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 //===----------------------------------------------------------------------===//
15 #include "ARMBaseInstrInfo.h"
16 #include "ARMBaseRegisterInfo.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMFeatures.h"
19 #include "ARMHazardRecognizer.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "MCTargetDesc/ARMAddressingModes.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/CodeGen/LiveVariables.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineJumpTableInfo.h"
28 #include "llvm/CodeGen/MachineMemOperand.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAGNodes.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/GlobalValue.h"
34 #include "llvm/MC/MCAsmInfo.h"
35 #include "llvm/MC/MCExpr.h"
36 #include "llvm/Support/BranchProbability.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
43 #define DEBUG_TYPE "arm-instrinfo"
45 #define GET_INSTRINFO_CTOR_DTOR
46 #include "ARMGenInstrInfo.inc"
49 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
50 cl::desc("Enable ARM 2-addr to 3-addr conv"));
53 WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true),
54 cl::desc("Widen ARM vmovs to vmovd when possible"));
56 static cl::opt<unsigned>
57 SwiftPartialUpdateClearance("swift-partial-update-clearance",
58 cl::Hidden, cl::init(12),
59 cl::desc("Clearance before partial register updates"));
61 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
63 uint16_t MLxOpc; // MLA / MLS opcode
64 uint16_t MulOpc; // Expanded multiplication opcode
65 uint16_t AddSubOpc; // Expanded add / sub opcode
66 bool NegAcc; // True if the acc is negated before the add / sub.
67 bool HasLane; // True if instruction has an extra "lane" operand.
70 static const ARM_MLxEntry ARM_MLxTable[] = {
71 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
73 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
74 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
75 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
76 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
77 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
78 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
79 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
80 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
83 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
84 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
85 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
86 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
87 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
88 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
89 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
90 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
93 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
94 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
96 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
97 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
98 assert(false && "Duplicated entries?");
99 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
100 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
104 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
105 // currently defaults to no prepass hazard recognizer.
106 ScheduleHazardRecognizer *
107 ARMBaseInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
108 const ScheduleDAG *DAG) const {
109 if (usePreRAHazardRecognizer()) {
110 const InstrItineraryData *II =
111 static_cast<const ARMSubtarget *>(STI)->getInstrItineraryData();
112 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
114 return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG);
117 ScheduleHazardRecognizer *ARMBaseInstrInfo::
118 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
119 const ScheduleDAG *DAG) const {
120 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
121 return (ScheduleHazardRecognizer *)new ARMHazardRecognizer(II, DAG);
122 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
126 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
127 MachineBasicBlock::iterator &MBBI,
128 LiveVariables *LV) const {
129 // FIXME: Thumb2 support.
134 MachineInstr *MI = MBBI;
135 MachineFunction &MF = *MI->getParent()->getParent();
136 uint64_t TSFlags = MI->getDesc().TSFlags;
138 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
139 default: return nullptr;
140 case ARMII::IndexModePre:
143 case ARMII::IndexModePost:
147 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
149 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
153 MachineInstr *UpdateMI = nullptr;
154 MachineInstr *MemMI = nullptr;
155 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
156 const MCInstrDesc &MCID = MI->getDesc();
157 unsigned NumOps = MCID.getNumOperands();
158 bool isLoad = !MI->mayStore();
159 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
160 const MachineOperand &Base = MI->getOperand(2);
161 const MachineOperand &Offset = MI->getOperand(NumOps-3);
162 unsigned WBReg = WB.getReg();
163 unsigned BaseReg = Base.getReg();
164 unsigned OffReg = Offset.getReg();
165 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
166 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
168 default: llvm_unreachable("Unknown indexed op!");
169 case ARMII::AddrMode2: {
170 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
171 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
173 if (ARM_AM::getSOImmVal(Amt) == -1)
174 // Can't encode it in a so_imm operand. This transformation will
175 // add more than 1 instruction. Abandon!
