1 //===-- ARMBaseInstrInfo.cpp - ARM Instruction Information ----------------===//
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 "ARMBaseRegisterInfo.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMHazardRecognizer.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "MCTargetDesc/ARMAddressingModes.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalValue.h"
24 #include "llvm/CodeGen/LiveVariables.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineJumpTableInfo.h"
29 #include "llvm/CodeGen/MachineMemOperand.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/SelectionDAGNodes.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"));
49 WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true),
50 cl::desc("Widen ARM vmovs to vmovd when possible"));
52 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
54 uint16_t MLxOpc; // MLA / MLS opcode
55 uint16_t MulOpc; // Expanded multiplication opcode
56 uint16_t AddSubOpc; // Expanded add / sub opcode
57 bool NegAcc; // True if the acc is negated before the add / sub.
58 bool HasLane; // True if instruction has an extra "lane" operand.
61 static const ARM_MLxEntry ARM_MLxTable[] = {
62 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
64 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
65 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
66 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
67 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
68 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
69 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
70 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
71 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
74 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
75 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
76 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
77 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
78 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
79 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
80 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
81 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
84 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
85 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
87 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
88 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
89 assert(false && "Duplicated entries?");
90 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
91 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
95 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
96 // currently defaults to no prepass hazard recognizer.
97 ScheduleHazardRecognizer *ARMBaseInstrInfo::
98 CreateTargetHazardRecognizer(const TargetMachine *TM,
99 const ScheduleDAG *DAG) const {
100 if (usePreRAHazardRecognizer()) {
101 const InstrItineraryData *II = TM->getInstrItineraryData();
102 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
104 return TargetInstrInfoImpl::CreateTargetHazardRecognizer(TM, DAG);
107 ScheduleHazardRecognizer *ARMBaseInstrInfo::
108 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
109 const ScheduleDAG *DAG) const {
110 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
111 return (ScheduleHazardRecognizer *)
112 new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget, DAG);
113 return TargetInstrInfoImpl::CreateTargetPostRAHazardRecognizer(II, DAG);
117 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
118 MachineBasicBlock::iterator &MBBI,
119 LiveVariables *LV) const {
120 // FIXME: Thumb2 support.
125 MachineInstr *MI = MBBI;
126 MachineFunction &MF = *MI->getParent()->getParent();
127 uint64_t TSFlags = MI->getDesc().TSFlags;
129 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
130 default: return NULL;
131 case ARMII::IndexModePre:
134 case ARMII::IndexModePost:
138 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
140 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
144 MachineInstr *UpdateMI = NULL;
145 MachineInstr *MemMI = NULL;
146 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
147 const MCInstrDesc &MCID = MI->getDesc();
148 unsigned NumOps = MCID.getNumOperands();
149 bool isLoad = !MI->mayStore();
150 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
151 const MachineOperand &Base = MI->getOperand(2);
152 const MachineOperand &Offset = MI->getOperand(NumOps-3);
153 unsigned WBReg = WB.getReg();
154 unsigned BaseReg = Base.getReg();
155 unsigned OffReg = Offset.getReg();
156 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
157 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
159 default: llvm_unreachable("Unknown indexed op!");
160 case ARMII::AddrMode2: {
161 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
162 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
164 if (ARM_AM::getSOImmVal(Amt) == -1)
165 // Can't encode it in a so_imm operand. This transformation will
166 // add more than 1 instruction. Abandon!
168 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
169 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
170 .addReg(BaseReg).addImm(Amt)
171 .addImm(Pred).addReg(0).addReg(0);
172 } else if (Amt != 0) {
173 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
174 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
175 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
176 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
177 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
178 .addImm(Pred).addReg(0).addReg(0);
180 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
181 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
182 .addReg(BaseReg).addReg(OffReg)
183 .addImm(Pred).addReg(0).addReg(0);
186 case ARMII::AddrMode3 : {
187 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
188 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
190 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
191 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
192 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
193 .addReg(BaseReg).addImm(Amt)
194 .addImm(Pred).addReg(0).addReg(0);
196 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
197 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
198 .addReg(BaseReg).addReg(OffReg)
199 .addImm(Pred).addReg(0).addReg(0);
204 std::vector<MachineInstr*> NewMIs;
207 MemMI = BuildMI(MF, MI->getDebugLoc(),
208 get(MemOpc), MI->getOperand(0).getReg())
209 .addReg(WBReg).addImm(0).addImm(Pred);
211 MemMI = BuildMI(MF, MI->getDebugLoc(),
212 get(MemOpc)).addReg(MI->getOperand(1).getReg())
213 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
214 NewMIs.push_back(MemMI);
215 NewMIs.push_back(UpdateMI);
218 MemMI = BuildMI(MF, MI->getDebugLoc(),
219 get(MemOpc), MI->getOperand(0).getReg())
220 .addReg(BaseReg).addImm(0).addImm(Pred);
222 MemMI = BuildMI(MF, MI->getDebugLoc(),
223 get(MemOpc)).addReg(MI->getOperand(1).getReg())
224 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
226 UpdateMI->getOperand(0).setIsDead();
227 NewMIs.push_back(UpdateMI);
228 NewMIs.push_back(MemMI);
231 // Transfer LiveVariables states, kill / dead info.
233 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
234 MachineOperand &MO = MI->getOperand(i);
235 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
236 unsigned Reg = MO.getReg();
238 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
240 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
242 LV->addVirtualRegisterDead(Reg, NewMI);
244 if (MO.isUse() && MO.isKill()) {
245 for (unsigned j = 0; j < 2; ++j) {
246 // Look at the two new MI's in reverse order.
247 MachineInstr *NewMI = NewMIs[j];
248 if (!NewMI->readsRegister(Reg))
250 LV->addVirtualRegisterKilled(Reg, NewMI);
251 if (VI.removeKill(MI))
252 VI.Kills.push_back(NewMI);
260 MFI->insert(MBBI, NewMIs[1]);
261 MFI->insert(MBBI, NewMIs[0]);
267 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
268 MachineBasicBlock *&FBB,
269 SmallVectorImpl<MachineOperand> &Cond,
270 bool AllowModify) const {
271 // If the block has no terminators, it just falls into the block after it.
272 MachineBasicBlock::iterator I = MBB.end();
273 if (I == MBB.begin())
276 while (I->isDebugValue()) {
277 if (I == MBB.begin())
281 if (!isUnpredicatedTerminator(I))
284 // Get the last instruction in the block.
285 MachineInstr *LastInst = I;
287 // If there is only one terminator instruction, process it.
288 unsigned LastOpc = LastInst->getOpcode();
289 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
290 if (isUncondBranchOpcode(LastOpc)) {
291 TBB = LastInst->getOperand(0).getMBB();
294 if (isCondBranchOpcode(LastOpc)) {
295 // Block ends with fall-through condbranch.
296 TBB = LastInst->getOperand(0).getMBB();
297 Cond.push_back(LastInst->getOperand(1));
298 Cond.push_back(LastInst->getOperand(2));
301 return true; // Can't handle indirect branch.
304 // Get the instruction before it if it is a terminator.
305 MachineInstr *SecondLastInst = I;
306 unsigned SecondLastOpc = SecondLastInst->getOpcode();
308 // If AllowModify is true and the block ends with two or more unconditional
309 // branches, delete all but the first unconditional branch.
310 if (AllowModify && isUncondBranchOpcode(LastOpc)) {
311 while (isUncondBranchOpcode(SecondLastOpc)) {
312 LastInst->eraseFromParent();
313 LastInst = SecondLastInst;
314 LastOpc = LastInst->getOpcode();
315 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
316 // Return now the only terminator is an unconditional branch.
317 TBB = LastInst->getOperand(0).getMBB();
321 SecondLastOpc = SecondLastInst->getOpcode();
326 // If there are three terminators, we don't know what sort of block this is.
327 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
330 // If the block ends with a B and a Bcc, handle it.
331 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
332 TBB = SecondLastInst->getOperand(0).getMBB();
333 Cond.push_back(SecondLastInst->getOperand(1));
334 Cond.push_back(SecondLastInst->getOperand(2));
335 FBB = LastInst->getOperand(0).getMBB();
339 // If the block ends with two unconditional branches, handle it. The second
340 // one is not executed, so remove it.
341 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
342 TBB = SecondLastInst->getOperand(0).getMBB();
345 I->eraseFromParent();
349 // ...likewise if it ends with a branch table followed by an unconditional
350 // branch. The branch folder can create these, and we must get rid of them for
351 // correctness of Thumb constant islands.
352 if ((isJumpTableBranchOpcode(SecondLastOpc) ||
353 isIndirectBranchOpcode(SecondLastOpc)) &&
354 isUncondBranchOpcode(LastOpc)) {
357 I->eraseFromParent();
361 // Otherwise, can't handle this.
366 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
367 MachineBasicBlock::iterator I = MBB.end();
368 if (I == MBB.begin()) return 0;
370 while (I->isDebugValue()) {
371 if (I == MBB.begin())
375 if (!isUncondBranchOpcode(I->getOpcode()) &&
376 !isCondBranchOpcode(I->getOpcode()))
379 // Remove the branch.
380 I->eraseFromParent();
384 if (I == MBB.begin()) return 1;
386 if (!isCondBranchOpcode(I->getOpcode()))
389 // Remove the branch.
390 I->eraseFromParent();
395 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
396 MachineBasicBlock *FBB,
397 const SmallVectorImpl<MachineOperand> &Cond,
399 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
400 int BOpc = !AFI->isThumbFunction()
401 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
402 int BccOpc = !AFI->isThumbFunction()
403 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
404 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
406 // Shouldn't be a fall through.
407 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
408 assert((Cond.size() == 2 || Cond.size() == 0) &&
409 "ARM branch conditions have two components!");
412 if (Cond.empty()) { // Unconditional branch?
414 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
416 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
418 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
419 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
423 // Two-way conditional branch.
