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/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/ADT/STLExtras.h"
38 #define GET_INSTRINFO_MC_DESC
39 #include "ARMGenInstrInfo.inc"
44 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
45 cl::desc("Enable ARM 2-addr to 3-addr conv"));
47 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
49 unsigned MLxOpc; // MLA / MLS opcode
50 unsigned MulOpc; // Expanded multiplication opcode
51 unsigned AddSubOpc; // Expanded add / sub opcode
52 bool NegAcc; // True if the acc is negated before the add / sub.
53 bool HasLane; // True if instruction has an extra "lane" operand.
56 static const ARM_MLxEntry ARM_MLxTable[] = {
57 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
59 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
60 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
61 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
62 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
63 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
64 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
65 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
66 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
69 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
70 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
71 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
72 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
73 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
74 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
75 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
76 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
79 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
80 : TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)),
82 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
83 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
84 assert(false && "Duplicated entries?");
85 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
86 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
90 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
91 // currently defaults to no prepass hazard recognizer.
92 ScheduleHazardRecognizer *ARMBaseInstrInfo::
93 CreateTargetHazardRecognizer(const TargetMachine *TM,
94 const ScheduleDAG *DAG) const {
95 if (usePreRAHazardRecognizer()) {
96 const InstrItineraryData *II = TM->getInstrItineraryData();
97 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
99 return TargetInstrInfoImpl::CreateTargetHazardRecognizer(TM, DAG);
102 ScheduleHazardRecognizer *ARMBaseInstrInfo::
103 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
104 const ScheduleDAG *DAG) const {
105 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
106 return (ScheduleHazardRecognizer *)
107 new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget, DAG);
108 return TargetInstrInfoImpl::CreateTargetPostRAHazardRecognizer(II, DAG);
112 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
113 MachineBasicBlock::iterator &MBBI,
114 LiveVariables *LV) const {
115 // FIXME: Thumb2 support.
120 MachineInstr *MI = MBBI;
121 MachineFunction &MF = *MI->getParent()->getParent();
122 uint64_t TSFlags = MI->getDesc().TSFlags;
124 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
125 default: return NULL;
126 case ARMII::IndexModePre:
129 case ARMII::IndexModePost:
133 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
135 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
139 MachineInstr *UpdateMI = NULL;
140 MachineInstr *MemMI = NULL;
141 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
142 const MCInstrDesc &MCID = MI->getDesc();
143 unsigned NumOps = MCID.getNumOperands();
144 bool isLoad = !MCID.mayStore();
145 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
146 const MachineOperand &Base = MI->getOperand(2);
147 const MachineOperand &Offset = MI->getOperand(NumOps-3);
148 unsigned WBReg = WB.getReg();
149 unsigned BaseReg = Base.getReg();
150 unsigned OffReg = Offset.getReg();
151 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
152 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
155 assert(false && "Unknown indexed op!");
157 case ARMII::AddrMode2: {
158 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
159 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
161 if (ARM_AM::getSOImmVal(Amt) == -1)
162 // Can't encode it in a so_imm operand. This transformation will
163 // add more than 1 instruction. Abandon!
165 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
166 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
167 .addReg(BaseReg).addImm(Amt)
168 .addImm(Pred).addReg(0).addReg(0);
169 } else if (Amt != 0) {
170 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
171 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
172 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
173 get(isSub ? ARM::SUBrs : ARM::ADDrs), WBReg)
174 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
175 .addImm(Pred).addReg(0).addReg(0);
177 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
178 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
179 .addReg(BaseReg).addReg(OffReg)
180 .addImm(Pred).addReg(0).addReg(0);
183 case ARMII::AddrMode3 : {
184 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
185 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
187 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
188 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
189 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
190 .addReg(BaseReg).addImm(Amt)
191 .addImm(Pred).addReg(0).addReg(0);
193 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
194 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
195 .addReg(BaseReg).addReg(OffReg)
196 .addImm(Pred).addReg(0).addReg(0);
201 std::vector<MachineInstr*> NewMIs;
204 MemMI = BuildMI(MF, MI->getDebugLoc(),
205 get(MemOpc), MI->getOperand(0).getReg())
206 .addReg(WBReg).addImm(0).addImm(Pred);
208 MemMI = BuildMI(MF, MI->getDebugLoc(),
209 get(MemOpc)).addReg(MI->getOperand(1).getReg())
210 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
211 NewMIs.push_back(MemMI);
212 NewMIs.push_back(UpdateMI);
215 MemMI = BuildMI(MF, MI->getDebugLoc(),
216 get(MemOpc), MI->getOperand(0).getReg())
217 .addReg(BaseReg).addImm(0).addImm(Pred);
219 MemMI = BuildMI(MF, MI->getDebugLoc(),
220 get(MemOpc)).addReg(MI->getOperand(1).getReg())
221 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
223 UpdateMI->getOperand(0).setIsDead();
224 NewMIs.push_back(UpdateMI);
225 NewMIs.push_back(MemMI);
228 // Transfer LiveVariables states, kill / dead info.
230 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
231 MachineOperand &MO = MI->getOperand(i);
232 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
233 unsigned Reg = MO.getReg();
235 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
237 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
239 LV->addVirtualRegisterDead(Reg, NewMI);
241 if (MO.isUse() && MO.isKill()) {
242 for (unsigned j = 0; j < 2; ++j) {
243 // Look at the two new MI's in reverse order.
244 MachineInstr *NewMI = NewMIs[j];
245 if (!NewMI->readsRegister(Reg))
247 LV->addVirtualRegisterKilled(Reg, NewMI);
248 if (VI.removeKill(MI))
249 VI.Kills.push_back(NewMI);
257 MFI->insert(MBBI, NewMIs[1]);
258 MFI->insert(MBBI, NewMIs[0]);
264 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
265 MachineBasicBlock *&FBB,
266 SmallVectorImpl<MachineOperand> &Cond,
267 bool AllowModify) const {
268 // If the block has no terminators, it just falls into the block after it.
269 MachineBasicBlock::iterator I = MBB.end();
270 if (I == MBB.begin())
273 while (I->isDebugValue()) {
274 if (I == MBB.begin())
278 if (!isUnpredicatedTerminator(I))
281 // Get the last instruction in the block.
282 MachineInstr *LastInst = I;
284 // If there is only one terminator instruction, process it.
285 unsigned LastOpc = LastInst->getOpcode();
286 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
287 if (isUncondBranchOpcode(LastOpc)) {
288 TBB = LastInst->getOperand(0).getMBB();
291 if (isCondBranchOpcode(LastOpc)) {
292 // Block ends with fall-through condbranch.
293 TBB = LastInst->getOperand(0).getMBB();
294 Cond.push_back(LastInst->getOperand(1));
295 Cond.push_back(LastInst->getOperand(2));
298 return true; // Can't handle indirect branch.
301 // Get the instruction before it if it is a terminator.
302 MachineInstr *SecondLastInst = I;
303 unsigned SecondLastOpc = SecondLastInst->getOpcode();
305 // If AllowModify is true and the block ends with two or more unconditional
306 // branches, delete all but the first unconditional branch.
307 if (AllowModify && isUncondBranchOpcode(LastOpc)) {
308 while (isUncondBranchOpcode(SecondLastOpc)) {
309 LastInst->eraseFromParent();
310 LastInst = SecondLastInst;
311 LastOpc = LastInst->getOpcode();
312 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
313 // Return now the only terminator is an unconditional branch.
314 TBB = LastInst->getOperand(0).getMBB();
318 SecondLastOpc = SecondLastInst->getOpcode();
323 // If there are three terminators, we don't know what sort of block this is.
324 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
327 // If the block ends with a B and a Bcc, handle it.
328 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
329 TBB = SecondLastInst->getOperand(0).getMBB();
330 Cond.push_back(SecondLastInst->getOperand(1));
331 Cond.push_back(SecondLastInst->getOperand(2));
332 FBB = LastInst->getOperand(0).getMBB();
336 // If the block ends with two unconditional branches, handle it. The second
337 // one is not executed, so remove it.
338 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
339 TBB = SecondLastInst->getOperand(0).getMBB();
342 I->eraseFromParent();
346 // ...likewise if it ends with a branch table followed by an unconditional
347 // branch. The branch folder can create these, and we must get rid of them for
348 // correctness of Thumb constant islands.