177 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
178 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
179 .addReg(BaseReg).addImm(Amt)
180 .addImm(Pred).addReg(0).addReg(0);
181 } else if (Amt != 0) {
182 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
183 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
184 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
185 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
186 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
187 .addImm(Pred).addReg(0).addReg(0);
189 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
190 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
191 .addReg(BaseReg).addReg(OffReg)
192 .addImm(Pred).addReg(0).addReg(0);
195 case ARMII::AddrMode3 : {
196 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
197 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
199 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
200 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
201 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
202 .addReg(BaseReg).addImm(Amt)
203 .addImm(Pred).addReg(0).addReg(0);
205 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
206 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
207 .addReg(BaseReg).addReg(OffReg)
208 .addImm(Pred).addReg(0).addReg(0);
213 std::vector<MachineInstr*> NewMIs;
216 MemMI = BuildMI(MF, MI->getDebugLoc(),
217 get(MemOpc), MI->getOperand(0).getReg())
218 .addReg(WBReg).addImm(0).addImm(Pred);
220 MemMI = BuildMI(MF, MI->getDebugLoc(),
221 get(MemOpc)).addReg(MI->getOperand(1).getReg())
222 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
223 NewMIs.push_back(MemMI);
224 NewMIs.push_back(UpdateMI);
227 MemMI = BuildMI(MF, MI->getDebugLoc(),
228 get(MemOpc), MI->getOperand(0).getReg())
229 .addReg(BaseReg).addImm(0).addImm(Pred);
231 MemMI = BuildMI(MF, MI->getDebugLoc(),
232 get(MemOpc)).addReg(MI->getOperand(1).getReg())
233 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
235 UpdateMI->getOperand(0).setIsDead();
236 NewMIs.push_back(UpdateMI);
237 NewMIs.push_back(MemMI);
240 // Transfer LiveVariables states, kill / dead info.
242 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
243 MachineOperand &MO = MI->getOperand(i);
244 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
245 unsigned Reg = MO.getReg();
247 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
249 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
251 LV->addVirtualRegisterDead(Reg, NewMI);
253 if (MO.isUse() && MO.isKill()) {
254 for (unsigned j = 0; j < 2; ++j) {
255 // Look at the two new MI's in reverse order.
256 MachineInstr *NewMI = NewMIs[j];
257 if (!NewMI->readsRegister(Reg))
259 LV->addVirtualRegisterKilled(Reg, NewMI);
260 if (VI.removeKill(MI))
261 VI.Kills.push_back(NewMI);
269 MFI->insert(MBBI, NewMIs[1]);
270 MFI->insert(MBBI, NewMIs[0]);
276 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
277 MachineBasicBlock *&FBB,
278 SmallVectorImpl<MachineOperand> &Cond,
279 bool AllowModify) const {
283 MachineBasicBlock::iterator I = MBB.end();
284 if (I == MBB.begin())
285 return false; // Empty blocks are easy.
288 // Walk backwards from the end of the basic block until the branch is
289 // analyzed or we give up.
290 while (isPredicated(I) || I->isTerminator() || I->isDebugValue()) {
292 // Flag to be raised on unanalyzeable instructions. This is useful in cases
293 // where we want to clean up on the end of the basic block before we bail
295 bool CantAnalyze = false;
297 // Skip over DEBUG values and predicated nonterminators.
298 while (I->isDebugValue() || !I->isTerminator()) {
299 if (I == MBB.begin())
304 if (isIndirectBranchOpcode(I->getOpcode()) ||
305 isJumpTableBranchOpcode(I->getOpcode())) {
306 // Indirect branches and jump tables can't be analyzed, but we still want
307 // to clean up any instructions at the tail of the basic block.
309 } else if (isUncondBranchOpcode(I->getOpcode())) {
310 TBB = I->getOperand(0).getMBB();
311 } else if (isCondBranchOpcode(I->getOpcode())) {
312 // Bail out if we encounter multiple conditional branches.
