->memoperands_begin())->getAlignment() < 8)
1899 // Assume the worst.
1907 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
1908 const MCInstrDesc &DefMCID,
1910 unsigned DefIdx, unsigned DefAlign) const {
1911 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1913 // Def is the address writeback.
1914 return ItinData->getOperandCycle(DefClass, DefIdx);
1917 if (Subtarget.isCortexA8()) {
1918 // (regno / 2) + (regno % 2) + 1
1919 DefCycle = RegNo / 2 + 1;
1922 } else if (Subtarget.isCortexA9()) {
1924 bool isSLoad = false;
1926 switch (DefMCID.getOpcode()) {
1929 case ARM::VLDMSIA_UPD:
1930 case ARM::VLDMSDB_UPD:
1935 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
1936 // then it takes an extra cycle.
1937 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
1940 // Assume the worst.
1941 DefCycle = RegNo + 2;
1948 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
1949 const MCInstrDesc &DefMCID,
1951 unsigned DefIdx, unsigned DefAlign) const {
1952 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1954 // Def is the address writeback.
1955 return ItinData->getOperandCycle(DefClass, DefIdx);
1958 if (Subtarget.isCortexA8()) {
1959 // 4 registers would be issued: 1, 2, 1.
1960 // 5 registers would be issued: 1, 2, 2.
1961 DefCycle = RegNo / 2;
1964 // Result latency is issue cycle + 2: E2.
1966 } else if (Subtarget.isCortexA9()) {
1967 DefCycle = (RegNo / 2);
1968 // If there are odd number of registers or if it's not 64-bit aligned,
1969 // then it takes an extra AGU (Address Generation Unit) cycle.
1970 if ((RegNo % 2) || DefAlign < 8)
1972 // Result latency is AGU cycles + 2.
1975 // Assume the worst.
1976 DefCycle = RegNo + 2;
1983 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
1984 const MCInstrDesc &UseMCID,
1986 unsigned UseIdx, unsigned UseAlign) const {
1987 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
1989 return ItinData->getOperandCycle(UseClass, UseIdx);
1992 if (Subtarget.isCortexA8()) {
1993 // (regno / 2) + (regno % 2) + 1
1994 UseCycle = RegNo / 2 + 1;
1997 } else if (Subtarget.isCortexA9()) {
1999 bool isSStore = false;
2001 switch (UseMCID.getOpcode()) {
2004 case ARM::VSTMSIA_UPD:
2005 case ARM::VSTMSDB_UPD:
2010 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2011 // then it takes an extra cycle.
2012 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2015 // Assume the worst.
2016 UseCycle = RegNo + 2;
2023 ARMBaseInstrInfo::getSTMUseCycle(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 UseCycle = RegNo / 2;
2038 } else if (Subtarget.isCortexA9()) {
2039 UseCycle = (RegNo / 2);
2040 // If there are odd number of registers or if it's not 64-bit aligned,
2041 // then it takes an extra AGU (Address Generation Unit) cycle.
2042 if ((RegNo % 2) || UseAlign < 8)
2045 // Assume the worst.
2052 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2053 const MCInstrDesc &DefMCID,
2054 unsigned DefIdx, unsigned DefAlign,
2055 const MCInstrDesc &UseMCID,
2056 unsigned UseIdx, unsigned UseAlign) const {
2057 unsigned DefClass = DefMCID.getSchedClass();
2058 unsigned UseClass = UseMCID.getSchedClass();
2060 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2061 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2063 // This may be a def / use of a variable_ops instruction, the operand
2064 // latency might be determinable dynamically. Let the target try to
2067 bool LdmBypass = false;
2068 switch (DefMCID.getOpcode()) {
2070 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2074 case ARM::VLDMDIA_UPD:
2075 case ARM::VLDMDDB_UPD:
2077 case ARM::VLDMSIA_UPD:
2078 case ARM::VLDMSDB_UPD:
2079 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2082 case ARM::LDMIA_RET:
2087 case ARM::LDMIA_UPD:
2088 case ARM::LDMDA_UPD:
2089 case ARM::LDMDB_UPD:
2090 case ARM::LDMIB_UPD:
2092 case ARM::tLDMIA_UPD:
2094 case ARM::t2LDMIA_RET:
2097 case ARM::t2LDMIA_UPD:
2098 case ARM::t2LDMDB_UPD:
2100 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2105 // We can't seem to determine the result latency of the def, assume it's 2.