349 if ((isJumpTableBranchOpcode(SecondLastOpc) ||
350 isIndirectBranchOpcode(SecondLastOpc)) &&
351 isUncondBranchOpcode(LastOpc)) {
354 I->eraseFromParent();
358 // Otherwise, can't handle this.
363 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
364 MachineBasicBlock::iterator I = MBB.end();
365 if (I == MBB.begin()) return 0;
367 while (I->isDebugValue()) {
368 if (I == MBB.begin())
372 if (!isUncondBranchOpcode(I->getOpcode()) &&
373 !isCondBranchOpcode(I->getOpcode()))
376 // Remove the branch.
377 I->eraseFromParent();
381 if (I == MBB.begin()) return 1;
383 if (!isCondBranchOpcode(I->getOpcode()))
386 // Remove the branch.
387 I->eraseFromParent();
392 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
393 MachineBasicBlock *FBB,
394 const SmallVectorImpl<MachineOperand> &Cond,
396 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
397 int BOpc = !AFI->isThumbFunction()
398 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
399 int BccOpc = !AFI->isThumbFunction()
400 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
402 // Shouldn't be a fall through.
403 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
404 assert((Cond.size() == 2 || Cond.size() == 0) &&
405 "ARM branch conditions have two components!");
408 if (Cond.empty()) // Unconditional branch?
409 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
411 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
412 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
416 // Two-way conditional branch.
417 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
418 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
419 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
423 bool ARMBaseInstrInfo::
424 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
425 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
426 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
430 bool ARMBaseInstrInfo::
431 PredicateInstruction(MachineInstr *MI,
432 const SmallVectorImpl<MachineOperand> &Pred) const {
433 unsigned Opc = MI->getOpcode();
434 if (isUncondBranchOpcode(Opc)) {
435 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
436 MI->addOperand(MachineOperand::CreateImm(Pred[0].getImm()));
437 MI->addOperand(MachineOperand::CreateReg(Pred[1].getReg(), false));
441 int PIdx = MI->findFirstPredOperandIdx();
443 MachineOperand &PMO = MI->getOperand(PIdx);
444 PMO.setImm(Pred[0].getImm());
445 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
451 bool ARMBaseInstrInfo::
452 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
453 const SmallVectorImpl<MachineOperand> &Pred2) const {
454 if (Pred1.size() > 2 || Pred2.size() > 2)
457 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
458 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
468 return CC2 == ARMCC::HI;
470 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
472 return CC2 == ARMCC::GT;
474 return CC2 == ARMCC::LT;
478 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
479 std::vector<MachineOperand> &Pred) const {
480 // FIXME: This confuses implicit_def with optional CPSR def.
481 const MCInstrDesc &MCID = MI->getDesc();
482 if (!MCID.getImplicitDefs() && !MCID.hasOptionalDef())
486 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
487 const MachineOperand &MO = MI->getOperand(i);
488 if (MO.isReg() && MO.getReg() == ARM::CPSR) {
497 /// isPredicable - Return true if the specified instruction can be predicated.
498 /// By default, this returns true for every instruction with a
499 /// PredicateOperand.
500 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
501 const MCInstrDesc &MCID = MI->getDesc();
502 if (!MCID.isPredicable())
505 if ((MCID.TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
506 ARMFunctionInfo *AFI =
507 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
508 return AFI->isThumb2Function();
513 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
514 LLVM_ATTRIBUTE_NOINLINE
515 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
517 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
519 assert(JTI < JT.size());
520 return JT[JTI].MBBs.size();
523 /// GetInstSize - Return the size of the specified MachineInstr.
525 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
526 const MachineBasicBlock &MBB = *MI->getParent();
527 const MachineFunction *MF = MBB.getParent();
528 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
530 // Basic size info comes from the TSFlags field.
531 const MCInstrDesc &MCID = MI->getDesc();
532 uint64_t TSFlags = MCID.TSFlags;
534 unsigned Opc = MI->getOpcode();
535 switch ((TSFlags & ARMII::SizeMask) >> ARMII::SizeShift) {
537 // If this machine instr is an inline asm, measure it.
538 if (MI->getOpcode() == ARM::INLINEASM)
539 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
544 llvm_unreachable("Unknown or unset size field for instr!");
545 case TargetOpcode::IMPLICIT_DEF:
546 case TargetOpcode::KILL:
547 case TargetOpcode::PROLOG_LABEL:
548 case TargetOpcode::EH_LABEL:
549 case TargetOpcode::DBG_VALUE:
554 case ARMII::Size8Bytes: return 8; // ARM instruction x 2.
555 case ARMII::Size4Bytes: return 4; // ARM / Thumb2 instruction.
556 case ARMII::Size2Bytes: return 2; // Thumb1 instruction.
557 case ARMII::SizeSpecial: {
559 case ARM::MOVi16_ga_pcrel:
560 case ARM::MOVTi16_ga_pcrel:
561 case ARM::t2MOVi16_ga_pcrel:
562 case ARM::t2MOVTi16_ga_pcrel:
565 case ARM::t2MOVi32imm:
567 case ARM::CONSTPOOL_ENTRY:
568 // If this machine instr is a constant pool entry, its size is recorded as
570 return MI->getOperand(2).getImm();
571 case ARM::Int_eh_sjlj_longjmp:
573 case ARM::tInt_eh_sjlj_longjmp:
575 case ARM::Int_eh_sjlj_setjmp:
576 case ARM::Int_eh_sjlj_setjmp_nofp:
578 case ARM::tInt_eh_sjlj_setjmp:
579 case ARM::t2Int_eh_sjlj_setjmp:
580 case ARM::t2Int_eh_sjlj_setjmp_nofp:
588 case ARM::t2TBH_JT: {
589 // These are jumptable branches, i.e. a branch followed by an inlined
590 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
591 // entry is one byte; TBH two byte each.
592 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
593 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
594 unsigned NumOps = MCID.getNumOperands();
595 MachineOperand JTOP =
596 MI->getOperand(NumOps - (MCID.isPredicable() ? 3 : 2));
597 unsigned JTI = JTOP.getIndex();
598 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
600 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
601 assert(JTI < JT.size());
602 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
603 // 4 aligned. The assembler / linker may add 2 byte padding just before
604 // the JT entries. The size does not include this padding; the
605 // constant islands pass does separate bookkeeping for it.
606 // FIXME: If we know the size of the function is less than (1 << 16) *2
607 // bytes, we can use 16-bit entries instead. Then there won't be an
609 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
610 unsigned NumEntries = getNumJTEntries(JT, JTI);
611 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
612 // Make sure the instruction that follows TBB is 2-byte aligned.
613 // FIXME: Constant island pass should insert an "ALIGN" instruction
616 return NumEntries * EntrySize + InstSize;
619 // Otherwise, pseudo-instruction sizes are zero.
624 return 0; // Not reached
627 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
628 MachineBasicBlock::iterator I, DebugLoc DL,
629 unsigned DestReg, unsigned SrcReg,
630 bool KillSrc) const {
631 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
632 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
634 if (GPRDest && GPRSrc) {
635 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
636 .addReg(SrcReg, getKillRegState(KillSrc))));
640 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
641 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
644 if (SPRDest && SPRSrc)
646 else if (GPRDest && SPRSrc)
648 else if (SPRDest && GPRSrc)
650 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
652 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
654 else if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
656 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
659 llvm_unreachable("Impossible reg-to-reg copy");
661 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
662 MIB.addReg(SrcReg, getKillRegState(KillSrc));
663 if (Opc != ARM::VMOVQQ && Opc != ARM::VMOVQQQQ)
668 MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB,
669 unsigned Reg, unsigned SubIdx, unsigned State,
670 const TargetRegisterInfo *TRI) {
672 return MIB.addReg(Reg, State);
674 if (TargetRegisterInfo::isPhysicalRegister(Reg))
675 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
676 return MIB.addReg(Reg, State, SubIdx);
679 void ARMBaseInstrInfo::
680 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
681 unsigned SrcReg, bool isKill, int FI,
682 const TargetRegisterClass *RC,
683 const TargetRegisterInfo *TRI) const {
685 if (I != MBB.end()) DL = I->getDebugLoc();
686 MachineFunction &MF = *MBB.getParent();
687 MachineFrameInfo &MFI = *MF.getFrameInfo();
688 unsigned Align = MFI.getObjectAlignment(FI);
690 MachineMemOperand *MMO =
691 MF.getMachineMemOperand(MachinePointerInfo(
692 PseudoSourceValue::getFixedStack(FI)),
693 MachineMemOperand::MOStore,
694 MFI.getObjectSize(FI),
697 // tGPR is used sometimes in ARM instructions that need to avoid using
698 // certain registers. Just treat it as GPR here. Likewise, rGPR.