316 assert(!FBB && "FBB should have been null.");
318 TBB = I->getOperand(0).getMBB();
319 Cond.push_back(I->getOperand(1));
320 Cond.push_back(I->getOperand(2));
321 } else if (I->isReturn()) {
322 // Returns can't be analyzed, but we should run cleanup.
323 CantAnalyze = !isPredicated(I);
325 // We encountered other unrecognized terminator. Bail out immediately.
329 // Cleanup code - to be run for unpredicated unconditional branches and
331 if (!isPredicated(I) &&
332 (isUncondBranchOpcode(I->getOpcode()) ||
333 isIndirectBranchOpcode(I->getOpcode()) ||
334 isJumpTableBranchOpcode(I->getOpcode()) ||
336 // Forget any previous condition branch information - it no longer applies.
340 // If we can modify the function, delete everything below this
341 // unconditional branch.
343 MachineBasicBlock::iterator DI = std::next(I);
344 while (DI != MBB.end()) {
345 MachineInstr *InstToDelete = DI;
347 InstToDelete->eraseFromParent();
355 if (I == MBB.begin())
361 // We made it past the terminators without bailing out - we must have
362 // analyzed this branch successfully.
367 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
368 MachineBasicBlock::iterator I = MBB.end();
369 if (I == MBB.begin()) return 0;
371 while (I->isDebugValue()) {
372 if (I == MBB.begin())
376 if (!isUncondBranchOpcode(I->getOpcode()) &&
377 !isCondBranchOpcode(I->getOpcode()))
380 // Remove the branch.
381 I->eraseFromParent();
385 if (I == MBB.begin()) return 1;
387 if (!isCondBranchOpcode(I->getOpcode()))
390 // Remove the branch.
391 I->eraseFromParent();
396 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
397 MachineBasicBlock *FBB,
398 const SmallVectorImpl<MachineOperand> &Cond,
400 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
401 int BOpc = !AFI->isThumbFunction()
402 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
403 int BccOpc = !AFI->isThumbFunction()
404 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
405 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
407 // Shouldn't be a fall through.
408 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
409 assert((Cond.size() == 2 || Cond.size() == 0) &&
410 "ARM branch conditions have two components!");
413 if (Cond.empty()) { // Unconditional branch?
415 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
417 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
419 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
420 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
424 // Two-way conditional branch.
425 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
426 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
428 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
430 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
434 bool ARMBaseInstrInfo::
435 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
436 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
437 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
441 bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
442 if (MI->isBundle()) {
443 MachineBasicBlock::const_instr_iterator I = MI;
444 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
445 while (++I != E && I->isInsideBundle()) {
446 int PIdx = I->findFirstPredOperandIdx();
447 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
453 int PIdx = MI->findFirstPredOperandIdx();
454 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
457 bool ARMBaseInstrInfo::
458 PredicateInstruction(MachineInstr *MI,
459 const SmallVectorImpl<MachineOperand> &Pred) const {
460 unsigned Opc = MI->getOpcode();
461 if (isUncondBranchOpcode(Opc)) {
462 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
463 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
464 .addImm(Pred[0].getImm())
465 .addReg(Pred[1].getReg());
469 int PIdx = MI->findFirstPredOperandIdx();
471 MachineOperand &PMO = MI->getOperand(PIdx);
472 PMO.setImm(Pred[0].getImm());
473 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
479 bool ARMBaseInstrInfo::
480 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
481 const SmallVectorImpl<MachineOperand> &Pred2) const {
482 if (Pred1.size() > 2 || Pred2.size() > 2)
485 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
486 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