424 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
425 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
427 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
429 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
433 bool ARMBaseInstrInfo::
434 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
435 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
436 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
440 bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
441 if (MI->isBundle()) {
442 MachineBasicBlock::const_instr_iterator I = MI;
443 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
444 while (++I != E && I->isInsideBundle()) {
445 int PIdx = I->findFirstPredOperandIdx();
446 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
452 int PIdx = MI->findFirstPredOperandIdx();
453 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
456 bool ARMBaseInstrInfo::
457 PredicateInstruction(MachineInstr *MI,
458 const SmallVectorImpl<MachineOperand> &Pred) const {
459 unsigned Opc = MI->getOpcode();
460 if (isUncondBranchOpcode(Opc)) {
461 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
462 MI->addOperand(MachineOperand::CreateImm(Pred[0].getImm()));
463 MI->addOperand(MachineOperand::CreateReg(Pred[1].getReg(), false));
467 int PIdx = MI->findFirstPredOperandIdx();
469 MachineOperand &PMO = MI->getOperand(PIdx);
470 PMO.setImm(Pred[0].getImm());
471 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
477 bool ARMBaseInstrInfo::
478 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
479 const SmallVectorImpl<MachineOperand> &Pred2) const {
480 if (Pred1.size() > 2 || Pred2.size() > 2)
483 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
484 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
494 return CC2 == ARMCC::HI;
496 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
498 return CC2 == ARMCC::GT;
500 return CC2 == ARMCC::LT;
504 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
505 std::vector<MachineOperand> &Pred) const {
507 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
508 const MachineOperand &MO = MI->getOperand(i);
509 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
510 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) {
519 /// isPredicable - Return true if the specified instruction can be predicated.
520 /// By default, this returns true for every instruction with a
521 /// PredicateOperand.
522 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
523 if (!MI->isPredicable())
526 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
527 ARMFunctionInfo *AFI =
528 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
529 return AFI->isThumb2Function();
534 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
535 LLVM_ATTRIBUTE_NOINLINE
536 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
538 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
540 assert(JTI < JT.size());
541 return JT[JTI].MBBs.size();
544 /// GetInstSize - Return the size of the specified MachineInstr.
546 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
547 const MachineBasicBlock &MBB = *MI->getParent();
548 const MachineFunction *MF = MBB.getParent();
549 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
551 const MCInstrDesc &MCID = MI->getDesc();
553 return MCID.getSize();
555 // If this machine instr is an inline asm, measure it.
556 if (MI->getOpcode() == ARM::INLINEASM)
557 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
560 unsigned Opc = MI->getOpcode();
562 case TargetOpcode::IMPLICIT_DEF:
563 case TargetOpcode::KILL:
564 case TargetOpcode::PROLOG_LABEL:
565 case TargetOpcode::EH_LABEL:
566 case TargetOpcode::DBG_VALUE:
568 case TargetOpcode::BUNDLE:
569 return getInstBundleLength(MI);
570 case ARM::MOVi16_ga_pcrel:
571 case ARM::MOVTi16_ga_pcrel:
572 case ARM::t2MOVi16_ga_pcrel:
573 case ARM::t2MOVTi16_ga_pcrel:
576 case ARM::t2MOVi32imm:
578 case ARM::CONSTPOOL_ENTRY:
579 // If this machine instr is a constant pool entry, its size is recorded as
581 return MI->getOperand(2).getImm();
582 case ARM::Int_eh_sjlj_longjmp:
584 case ARM::tInt_eh_sjlj_longjmp:
586 case ARM::Int_eh_sjlj_setjmp:
587 case ARM::Int_eh_sjlj_setjmp_nofp:
589 case ARM::tInt_eh_sjlj_setjmp:
590 case ARM::t2Int_eh_sjlj_setjmp:
591 case ARM::t2Int_eh_sjlj_setjmp_nofp:
599 case ARM::t2TBH_JT: {
600 // These are jumptable branches, i.e. a branch followed by an inlined
601 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
602 // entry is one byte; TBH two byte each.
603 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
604 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
605 unsigned NumOps = MCID.getNumOperands();
606 MachineOperand JTOP =
607 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
608 unsigned JTI = JTOP.getIndex();
609 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
611 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
612 assert(JTI < JT.size());
613 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
614 // 4 aligned. The assembler / linker may add 2 byte padding just before
615 // the JT entries. The size does not include this padding; the
616 // constant islands pass does separate bookkeeping for it.
617 // FIXME: If we know the size of the function is less than (1 << 16) *2
618 // bytes, we can use 16-bit entries instead. Then there won't be an
620 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
621 unsigned NumEntries = getNumJTEntries(JT, JTI);
622 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
623 // Make sure the instruction that follows TBB is 2-byte aligned.
624 // FIXME: Constant island pass should insert an "ALIGN" instruction
627 return NumEntries * EntrySize + InstSize;
630 // Otherwise, pseudo-instruction sizes are zero.
635 unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
637 MachineBasicBlock::const_instr_iterator I = MI;
638 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
639 while (++I != E && I->isInsideBundle()) {
640 assert(!I->isBundle() && "No nested bundle!");
641 Size += GetInstSizeInBytes(&*I);
646 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
647 MachineBasicBlock::iterator I, DebugLoc DL,
648 unsigned DestReg, unsigned SrcReg,
649 bool KillSrc) const {
650 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
651 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
653 if (GPRDest && GPRSrc) {
654 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
655 .addReg(SrcReg, getKillRegState(KillSrc))));
659 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
660 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
663 if (SPRDest && SPRSrc)
665 else if (GPRDest && SPRSrc)
667 else if (SPRDest && GPRSrc)
669 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
671 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
675 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
676 MIB.addReg(SrcReg, getKillRegState(KillSrc));
677 if (Opc == ARM::VORRq)
678 MIB.addReg(SrcReg, getKillRegState(KillSrc));
683 // Handle register classes that require multiple instructions.
684 unsigned BeginIdx = 0;
685 unsigned SubRegs = 0;
686 unsigned Spacing = 1;
688 // Use VORRq when possible.
689 if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
690 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 2;
691 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
692 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 4;
693 // Fall back to VMOVD.
694 else if (ARM::DPairRegClass.contains(DestReg, SrcReg))
695 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2;
696 else if (ARM::DTripleRegClass.contains(DestReg, SrcReg))
697 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3;
698 else if (ARM::DQuadRegClass.contains(DestReg, SrcReg))
699 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4;
701 else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg))
702 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2, Spacing = 2;
703 else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg))
704 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3, Spacing = 2;
705 else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg))
706 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4, Spacing = 2;
709 const TargetRegisterInfo *TRI = &getRegisterInfo();
710 MachineInstrBuilder Mov;
711 for (unsigned i = 0; i != SubRegs; ++i) {
712 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i*Spacing);
713 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i*Spacing);
714 assert(Dst && Src && "Bad sub-register");
715 Mov = AddDefaultPred(BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst)
717 // VORR takes two source operands.
718 if (Opc == ARM::VORRq)
721 // Add implicit super-register defs and kills to the last instruction.
722 Mov->addRegisterDefined(DestReg, TRI);
724 Mov->addRegisterKilled(SrcReg, TRI);
728 llvm_unreachable("Impossible reg-to-reg copy");
732 MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB,
733 unsigned Reg, unsigned SubIdx, unsigned State,
734 const TargetRegisterInfo *TRI) {
736 return MIB.addReg(Reg, State);
738 if (TargetRegisterInfo::isPhysicalRegister(Reg))
739 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
740 return MIB.addReg(Reg, State, SubIdx);
743 void ARMBaseInstrInfo::
744 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
745 unsigned SrcReg, bool isKill, int FI,
746 const TargetRegisterClass *RC,
747 const TargetRegisterInfo *TRI) const {
749 if (I != MBB.end()) DL = I->getDebugLoc();
750 MachineFunction &MF = *MBB.getParent();
751 MachineFrameInfo &MFI = *MF.getFrameInfo();
752 unsigned Align = MFI.getObjectAlignment(FI);
754 MachineMemOperand *MMO =
755 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
756 MachineMemOperand::MOStore,
757 MFI.getObjectSize(FI),
760 switch (RC->getSize()) {
762 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
763 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
764 .addReg(SrcReg, getKillRegState(isKill))
765 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
766 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
767 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
768 .addReg(SrcReg, getKillRegState(isKill))
769 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
771 llvm_unreachable("Unknown reg class!");
774 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
775 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
776 .addReg(SrcReg, getKillRegState(isKill))
777 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
779 llvm_unreachable("Unknown reg class!");
782 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
783 // Use aligned spills if the stack can be realigned.
784 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
785 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
786 .addFrameIndex(FI).addImm(16)
787 .addReg(SrcReg, getKillRegState(isKill))
788 .addMemOperand(MMO));
790 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
791 .addReg(SrcReg, getKillRegState(isKill))
793 .addMemOperand(MMO));
796 llvm_unreachable("Unknown reg class!");
799 if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
800 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
801 // FIXME: It's possible to only store part of the QQ register if the
802 // spilled def has a sub-register index.