699 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
700 || RC == ARM::rGPRRegisterClass)
701 RC = ARM::GPRRegisterClass;
703 switch (RC->getID()) {
704 case ARM::GPRRegClassID:
705 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
706 .addReg(SrcReg, getKillRegState(isKill))
707 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
709 case ARM::SPRRegClassID:
710 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
711 .addReg(SrcReg, getKillRegState(isKill))
712 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
714 case ARM::DPRRegClassID:
715 case ARM::DPR_VFP2RegClassID:
716 case ARM::DPR_8RegClassID:
717 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
718 .addReg(SrcReg, getKillRegState(isKill))
719 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
721 case ARM::QPRRegClassID:
722 case ARM::QPR_VFP2RegClassID:
723 case ARM::QPR_8RegClassID:
724 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
725 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64Pseudo))
726 .addFrameIndex(FI).addImm(16)
727 .addReg(SrcReg, getKillRegState(isKill))
728 .addMemOperand(MMO));
730 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
731 .addReg(SrcReg, getKillRegState(isKill))
733 .addMemOperand(MMO));
736 case ARM::QQPRRegClassID:
737 case ARM::QQPR_VFP2RegClassID:
738 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
739 // FIXME: It's possible to only store part of the QQ register if the
740 // spilled def has a sub-register index.
741 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
742 .addFrameIndex(FI).addImm(16)
743 .addReg(SrcReg, getKillRegState(isKill))
744 .addMemOperand(MMO));
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 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
756 case ARM::QQQQPRRegClassID: {
757 MachineInstrBuilder MIB =
758 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
761 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
762 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
763 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
764 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
765 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
766 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
767 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
768 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
772 llvm_unreachable("Unknown regclass!");
777 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
778 int &FrameIndex) const {
779 switch (MI->getOpcode()) {
782 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
783 if (MI->getOperand(1).isFI() &&
784 MI->getOperand(2).isReg() &&
785 MI->getOperand(3).isImm() &&
786 MI->getOperand(2).getReg() == 0 &&
787 MI->getOperand(3).getImm() == 0) {
788 FrameIndex = MI->getOperand(1).getIndex();
789 return MI->getOperand(0).getReg();
797 if (MI->getOperand(1).isFI() &&
798 MI->getOperand(2).isImm() &&
799 MI->getOperand(2).getImm() == 0) {
800 FrameIndex = MI->getOperand(1).getIndex();
801 return MI->getOperand(0).getReg();
804 case ARM::VST1q64Pseudo:
805 if (MI->getOperand(0).isFI() &&
806 MI->getOperand(2).getSubReg() == 0) {
807 FrameIndex = MI->getOperand(0).getIndex();
808 return MI->getOperand(2).getReg();
812 if (MI->getOperand(1).isFI() &&
813 MI->getOperand(0).getSubReg() == 0) {
814 FrameIndex = MI->getOperand(1).getIndex();
815 return MI->getOperand(0).getReg();
823 void ARMBaseInstrInfo::
824 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
825 unsigned DestReg, int FI,
826 const TargetRegisterClass *RC,
827 const TargetRegisterInfo *TRI) const {
829 if (I != MBB.end()) DL = I->getDebugLoc();
830 MachineFunction &MF = *MBB.getParent();
831 MachineFrameInfo &MFI = *MF.getFrameInfo();
832 unsigned Align = MFI.getObjectAlignment(FI);
833 MachineMemOperand *MMO =
834 MF.getMachineMemOperand(
835 MachinePointerInfo(PseudoSourceValue::getFixedStack(FI)),
836 MachineMemOperand::MOLoad,
837 MFI.getObjectSize(FI),
840 // tGPR is used sometimes in ARM instructions that need to avoid using
841 // certain registers. Just treat it as GPR here.
842 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass
843 || RC == ARM::rGPRRegisterClass)
844 RC = ARM::GPRRegisterClass;
846 switch (RC->getID()) {
847 case ARM::GPRRegClassID:
848 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
849 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
851 case ARM::SPRRegClassID:
852 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
853 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
855 case ARM::DPRRegClassID:
856 case ARM::DPR_VFP2RegClassID:
857 case ARM::DPR_8RegClassID:
858 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
859 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
861 case ARM::QPRRegClassID:
862 case ARM::QPR_VFP2RegClassID:
863 case ARM::QPR_8RegClassID:
864 if (Align >= 16 && getRegisterInfo().needsStackRealignment(MF)) {
865 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64Pseudo), DestReg)
866 .addFrameIndex(FI).addImm(16)
867 .addMemOperand(MMO));
869 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
871 .addMemOperand(MMO));
874 case ARM::QQPRRegClassID:
875 case ARM::QQPR_VFP2RegClassID:
876 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
877 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
878 .addFrameIndex(FI).addImm(16)
879 .addMemOperand(MMO));
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 AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
891 case ARM::QQQQPRRegClassID: {
892 MachineInstrBuilder MIB =
893 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
896 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
897 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
898 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
899 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
900 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI);
901 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI);
902 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI);
903 AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
907 llvm_unreachable("Unknown regclass!");
912 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
913 int &FrameIndex) const {
914 switch (MI->getOpcode()) {
917 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
918 if (MI->getOperand(1).isFI() &&
919 MI->getOperand(2).isReg() &&
920 MI->getOperand(3).isImm() &&
921 MI->getOperand(2).getReg() == 0 &&
922 MI->getOperand(3).getImm() == 0) {
923 FrameIndex = MI->getOperand(1).getIndex();
924 return MI->getOperand(0).getReg();
932 if (MI->getOperand(1).isFI() &&
933 MI->getOperand(2).isImm() &&
934 MI->getOperand(2).getImm() == 0) {
935 FrameIndex = MI->getOperand(1).getIndex();
936 return MI->getOperand(0).getReg();
939 case ARM::VLD1q64Pseudo:
940 if (MI->getOperand(1).isFI() &&
941 MI->getOperand(0).getSubReg() == 0) {
942 FrameIndex = MI->getOperand(1).getIndex();
943 return MI->getOperand(0).getReg();
947 if (MI->getOperand(1).isFI() &&
948 MI->getOperand(0).getSubReg() == 0) {
949 FrameIndex = MI->getOperand(1).getIndex();
950 return MI->getOperand(0).getReg();
959 ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
960 int FrameIx, uint64_t Offset,
963 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
964 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
968 /// Create a copy of a const pool value. Update CPI to the new index and return
970 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
971 MachineConstantPool *MCP = MF.getConstantPool();
972 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
974 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
975 assert(MCPE.isMachineConstantPoolEntry() &&
976 "Expecting a machine constantpool entry!");
977 ARMConstantPoolValue *ACPV =
978 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
980 unsigned PCLabelId = AFI->createPICLabelUId();
981 ARMConstantPoolValue *NewCPV = 0;
982 // FIXME: The below assumes PIC relocation model and that the function
983 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
984 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
985 // instructions, so that's probably OK, but is PIC always correct when
987 if (ACPV->isGlobalValue())
988 NewCPV = new ARMConstantPoolValue(ACPV->getGV(), PCLabelId,
990 else if (ACPV->isExtSymbol())
991 NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(),
992 ACPV->getSymbol(), PCLabelId, 4);
993 else if (ACPV->isBlockAddress())
994 NewCPV = new ARMConstantPoolValue(ACPV->getBlockAddress(), PCLabelId,
995 ARMCP::CPBlockAddress, 4);
996 else if (ACPV->isLSDA())
997 NewCPV = new ARMConstantPoolValue(MF.getFunction(), PCLabelId,
1000 llvm_unreachable("Unexpected ARM constantpool value type!!");