496 return CC2 == ARMCC::HI;
498 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
500 return CC2 == ARMCC::GT;
502 return CC2 == ARMCC::LT;
506 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
507 std::vector<MachineOperand> &Pred) const {
509 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
510 const MachineOperand &MO = MI->getOperand(i);
511 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
512 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) {
521 static bool isCPSRDefined(const MachineInstr *MI) {
522 for (const auto &MO : MI->operands())
523 if (MO.isReg() && MO.getReg() == ARM::CPSR && MO.isDef())
528 static bool isEligibleForITBlock(const MachineInstr *MI) {
529 switch (MI->getOpcode()) {
530 default: return true;
531 case ARM::tADC: // ADC (register) T1
532 case ARM::tADDi3: // ADD (immediate) T1
533 case ARM::tADDi8: // ADD (immediate) T2
534 case ARM::tADDrr: // ADD (register) T1
535 case ARM::tAND: // AND (register) T1
536 case ARM::tASRri: // ASR (immediate) T1
537 case ARM::tASRrr: // ASR (register) T1
538 case ARM::tBIC: // BIC (register) T1
539 case ARM::tEOR: // EOR (register) T1
540 case ARM::tLSLri: // LSL (immediate) T1
541 case ARM::tLSLrr: // LSL (register) T1
542 case ARM::tLSRri: // LSR (immediate) T1
543 case ARM::tLSRrr: // LSR (register) T1
544 case ARM::tMUL: // MUL T1
545 case ARM::tMVN: // MVN (register) T1
546 case ARM::tORR: // ORR (register) T1
547 case ARM::tROR: // ROR (register) T1
548 case ARM::tRSB: // RSB (immediate) T1
549 case ARM::tSBC: // SBC (register) T1
550 case ARM::tSUBi3: // SUB (immediate) T1
551 case ARM::tSUBi8: // SUB (immediate) T2
552 case ARM::tSUBrr: // SUB (register) T1
553 return !isCPSRDefined(MI);
557 /// isPredicable - Return true if the specified instruction can be predicated.
558 /// By default, this returns true for every instruction with a
559 /// PredicateOperand.
560 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
561 if (!MI->isPredicable())
564 if (!isEligibleForITBlock(MI))
567 ARMFunctionInfo *AFI =
568 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
570 if (AFI->isThumb2Function()) {
571 if (getSubtarget().restrictIT())
572 return isV8EligibleForIT(MI);
573 } else { // non-Thumb
574 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON)
582 template <> bool IsCPSRDead<MachineInstr>(MachineInstr *MI) {
583 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
584 const MachineOperand &MO = MI->getOperand(i);
585 if (!MO.isReg() || MO.isUndef() || MO.isUse())
587 if (MO.getReg() != ARM::CPSR)
592 // all definitions of CPSR are dead
597 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
598 LLVM_ATTRIBUTE_NOINLINE
599 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
601 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
603 assert(JTI < JT.size());
604 return JT[JTI].MBBs.size();
607 /// GetInstSize - Return the size of the specified MachineInstr.
609 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
610 const MachineBasicBlock &MBB = *MI->getParent();
611 const MachineFunction *MF = MBB.getParent();
612 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
614 const MCInstrDesc &MCID = MI->getDesc();
616 return MCID.getSize();
618 // If this machine instr is an inline asm, measure it.
619 if (MI->getOpcode() == ARM::INLINEASM)
620 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
621 unsigned Opc = MI->getOpcode();
624 // pseudo-instruction sizes are zero.
626 case TargetOpcode::BUNDLE:
627 return getInstBundleLength(MI);
628 case ARM::MOVi16_ga_pcrel:
629 case ARM::MOVTi16_ga_pcrel:
630 case ARM::t2MOVi16_ga_pcrel:
631 case ARM::t2MOVTi16_ga_pcrel:
634 case ARM::t2MOVi32imm:
636 case ARM::CONSTPOOL_ENTRY:
637 // If this machine instr is a constant pool entry, its size is recorded as
639 return MI->getOperand(2).getImm();
640 case ARM::Int_eh_sjlj_longjmp:
642 case ARM::tInt_eh_sjlj_longjmp:
644 case ARM::Int_eh_sjlj_setjmp:
645 case ARM::Int_eh_sjlj_setjmp_nofp:
647 case ARM::tInt_eh_sjlj_setjmp:
648 case ARM::t2Int_eh_sjlj_setjmp:
649 case ARM::t2Int_eh_sjlj_setjmp_nofp:
657 case ARM::t2TBH_JT: {
658 // These are jumptable branches, i.e. a branch followed by an inlined
659 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
660 // entry is one byte; TBH two byte each.