803 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
804 .addFrameIndex(FI).addImm(16)
805 .addReg(SrcReg, getKillRegState(isKill))
806 .addMemOperand(MMO));
808 MachineInstrBuilder MIB =
809 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
812 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
813 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
814 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
815 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
818 llvm_unreachable("Unknown reg class!");
821 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
822 MachineInstrBuilder MIB =
823 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
826 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
827 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
828 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
829 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
830 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
831 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
832 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
833 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
835 llvm_unreachable("Unknown reg class!");
838 llvm_unreachable("Unknown reg class!");
843 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
844 int &FrameIndex) const {
845 switch (MI->getOpcode()) {
848 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
849 if (MI->getOperand(1).isFI() &&
850 MI->getOperand(2).isReg() &&
851 MI->getOperand(3).isImm() &&
852 MI->getOperand(2).getReg() == 0 &&
853 MI->getOperand(3).getImm() == 0) {
854 FrameIndex = MI->getOperand(1).getIndex();
855 return MI->getOperand(0).getReg();
863 if (MI->getOperand(1).isFI() &&
864 MI->getOperand(2).isImm() &&
865 MI->getOperand(2).getImm() == 0) {
866 FrameIndex = MI->getOperand(1).getIndex();
867 return MI->getOperand(0).getReg();
871 if (MI->getOperand(0).isFI() &&
872 MI->getOperand(2).getSubReg() == 0) {
873 FrameIndex = MI->getOperand(0).getIndex();
874 return MI->getOperand(2).getReg();
878 if (MI->getOperand(1).isFI() &&
879 MI->getOperand(0).getSubReg() == 0) {
880 FrameIndex = MI->getOperand(1).getIndex();
881 return MI->getOperand(0).getReg();
889 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
890 int &FrameIndex) const {
891 const MachineMemOperand *Dummy;
892 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
895 void ARMBaseInstrInfo::
896 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
897 unsigned DestReg, int FI,
898 const TargetRegisterClass *RC,
899 const TargetRegisterInfo *TRI) const {
901 if (I != MBB.end()) DL = I->getDebugLoc();
902 MachineFunction &MF = *MBB.getParent();
903 MachineFrameInfo &MFI = *MF.getFrameInfo();
904 unsigned Align = MFI.getObjectAlignment(FI);
905 MachineMemOperand *MMO =
906 MF.getMachineMemOperand(
907 MachinePointerInfo::getFixedStack(FI),
908 MachineMemOperand::MOLoad,
909 MFI.getObjectSize(FI),
912 switch (RC->getSize()) {
914 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
915 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
916 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
918 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
919 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
920 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
922 llvm_unreachable("Unknown reg class!");
925 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
926 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
927 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
929 llvm_unreachable("Unknown reg class!");
932 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
933 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
934 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
935 .addFrameIndex(FI).addImm(16)
936 .addMemOperand(MMO));
938 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
940 .addMemOperand(MMO));
943 llvm_unreachable("Unknown reg class!");
946 if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
947 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
948 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
949 .addFrameIndex(FI).addImm(16)
950 .addMemOperand(MMO));
952 MachineInstrBuilder MIB =
953 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
956 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
957 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
958 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
959 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
960 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
961 MIB.addReg(DestReg, RegState::ImplicitDefine);
964 llvm_unreachable("Unknown reg class!");
967 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
968 MachineInstrBuilder MIB =
969 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
972 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
973 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
974 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
975 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
976 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
977 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
978 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
979 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
980 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
981 MIB.addReg(DestReg, RegState::ImplicitDefine);
983 llvm_unreachable("Unknown reg class!");
986 llvm_unreachable("Unknown regclass!");
991 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
992 int &FrameIndex) const {
993 switch (MI->getOpcode()) {
996 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
997 if (MI->getOperand(1).isFI() &&
998 MI->getOperand(2).isReg() &&
999 MI->getOperand(3).isImm() &&
1000 MI->getOperand(2).getReg() == 0 &&
1001 MI->getOperand(3).getImm() == 0) {
1002 FrameIndex = MI->getOperand(1).getIndex();
1003 return MI->getOperand(0).getReg();
1011 if (MI->getOperand(1).isFI() &&
1012 MI->getOperand(2).isImm() &&
1013 MI->getOperand(2).getImm() == 0) {
1014 FrameIndex = MI->getOperand(1).getIndex();
1015 return MI->getOperand(0).getReg();
1019 if (MI->getOperand(1).isFI() &&
1020 MI->getOperand(0).getSubReg() == 0) {
1021 FrameIndex = MI->getOperand(1).getIndex();
1022 return MI->getOperand(0).getReg();
1026 if (MI->getOperand(1).isFI() &&
1027 MI->getOperand(0).getSubReg() == 0) {
1028 FrameIndex = MI->getOperand(1).getIndex();
1029 return MI->getOperand(0).getReg();
1037 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1038 int &FrameIndex) const {
1039 const MachineMemOperand *Dummy;
1040 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1043 bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
1044 // This hook gets to expand COPY instructions before they become
1045 // copyPhysReg() calls. Look for VMOVS instructions that can legally be
1046 // widened to VMOVD. We prefer the VMOVD when possible because it may be
1047 // changed into a VORR that can go down the NEON pipeline.
1048 if (!WidenVMOVS || !MI->isCopy())
1051 // Look for a copy between even S-registers. That is where we keep floats
1052 // when using NEON v2f32 instructions for f32 arithmetic.
1053 unsigned DstRegS = MI->getOperand(0).getReg();
1054 unsigned SrcRegS = MI->getOperand(1).getReg();
1055 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
1058 const TargetRegisterInfo *TRI = &getRegisterInfo();
1059 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
1061 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
1063 if (!DstRegD || !SrcRegD)
1066 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
1067 // legal if the COPY already defines the full DstRegD, and it isn't a
1068 // sub-register insertion.
1069 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
1072 // A dead copy shouldn't show up here, but reject it just in case.
1073 if (MI->getOperand(0).isDead())
1076 // All clear, widen the COPY.
1077 DEBUG(dbgs() << "widening: " << *MI);
1079 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
1080 // or some other super-register.
1081 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
1082 if (ImpDefIdx != -1)
1083 MI->RemoveOperand(ImpDefIdx);
1085 // Change the opcode and operands.
1086 MI->setDesc(get(ARM::VMOVD));
1087 MI->getOperand(0).setReg(DstRegD);
1088 MI->getOperand(1).setReg(SrcRegD);
1089 AddDefaultPred(MachineInstrBuilder(MI));
1091 // We are now reading SrcRegD instead of SrcRegS. This may upset the
1092 // register scavenger and machine verifier, so we need to indicate that we
1093 // are reading an undefined value from SrcRegD, but a proper value from
1095 MI->getOperand(1).setIsUndef();
1096 MachineInstrBuilder(MI).addReg(SrcRegS, RegState::Implicit);
1098 // SrcRegD may actually contain an unrelated value in the ssub_1
1099 // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
1100 if (MI->getOperand(1).isKill()) {
1101 MI->getOperand(1).setIsKill(false);
1102 MI->addRegisterKilled(SrcRegS, TRI, true);
1105 DEBUG(dbgs() << "replaced by: " << *MI);
1110 ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
1111 int FrameIx, uint64_t Offset,
1112 const MDNode *MDPtr,
1113 DebugLoc DL) const {
1114 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
1115 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
1119 /// Create a copy of a const pool value. Update CPI to the new index and return
1121 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1122 MachineConstantPool *MCP = MF.getConstantPool();
1123 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1125 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1126 assert(MCPE.isMachineConstantPoolEntry() &&
1127 "Expecting a machine constantpool entry!");
1128 ARMConstantPoolValue *ACPV =
1129 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1131 unsigned PCLabelId = AFI->createPICLabelUId();
1132 ARMConstantPoolValue *NewCPV = 0;
1133 // FIXME: The below assumes PIC relocation model and that the function
1134 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1135 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1136 // instructions, so that's probably OK, but is PIC always correct when
1138 if (ACPV->isGlobalValue())
1139 NewCPV = ARMConstantPoolConstant::
1140 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
1142 else if (ACPV->isExtSymbol())
1143 NewCPV = ARMConstantPoolSymbol::
1144 Create(MF.getFunction()->getContext(),
1145 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
1146 else if (ACPV->isBlockAddress())
1147 NewCPV = ARMConstantPoolConstant::
1148 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
1149 ARMCP::CPBlockAddress, 4);
1150 else if (ACPV->isLSDA())
1151 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
1153 else if (ACPV->isMachineBasicBlock())
1154 NewCPV = ARMConstantPoolMBB::
1155 Create(MF.getFunction()->getContext(),
1156 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
1158 llvm_unreachable("Unexpected ARM constantpool value type!!");
1159 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1163 void ARMBaseInstrInfo::
1164 reMaterialize(MachineBasicBlock &MBB,
1165 MachineBasicBlock::iterator I,
1166 unsigned DestReg, unsigned SubIdx,
1167 const MachineInstr *Orig,
1168 const TargetRegisterInfo &TRI) const {
1169 unsigned Opcode = Orig->getOpcode();
1172 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1173 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1177 case ARM::tLDRpci_pic:
1178 case ARM::t2LDRpci_pic: {
1179 MachineFunction &MF = *MBB.getParent();
1180 unsigned CPI = Orig->getOperand(1).getIndex();
1181 unsigned PCLabelId = duplicateCPV(MF, CPI);
1182 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1184 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1185 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1192 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1193 MachineInstr *MI = TargetInstrInfoImpl::duplicate(Orig, MF);
1194 switch(Orig->getOpcode()) {
1195 case ARM::tLDRpci_pic:
1196 case ARM::t2LDRpci_pic: {
1197 unsigned CPI = Orig->getOperand(1).getIndex();
1198 unsigned PCLabelId = duplicateCPV(MF, CPI);
1199 Orig->getOperand(1).setIndex(CPI);
1200 Orig->getOperand(2).setImm(PCLabelId);
1207 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1208 const MachineInstr *MI1,
1209 const MachineRegisterInfo *MRI) const {
1210 int Opcode = MI0->getOpcode();
1211 if (Opcode == ARM::t2LDRpci ||
1212 Opcode == ARM::t2LDRpci_pic ||
1213 Opcode == ARM::tLDRpci ||
1214 Opcode == ARM::tLDRpci_pic ||
1215 Opcode == ARM::MOV_ga_dyn ||
1216 Opcode == ARM::MOV_ga_pcrel ||
1217 Opcode == ARM::MOV_ga_pcrel_ldr ||
1218 Opcode == ARM::t2MOV_ga_dyn ||
1219 Opcode == ARM::t2MOV_ga_pcrel) {
1220 if (MI1->getOpcode() != Opcode)
1222 if (MI0->getNumOperands() != MI1->getNumOperands())
1225 const MachineOperand &MO0 = MI0->getOperand(1);
1226 const MachineOperand &MO1 = MI1->getOperand(1);
1227 if (MO0.getOffset() != MO1.getOffset())
1230 if (Opcode == ARM::MOV_ga_dyn ||
1231 Opcode == ARM::MOV_ga_pcrel ||
1232 Opcode == ARM::MOV_ga_pcrel_ldr ||
1233 Opcode == ARM::t2MOV_ga_dyn ||
1234 Opcode == ARM::t2MOV_ga_pcrel)
1235 // Ignore the PC labels.
1236 return MO0.getGlobal() == MO1.getGlobal();
1238 const MachineFunction *MF = MI0->getParent()->getParent();
1239 const MachineConstantPool *MCP = MF->getConstantPool();
1240 int CPI0 = MO0.getIndex();
1241 int CPI1 = MO1.getIndex();
1242 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1243 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1244 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1245 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1246 if (isARMCP0 && isARMCP1) {
1247 ARMConstantPoolValue *ACPV0 =
1248 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1249 ARMConstantPoolValue *ACPV1 =
1250 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1251 return ACPV0->hasSameValue(ACPV1);
1252 } else if (!isARMCP0 && !isARMCP1) {
1253 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1256 } else if (Opcode == ARM::PICLDR) {
1257 if (MI1->getOpcode() != Opcode)
1259 if (MI0->getNumOperands() != MI1->getNumOperands())
1262 unsigned Addr0 = MI0->getOperand(1).getReg();
1263 unsigned Addr1 = MI1->getOperand(1).getReg();
1264 if (Addr0 != Addr1) {
1266 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1267 !TargetRegisterInfo::isVirtualRegister(Addr1))
1270 // This assumes SSA form.