1001 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1005 void ARMBaseInstrInfo::
1006 reMaterialize(MachineBasicBlock &MBB,
1007 MachineBasicBlock::iterator I,
1008 unsigned DestReg, unsigned SubIdx,
1009 const MachineInstr *Orig,
1010 const TargetRegisterInfo &TRI) const {
1011 unsigned Opcode = Orig->getOpcode();
1014 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1015 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1019 case ARM::tLDRpci_pic:
1020 case ARM::t2LDRpci_pic: {
1021 MachineFunction &MF = *MBB.getParent();
1022 unsigned CPI = Orig->getOperand(1).getIndex();
1023 unsigned PCLabelId = duplicateCPV(MF, CPI);
1024 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1026 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1027 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1034 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1035 MachineInstr *MI = TargetInstrInfoImpl::duplicate(Orig, MF);
1036 switch(Orig->getOpcode()) {
1037 case ARM::tLDRpci_pic:
1038 case ARM::t2LDRpci_pic: {
1039 unsigned CPI = Orig->getOperand(1).getIndex();
1040 unsigned PCLabelId = duplicateCPV(MF, CPI);
1041 Orig->getOperand(1).setIndex(CPI);
1042 Orig->getOperand(2).setImm(PCLabelId);
1049 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1050 const MachineInstr *MI1,
1051 const MachineRegisterInfo *MRI) const {
1052 int Opcode = MI0->getOpcode();
1053 if (Opcode == ARM::t2LDRpci ||
1054 Opcode == ARM::t2LDRpci_pic ||
1055 Opcode == ARM::tLDRpci ||
1056 Opcode == ARM::tLDRpci_pic ||
1057 Opcode == ARM::MOV_ga_dyn ||
1058 Opcode == ARM::MOV_ga_pcrel ||
1059 Opcode == ARM::MOV_ga_pcrel_ldr ||
1060 Opcode == ARM::t2MOV_ga_dyn ||
1061 Opcode == ARM::t2MOV_ga_pcrel) {
1062 if (MI1->getOpcode() != Opcode)
1064 if (MI0->getNumOperands() != MI1->getNumOperands())
1067 const MachineOperand &MO0 = MI0->getOperand(1);
1068 const MachineOperand &MO1 = MI1->getOperand(1);
1069 if (MO0.getOffset() != MO1.getOffset())
1072 if (Opcode == ARM::MOV_ga_dyn ||
1073 Opcode == ARM::MOV_ga_pcrel ||
1074 Opcode == ARM::MOV_ga_pcrel_ldr ||
1075 Opcode == ARM::t2MOV_ga_dyn ||
1076 Opcode == ARM::t2MOV_ga_pcrel)
1077 // Ignore the PC labels.
1078 return MO0.getGlobal() == MO1.getGlobal();
1080 const MachineFunction *MF = MI0->getParent()->getParent();
1081 const MachineConstantPool *MCP = MF->getConstantPool();
1082 int CPI0 = MO0.getIndex();
1083 int CPI1 = MO1.getIndex();
1084 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1085 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1086 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1087 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1088 if (isARMCP0 && isARMCP1) {
1089 ARMConstantPoolValue *ACPV0 =
1090 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1091 ARMConstantPoolValue *ACPV1 =
1092 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1093 return ACPV0->hasSameValue(ACPV1);
1094 } else if (!isARMCP0 && !isARMCP1) {
1095 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1098 } else if (Opcode == ARM::PICLDR) {
1099 if (MI1->getOpcode() != Opcode)
1101 if (MI0->getNumOperands() != MI1->getNumOperands())
1104 unsigned Addr0 = MI0->getOperand(1).getReg();
1105 unsigned Addr1 = MI1->getOperand(1).getReg();
1106 if (Addr0 != Addr1) {
1108 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1109 !TargetRegisterInfo::isVirtualRegister(Addr1))
1112 // This assumes SSA form.
1113 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1114 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1115 // Check if the loaded value, e.g. a constantpool of a global address, are
1117 if (!produceSameValue(Def0, Def1, MRI))
1121 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1122 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1123 const MachineOperand &MO0 = MI0->getOperand(i);
1124 const MachineOperand &MO1 = MI1->getOperand(i);
1125 if (!MO0.isIdenticalTo(MO1))
1131 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1134 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1135 /// determine if two loads are loading from the same base address. It should
1136 /// only return true if the base pointers are the same and the only differences
1137 /// between the two addresses is the offset. It also returns the offsets by
1139 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1141 int64_t &Offset2) const {
1142 // Don't worry about Thumb: just ARM and Thumb2.
1143 if (Subtarget.isThumb1Only()) return false;
1145 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1148 switch (Load1->getMachineOpcode()) {
1161 case ARM::t2LDRSHi8:
1163 case ARM::t2LDRSHi12:
1167 switch (Load2->getMachineOpcode()) {
1180 case ARM::t2LDRSHi8:
1182 case ARM::t2LDRSHi12:
1186 // Check if base addresses and chain operands match.
1187 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1188 Load1->getOperand(4) != Load2->getOperand(4))
1191 // Index should be Reg0.
1192 if (Load1->getOperand(3) != Load2->getOperand(3))
1195 // Determine the offsets.
1196 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1197 isa<ConstantSDNode>(Load2->getOperand(1))) {
1198 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1199 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1206 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1207 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1208 /// be scheduled togther. On some targets if two loads are loading from
1209 /// addresses in the same cache line, it's better if they are scheduled
1210 /// together. This function takes two integers that represent the load offsets
1211 /// from the common base address. It returns true if it decides it's desirable
1212 /// to schedule the two loads together. "NumLoads" is the number of loads that
1213 /// have already been scheduled after Load1.
1214 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1215 int64_t Offset1, int64_t Offset2,
1216 unsigned NumLoads) const {
1217 // Don't worry about Thumb: just ARM and Thumb2.
1218 if (Subtarget.isThumb1Only()) return false;
1220 assert(Offset2 > Offset1);
1222 if ((Offset2 - Offset1) / 8 > 64)
1225 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1226 return false; // FIXME: overly conservative?
1228 // Four loads in a row should be sufficient.
1235 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1236 const MachineBasicBlock *MBB,
1237 const MachineFunction &MF) const {
1238 // Debug info is never a scheduling boundary. It's necessary to be explicit
1239 // due to the special treatment of IT instructions below, otherwise a
1240 // dbg_value followed by an IT will result in the IT instruction being
1241 // considered a scheduling hazard, which is wrong. It should be the actual
1242 // instruction preceding the dbg_value instruction(s), just like it is
1243 // when debug info is not present.
1244 if (MI->isDebugValue())
1247 // Terminators and labels can't be scheduled around.
1248 if (MI->getDesc().isTerminator() || MI->isLabel())
1251 // Treat the start of the IT block as a scheduling boundary, but schedule
1252 // t2IT along with all instructions following it.
1253 // FIXME: This is a big hammer. But the alternative is to add all potential
1254 // true and anti dependencies to IT block instructions as implicit operands
1255 // to the t2IT instruction. The added compile time and complexity does not
1257 MachineBasicBlock::const_iterator I = MI;
1258 // Make sure to skip any dbg_value instructions
1259 while (++I != MBB->end() && I->isDebugValue())
1261 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1264 // Don't attempt to schedule around any instruction that defines
1265 // a stack-oriented pointer, as it's unlikely to be profitable. This
1266 // saves compile time, because it doesn't require every single
1267 // stack slot reference to depend on the instruction that does the
1269 if (MI->definesRegister(ARM::SP))
1275 bool ARMBaseInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
1277 unsigned ExtraPredCycles,
1279 float Confidence) const {
1283 // Attempt to estimate the relative costs of predication versus branching.
1284 float UnpredCost = Probability * NumCycles;
1285 UnpredCost += 1.0; // The branch itself
1286 UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
1288 return (float)(NumCycles + ExtraPredCycles) < UnpredCost;
1291 bool ARMBaseInstrInfo::
1292 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1293 unsigned TCycles, unsigned TExtra,
1294 MachineBasicBlock &FMBB,
1295 unsigned FCycles, unsigned FExtra,
1296 float Probability, float Confidence) const {
1297 if (!TCycles || !FCycles)
1300 // Attempt to estimate the relative costs of predication versus branching.