2109 switch (UseMCID.getOpcode()) {
2111 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
2115 case ARM::VSTMDIA_UPD:
2116 case ARM::VSTMDDB_UPD:
2118 case ARM::VSTMSIA_UPD:
2119 case ARM::VSTMSDB_UPD:
2120 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2127 case ARM::STMIA_UPD:
2128 case ARM::STMDA_UPD:
2129 case ARM::STMDB_UPD:
2130 case ARM::STMIB_UPD:
2132 case ARM::tSTMIA_UPD:
2137 case ARM::t2STMIA_UPD:
2138 case ARM::t2STMDB_UPD:
2139 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2144 // Assume it's read in the first stage.
2147 UseCycle = DefCycle - UseCycle + 1;
2150 // It's a variable_ops instruction so we can't use DefIdx here. Just use
2151 // first def operand.
2152 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
2155 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
2156 UseClass, UseIdx)) {
2165 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2166 const MachineInstr *DefMI, unsigned DefIdx,
2167 const MachineInstr *UseMI, unsigned UseIdx) const {
2168 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
2169 DefMI->isRegSequence() || DefMI->isImplicitDef())
2172 const MCInstrDesc &DefMCID = DefMI->getDesc();
2173 if (!ItinData || ItinData->isEmpty())
2174 return DefMCID.mayLoad() ? 3 : 1;
2176 const MCInstrDesc &UseMCID = UseMI->getDesc();
2177 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
2178 if (DefMO.getReg() == ARM::CPSR) {
2179 if (DefMI->getOpcode() == ARM::FMSTAT) {
2180 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
2181 return Subtarget.isCortexA9() ? 1 : 20;
2184 // CPSR set and branch can be paired in the same cycle.
2185 if (UseMCID.isBranch())
2189 unsigned DefAlign = DefMI->hasOneMemOperand()
2190 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
2191 unsigned UseAlign = UseMI->hasOneMemOperand()
2192 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
2193 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2194 UseMCID, UseIdx, UseAlign);
2197 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2198 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2199 // variants are one cycle cheaper.
2200 switch (DefMCID.getOpcode()) {
2204 unsigned ShOpVal = DefMI->getOperand(3).getImm();
2205 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2207 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2214 case ARM::t2LDRSHs: {
2215 // Thumb2 mode: lsl only.
2216 unsigned ShAmt = DefMI->getOperand(3).getImm();
2217 if (ShAmt == 0 || ShAmt == 2)
2224 if (DefAlign < 8 && Subtarget.isCortexA9())
2225 switch (DefMCID.getOpcode()) {
2231 case ARM::VLD1q8_UPD:
2232 case ARM::VLD1q16_UPD:
2233 case ARM::VLD1q32_UPD:
2234 case ARM::VLD1q64_UPD:
2241 case ARM::VLD2d8_UPD:
2242 case ARM::VLD2d16_UPD:
2243 case ARM::VLD2d32_UPD:
2244 case ARM::VLD2q8_UPD:
2245 case ARM::VLD2q16_UPD:
2246 case ARM::VLD2q32_UPD:
2251 case ARM::VLD3d8_UPD:
2252 case ARM::VLD3d16_UPD:
2253 case ARM::VLD3d32_UPD:
2254 case ARM::VLD1d64T_UPD:
2255 case ARM::VLD3q8_UPD:
2256 case ARM::VLD3q16_UPD:
2257 case ARM::VLD3q32_UPD:
2262 case ARM::VLD4d8_UPD:
2263 case ARM::VLD4d16_UPD:
2264 case ARM::VLD4d32_UPD:
2265 case ARM::VLD1d64Q_UPD:
2266 case ARM::VLD4q8_UPD:
2267 case ARM::VLD4q16_UPD:
2268 case ARM::VLD4q32_UPD:
2269 case ARM::VLD1DUPq8:
2270 case ARM::VLD1DUPq16:
2271 case ARM::VLD1DUPq32:
2272 case ARM::VLD1DUPq8_UPD:
2273 case ARM::VLD1DUPq16_UPD:
2274 case ARM::VLD1DUPq32_UPD:
2275 case ARM::VLD2DUPd8:
2276 case ARM::VLD2DUPd16:
2277 case ARM::VLD2DUPd32:
2278 case ARM::VLD2DUPd8_UPD:
2279 case ARM::VLD2DUPd16_UPD:
2280 case ARM::VLD2DUPd32_UPD:
2281 case ARM::VLD4DUPd8:
2282 case ARM::VLD4DUPd16:
2283 case ARM::VLD4DUPd32:
2284 case ARM::VLD4DUPd8_UPD:
2285 case ARM::VLD4DUPd16_UPD:
2286 case ARM::VLD4DUPd32_UPD:
2288 case ARM::VLD1LNd16:
2289 case ARM::VLD1LNd32:
2290 case ARM::VLD1LNd8_UPD:
2291 case ARM::VLD1LNd16_UPD:
2292 case ARM::VLD1LNd32_UPD:
2294 case ARM::VLD2LNd16:
2295 case ARM::VLD2LNd32:
2296 case ARM::VLD2LNq16:
2297 case ARM::VLD2LNq32:
2298 case ARM::VLD2LNd8_UPD:
2299 case ARM::VLD2LNd16_UPD:
2300 case ARM::VLD2LNd32_UPD:
2301 case ARM::VLD2LNq16_UPD:
2302 case ARM::VLD2LNq32_UPD:
2304 case ARM::VLD4LNd16:
2305 case ARM::VLD4LNd32:
2306 case ARM::VLD4LNq16:
2307 case ARM::VLD4LNq32:
2308 case ARM::VLD4LNd8_UPD:
2309 case ARM::VLD4LNd16_UPD:
2310 case ARM::VLD4LNd32_UPD:
2311 case ARM::VLD4LNq16_UPD:
2312 case ARM::VLD4LNq32_UPD:
2313 // If the address is not 64-bit aligned, the latencies of these
2314 // instructions increases by one.
2323 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2324 SDNode *DefNode, unsigned DefIdx,
2325 SDNode *UseNode, unsigned UseIdx) const {
2326 if (!DefNode->isMachineOpcode())
2329 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
2331 if (isZeroCost(DefMCID.Opcode))
2334 if (!ItinData || ItinData->isEmpty())
2335 return DefMCID.mayLoad() ? 3 : 1;
2337 if (!UseNode->isMachineOpcode()) {
2338 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
2339 if (Subtarget.isCortexA9())
2340 return Latency <= 2 ? 1 : Latency - 1;
2342 return Latency <= 3 ? 1 : Latency - 2;
2345 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
2346 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
2347 unsigned DefAlign = !DefMN->memoperands_empty()
2348 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
2349 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
2350 unsigned UseAlign = !UseMN->memoperands_empty()
2351 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
2352 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2353 UseMCID, UseIdx, UseAlign);
2356 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2357 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2358 // variants are one cycle cheaper.
2359 switch (DefMCID.getOpcode()) {
2364 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2365 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2367 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2374 case ARM::t2LDRSHs: {
2375 // Thumb2 mode: lsl only.