1271 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1272 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1273 // Check if the loaded value, e.g. a constantpool of a global address, are
1275 if (!produceSameValue(Def0, Def1, MRI))
1279 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1280 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1281 const MachineOperand &MO0 = MI0->getOperand(i);
1282 const MachineOperand &MO1 = MI1->getOperand(i);
1283 if (!MO0.isIdenticalTo(MO1))
1289 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1292 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1293 /// determine if two loads are loading from the same base address. It should
1294 /// only return true if the base pointers are the same and the only differences
1295 /// between the two addresses is the offset. It also returns the offsets by
1297 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1299 int64_t &Offset2) const {
1300 // Don't worry about Thumb: just ARM and Thumb2.
1301 if (Subtarget.isThumb1Only()) return false;
1303 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1306 switch (Load1->getMachineOpcode()) {
1319 case ARM::t2LDRSHi8:
1321 case ARM::t2LDRSHi12:
1325 switch (Load2->getMachineOpcode()) {
1338 case ARM::t2LDRSHi8:
1340 case ARM::t2LDRSHi12:
1344 // Check if base addresses and chain operands match.
1345 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1346 Load1->getOperand(4) != Load2->getOperand(4))
1349 // Index should be Reg0.
1350 if (Load1->getOperand(3) != Load2->getOperand(3))
1353 // Determine the offsets.
1354 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1355 isa<ConstantSDNode>(Load2->getOperand(1))) {
1356 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1357 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1364 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1365 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1366 /// be scheduled togther. On some targets if two loads are loading from
1367 /// addresses in the same cache line, it's better if they are scheduled
1368 /// together. This function takes two integers that represent the load offsets
1369 /// from the common base address. It returns true if it decides it's desirable
1370 /// to schedule the two loads together. "NumLoads" is the number of loads that
1371 /// have already been scheduled after Load1.
1372 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1373 int64_t Offset1, int64_t Offset2,
1374 unsigned NumLoads) const {
1375 // Don't worry about Thumb: just ARM and Thumb2.
1376 if (Subtarget.isThumb1Only()) return false;
1378 assert(Offset2 > Offset1);
1380 if ((Offset2 - Offset1) / 8 > 64)
1383 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1384 return false; // FIXME: overly conservative?
1386 // Four loads in a row should be sufficient.
1393 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1394 const MachineBasicBlock *MBB,
1395 const MachineFunction &MF) const {
1396 // Debug info is never a scheduling boundary. It's necessary to be explicit
1397 // due to the special treatment of IT instructions below, otherwise a
1398 // dbg_value followed by an IT will result in the IT instruction being
1399 // considered a scheduling hazard, which is wrong. It should be the actual
1400 // instruction preceding the dbg_value instruction(s), just like it is
1401 // when debug info is not present.
1402 if (MI->isDebugValue())
1405 // Terminators and labels can't be scheduled around.
1406 if (MI->isTerminator() || MI->isLabel())
1409 // Treat the start of the IT block as a scheduling boundary, but schedule
1410 // t2IT along with all instructions following it.
1411 // FIXME: This is a big hammer. But the alternative is to add all potential
1412 // true and anti dependencies to IT block instructions as implicit operands
1413 // to the t2IT instruction. The added compile time and complexity does not
1415 MachineBasicBlock::const_iterator I = MI;
1416 // Make sure to skip any dbg_value instructions
1417 while (++I != MBB->end() && I->isDebugValue())
1419 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1422 // Don't attempt to schedule around any instruction that defines
1423 // a stack-oriented pointer, as it's unlikely to be profitable. This
1424 // saves compile time, because it doesn't require every single
1425 // stack slot reference to depend on the instruction that does the
1427 // Calls don't actually change the stack pointer, even if they have imp-defs.
1428 // No ARM calling conventions change the stack pointer. (X86 calling
1429 // conventions sometimes do).
1430 if (!MI->isCall() && MI->definesRegister(ARM::SP))
1436 bool ARMBaseInstrInfo::
1437 isProfitableToIfCvt(MachineBasicBlock &MBB,
1438 unsigned NumCycles, unsigned ExtraPredCycles,
1439 const BranchProbability &Probability) const {
1443 // Attempt to estimate the relative costs of predication versus branching.
1444 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1445 UnpredCost /= Probability.getDenominator();
1446 UnpredCost += 1; // The branch itself
1447 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1449 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1452 bool ARMBaseInstrInfo::
1453 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1454 unsigned TCycles, unsigned TExtra,
1455 MachineBasicBlock &FMBB,
1456 unsigned FCycles, unsigned FExtra,
1457 const BranchProbability &Probability) const {
1458 if (!TCycles || !FCycles)
1461 // Attempt to estimate the relative costs of predication versus branching.
1462 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1463 TUnpredCost /= Probability.getDenominator();
1465 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1466 unsigned FUnpredCost = Comp * FCycles;
1467 FUnpredCost /= Probability.getDenominator();
1469 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1470 UnpredCost += 1; // The branch itself
1471 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1473 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1476 /// getInstrPredicate - If instruction is predicated, returns its predicate
1477 /// condition, otherwise returns AL. It also returns the condition code
1478 /// register by reference.
1480 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1481 int PIdx = MI->findFirstPredOperandIdx();
1487 PredReg = MI->getOperand(PIdx+1).getReg();
1488 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1492 int llvm::getMatchingCondBranchOpcode(int Opc) {
1497 if (Opc == ARM::t2B)
1500 llvm_unreachable("Unknown unconditional branch opcode!");
1503 /// commuteInstruction - Handle commutable instructions.
1505 ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1506 switch (MI->getOpcode()) {
1508 case ARM::t2MOVCCr: {
1509 // MOVCC can be commuted by inverting the condition.
1510 unsigned PredReg = 0;
1511 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
1512 // MOVCC AL can't be inverted. Shouldn't happen.
1513 if (CC == ARMCC::AL || PredReg != ARM::CPSR)
1515 MI = TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1518 // After swapping the MOVCC operands, also invert the condition.
1519 MI->getOperand(MI->findFirstPredOperandIdx())
1520 .setImm(ARMCC::getOppositeCondition(CC));
1524 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1527 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
1528 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
1531 /// This will go away once we can teach tblgen how to set the optional CPSR def
1533 struct AddSubFlagsOpcodePair {
1535 uint16_t MachineOpc;
1538 static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
1539 {ARM::ADDSri, ARM::ADDri},
1540 {ARM::ADDSrr, ARM::ADDrr},
1541 {ARM::ADDSrsi, ARM::ADDrsi},
1542 {ARM::ADDSrsr, ARM::ADDrsr},
1544 {ARM::SUBSri, ARM::SUBri},
1545 {ARM::SUBSrr, ARM::SUBrr},
1546 {ARM::SUBSrsi, ARM::SUBrsi},
1547 {ARM::SUBSrsr, ARM::SUBrsr},
1549 {ARM::RSBSri, ARM::RSBri},
1550 {ARM::RSBSrsi, ARM::RSBrsi},
1551 {ARM::RSBSrsr, ARM::RSBrsr},
1553 {ARM::t2ADDSri, ARM::t2ADDri},
1554 {ARM::t2ADDSrr, ARM::t2ADDrr},
1555 {ARM::t2ADDSrs, ARM::t2ADDrs},
1557 {ARM::t2SUBSri, ARM::t2SUBri},
1558 {ARM::t2SUBSrr, ARM::t2SUBrr},
1559 {ARM::t2SUBSrs, ARM::t2SUBrs},
1561 {ARM::t2RSBSri, ARM::t2RSBri},
1562 {ARM::t2RSBSrs, ARM::t2RSBrs},
1565 unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
1566 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
1567 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
1568 return AddSubFlagsOpcodeMap[i].MachineOpc;
1572 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1573 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1574 unsigned DestReg, unsigned BaseReg, int NumBytes,
1575 ARMCC::CondCodes Pred, unsigned PredReg,
1576 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1577 bool isSub = NumBytes < 0;
1578 if (isSub) NumBytes = -NumBytes;
1581 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1582 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1583 assert(ThisVal && "Didn't extract field correctly");
1585 // We will handle these bits from offset, clear them.
1586 NumBytes &= ~ThisVal;
1588 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1590 // Build the new ADD / SUB.
1591 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1592 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1593 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1594 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1595 .setMIFlags(MIFlags);
1600 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1601 unsigned FrameReg, int &Offset,
1602 const ARMBaseInstrInfo &TII) {
1603 unsigned Opcode = MI.getOpcode();
1604 const MCInstrDesc &Desc = MI.getDesc();
1605 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1608 // Memory operands in inline assembly always use AddrMode2.
1609 if (Opcode == ARM::INLINEASM)
1610 AddrMode = ARMII::AddrMode2;
1612 if (Opcode == ARM::ADDri) {
1613 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1615 // Turn it into a move.
1616 MI.setDesc(TII.get(ARM::MOVr));
1617 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1618 MI.RemoveOperand(FrameRegIdx+1);
1621 } else if (Offset < 0) {
1624 MI.setDesc(TII.get(ARM::SUBri));
1627 // Common case: small offset, fits into instruction.
1628 if (ARM_AM::getSOImmVal(Offset) != -1) {
1629 // Replace the FrameIndex with sp / fp
1630 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1631 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1636 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1638 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1639 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1641 // We will handle these bits from offset, clear them.
1642 Offset &= ~ThisImmVal;
1644 // Get the properly encoded SOImmVal field.
1645 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1646 "Bit extraction didn't work?");
1647 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1649 unsigned ImmIdx = 0;
1651 unsigned NumBits = 0;
1654 case ARMII::AddrMode_i12: {
1655 ImmIdx = FrameRegIdx + 1;
1656 InstrOffs = MI.getOperand(ImmIdx).getImm();
1660 case ARMII::AddrMode2: {
1661 ImmIdx = FrameRegIdx+2;
1662 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1663 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1668 case ARMII::AddrMode3: {
1669 ImmIdx = FrameRegIdx+2;
1670 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1671 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1676 case ARMII::AddrMode4:
1677 case ARMII::AddrMode6:
1678 // Can't fold any offset even if it's zero.
1680 case ARMII::AddrMode5: {
1681 ImmIdx = FrameRegIdx+1;
1682 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1683 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1690 llvm_unreachable("Unsupported addressing mode!");
1693 Offset += InstrOffs * Scale;
1694 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1700 // Attempt to fold address comp. if opcode has offset bits
1702 // Common case: small offset, fits into instruction.