1301 float UnpredCost = Probability * TCycles + (1.0 - Probability) * FCycles;
1302 UnpredCost += 1.0; // The branch itself
1303 UnpredCost += (1.0 - Confidence) * Subtarget.getMispredictionPenalty();
1305 return (float)(TCycles + FCycles + TExtra + FExtra) < UnpredCost;
1308 /// getInstrPredicate - If instruction is predicated, returns its predicate
1309 /// condition, otherwise returns AL. It also returns the condition code
1310 /// register by reference.
1312 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1313 int PIdx = MI->findFirstPredOperandIdx();
1319 PredReg = MI->getOperand(PIdx+1).getReg();
1320 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1324 int llvm::getMatchingCondBranchOpcode(int Opc) {
1327 else if (Opc == ARM::tB)
1329 else if (Opc == ARM::t2B)
1332 llvm_unreachable("Unknown unconditional branch opcode!");
1337 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1338 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1339 unsigned DestReg, unsigned BaseReg, int NumBytes,
1340 ARMCC::CondCodes Pred, unsigned PredReg,
1341 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1342 bool isSub = NumBytes < 0;
1343 if (isSub) NumBytes = -NumBytes;
1346 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1347 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1348 assert(ThisVal && "Didn't extract field correctly");
1350 // We will handle these bits from offset, clear them.
1351 NumBytes &= ~ThisVal;
1353 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1355 // Build the new ADD / SUB.
1356 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1357 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1358 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1359 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1360 .setMIFlags(MIFlags);
1365 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1366 unsigned FrameReg, int &Offset,
1367 const ARMBaseInstrInfo &TII) {
1368 unsigned Opcode = MI.getOpcode();
1369 const MCInstrDesc &Desc = MI.getDesc();
1370 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1373 // Memory operands in inline assembly always use AddrMode2.
1374 if (Opcode == ARM::INLINEASM)
1375 AddrMode = ARMII::AddrMode2;
1377 if (Opcode == ARM::ADDri) {
1378 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1380 // Turn it into a move.
1381 MI.setDesc(TII.get(ARM::MOVr));
1382 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1383 MI.RemoveOperand(FrameRegIdx+1);
1386 } else if (Offset < 0) {
1389 MI.setDesc(TII.get(ARM::SUBri));
1392 // Common case: small offset, fits into instruction.
1393 if (ARM_AM::getSOImmVal(Offset) != -1) {
1394 // Replace the FrameIndex with sp / fp
1395 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1396 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1401 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1403 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1404 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1406 // We will handle these bits from offset, clear them.
1407 Offset &= ~ThisImmVal;
1409 // Get the properly encoded SOImmVal field.
1410 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1411 "Bit extraction didn't work?");
1412 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1414 unsigned ImmIdx = 0;
1416 unsigned NumBits = 0;
1419 case ARMII::AddrMode_i12: {
1420 ImmIdx = FrameRegIdx + 1;
1421 InstrOffs = MI.getOperand(ImmIdx).getImm();
1425 case ARMII::AddrMode2: {
1426 ImmIdx = FrameRegIdx+2;
1427 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1428 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1433 case ARMII::AddrMode3: {
1434 ImmIdx = FrameRegIdx+2;
1435 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1436 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1441 case ARMII::AddrMode4:
1442 case ARMII::AddrMode6:
1443 // Can't fold any offset even if it's zero.
1445 case ARMII::AddrMode5: {
1446 ImmIdx = FrameRegIdx+1;
1447 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1448 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1455 llvm_unreachable("Unsupported addressing mode!");
1459 Offset += InstrOffs * Scale;
1460 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1466 // Attempt to fold address comp. if opcode has offset bits
1468 // Common case: small offset, fits into instruction.
1469 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1470 int ImmedOffset = Offset / Scale;
1471 unsigned Mask = (1 << NumBits) - 1;
1472 if ((unsigned)Offset <= Mask * Scale) {
1473 // Replace the FrameIndex with sp
1474 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1475 // FIXME: When addrmode2 goes away, this will simplify (like the
1476 // T2 version), as the LDR.i12 versions don't need the encoding
1477 // tricks for the offset value.
1479 if (AddrMode == ARMII::AddrMode_i12)
1480 ImmedOffset = -ImmedOffset;
1482 ImmedOffset |= 1 << NumBits;
1484 ImmOp.ChangeToImmediate(ImmedOffset);
1489 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1490 ImmedOffset = ImmedOffset & Mask;
1492 if (AddrMode == ARMII::AddrMode_i12)
1493 ImmedOffset = -ImmedOffset;
1495 ImmedOffset |= 1 << NumBits;
1497 ImmOp.ChangeToImmediate(ImmedOffset);
1498 Offset &= ~(Mask*Scale);
1502 Offset = (isSub) ? -Offset : Offset;
1506 bool ARMBaseInstrInfo::
1507 AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpMask,
1508 int &CmpValue) const {
1509 switch (MI->getOpcode()) {
1513 SrcReg = MI->getOperand(0).getReg();
1515 CmpValue = MI->getOperand(1).getImm();
1519 SrcReg = MI->getOperand(0).getReg();
1520 CmpMask = MI->getOperand(1).getImm();
1528 /// isSuitableForMask - Identify a suitable 'and' instruction that
1529 /// operates on the given source register and applies the same mask
1530 /// as a 'tst' instruction. Provide a limited look-through for copies.
1531 /// When successful, MI will hold the found instruction.
1532 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
1533 int CmpMask, bool CommonUse) {
1534 switch (MI->getOpcode()) {
1537 if (CmpMask != MI->getOperand(2).getImm())
1539 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
1543 // Walk down one instruction which is potentially an 'and'.
1544 const MachineInstr &Copy = *MI;
1545 MachineBasicBlock::iterator AND(
1546 llvm::next(MachineBasicBlock::iterator(MI)));
1547 if (AND == MI->getParent()->end()) return false;
1549 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
1557 /// OptimizeCompareInstr - Convert the instruction supplying the argument to the
1558 /// comparison into one that sets the zero bit in the flags register.
1559 bool ARMBaseInstrInfo::
1560 OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
1561 int CmpValue, const MachineRegisterInfo *MRI) const {
1565 MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
1566 if (llvm::next(DI) != MRI->def_end())
1567 // Only support one definition.
1570 MachineInstr *MI = &*DI;
1572 // Masked compares sometimes use the same register as the corresponding 'and'.
1573 if (CmpMask != ~0) {
1574 if (!isSuitableForMask(MI, SrcReg, CmpMask, false)) {
1576 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
1577 UE = MRI->use_end(); UI != UE; ++UI) {
1578 if (UI->getParent() != CmpInstr->getParent()) continue;
1579 MachineInstr *PotentialAND = &*UI;
1580 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true))
1585 if (!MI) return false;
1589 // Conservatively refuse to convert an instruction which isn't in the same BB
1590 // as the comparison.
1591 if (MI->getParent() != CmpInstr->getParent())
1594 // Check that CPSR isn't set between the comparison instruction and the one we
1596 MachineBasicBlock::const_iterator I = CmpInstr, E = MI,
1597 B = MI->getParent()->begin();
1599 // Early exit if CmpInstr is at the beginning of the BB.
1600 if (I == B) return false;
1603 for (; I != E; --I) {
1604 const MachineInstr &Instr = *I;
1606 for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) {
1607 const MachineOperand &MO = Instr.getOperand(IO);
1608 if (!MO.isReg()) continue;
1610 // This instruction modifies or uses CPSR after the one we want to
1611 // change. We can't do this transformation.
1612 if (MO.getReg() == ARM::CPSR)
1617 // The 'and' is below the comparison instruction.
1621 // Set the "zero" bit in CPSR.
1622 switch (MI->getOpcode()) {
1656 case ARM::t2EORri: {
1657 // Scan forward for the use of CPSR, if it's a conditional code requires
1658 // checking of V bit, then this is not safe to do. If we can't find the
1659 // CPSR use (i.e. used in another block), then it's not safe to perform
1660 // the optimization.
1661 bool isSafe = false;
1663 E = MI->getParent()->end();
1664 while (!isSafe && ++I != E) {
1665 const MachineInstr &Instr = *I;
1666 for (unsigned IO = 0, EO = Instr.getNumOperands();
1667 !isSafe && IO != EO; ++IO) {
1668 const MachineOperand &MO = Instr.getOperand(IO);
1669 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
1675 // Condition code is after the operand before CPSR.