2377 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2378 if (ShAmt == 0 || ShAmt == 2)
2385 if (DefAlign < 8 && Subtarget.isCortexA9())
2386 switch (DefMCID.getOpcode()) {
2388 case ARM::VLD1q8Pseudo:
2389 case ARM::VLD1q16Pseudo:
2390 case ARM::VLD1q32Pseudo:
2391 case ARM::VLD1q64Pseudo:
2392 case ARM::VLD1q8Pseudo_UPD:
2393 case ARM::VLD1q16Pseudo_UPD:
2394 case ARM::VLD1q32Pseudo_UPD:
2395 case ARM::VLD1q64Pseudo_UPD:
2396 case ARM::VLD2d8Pseudo:
2397 case ARM::VLD2d16Pseudo:
2398 case ARM::VLD2d32Pseudo:
2399 case ARM::VLD2q8Pseudo:
2400 case ARM::VLD2q16Pseudo:
2401 case ARM::VLD2q32Pseudo:
2402 case ARM::VLD2d8Pseudo_UPD:
2403 case ARM::VLD2d16Pseudo_UPD:
2404 case ARM::VLD2d32Pseudo_UPD:
2405 case ARM::VLD2q8Pseudo_UPD:
2406 case ARM::VLD2q16Pseudo_UPD:
2407 case ARM::VLD2q32Pseudo_UPD:
2408 case ARM::VLD3d8Pseudo:
2409 case ARM::VLD3d16Pseudo:
2410 case ARM::VLD3d32Pseudo:
2411 case ARM::VLD1d64TPseudo:
2412 case ARM::VLD3d8Pseudo_UPD:
2413 case ARM::VLD3d16Pseudo_UPD:
2414 case ARM::VLD3d32Pseudo_UPD:
2415 case ARM::VLD1d64TPseudo_UPD:
2416 case ARM::VLD3q8Pseudo_UPD:
2417 case ARM::VLD3q16Pseudo_UPD:
2418 case ARM::VLD3q32Pseudo_UPD:
2419 case ARM::VLD3q8oddPseudo:
2420 case ARM::VLD3q16oddPseudo:
2421 case ARM::VLD3q32oddPseudo:
2422 case ARM::VLD3q8oddPseudo_UPD:
2423 case ARM::VLD3q16oddPseudo_UPD:
2424 case ARM::VLD3q32oddPseudo_UPD:
2425 case ARM::VLD4d8Pseudo:
2426 case ARM::VLD4d16Pseudo:
2427 case ARM::VLD4d32Pseudo:
2428 case ARM::VLD1d64QPseudo:
2429 case ARM::VLD4d8Pseudo_UPD:
2430 case ARM::VLD4d16Pseudo_UPD:
2431 case ARM::VLD4d32Pseudo_UPD:
2432 case ARM::VLD1d64QPseudo_UPD:
2433 case ARM::VLD4q8Pseudo_UPD:
2434 case ARM::VLD4q16Pseudo_UPD:
2435 case ARM::VLD4q32Pseudo_UPD:
2436 case ARM::VLD4q8oddPseudo:
2437 case ARM::VLD4q16oddPseudo:
2438 case ARM::VLD4q32oddPseudo:
2439 case ARM::VLD4q8oddPseudo_UPD:
2440 case ARM::VLD4q16oddPseudo_UPD:
2441 case ARM::VLD4q32oddPseudo_UPD:
2442 case ARM::VLD1DUPq8Pseudo:
2443 case ARM::VLD1DUPq16Pseudo:
2444 case ARM::VLD1DUPq32Pseudo:
2445 case ARM::VLD1DUPq8Pseudo_UPD:
2446 case ARM::VLD1DUPq16Pseudo_UPD:
2447 case ARM::VLD1DUPq32Pseudo_UPD:
2448 case ARM::VLD2DUPd8Pseudo:
2449 case ARM::VLD2DUPd16Pseudo:
2450 case ARM::VLD2DUPd32Pseudo:
2451 case ARM::VLD2DUPd8Pseudo_UPD:
2452 case ARM::VLD2DUPd16Pseudo_UPD:
2453 case ARM::VLD2DUPd32Pseudo_UPD:
2454 case ARM::VLD4DUPd8Pseudo:
2455 case ARM::VLD4DUPd16Pseudo:
2456 case ARM::VLD4DUPd32Pseudo:
2457 case ARM::VLD4DUPd8Pseudo_UPD:
2458 case ARM::VLD4DUPd16Pseudo_UPD:
2459 case ARM::VLD4DUPd32Pseudo_UPD:
2460 case ARM::VLD1LNq8Pseudo:
2461 case ARM::VLD1LNq16Pseudo:
2462 case ARM::VLD1LNq32Pseudo:
2463 case ARM::VLD1LNq8Pseudo_UPD:
2464 case ARM::VLD1LNq16Pseudo_UPD:
2465 case ARM::VLD1LNq32Pseudo_UPD:
2466 case ARM::VLD2LNd8Pseudo:
2467 case ARM::VLD2LNd16Pseudo:
2468 case ARM::VLD2LNd32Pseudo:
2469 case ARM::VLD2LNq16Pseudo:
2470 case ARM::VLD2LNq32Pseudo:
2471 case ARM::VLD2LNd8Pseudo_UPD:
2472 case ARM::VLD2LNd16Pseudo_UPD:
2473 case ARM::VLD2LNd32Pseudo_UPD:
2474 case ARM::VLD2LNq16Pseudo_UPD:
2475 case ARM::VLD2LNq32Pseudo_UPD:
2476 case ARM::VLD4LNd8Pseudo:
2477 case ARM::VLD4LNd16Pseudo:
2478 case ARM::VLD4LNd32Pseudo:
2479 case ARM::VLD4LNq16Pseudo:
2480 case ARM::VLD4LNq32Pseudo:
2481 case ARM::VLD4LNd8Pseudo_UPD:
2482 case ARM::VLD4LNd16Pseudo_UPD:
2483 case ARM::VLD4LNd32Pseudo_UPD:
2484 case ARM::VLD4LNq16Pseudo_UPD:
2485 case ARM::VLD4LNq32Pseudo_UPD:
2486 // If the address is not 64-bit aligned, the latencies of these
2487 // instructions increases by one.
2495 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2496 const MachineInstr *MI,
2497 unsigned *PredCost) const {
2498 if (MI->isCopyLike() || MI->isInsertSubreg() ||
2499 MI->isRegSequence() || MI->isImplicitDef())
2502 if (!ItinData || ItinData->isEmpty())
2505 const MCInstrDesc &MCID = MI->getDesc();
2506 unsigned Class = MCID.getSchedClass();
2507 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
2508 if (PredCost && MCID.hasImplicitDefOfPhysReg(ARM::CPSR))
2509 // When predicated, CPSR is an additional source operand for CPSR updating
2510 // instructions, this apparently increases their latencies.
2513 return ItinData->getStageLatency(Class);
2514 return getNumMicroOps(ItinData, MI);
2517 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2518 SDNode *Node) const {
2519 if (!Node->isMachineOpcode())
2522 if (!ItinData || ItinData->isEmpty())
2525 unsigned Opcode = Node->getMachineOpcode();
2528 return ItinData->getStageLatency(get(Opcode).getSchedClass());
2535 bool ARMBaseInstrInfo::
2536 hasHighOperandLatency(const InstrItineraryData *ItinData,
2537 const MachineRegisterInfo *MRI,
2538 const MachineInstr *DefMI, unsigned DefIdx,
2539 const MachineInstr *UseMI, unsigned UseIdx) const {
2540 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2541 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
2542 if (Subtarget.isCortexA8() &&
2543 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
2544 // CortexA8 VFP instructions are not pipelined.
2547 // Hoist VFP / NEON instructions with 4 or higher latency.
2548 int Latency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
2551 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
2552 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
2555 bool ARMBaseInstrInfo::
2556 hasLowDefLatency(const InstrItineraryData *ItinData,
2557 const MachineInstr *DefMI, unsigned DefIdx) const {
2558 if (!ItinData || ItinData->isEmpty())
2561 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2562 if (DDomain == ARMII::DomainGeneral) {
2563 unsigned DefClass = DefMI->getDesc().getSchedClass();
2564 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2565 return (DefCycle != -1 && DefCycle <= 2);
2571 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
2572 unsigned &AddSubOpc,
2573 bool &NegAcc, bool &HasLane) const {
2574 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
2575 if (I == MLxEntryMap.end())
2578 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
2579 MulOpc = Entry.MulOpc;
2580 AddSubOpc = Entry.AddSubOpc;
2581 NegAcc = Entry.NegAcc;
2582 HasLane = Entry.HasLane;