1703 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1704 int ImmedOffset = Offset / Scale;
1705 unsigned Mask = (1 << NumBits) - 1;
1706 if ((unsigned)Offset <= Mask * Scale) {
1707 // Replace the FrameIndex with sp
1708 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1709 // FIXME: When addrmode2 goes away, this will simplify (like the
1710 // T2 version), as the LDR.i12 versions don't need the encoding
1711 // tricks for the offset value.
1713 if (AddrMode == ARMII::AddrMode_i12)
1714 ImmedOffset = -ImmedOffset;
1716 ImmedOffset |= 1 << NumBits;
1718 ImmOp.ChangeToImmediate(ImmedOffset);
1723 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1724 ImmedOffset = ImmedOffset & Mask;
1726 if (AddrMode == ARMII::AddrMode_i12)
1727 ImmedOffset = -ImmedOffset;
1729 ImmedOffset |= 1 << NumBits;
1731 ImmOp.ChangeToImmediate(ImmedOffset);
1732 Offset &= ~(Mask*Scale);
1736 Offset = (isSub) ? -Offset : Offset;
1740 bool ARMBaseInstrInfo::
1741 AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpMask,
1742 int &CmpValue) const {
1743 switch (MI->getOpcode()) {
1747 SrcReg = MI->getOperand(0).getReg();
1749 CmpValue = MI->getOperand(1).getImm();
1753 SrcReg = MI->getOperand(0).getReg();
1759 SrcReg = MI->getOperand(0).getReg();
1760 CmpMask = MI->getOperand(1).getImm();
1768 /// isSuitableForMask - Identify a suitable 'and' instruction that
1769 /// operates on the given source register and applies the same mask
1770 /// as a 'tst' instruction. Provide a limited look-through for copies.
1771 /// When successful, MI will hold the found instruction.
1772 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
1773 int CmpMask, bool CommonUse) {
1774 switch (MI->getOpcode()) {
1777 if (CmpMask != MI->getOperand(2).getImm())
1779 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
1783 // Walk down one instruction which is potentially an 'and'.
1784 const MachineInstr &Copy = *MI;
1785 MachineBasicBlock::iterator AND(
1786 llvm::next(MachineBasicBlock::iterator(MI)));
1787 if (AND == MI->getParent()->end()) return false;
1789 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
1797 /// OptimizeCompareInstr - Convert the instruction supplying the argument to the
1798 /// comparison into one that sets the zero bit in the flags register. Convert
1799 /// the SUBrr(r1,r2)|Subri(r1,CmpValue) instruction into one that sets the flags
1800 /// register and remove the CMPrr(r1,r2)|CMPrr(r2,r1)|CMPri(r1,CmpValue)
1802 bool ARMBaseInstrInfo::
1803 OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
1804 int CmpValue, const MachineRegisterInfo *MRI) const {
1805 if (MRI->def_empty(SrcReg))
1808 MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
1809 if (llvm::next(DI) != MRI->def_end())
1810 // Only support one definition.
1813 MachineInstr *MI = &*DI;
1815 // Masked compares sometimes use the same register as the corresponding 'and'.
1816 if (CmpMask != ~0) {
1817 if (!isSuitableForMask(MI, SrcReg, CmpMask, false)) {
1819 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
1820 UE = MRI->use_end(); UI != UE; ++UI) {
1821 if (UI->getParent() != CmpInstr->getParent()) continue;
1822 MachineInstr *PotentialAND = &*UI;
1823 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true))
1828 if (!MI) return false;
1832 // Get ready to iterate backward from CmpInstr.
1833 MachineBasicBlock::iterator I = CmpInstr, E = MI,
1834 B = CmpInstr->getParent()->begin();
1836 // Early exit if CmpInstr is at the beginning of the BB.
1837 if (I == B) return false;
1839 // There are two possible candidates which can be changed to set CPSR:
1840 // One is MI, the other is a SUB instruction.
1841 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
1842 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
1843 MachineInstr *Sub = NULL;
1844 unsigned SrcReg2 = 0;
1845 if (CmpInstr->getOpcode() == ARM::CMPrr ||
1846 CmpInstr->getOpcode() == ARM::t2CMPrr) {
1847 SrcReg2 = CmpInstr->getOperand(1).getReg();
1848 // MI is not a candidate for CMPrr.
1850 } else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
1851 // Conservatively refuse to convert an instruction which isn't in the same
1852 // BB as the comparison.
1853 // For CMPri, we need to check Sub, thus we can't return here.
1854 if (CmpInstr->getOpcode() == ARM::CMPri ||
1855 CmpInstr->getOpcode() == ARM::t2CMPri)
1861 // Check that CPSR isn't set between the comparison instruction and the one we
1862 // want to change. At the same time, search for Sub.
1864 for (; I != E; --I) {
1865 const MachineInstr &Instr = *I;
1867 for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) {
1868 const MachineOperand &MO = Instr.getOperand(IO);
1869 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR))
1871 if (!MO.isReg()) continue;
1873 // This instruction modifies or uses CPSR after the one we want to
1874 // change. We can't do this transformation.
1875 if (MO.getReg() == ARM::CPSR)
1879 // Check whether the current instruction is SUB(r1, r2) or SUB(r2, r1).
1881 (Instr.getOpcode() == ARM::SUBrr ||
1882 Instr.getOpcode() == ARM::t2SUBrr) &&
1883 ((Instr.getOperand(1).getReg() == SrcReg &&
1884 Instr.getOperand(2).getReg() == SrcReg2) ||
1885 (Instr.getOperand(1).getReg() == SrcReg2 &&
1886 Instr.getOperand(2).getReg() == SrcReg))) {
1891 // Check whether the current instruction is SUBri(r1, CmpValue).
1892 if ((CmpInstr->getOpcode() == ARM::CMPri ||
1893 CmpInstr->getOpcode() == ARM::t2CMPri) &&
1894 Instr.getOpcode() == ARM::SUBri && CmpValue != 0 &&
1895 Instr.getOperand(1).getReg() == SrcReg &&
1896 Instr.getOperand(2).getImm() == CmpValue) {
1902 // The 'and' is below the comparison instruction.
1906 // Return false if no candidates exist.
1910 // The single candidate is called MI.
1913 switch (MI->getOpcode()) {
1947 case ARM::t2EORri: {
1948 // Scan forward for the use of CPSR
1949 // When checking against MI: if it's a conditional code requires
1950 // checking of V bit, then this is not safe to do. If we can't find the
1951 // CPSR use (i.e. used in another block), then it's not safe to perform
1952 // the optimization.
1953 // When checking against Sub, we handle the condition codes GE, LT, GT, LE.
1954 SmallVector<MachineOperand*, 4> OperandsToUpdate;
1955 bool isSafe = false;
1957 E = CmpInstr->getParent()->end();
1958 while (!isSafe && ++I != E) {
1959 const MachineInstr &Instr = *I;
1960 for (unsigned IO = 0, EO = Instr.getNumOperands();
1961 !isSafe && IO != EO; ++IO) {
1962 const MachineOperand &MO = Instr.getOperand(IO);
1963 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
1967 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
1973 // Condition code is after the operand before CPSR.
1974 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
1989 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
1990 // on CMP needs to be updated to be based on SUB.
1991 // Push the condition code operands to OperandsToUpdate.
1992 // If it is safe to remove CmpInstr, the condition code of these
1993 // operands will be modified.
1994 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
1995 Sub->getOperand(2).getReg() == SrcReg)
1996 OperandsToUpdate.push_back(&((*I).getOperand(IO-1)));
2015 // If the candidate is Sub, we may exit the loop at end of the basic block.
2016 // In that case, it is still safe to remove CmpInstr.
2017 if (!isSafe && !Sub)
2020 // Toggle the optional operand to CPSR.
2021 MI->getOperand(5).setReg(ARM::CPSR);
2022 MI->getOperand(5).setIsDef(true);
2023 CmpInstr->eraseFromParent();
2025 // Modify the condition code of operands in OperandsToUpdate.
2026 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2027 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2028 for (unsigned i = 0; i < OperandsToUpdate.size(); i++) {
2029 ARMCC::CondCodes CC = (ARMCC::CondCodes)OperandsToUpdate[i]->getImm();
2030 ARMCC::CondCodes NewCC;
2032 default: llvm_unreachable("only expecting less/greater comparisons here");
2033 case ARMCC::GE: NewCC = ARMCC::LE; break;
2034 case ARMCC::LT: NewCC = ARMCC::GT; break;
2035 case ARMCC::GT: NewCC = ARMCC::LT; break;
2036 case ARMCC::LE: NewCC = ARMCC::GE; break;
2037 case ARMCC::HS: NewCC = ARMCC::LS; break;
2038 case ARMCC::LS: NewCC = ARMCC::HS; break;
2039 case ARMCC::HI: NewCC = ARMCC::LO; break;
2040 case ARMCC::LO: NewCC = ARMCC::HI; break;
2046 OperandsToUpdate[i]->setImm(NewCC);
2055 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
2056 MachineInstr *DefMI, unsigned Reg,
2057 MachineRegisterInfo *MRI) const {
2058 // Fold large immediates into add, sub, or, xor.
2059 unsigned DefOpc = DefMI->getOpcode();
2060 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
2062 if (!DefMI->getOperand(1).isImm())
2063 // Could be t2MOVi32imm <ga:xx>
2066 if (!MRI->hasOneNonDBGUse(Reg))
2069 const MCInstrDesc &DefMCID = DefMI->getDesc();
2070 if (DefMCID.hasOptionalDef()) {
2071 unsigned NumOps = DefMCID.getNumOperands();
2072 const MachineOperand &MO = DefMI->getOperand(NumOps-1);
2073 if (MO.getReg() == ARM::CPSR && !MO.isDead())
2074 // If DefMI defines CPSR and it is not dead, it's obviously not safe
2079 const MCInstrDesc &UseMCID = UseMI->getDesc();
2080 if (UseMCID.hasOptionalDef()) {
2081 unsigned NumOps = UseMCID.getNumOperands();
2082 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
2083 // If the instruction sets the flag, do not attempt this optimization
2084 // since it may change the semantics of the code.