1676 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
1695 // Toggle the optional operand to CPSR.
1696 MI->getOperand(5).setReg(ARM::CPSR);
1697 MI->getOperand(5).setIsDef(true);
1698 CmpInstr->eraseFromParent();
1706 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
1707 MachineInstr *DefMI, unsigned Reg,
1708 MachineRegisterInfo *MRI) const {
1709 // Fold large immediates into add, sub, or, xor.
1710 unsigned DefOpc = DefMI->getOpcode();
1711 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
1713 if (!DefMI->getOperand(1).isImm())
1714 // Could be t2MOVi32imm <ga:xx>
1717 if (!MRI->hasOneNonDBGUse(Reg))
1720 unsigned UseOpc = UseMI->getOpcode();
1721 unsigned NewUseOpc = 0;
1722 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
1723 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
1724 bool Commute = false;
1726 default: return false;
1734 case ARM::t2EORrr: {
1735 Commute = UseMI->getOperand(2).getReg() != Reg;
1742 NewUseOpc = ARM::SUBri;
1748 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
1750 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
1751 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
1754 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
1755 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
1756 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
1760 case ARM::t2SUBrr: {
1764 NewUseOpc = ARM::t2SUBri;
1769 case ARM::t2EORrr: {
1770 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
1772 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
1773 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
1776 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
1777 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
1778 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
1786 unsigned OpIdx = Commute ? 2 : 1;
1787 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
1788 bool isKill = UseMI->getOperand(OpIdx).isKill();
1789 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
1790 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
1791 *UseMI, UseMI->getDebugLoc(),
1792 get(NewUseOpc), NewReg)
1793 .addReg(Reg1, getKillRegState(isKill))
1794 .addImm(SOImmValV1)));
1795 UseMI->setDesc(get(NewUseOpc));
1796 UseMI->getOperand(1).setReg(NewReg);
1797 UseMI->getOperand(1).setIsKill();
1798 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
1799 DefMI->eraseFromParent();
1804 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1805 const MachineInstr *MI) const {
1806 if (!ItinData || ItinData->isEmpty())
1809 const MCInstrDesc &Desc = MI->getDesc();
1810 unsigned Class = Desc.getSchedClass();
1811 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
1815 unsigned Opc = MI->getOpcode();
1818 llvm_unreachable("Unexpected multi-uops instruction!");
1824 // The number of uOps for load / store multiple are determined by the number
1827 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
1828 // same cycle. The scheduling for the first load / store must be done
1829 // separately by assuming the the address is not 64-bit aligned.
1831 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
1832 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
1833 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
1835 case ARM::VLDMDIA_UPD:
1836 case ARM::VLDMDDB_UPD:
1838 case ARM::VLDMSIA_UPD:
1839 case ARM::VLDMSDB_UPD:
1841 case ARM::VSTMDIA_UPD:
1842 case ARM::VSTMDDB_UPD:
1844 case ARM::VSTMSIA_UPD:
1845 case ARM::VSTMSDB_UPD: {
1846 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
1847 return (NumRegs / 2) + (NumRegs % 2) + 1;
1850 case ARM::LDMIA_RET:
1855 case ARM::LDMIA_UPD:
1856 case ARM::LDMDA_UPD:
1857 case ARM::LDMDB_UPD:
1858 case ARM::LDMIB_UPD:
1863 case ARM::STMIA_UPD:
1864 case ARM::STMDA_UPD:
1865 case ARM::STMDB_UPD:
1866 case ARM::STMIB_UPD:
1868 case ARM::tLDMIA_UPD:
1870 case ARM::tSTMIA_UPD:
1874 case ARM::t2LDMIA_RET:
1877 case ARM::t2LDMIA_UPD:
1878 case ARM::t2LDMDB_UPD:
1881 case ARM::t2STMIA_UPD:
1882 case ARM::t2STMDB_UPD: {
1883 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
1884 if (Subtarget.isCortexA8()) {
1887 // 4 registers would be issued: 2, 2.
1888 // 5 registers would be issued: 2, 2, 1.
1889 UOps = (NumRegs / 2);
1893 } else if (Subtarget.isCortexA9()) {
1894 UOps = (NumRegs / 2);
1895 // If there are odd number of registers or if it's not 64-bit aligned,
1896 // then it takes an extra AGU (Address Generation Unit) cycle.
1897 if ((NumRegs % 2) ||
1898 !MI->hasOneMemOperand() ||
1899 (*MI->memoperands_begin())->getAlignment() < 8)
1903 // Assume the worst.
1911 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
1912 const MCInstrDesc &DefMCID,
1914 unsigned DefIdx, unsigned DefAlign) const {
1915 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1917 // Def is the address writeback.
1918 return ItinData->getOperandCycle(DefClass, DefIdx);
1921 if (Subtarget.isCortexA8()) {
1922 // (regno / 2) + (regno % 2) + 1
1923 DefCycle = RegNo / 2 + 1;
1926 } else if (Subtarget.isCortexA9()) {
1928 bool isSLoad = false;
1930 switch (DefMCID.getOpcode()) {
1933 case ARM::VLDMSIA_UPD:
1934 case ARM::VLDMSDB_UPD:
1939 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
1940 // then it takes an extra cycle.
1941 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
1944 // Assume the worst.
1945 DefCycle = RegNo + 2;
1952 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
1953 const MCInstrDesc &DefMCID,
1955 unsigned DefIdx, unsigned DefAlign) const {
1956 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
1958 // Def is the address writeback.
1959 return ItinData->getOperandCycle(DefClass, DefIdx);
1962 if (Subtarget.isCortexA8()) {
1963 // 4 registers would be issued: 1, 2, 1.
1964 // 5 registers would be issued: 1, 2, 2.
1965 DefCycle = RegNo / 2;
1968 // Result latency is issue cycle + 2: E2.
1970 } else if (Subtarget.isCortexA9()) {
1971 DefCycle = (RegNo / 2);
1972 // If there are odd number of registers or if it's not 64-bit aligned,
1973 // then it takes an extra AGU (Address Generation Unit) cycle.
1974 if ((RegNo % 2) || DefAlign < 8)
1976 // Result latency is AGU cycles + 2.
1979 // Assume the worst.
1980 DefCycle = RegNo + 2;
1987 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
1988 const MCInstrDesc &UseMCID,
1990 unsigned UseIdx, unsigned UseAlign) const {
1991 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
1993 return ItinData->getOperandCycle(UseClass, UseIdx);
1996 if (Subtarget.isCortexA8()) {
1997 // (regno / 2) + (regno % 2) + 1
1998 UseCycle = RegNo / 2 + 1;
2001 } else if (Subtarget.isCortexA9()) {
2003 bool isSStore = false;
2005 switch (UseMCID.getOpcode()) {
2008 case ARM::VSTMSIA_UPD:
2009 case ARM::VSTMSDB_UPD:
2014 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2015 // then it takes an extra cycle.
2016 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2019 // Assume the worst.
2020 UseCycle = RegNo + 2;
2027 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2028 const MCInstrDesc &UseMCID,
2030 unsigned UseIdx, unsigned UseAlign) const {
2031 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2033 return ItinData->getOperandCycle(UseClass, UseIdx);
2036 if (Subtarget.isCortexA8()) {
2037 UseCycle = RegNo / 2;
2042 } else if (Subtarget.isCortexA9()) {
2043 UseCycle = (RegNo / 2);
2044 // If there are odd number of registers or if it's not 64-bit aligned,
2045 // then it takes an extra AGU (Address Generation Unit) cycle.
2046 if ((RegNo % 2) || UseAlign < 8)
2049 // Assume the worst.