661 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
662 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
663 unsigned NumOps = MCID.getNumOperands();
664 MachineOperand JTOP =
665 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
666 unsigned JTI = JTOP.getIndex();
667 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
668 assert(MJTI != nullptr);
669 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
670 assert(JTI < JT.size());
671 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
672 // 4 aligned. The assembler / linker may add 2 byte padding just before
673 // the JT entries. The size does not include this padding; the
674 // constant islands pass does separate bookkeeping for it.
675 // FIXME: If we know the size of the function is less than (1 << 16) *2
676 // bytes, we can use 16-bit entries instead. Then there won't be an
678 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
679 unsigned NumEntries = getNumJTEntries(JT, JTI);
680 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
681 // Make sure the instruction that follows TBB is 2-byte aligned.
682 // FIXME: Constant island pass should insert an "ALIGN" instruction
685 return NumEntries * EntrySize + InstSize;
690 unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
692 MachineBasicBlock::const_instr_iterator I = MI;
693 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
694 while (++I != E && I->isInsideBundle()) {
695 assert(!I->isBundle() && "No nested bundle!");
696 Size += GetInstSizeInBytes(&*I);
701 void ARMBaseInstrInfo::copyFromCPSR(MachineBasicBlock &MBB,
702 MachineBasicBlock::iterator I,
703 unsigned DestReg, bool KillSrc,
704 const ARMSubtarget &Subtarget) const {
705 unsigned Opc = Subtarget.isThumb()
706 ? (Subtarget.isMClass() ? ARM::t2MRS_M : ARM::t2MRS_AR)
709 MachineInstrBuilder MIB =
710 BuildMI(MBB, I, I->getDebugLoc(), get(Opc), DestReg);
712 // There is only 1 A/R class MRS instruction, and it always refers to
713 // APSR. However, there are lots of other possibilities on M-class cores.
714 if (Subtarget.isMClass())
719 MIB.addReg(ARM::CPSR, RegState::Implicit | getKillRegState(KillSrc));
722 void ARMBaseInstrInfo::copyToCPSR(MachineBasicBlock &MBB,
723 MachineBasicBlock::iterator I,
724 unsigned SrcReg, bool KillSrc,
725 const ARMSubtarget &Subtarget) const {
726 unsigned Opc = Subtarget.isThumb()
727 ? (Subtarget.isMClass() ? ARM::t2MSR_M : ARM::t2MSR_AR)
730 MachineInstrBuilder MIB = BuildMI(MBB, I, I->getDebugLoc(), get(Opc));
732 if (Subtarget.isMClass())
737 MIB.addReg(SrcReg, getKillRegState(KillSrc));
741 MIB.addReg(ARM::CPSR, RegState::Implicit | RegState::Define);
744 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
745 MachineBasicBlock::iterator I, DebugLoc DL,
746 unsigned DestReg, unsigned SrcReg,
747 bool KillSrc) const {
748 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
749 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
751 if (GPRDest && GPRSrc) {
752 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
753 .addReg(SrcReg, getKillRegState(KillSrc))));
757 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
758 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
761 if (SPRDest && SPRSrc)
763 else if (GPRDest && SPRSrc)
765 else if (SPRDest && GPRSrc)
767 else if (ARM::DPRRegClass.contains(DestReg, SrcReg) && !Subtarget.isFPOnlySP())
769 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
773 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
774 MIB.addReg(SrcReg, getKillRegState(KillSrc));
775 if (Opc == ARM::VORRq)
776 MIB.addReg(SrcReg, getKillRegState(KillSrc));
781 // Handle register classes that require multiple instructions.