2088 unsigned UseOpc = UseMI->getOpcode();
2089 unsigned NewUseOpc = 0;
2090 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
2091 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
2092 bool Commute = false;
2094 default: return false;
2102 case ARM::t2EORrr: {
2103 Commute = UseMI->getOperand(2).getReg() != Reg;
2110 NewUseOpc = ARM::SUBri;
2116 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
2118 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
2119 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
2122 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
2123 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
2124 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
2128 case ARM::t2SUBrr: {
2132 NewUseOpc = ARM::t2SUBri;
2137 case ARM::t2EORrr: {
2138 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
2140 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
2141 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
2144 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
2145 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
2146 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
2154 unsigned OpIdx = Commute ? 2 : 1;
2155 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
2156 bool isKill = UseMI->getOperand(OpIdx).isKill();
2157 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
2158 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
2159 UseMI, UseMI->getDebugLoc(),
2160 get(NewUseOpc), NewReg)
2161 .addReg(Reg1, getKillRegState(isKill))
2162 .addImm(SOImmValV1)));
2163 UseMI->setDesc(get(NewUseOpc));
2164 UseMI->getOperand(1).setReg(NewReg);
2165 UseMI->getOperand(1).setIsKill();
2166 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
2167 DefMI->eraseFromParent();
2172 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
2173 const MachineInstr *MI) const {
2174 if (!ItinData || ItinData->isEmpty())
2177 const MCInstrDesc &Desc = MI->getDesc();
2178 unsigned Class = Desc.getSchedClass();
2179 unsigned UOps = ItinData->Itineraries[Class].NumMicroOps;
2183 unsigned Opc = MI->getOpcode();
2186 llvm_unreachable("Unexpected multi-uops instruction!");
2191 // The number of uOps for load / store multiple are determined by the number
2194 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
2195 // same cycle. The scheduling for the first load / store must be done
2196 // separately by assuming the the address is not 64-bit aligned.
2198 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
2199 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
2200 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
2202 case ARM::VLDMDIA_UPD:
2203 case ARM::VLDMDDB_UPD:
2205 case ARM::VLDMSIA_UPD:
2206 case ARM::VLDMSDB_UPD:
2208 case ARM::VSTMDIA_UPD:
2209 case ARM::VSTMDDB_UPD:
2211 case ARM::VSTMSIA_UPD:
2212 case ARM::VSTMSDB_UPD: {
2213 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
2214 return (NumRegs / 2) + (NumRegs % 2) + 1;
2217 case ARM::LDMIA_RET:
2222 case ARM::LDMIA_UPD:
2223 case ARM::LDMDA_UPD:
2224 case ARM::LDMDB_UPD:
2225 case ARM::LDMIB_UPD:
2230 case ARM::STMIA_UPD:
2231 case ARM::STMDA_UPD:
2232 case ARM::STMDB_UPD:
2233 case ARM::STMIB_UPD:
2235 case ARM::tLDMIA_UPD:
2236 case ARM::tSTMIA_UPD:
2240 case ARM::t2LDMIA_RET:
2243 case ARM::t2LDMIA_UPD:
2244 case ARM::t2LDMDB_UPD:
2247 case ARM::t2STMIA_UPD:
2248 case ARM::t2STMDB_UPD: {
2249 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
2250 if (Subtarget.isCortexA8()) {
2253 // 4 registers would be issued: 2, 2.
2254 // 5 registers would be issued: 2, 2, 1.
2255 UOps = (NumRegs / 2);
2259 } else if (Subtarget.isCortexA9()) {
2260 UOps = (NumRegs / 2);
2261 // If there are odd number of registers or if it's not 64-bit aligned,
2262 // then it takes an extra AGU (Address Generation Unit) cycle.
2263 if ((NumRegs % 2) ||
2264 !MI->hasOneMemOperand() ||
2265 (*MI->memoperands_begin())->getAlignment() < 8)
2269 // Assume the worst.
2277 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
2278 const MCInstrDesc &DefMCID,
2280 unsigned DefIdx, unsigned DefAlign) const {
2281 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2283 // Def is the address writeback.
2284 return ItinData->getOperandCycle(DefClass, DefIdx);
2287 if (Subtarget.isCortexA8()) {
2288 // (regno / 2) + (regno % 2) + 1
2289 DefCycle = RegNo / 2 + 1;
2292 } else if (Subtarget.isCortexA9()) {
2294 bool isSLoad = false;
2296 switch (DefMCID.getOpcode()) {
2299 case ARM::VLDMSIA_UPD:
2300 case ARM::VLDMSDB_UPD:
2305 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2306 // then it takes an extra cycle.
2307 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
2310 // Assume the worst.
2311 DefCycle = RegNo + 2;
2318 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
2319 const MCInstrDesc &DefMCID,
2321 unsigned DefIdx, unsigned DefAlign) const {
2322 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2324 // Def is the address writeback.
2325 return ItinData->getOperandCycle(DefClass, DefIdx);
2328 if (Subtarget.isCortexA8()) {
2329 // 4 registers would be issued: 1, 2, 1.
2330 // 5 registers would be issued: 1, 2, 2.
2331 DefCycle = RegNo / 2;
2334 // Result latency is issue cycle + 2: E2.
2336 } else if (Subtarget.isCortexA9()) {
2337 DefCycle = (RegNo / 2);
2338 // If there are odd number of registers or if it's not 64-bit aligned,
2339 // then it takes an extra AGU (Address Generation Unit) cycle.
2340 if ((RegNo % 2) || DefAlign < 8)
2342 // Result latency is AGU cycles + 2.
2345 // Assume the worst.
2346 DefCycle = RegNo + 2;
2353 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
2354 const MCInstrDesc &UseMCID,
2356 unsigned UseIdx, unsigned UseAlign) const {
2357 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2359 return ItinData->getOperandCycle(UseClass, UseIdx);
2362 if (Subtarget.isCortexA8()) {
2363 // (regno / 2) + (regno % 2) + 1
2364 UseCycle = RegNo / 2 + 1;
2367 } else if (Subtarget.isCortexA9()) {
2369 bool isSStore = false;
2371 switch (UseMCID.getOpcode()) {
2374 case ARM::VSTMSIA_UPD:
2375 case ARM::VSTMSDB_UPD:
2380 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2381 // then it takes an extra cycle.
2382 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2385 // Assume the worst.
2386 UseCycle = RegNo + 2;
2393 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2394 const MCInstrDesc &UseMCID,
2396 unsigned UseIdx, unsigned UseAlign) const {
2397 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2399 return ItinData->getOperandCycle(UseClass, UseIdx);
2402 if (Subtarget.isCortexA8()) {
2403 UseCycle = RegNo / 2;
2408 } else if (Subtarget.isCortexA9()) {
2409 UseCycle = (RegNo / 2);
2410 // If there are odd number of registers or if it's not 64-bit aligned,
2411 // then it takes an extra AGU (Address Generation Unit) cycle.
2412 if ((RegNo % 2) || UseAlign < 8)
2415 // Assume the worst.
2422 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2423 const MCInstrDesc &DefMCID,
2424 unsigned DefIdx, unsigned DefAlign,
2425 const MCInstrDesc &UseMCID,
2426 unsigned UseIdx, unsigned UseAlign) const {
2427 unsigned DefClass = DefMCID.getSchedClass();
2428 unsigned UseClass = UseMCID.getSchedClass();
2430 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2431 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2433 // This may be a def / use of a variable_ops instruction, the operand
2434 // latency might be determinable dynamically. Let the target try to
2437 bool LdmBypass = false;
2438 switch (DefMCID.getOpcode()) {
2440 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2444 case ARM::VLDMDIA_UPD:
2445 case ARM::VLDMDDB_UPD:
2447 case ARM::VLDMSIA_UPD:
2448 case ARM::VLDMSDB_UPD:
2449 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2452 case ARM::LDMIA_RET:
2457 case ARM::LDMIA_UPD:
2458 case ARM::LDMDA_UPD:
2459 case ARM::LDMDB_UPD:
2460 case ARM::LDMIB_UPD:
2462 case ARM::tLDMIA_UPD:
2464 case ARM::t2LDMIA_RET:
2467 case ARM::t2LDMIA_UPD:
2468 case ARM::t2LDMDB_UPD:
2470 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2475 // We can't seem to determine the result latency of the def, assume it's 2.
2479 switch (UseMCID.getOpcode()) {
2481 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
2485 case ARM::VSTMDIA_UPD:
2486 case ARM::VSTMDDB_UPD:
2488 case ARM::VSTMSIA_UPD:
2489 case ARM::VSTMSDB_UPD:
2490 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2497 case ARM::STMIA_UPD:
2498 case ARM::STMDA_UPD:
2499 case ARM::STMDB_UPD:
2500 case ARM::STMIB_UPD:
2501 case ARM::tSTMIA_UPD:
2506 case ARM::t2STMIA_UPD:
2507 case ARM::t2STMDB_UPD:
2508 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
2513 // Assume it's read in the first stage.
2516 UseCycle = DefCycle - UseCycle + 1;
2519 // It's a variable_ops instruction so we can't use DefIdx here. Just use
2520 // first def operand.
2521 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
2524 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
2525 UseClass, UseIdx)) {
2533 static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
2534 const MachineInstr *MI, unsigned Reg,
2535 unsigned &DefIdx, unsigned &Dist) {
2538 MachineBasicBlock::const_iterator I = MI; ++I;
2539 MachineBasicBlock::const_instr_iterator II =
2540 llvm::prior(I.getInstrIterator());
2541 assert(II->isInsideBundle() && "Empty bundle?");
2544 while (II->isInsideBundle()) {
2545 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
2552 assert(Idx != -1 && "Cannot find bundled definition!");
2557 static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
2558 const MachineInstr *MI, unsigned Reg,
2559 unsigned &UseIdx, unsigned &Dist) {
2562 MachineBasicBlock::const_instr_iterator II = MI; ++II;
2563 assert(II->isInsideBundle() && "Empty bundle?");
2564 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
2566 // FIXME: This doesn't properly handle multiple uses.
2568 while (II != E && II->isInsideBundle()) {
2569 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
2572 if (II->getOpcode() != ARM::t2IT)
2586 /// Return the number of cycles to add to (or subtract from) the static
2587 /// itinerary based on the def opcode and alignment. The caller will ensure that
2588 /// adjusted latency is at least one cycle.
2589 static int adjustDefLatency(const ARMSubtarget &Subtarget,
2590 const MachineInstr *DefMI,
2591 const MCInstrDesc *DefMCID, unsigned DefAlign) {
2593 if (Subtarget.isCortexA8() || Subtarget.isCortexA9()) {
2594 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2595 // variants are one cycle cheaper.