2056 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2057 const MCInstrDesc &DefMCID,
2058 unsigned DefIdx, unsigned DefAlign,
2059 const MCInstrDesc &UseMCID,
2060 unsigned UseIdx, unsigned UseAlign) const {
2061 unsigned DefClass = DefMCID.getSchedClass();
2062 unsigned UseClass = UseMCID.getSchedClass();
2064 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2065 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2067 // This may be a def / use of a variable_ops instruction, the operand
2068 // latency might be determinable dynamically. Let the target try to
2071 bool LdmBypass = false;
2072 switch (DefMCID.getOpcode()) {
2074 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2078 case ARM::VLDMDIA_UPD:
2079 case ARM::VLDMDDB_UPD:
2081 case ARM::VLDMSIA_UPD:
2082 case ARM::VLDMSDB_UPD:
2083 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2086 case ARM::LDMIA_RET:
2091 case ARM::LDMIA_UPD:
2092 case ARM::LDMDA_UPD:
2093 case ARM::LDMDB_UPD:
2094 case ARM::LDMIB_UPD:
2096 case ARM::tLDMIA_UPD:
2098 case ARM::t2LDMIA_RET:
2101 case ARM::t2LDMIA_UPD:
2102 case ARM::t2LDMDB_UPD:
2104 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2109 // We can't seem to determine the result latency of the def, assume it's 2.
2113 switch (UseMCID.getOpcode()) {
2115 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
2119 case ARM::VSTMDIA_UPD:
2120 case ARM::VSTMDDB_UPD:
2122 case ARM::VSTMSIA_UPD:
2123 case ARM::VSTMSDB_UPD:
2124 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2131 case ARM::STMIA_UPD:
2132 case ARM::STMDA_UPD:
2133 case ARM::STMDB_UPD:
2134 case ARM::STMIB_UPD:
2136 case ARM::tSTMIA_UPD:
2141 case ARM::t2STMIA_UPD:
2142 case ARM::t2STMDB_UPD:
2143 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2148 // Assume it's read in the first stage.
2151 UseCycle = DefCycle - UseCycle + 1;
2154 // It's a variable_ops instruction so we can't use DefIdx here. Just use
2155 // first def operand.
2156 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
2159 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
2160 UseClass, UseIdx)) {
2169 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2170 const MachineInstr *DefMI, unsigned DefIdx,
2171 const MachineInstr *UseMI, unsigned UseIdx) const {
2172 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
2173 DefMI->isRegSequence() || DefMI->isImplicitDef())
2176 const MCInstrDesc &DefMCID = DefMI->getDesc();
2177 if (!ItinData || ItinData->isEmpty())
2178 return DefMCID.mayLoad() ? 3 : 1;
2180 const MCInstrDesc &UseMCID = UseMI->getDesc();
2181 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
2182 if (DefMO.getReg() == ARM::CPSR) {
2183 if (DefMI->getOpcode() == ARM::FMSTAT) {
2184 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
2185 return Subtarget.isCortexA9() ? 1 : 20;
2188 // CPSR set and branch can be paired in the same cycle.
2189 if (UseMCID.isBranch())
2193 unsigned DefAlign = DefMI->hasOneMemOperand()
2194 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
2195 unsigned UseAlign = UseMI->hasOneMemOperand()
2196 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
2197 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2198 UseMCID, UseIdx, UseAlign);
2201 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2202 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2203 // variants are one cycle cheaper.
2204 switch (DefMCID.getOpcode()) {
2208 unsigned ShOpVal = DefMI->getOperand(3).getImm();
2209 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2211 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2218 case ARM::t2LDRSHs: {
2219 // Thumb2 mode: lsl only.
2220 unsigned ShAmt = DefMI->getOperand(3).getImm();
2221 if (ShAmt == 0 || ShAmt == 2)
2228 if (DefAlign < 8 && Subtarget.isCortexA9())
2229 switch (DefMCID.getOpcode()) {
2235 case ARM::VLD1q8_UPD:
2236 case ARM::VLD1q16_UPD:
2237 case ARM::VLD1q32_UPD:
2238 case ARM::VLD1q64_UPD:
2245 case ARM::VLD2d8_UPD:
2246 case ARM::VLD2d16_UPD:
2247 case ARM::VLD2d32_UPD:
2248 case ARM::VLD2q8_UPD:
2249 case ARM::VLD2q16_UPD:
2250 case ARM::VLD2q32_UPD:
2255 case ARM::VLD3d8_UPD:
2256 case ARM::VLD3d16_UPD:
2257 case ARM::VLD3d32_UPD:
2258 case ARM::VLD1d64T_UPD:
2259 case ARM::VLD3q8_UPD:
2260 case ARM::VLD3q16_UPD:
2261 case ARM::VLD3q32_UPD:
2266 case ARM::VLD4d8_UPD:
2267 case ARM::VLD4d16_UPD:
2268 case ARM::VLD4d32_UPD:
2269 case ARM::VLD1d64Q_UPD:
2270 case ARM::VLD4q8_UPD:
2271 case ARM::VLD4q16_UPD:
2272 case ARM::VLD4q32_UPD:
2273 case ARM::VLD1DUPq8:
2274 case ARM::VLD1DUPq16:
2275 case ARM::VLD1DUPq32:
2276 case ARM::VLD1DUPq8_UPD:
2277 case ARM::VLD1DUPq16_UPD:
2278 case ARM::VLD1DUPq32_UPD:
2279 case ARM::VLD2DUPd8:
2280 case ARM::VLD2DUPd16:
2281 case ARM::VLD2DUPd32:
2282 case ARM::VLD2DUPd8_UPD:
2283 case ARM::VLD2DUPd16_UPD:
2284 case ARM::VLD2DUPd32_UPD:
2285 case ARM::VLD4DUPd8:
2286 case ARM::VLD4DUPd16:
2287 case ARM::VLD4DUPd32:
2288 case ARM::VLD4DUPd8_UPD:
2289 case ARM::VLD4DUPd16_UPD:
2290 case ARM::VLD4DUPd32_UPD:
2292 case ARM::VLD1LNd16:
2293 case ARM::VLD1LNd32:
2294 case ARM::VLD1LNd8_UPD:
2295 case ARM::VLD1LNd16_UPD:
2296 case ARM::VLD1LNd32_UPD:
2298 case ARM::VLD2LNd16:
2299 case ARM::VLD2LNd32:
2300 case ARM::VLD2LNq16:
2301 case ARM::VLD2LNq32:
2302 case ARM::VLD2LNd8_UPD:
2303 case ARM::VLD2LNd16_UPD:
2304 case ARM::VLD2LNd32_UPD:
2305 case ARM::VLD2LNq16_UPD:
2306 case ARM::VLD2LNq32_UPD:
2308 case ARM::VLD4LNd16:
2309 case ARM::VLD4LNd32:
2310 case ARM::VLD4LNq16:
2311 case ARM::VLD4LNq32:
2312 case ARM::VLD4LNd8_UPD:
2313 case ARM::VLD4LNd16_UPD:
2314 case ARM::VLD4LNd32_UPD:
2315 case ARM::VLD4LNq16_UPD:
2316 case ARM::VLD4LNq32_UPD:
2317 // If the address is not 64-bit aligned, the latencies of these
2318 // instructions increases by one.
2327 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2328 SDNode *DefNode, unsigned DefIdx,
2329 SDNode *UseNode, unsigned UseIdx) const {
2330 if (!DefNode->isMachineOpcode())
2333 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
2335 if (isZeroCost(DefMCID.Opcode))
2338 if (!ItinData || ItinData->isEmpty())
2339 return DefMCID.mayLoad() ? 3 : 1;
2341 if (!UseNode->isMachineOpcode()) {
2342 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
2343 if (Subtarget.isCortexA9())
2344 return Latency <= 2 ? 1 : Latency - 1;
2346 return Latency <= 3 ? 1 : Latency - 2;
2349 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
2350 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
2351 unsigned DefAlign = !DefMN->memoperands_empty()
2352 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
2353 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
2354 unsigned UseAlign = !UseMN->memoperands_empty()
2355 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
2356 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2357 UseMCID, UseIdx, UseAlign);
2360 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2361 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2362 // variants are one cycle cheaper.
2363 switch (DefMCID.getOpcode()) {
2368 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2369 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2371 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2378 case ARM::t2LDRSHs: {
2379 // Thumb2 mode: lsl only.