782 unsigned BeginIdx = 0;
783 unsigned SubRegs = 0;
786 // Use VORRq when possible.
787 if (ARM::QQPRRegClass.contains(DestReg, SrcReg)) {
789 BeginIdx = ARM::qsub_0;
791 } else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg)) {
793 BeginIdx = ARM::qsub_0;
795 // Fall back to VMOVD.
796 } else if (ARM::DPairRegClass.contains(DestReg, SrcReg)) {
798 BeginIdx = ARM::dsub_0;
800 } else if (ARM::DTripleRegClass.contains(DestReg, SrcReg)) {
802 BeginIdx = ARM::dsub_0;
804 } else if (ARM::DQuadRegClass.contains(DestReg, SrcReg)) {
806 BeginIdx = ARM::dsub_0;
808 } else if (ARM::GPRPairRegClass.contains(DestReg, SrcReg)) {
809 Opc = Subtarget.isThumb2() ? ARM::tMOVr : ARM::MOVr;
810 BeginIdx = ARM::gsub_0;
812 } else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg)) {
814 BeginIdx = ARM::dsub_0;
817 } else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg)) {
819 BeginIdx = ARM::dsub_0;
822 } else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg)) {
824 BeginIdx = ARM::dsub_0;
827 } else if (ARM::DPRRegClass.contains(DestReg, SrcReg) && Subtarget.isFPOnlySP()) {
829 BeginIdx = ARM::ssub_0;
831 } else if (SrcReg == ARM::CPSR) {
832 copyFromCPSR(MBB, I, DestReg, KillSrc, Subtarget);
834 } else if (DestReg == ARM::CPSR) {
835 copyToCPSR(MBB, I, SrcReg, KillSrc, Subtarget);
839 assert(Opc && "Impossible reg-to-reg copy");
841 const TargetRegisterInfo *TRI = &getRegisterInfo();
842 MachineInstrBuilder Mov;
844 // Copy register tuples backward when the first Dest reg overlaps with SrcReg.
845 if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) {
846 BeginIdx = BeginIdx + ((SubRegs - 1) * Spacing);
850 SmallSet<unsigned, 4> DstRegs;
852 for (unsigned i = 0; i != SubRegs; ++i) {
853 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i * Spacing);
854 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i * Spacing);
855 assert(Dst && Src && "Bad sub-register");
857 assert(!DstRegs.count(Src) && "destructive vector copy");
860 Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst).addReg(Src);
861 // VORR takes two source operands.
862 if (Opc == ARM::VORRq)
864 Mov = AddDefaultPred(Mov);
866 if (Opc == ARM::MOVr)
867 Mov = AddDefaultCC(Mov);
869 // Add implicit super-register defs and kills to the last instruction.
870 Mov->addRegisterDefined(DestReg, TRI);
872 Mov->addRegisterKilled(SrcReg, TRI);
875 const MachineInstrBuilder &
876 ARMBaseInstrInfo::AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
877 unsigned SubIdx, unsigned State,
878 const TargetRegisterInfo *TRI) const {
880 return MIB.addReg(Reg, State);
882 if (TargetRegisterInfo::isPhysicalRegister(Reg))
883 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
884 return MIB.addReg(Reg, State, SubIdx);
887 void ARMBaseInstrInfo::
888 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
889 unsigned SrcReg, bool isKill, int FI,
890 const TargetRegisterClass *RC,
891 const TargetRegisterInfo *TRI) const {
893 if (I != MBB.end()) DL = I->getDebugLoc();
894 MachineFunction &MF = *MBB.getParent();
895 MachineFrameInfo &MFI = *MF.getFrameInfo();
896 unsigned Align = MFI.getObjectAlignment(FI);
898 MachineMemOperand *MMO =
899 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
900 MachineMemOperand::MOStore,
901 MFI.getObjectSize(FI),
904 switch (RC->getSize()) {
906 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
907 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
908 .addReg(SrcReg, getKillRegState(isKill))
909 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
910 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
911 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
912 .addReg(SrcReg, getKillRegState(isKill))
913 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
915 llvm_unreachable("Unknown reg class!");
918 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
919 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
920 .addReg(SrcReg, getKillRegState(isKill))
921 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
922 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
923 if (Subtarget.hasV5TEOps()) {
924 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::STRD));
925 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
926 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
927 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
931 // Fallback to STM instruction, which has existed since the dawn of
933 MachineInstrBuilder MIB =
934 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STMIA))
935 .addFrameIndex(FI).addMemOperand(MMO));
936 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
937 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
940 llvm_unreachable("Unknown reg class!");
943 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
944 // Use aligned spills if the stack can be realigned.