2596 switch (DefMCID->getOpcode()) {
2600 unsigned ShOpVal = DefMI->getOperand(3).getImm();
2601 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2603 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2610 case ARM::t2LDRSHs: {
2611 // Thumb2 mode: lsl only.
2612 unsigned ShAmt = DefMI->getOperand(3).getImm();
2613 if (ShAmt == 0 || ShAmt == 2)
2620 if (DefAlign < 8 && Subtarget.isCortexA9()) {
2621 switch (DefMCID->getOpcode()) {
2627 case ARM::VLD1q8wb_fixed:
2628 case ARM::VLD1q16wb_fixed:
2629 case ARM::VLD1q32wb_fixed:
2630 case ARM::VLD1q64wb_fixed:
2631 case ARM::VLD1q8wb_register:
2632 case ARM::VLD1q16wb_register:
2633 case ARM::VLD1q32wb_register:
2634 case ARM::VLD1q64wb_register:
2641 case ARM::VLD2d8wb_fixed:
2642 case ARM::VLD2d16wb_fixed:
2643 case ARM::VLD2d32wb_fixed:
2644 case ARM::VLD2q8wb_fixed:
2645 case ARM::VLD2q16wb_fixed:
2646 case ARM::VLD2q32wb_fixed:
2647 case ARM::VLD2d8wb_register:
2648 case ARM::VLD2d16wb_register:
2649 case ARM::VLD2d32wb_register:
2650 case ARM::VLD2q8wb_register:
2651 case ARM::VLD2q16wb_register:
2652 case ARM::VLD2q32wb_register:
2657 case ARM::VLD3d8_UPD:
2658 case ARM::VLD3d16_UPD:
2659 case ARM::VLD3d32_UPD:
2660 case ARM::VLD1d64Twb_fixed:
2661 case ARM::VLD1d64Twb_register:
2662 case ARM::VLD3q8_UPD:
2663 case ARM::VLD3q16_UPD:
2664 case ARM::VLD3q32_UPD:
2669 case ARM::VLD4d8_UPD:
2670 case ARM::VLD4d16_UPD:
2671 case ARM::VLD4d32_UPD:
2672 case ARM::VLD1d64Qwb_fixed:
2673 case ARM::VLD1d64Qwb_register:
2674 case ARM::VLD4q8_UPD:
2675 case ARM::VLD4q16_UPD:
2676 case ARM::VLD4q32_UPD:
2677 case ARM::VLD1DUPq8:
2678 case ARM::VLD1DUPq16:
2679 case ARM::VLD1DUPq32:
2680 case ARM::VLD1DUPq8wb_fixed:
2681 case ARM::VLD1DUPq16wb_fixed:
2682 case ARM::VLD1DUPq32wb_fixed:
2683 case ARM::VLD1DUPq8wb_register:
2684 case ARM::VLD1DUPq16wb_register:
2685 case ARM::VLD1DUPq32wb_register:
2686 case ARM::VLD2DUPd8:
2687 case ARM::VLD2DUPd16:
2688 case ARM::VLD2DUPd32:
2689 case ARM::VLD2DUPd8wb_fixed:
2690 case ARM::VLD2DUPd16wb_fixed:
2691 case ARM::VLD2DUPd32wb_fixed:
2692 case ARM::VLD2DUPd8wb_register:
2693 case ARM::VLD2DUPd16wb_register:
2694 case ARM::VLD2DUPd32wb_register:
2695 case ARM::VLD4DUPd8:
2696 case ARM::VLD4DUPd16:
2697 case ARM::VLD4DUPd32:
2698 case ARM::VLD4DUPd8_UPD:
2699 case ARM::VLD4DUPd16_UPD:
2700 case ARM::VLD4DUPd32_UPD:
2702 case ARM::VLD1LNd16:
2703 case ARM::VLD1LNd32:
2704 case ARM::VLD1LNd8_UPD:
2705 case ARM::VLD1LNd16_UPD:
2706 case ARM::VLD1LNd32_UPD:
2708 case ARM::VLD2LNd16:
2709 case ARM::VLD2LNd32:
2710 case ARM::VLD2LNq16:
2711 case ARM::VLD2LNq32:
2712 case ARM::VLD2LNd8_UPD:
2713 case ARM::VLD2LNd16_UPD:
2714 case ARM::VLD2LNd32_UPD:
2715 case ARM::VLD2LNq16_UPD:
2716 case ARM::VLD2LNq32_UPD:
2718 case ARM::VLD4LNd16:
2719 case ARM::VLD4LNd32:
2720 case ARM::VLD4LNq16:
2721 case ARM::VLD4LNq32:
2722 case ARM::VLD4LNd8_UPD:
2723 case ARM::VLD4LNd16_UPD:
2724 case ARM::VLD4LNd32_UPD:
2725 case ARM::VLD4LNq16_UPD:
2726 case ARM::VLD4LNq32_UPD:
2727 // If the address is not 64-bit aligned, the latencies of these
2728 // instructions increases by one.
2739 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2740 const MachineInstr *DefMI, unsigned DefIdx,
2741 const MachineInstr *UseMI,
2742 unsigned UseIdx) const {
2743 // No operand latency. The caller may fall back to getInstrLatency.
2744 if (!ItinData || ItinData->isEmpty())
2747 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
2748 unsigned Reg = DefMO.getReg();
2749 const MCInstrDesc *DefMCID = &DefMI->getDesc();
2750 const MCInstrDesc *UseMCID = &UseMI->getDesc();
2752 unsigned DefAdj = 0;
2753 if (DefMI->isBundle()) {
2754 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
2755 DefMCID = &DefMI->getDesc();
2757 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
2758 DefMI->isRegSequence() || DefMI->isImplicitDef()) {
2762 unsigned UseAdj = 0;
2763 if (UseMI->isBundle()) {
2765 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
2766 Reg, NewUseIdx, UseAdj);
2772 UseMCID = &UseMI->getDesc();
2775 if (Reg == ARM::CPSR) {
2776 if (DefMI->getOpcode() == ARM::FMSTAT) {
2777 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
2778 return Subtarget.isCortexA9() ? 1 : 20;
2781 // CPSR set and branch can be paired in the same cycle.
2782 if (UseMI->isBranch())
2785 // Otherwise it takes the instruction latency (generally one).
2786 unsigned Latency = getInstrLatency(ItinData, DefMI);
2788 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
2789 // its uses. Instructions which are otherwise scheduled between them may
2790 // incur a code size penalty (not able to use the CPSR setting 16-bit
2792 if (Latency > 0 && Subtarget.isThumb2()) {
2793 const MachineFunction *MF = DefMI->getParent()->getParent();
2794 if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize))
2800 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
2803 unsigned DefAlign = DefMI->hasOneMemOperand()
2804 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
2805 unsigned UseAlign = UseMI->hasOneMemOperand()
2806 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
2808 // Get the itinerary's latency if possible, and handle variable_ops.
2809 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
2810 *UseMCID, UseIdx, UseAlign);
2811 // Unable to find operand latency. The caller may resort to getInstrLatency.
2815 // Adjust for IT block position.
2816 int Adj = DefAdj + UseAdj;
2818 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
2819 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
2820 if (Adj >= 0 || (int)Latency > -Adj) {
2821 return Latency + Adj;
2823 // Return the itinerary latency, which may be zero but not less than zero.
2828 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2829 SDNode *DefNode, unsigned DefIdx,
2830 SDNode *UseNode, unsigned UseIdx) const {
2831 if (!DefNode->isMachineOpcode())
2834 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
2836 if (isZeroCost(DefMCID.Opcode))
2839 if (!ItinData || ItinData->isEmpty())
2840 return DefMCID.mayLoad() ? 3 : 1;
2842 if (!UseNode->isMachineOpcode()) {
2843 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
2844 if (Subtarget.isCortexA9())
2845 return Latency <= 2 ? 1 : Latency - 1;
2847 return Latency <= 3 ? 1 : Latency - 2;
2850 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
2851 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
2852 unsigned DefAlign = !DefMN->memoperands_empty()
2853 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
2854 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
2855 unsigned UseAlign = !UseMN->memoperands_empty()
2856 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
2857 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
2858 UseMCID, UseIdx, UseAlign);
2861 (Subtarget.isCortexA8() || Subtarget.isCortexA9())) {
2862 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
2863 // variants are one cycle cheaper.
2864 switch (DefMCID.getOpcode()) {
2869 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2870 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2872 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
2879 case ARM::t2LDRSHs: {
2880 // Thumb2 mode: lsl only.