2381 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2382 if (ShAmt == 0 || ShAmt == 2)
2389 if (DefAlign < 8 && Subtarget.isCortexA9())
2390 switch (DefMCID.getOpcode()) {
2392 case ARM::VLD1q8Pseudo:
2393 case ARM::VLD1q16Pseudo:
2394 case ARM::VLD1q32Pseudo:
2395 case ARM::VLD1q64Pseudo:
2396 case ARM::VLD1q8Pseudo_UPD:
2397 case ARM::VLD1q16Pseudo_UPD:
2398 case ARM::VLD1q32Pseudo_UPD:
2399 case ARM::VLD1q64Pseudo_UPD:
2400 case ARM::VLD2d8Pseudo:
2401 case ARM::VLD2d16Pseudo:
2402 case ARM::VLD2d32Pseudo:
2403 case ARM::VLD2q8Pseudo:
2404 case ARM::VLD2q16Pseudo:
2405 case ARM::VLD2q32Pseudo:
2406 case ARM::VLD2d8Pseudo_UPD:
2407 case ARM::VLD2d16Pseudo_UPD:
2408 case ARM::VLD2d32Pseudo_UPD:
2409 case ARM::VLD2q8Pseudo_UPD:
2410 case ARM::VLD2q16Pseudo_UPD:
2411 case ARM::VLD2q32Pseudo_UPD:
2412 case ARM::VLD3d8Pseudo:
2413 case ARM::VLD3d16Pseudo:
2414 case ARM::VLD3d32Pseudo:
2415 case ARM::VLD1d64TPseudo:
2416 case ARM::VLD3d8Pseudo_UPD:
2417 case ARM::VLD3d16Pseudo_UPD:
2418 case ARM::VLD3d32Pseudo_UPD:
2419 case ARM::VLD1d64TPseudo_UPD:
2420 case ARM::VLD3q8Pseudo_UPD:
2421 case ARM::VLD3q16Pseudo_UPD:
2422 case ARM::VLD3q32Pseudo_UPD:
2423 case ARM::VLD3q8oddPseudo:
2424 case ARM::VLD3q16oddPseudo:
2425 case ARM::VLD3q32oddPseudo:
2426 case ARM::VLD3q8oddPseudo_UPD:
2427 case ARM::VLD3q16oddPseudo_UPD:
2428 case ARM::VLD3q32oddPseudo_UPD:
2429 case ARM::VLD4d8Pseudo:
2430 case ARM::VLD4d16Pseudo:
2431 case ARM::VLD4d32Pseudo:
2432 case ARM::VLD1d64QPseudo:
2433 case ARM::VLD4d8Pseudo_UPD:
2434 case ARM::VLD4d16Pseudo_UPD:
2435 case ARM::VLD4d32Pseudo_UPD:
2436 case ARM::VLD1d64QPseudo_UPD:
2437 case ARM::VLD4q8Pseudo_UPD:
2438 case ARM::VLD4q16Pseudo_UPD:
2439 case ARM::VLD4q32Pseudo_UPD:
2440 case ARM::VLD4q8oddPseudo:
2441 case ARM::VLD4q16oddPseudo:
2442 case ARM::VLD4q32oddPseudo:
2443 case ARM::VLD4q8oddPseudo_UPD:
2444 case ARM::VLD4q16oddPseudo_UPD:
2445 case ARM::VLD4q32oddPseudo_UPD:
2446 case ARM::VLD1DUPq8Pseudo:
2447 case ARM::VLD1DUPq16Pseudo:
2448 case ARM::VLD1DUPq32Pseudo:
2449 case ARM::VLD1DUPq8Pseudo_UPD:
2450 case ARM::VLD1DUPq16Pseudo_UPD:
2451 case ARM::VLD1DUPq32Pseudo_UPD:
2452 case ARM::VLD2DUPd8Pseudo:
2453 case ARM::VLD2DUPd16Pseudo:
2454 case ARM::VLD2DUPd32Pseudo:
2455 case ARM::VLD2DUPd8Pseudo_UPD:
2456 case ARM::VLD2DUPd16Pseudo_UPD:
2457 case ARM::VLD2DUPd32Pseudo_UPD:
2458 case ARM::VLD4DUPd8Pseudo:
2459 case ARM::VLD4DUPd16Pseudo:
2460 case ARM::VLD4DUPd32Pseudo:
2461 case ARM::VLD4DUPd8Pseudo_UPD:
2462 case ARM::VLD4DUPd16Pseudo_UPD:
2463 case ARM::VLD4DUPd32Pseudo_UPD:
2464 case ARM::VLD1LNq8Pseudo:
2465 case ARM::VLD1LNq16Pseudo:
2466 case ARM::VLD1LNq32Pseudo:
2467 case ARM::VLD1LNq8Pseudo_UPD:
2468 case ARM::VLD1LNq16Pseudo_UPD:
2469 case ARM::VLD1LNq32Pseudo_UPD:
2470 case ARM::VLD2LNd8Pseudo:
2471 case ARM::VLD2LNd16Pseudo:
2472 case ARM::VLD2LNd32Pseudo:
2473 case ARM::VLD2LNq16Pseudo:
2474 case ARM::VLD2LNq32Pseudo:
2475 case ARM::VLD2LNd8Pseudo_UPD:
2476 case ARM::VLD2LNd16Pseudo_UPD:
2477 case ARM::VLD2LNd32Pseudo_UPD:
2478 case ARM::VLD2LNq16Pseudo_UPD:
2479 case ARM::VLD2LNq32Pseudo_UPD:
2480 case ARM::VLD4LNd8Pseudo:
2481 case ARM::VLD4LNd16Pseudo:
2482 case ARM::VLD4LNd32Pseudo:
2483 case ARM::VLD4LNq16Pseudo:
2484 case ARM::VLD4LNq32Pseudo:
2485 case ARM::VLD4LNd8Pseudo_UPD:
2486 case ARM::VLD4LNd16Pseudo_UPD:
2487 case ARM::VLD4LNd32Pseudo_UPD:
2488 case ARM::VLD4LNq16Pseudo_UPD:
2489 case ARM::VLD4LNq32Pseudo_UPD:
2490 // If the address is not 64-bit aligned, the latencies of these
2491 // instructions increases by one.
2499 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2500 const MachineInstr *MI,
2501 unsigned *PredCost) const {
2502 if (MI->isCopyLike() || MI->isInsertSubreg() ||
2503 MI->isRegSequence() || MI->isImplicitDef())
2506 if (!ItinData || ItinData->isEmpty())
2509 const MCInstrDesc &MCID = MI->getDesc();
2510 unsigned Class = MCID.getSchedClass();
2511 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
2512 if (PredCost && MCID.hasImplicitDefOfPhysReg(ARM::CPSR))
2513 // When predicated, CPSR is an additional source operand for CPSR updating
2514 // instructions, this apparently increases their latencies.
2517 return ItinData->getStageLatency(Class);
2518 return getNumMicroOps(ItinData, MI);
2521 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
2522 SDNode *Node) const {
2523 if (!Node->isMachineOpcode())
2526 if (!ItinData || ItinData->isEmpty())
2529 unsigned Opcode = Node->getMachineOpcode();
2532 return ItinData->getStageLatency(get(Opcode).getSchedClass());
2539 bool ARMBaseInstrInfo::
2540 hasHighOperandLatency(const InstrItineraryData *ItinData,
2541 const MachineRegisterInfo *MRI,
2542 const MachineInstr *DefMI, unsigned DefIdx,
2543 const MachineInstr *UseMI, unsigned UseIdx) const {
2544 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2545 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
2546 if (Subtarget.isCortexA8() &&
2547 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
2548 // CortexA8 VFP instructions are not pipelined.
2551 // Hoist VFP / NEON instructions with 4 or higher latency.
2552 int Latency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
2555 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
2556 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
2559 bool ARMBaseInstrInfo::
2560 hasLowDefLatency(const InstrItineraryData *ItinData,
2561 const MachineInstr *DefMI, unsigned DefIdx) const {
2562 if (!ItinData || ItinData->isEmpty())
2565 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
2566 if (DDomain == ARMII::DomainGeneral) {
2567 unsigned DefClass = DefMI->getDesc().getSchedClass();
2568 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2569 return (DefCycle != -1 && DefCycle <= 2);
2575 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
2576 unsigned &AddSubOpc,
2577 bool &NegAcc, bool &HasLane) const {
2578 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
2579 if (I == MLxEntryMap.end())
2582 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
2583 MulOpc = Entry.MulOpc;
2584 AddSubOpc = Entry.AddSubOpc;
2585 NegAcc = Entry.NegAcc;
2586 HasLane = Entry.HasLane;