2882 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
2883 if (ShAmt == 0 || ShAmt == 2)
2890 if (DefAlign < 8 && Subtarget.isCortexA9())
2891 switch (DefMCID.getOpcode()) {
2897 case ARM::VLD1q8wb_register:
2898 case ARM::VLD1q16wb_register:
2899 case ARM::VLD1q32wb_register:
2900 case ARM::VLD1q64wb_register:
2901 case ARM::VLD1q8wb_fixed:
2902 case ARM::VLD1q16wb_fixed:
2903 case ARM::VLD1q32wb_fixed:
2904 case ARM::VLD1q64wb_fixed:
2908 case ARM::VLD2q8Pseudo:
2909 case ARM::VLD2q16Pseudo:
2910 case ARM::VLD2q32Pseudo:
2911 case ARM::VLD2d8wb_fixed:
2912 case ARM::VLD2d16wb_fixed:
2913 case ARM::VLD2d32wb_fixed:
2914 case ARM::VLD2q8PseudoWB_fixed:
2915 case ARM::VLD2q16PseudoWB_fixed:
2916 case ARM::VLD2q32PseudoWB_fixed:
2917 case ARM::VLD2d8wb_register:
2918 case ARM::VLD2d16wb_register:
2919 case ARM::VLD2d32wb_register:
2920 case ARM::VLD2q8PseudoWB_register:
2921 case ARM::VLD2q16PseudoWB_register:
2922 case ARM::VLD2q32PseudoWB_register:
2923 case ARM::VLD3d8Pseudo:
2924 case ARM::VLD3d16Pseudo:
2925 case ARM::VLD3d32Pseudo:
2926 case ARM::VLD1d64TPseudo:
2927 case ARM::VLD3d8Pseudo_UPD:
2928 case ARM::VLD3d16Pseudo_UPD:
2929 case ARM::VLD3d32Pseudo_UPD:
2930 case ARM::VLD3q8Pseudo_UPD:
2931 case ARM::VLD3q16Pseudo_UPD:
2932 case ARM::VLD3q32Pseudo_UPD:
2933 case ARM::VLD3q8oddPseudo:
2934 case ARM::VLD3q16oddPseudo:
2935 case ARM::VLD3q32oddPseudo:
2936 case ARM::VLD3q8oddPseudo_UPD:
2937 case ARM::VLD3q16oddPseudo_UPD:
2938 case ARM::VLD3q32oddPseudo_UPD:
2939 case ARM::VLD4d8Pseudo:
2940 case ARM::VLD4d16Pseudo:
2941 case ARM::VLD4d32Pseudo:
2942 case ARM::VLD1d64QPseudo:
2943 case ARM::VLD4d8Pseudo_UPD:
2944 case ARM::VLD4d16Pseudo_UPD:
2945 case ARM::VLD4d32Pseudo_UPD:
2946 case ARM::VLD4q8Pseudo_UPD:
2947 case ARM::VLD4q16Pseudo_UPD:
2948 case ARM::VLD4q32Pseudo_UPD:
2949 case ARM::VLD4q8oddPseudo:
2950 case ARM::VLD4q16oddPseudo:
2951 case ARM::VLD4q32oddPseudo:
2952 case ARM::VLD4q8oddPseudo_UPD:
2953 case ARM::VLD4q16oddPseudo_UPD:
2954 case ARM::VLD4q32oddPseudo_UPD:
2955 case ARM::VLD1DUPq8:
2956 case ARM::VLD1DUPq16:
2957 case ARM::VLD1DUPq32:
2958 case ARM::VLD1DUPq8wb_fixed:
2959 case ARM::VLD1DUPq16wb_fixed:
2960 case ARM::VLD1DUPq32wb_fixed:
2961 case ARM::VLD1DUPq8wb_register:
2962 case ARM::VLD1DUPq16wb_register:
2963 case ARM::VLD1DUPq32wb_register:
2964 case ARM::VLD2DUPd8:
2965 case ARM::VLD2DUPd16:
2966 case ARM::VLD2DUPd32:
2967 case ARM::VLD2DUPd8wb_fixed:
2968 case ARM::VLD2DUPd16wb_fixed:
2969 case ARM::VLD2DUPd32wb_fixed:
2970 case ARM::VLD2DUPd8wb_register:
2971 case ARM::VLD2DUPd16wb_register:
2972 case ARM::VLD2DUPd32wb_register:
2973 case ARM::VLD4DUPd8Pseudo:
2974 case ARM::VLD4DUPd16Pseudo:
2975 case ARM::VLD4DUPd32Pseudo:
2976 case ARM::VLD4DUPd8Pseudo_UPD:
2977 case ARM::VLD4DUPd16Pseudo_UPD:
2978 case ARM::VLD4DUPd32Pseudo_UPD:
2979 case ARM::VLD1LNq8Pseudo:
2980 case ARM::VLD1LNq16Pseudo:
2981 case ARM::VLD1LNq32Pseudo:
2982 case ARM::VLD1LNq8Pseudo_UPD:
2983 case ARM::VLD1LNq16Pseudo_UPD:
2984 case ARM::VLD1LNq32Pseudo_UPD:
2985 case ARM::VLD2LNd8Pseudo:
2986 case ARM::VLD2LNd16Pseudo:
2987 case ARM::VLD2LNd32Pseudo:
2988 case ARM::VLD2LNq16Pseudo:
2989 case ARM::VLD2LNq32Pseudo:
2990 case ARM::VLD2LNd8Pseudo_UPD:
2991 case ARM::VLD2LNd16Pseudo_UPD:
2992 case ARM::VLD2LNd32Pseudo_UPD:
2993 case ARM::VLD2LNq16Pseudo_UPD:
2994 case ARM::VLD2LNq32Pseudo_UPD:
2995 case ARM::VLD4LNd8Pseudo:
2996 case ARM::VLD4LNd16Pseudo:
2997 case ARM::VLD4LNd32Pseudo:
2998 case ARM::VLD4LNq16Pseudo:
2999 case ARM::VLD4LNq32Pseudo:
3000 case ARM::VLD4LNd8Pseudo_UPD:
3001 case ARM::VLD4LNd16Pseudo_UPD:
3002 case ARM::VLD4LNd32Pseudo_UPD:
3003 case ARM::VLD4LNq16Pseudo_UPD:
3004 case ARM::VLD4LNq32Pseudo_UPD:
3005 // If the address is not 64-bit aligned, the latencies of these
3006 // instructions increases by one.
3015 ARMBaseInstrInfo::getOutputLatency(const InstrItineraryData *ItinData,
3016 const MachineInstr *DefMI, unsigned DefIdx,
3017 const MachineInstr *DepMI) const {
3018 unsigned Reg = DefMI->getOperand(DefIdx).getReg();
3019 if (DepMI->readsRegister(Reg, &getRegisterInfo()) || !isPredicated(DepMI))
3022 // If the second MI is predicated, then there is an implicit use dependency.
3023 return getInstrLatency(ItinData, DefMI);
3026 unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3027 const MachineInstr *MI,
3028 unsigned *PredCost) const {
3029 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3030 MI->isRegSequence() || MI->isImplicitDef())
3033 // An instruction scheduler typically runs on unbundled instructions, however
3034 // other passes may query the latency of a bundled instruction.
3035 if (MI->isBundle()) {
3036 unsigned Latency = 0;
3037 MachineBasicBlock::const_instr_iterator I = MI;
3038 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3039 while (++I != E && I->isInsideBundle()) {
3040 if (I->getOpcode() != ARM::t2IT)
3041 Latency += getInstrLatency(ItinData, I, PredCost);
3046 const MCInstrDesc &MCID = MI->getDesc();
3047 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) {
3048 // When predicated, CPSR is an additional source operand for CPSR updating
3049 // instructions, this apparently increases their latencies.
3052 // Be sure to call getStageLatency for an empty itinerary in case it has a
3053 // valid MinLatency property.
3055 return MI->mayLoad() ? 3 : 1;
3057 unsigned Class = MCID.getSchedClass();
3059 // For instructions with variable uops, use uops as latency.
3060 if (!ItinData->isEmpty() && !ItinData->Itineraries[Class].NumMicroOps) {
3061 return getNumMicroOps(ItinData, MI);
3063 // For the common case, fall back on the itinerary's latency.
3064 unsigned Latency = ItinData->getStageLatency(Class);
3066 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3067 unsigned DefAlign = MI->hasOneMemOperand()
3068 ? (*MI->memoperands_begin())->getAlignment() : 0;
3069 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
3070 if (Adj >= 0 || (int)Latency > -Adj) {
3071 return Latency + Adj;
3076 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3077 SDNode *Node) const {
3078 if (!Node->isMachineOpcode())
3081 if (!ItinData || ItinData->isEmpty())
3084 unsigned Opcode = Node->getMachineOpcode();
3087 return ItinData->getStageLatency(get(Opcode).getSchedClass());
3094 bool ARMBaseInstrInfo::
3095 hasHighOperandLatency(const InstrItineraryData *ItinData,
3096 const MachineRegisterInfo *MRI,
3097 const MachineInstr *DefMI, unsigned DefIdx,
3098 const MachineInstr *UseMI, unsigned UseIdx) const {
3099 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3100 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
3101 if (Subtarget.isCortexA8() &&
3102 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
3103 // CortexA8 VFP instructions are not pipelined.
3106 // Hoist VFP / NEON instructions with 4 or higher latency.
3107 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx,
3110 Latency = getInstrLatency(ItinData, DefMI);
3113 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
3114 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
3117 bool ARMBaseInstrInfo::
3118 hasLowDefLatency(const InstrItineraryData *ItinData,
3119 const MachineInstr *DefMI, unsigned DefIdx) const {
3120 if (!ItinData || ItinData->isEmpty())
3123 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3124 if (DDomain == ARMII::DomainGeneral) {
3125 unsigned DefClass = DefMI->getDesc().getSchedClass();
3126 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3127 return (DefCycle != -1 && DefCycle <= 2);
3132 bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
3133 StringRef &ErrInfo) const {
3134 if (convertAddSubFlagsOpcode(MI->getOpcode())) {
3135 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
3142 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
3143 unsigned &AddSubOpc,
3144 bool &NegAcc, bool &HasLane) const {
3145 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
3146 if (I == MLxEntryMap.end())
3149 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
3150 MulOpc = Entry.MulOpc;
3151 AddSubOpc = Entry.AddSubOpc;
3152 NegAcc = Entry.NegAcc;
3153 HasLane = Entry.HasLane;
3157 //===----------------------------------------------------------------------===//
3158 // Execution domains.
3159 //===----------------------------------------------------------------------===//
3161 // Some instructions go down the NEON pipeline, some go down the VFP pipeline,
3162 // and some can go down both. The vmov instructions go down the VFP pipeline,
3163 // but they can be changed to vorr equivalents that are executed by the NEON
3166 // We use the following execution domain numbering:
3174 // Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
3176 std::pair<uint16_t, uint16_t>
3177 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
3178 // VMOVD is a VFP instruction, but can be changed to NEON if it isn't
3180 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
3181 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3183 // No other instructions can be swizzled, so just determine their domain.
3184 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
3186 if (Domain & ARMII::DomainNEON)
3187 return std::make_pair(ExeNEON, 0);
3189 // Certain instructions can go either way on Cortex-A8.
3190 // Treat them as NEON instructions.
3191 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
3192 return std::make_pair(ExeNEON, 0);
3194 if (Domain & ARMII::DomainVFP)
3195 return std::make_pair(ExeVFP, 0);
3197 return std::make_pair(ExeGeneric, 0);
3201 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
3202 // We only know how to change VMOVD into VORR.
3203 assert(MI->getOpcode() == ARM::VMOVD && "Can only swizzle VMOVD");
3204 if (Domain != ExeNEON)
3207 // Zap the predicate operands.
3208 assert(!isPredicated(MI) && "Cannot predicate a VORRd");
3209 MI->RemoveOperand(3);
3210 MI->RemoveOperand(2);
3212 // Change to a VORRd which requires two identical use operands.
3213 MI->setDesc(get(ARM::VORRd));
3215 // Add the extra source operand and new predicates.
3216 // This will go before any implicit ops.
3217 AddDefaultPred(MachineInstrBuilder(MI).addOperand(MI->getOperand(1)));
3220 bool ARMBaseInstrInfo::hasNOP() const {
3221 return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;