945 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
946 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
947 .addFrameIndex(FI).addImm(16)
948 .addReg(SrcReg, getKillRegState(isKill))
949 .addMemOperand(MMO));
951 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
952 .addReg(SrcReg, getKillRegState(isKill))
954 .addMemOperand(MMO));
957 llvm_unreachable("Unknown reg class!");
960 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
961 // Use aligned spills if the stack can be realigned.
962 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
963 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
964 .addFrameIndex(FI).addImm(16)
965 .addReg(SrcReg, getKillRegState(isKill))
966 .addMemOperand(MMO));
968 MachineInstrBuilder MIB =
969 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
972 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
973 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
974 AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
977 llvm_unreachable("Unknown reg class!");
980 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
981 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
982 // FIXME: It's possible to only store part of the QQ register if the
983 // spilled def has a sub-register index.
984 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
985 .addFrameIndex(FI).addImm(16)
986 .addReg(SrcReg, getKillRegState(isKill))
987 .addMemOperand(MMO));
989 MachineInstrBuilder MIB =
990 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
993 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
994 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
995 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
996 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
999 llvm_unreachable("Unknown reg class!");
1002 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
1003 MachineInstrBuilder MIB =
1004 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
1006 .addMemOperand(MMO);
1007 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
1008 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
1009 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
1010 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
1011 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
1012 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
1013 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
1014 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
1016 llvm_unreachable("Unknown reg class!");
1019 llvm_unreachable("Unknown reg class!");
1024 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1025 int &FrameIndex) const {
1026 switch (MI->getOpcode()) {
1029 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
1030 if (MI->getOperand(1).isFI() &&
1031 MI->getOperand(2).isReg() &&
1032 MI->getOperand(3).isImm() &&
1033 MI->getOperand(2).getReg() == 0 &&
1034 MI->getOperand(3).getImm() == 0) {
1035 FrameIndex = MI->getOperand(1).getIndex();
1036 return MI->getOperand(0).getReg();
1044 if (MI->getOperand(1).isFI() &&
1045 MI->getOperand(2).isImm() &&
1046 MI->getOperand(2).getImm() == 0) {
1047 FrameIndex = MI->getOperand(1).getIndex();
1048 return MI->getOperand(0).getReg();
1052 case ARM::VST1d64TPseudo:
1053 case ARM::VST1d64QPseudo:
1054 if (MI->getOperand(0).isFI() &&
1055 MI->getOperand(2).getSubReg() == 0) {
1056 FrameIndex = MI->getOperand(0).getIndex();
1057 return MI->getOperand(2).getReg();
1061 if (MI->getOperand(1).isFI() &&
1062 MI->getOperand(0).getSubReg() == 0) {
1063 FrameIndex = MI->getOperand(1).getIndex();
1064 return MI->getOperand(0).getReg();
1072 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1073 int &FrameIndex) const {
1074 const MachineMemOperand *Dummy;
1075 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
1078 void ARMBaseInstrInfo::
1079 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
1080 unsigned DestReg, int FI,
1081 const TargetRegisterClass *RC,
1082 const TargetRegisterInfo *TRI) const {
1084 if (I != MBB.end()) DL = I->getDebugLoc();
1085 MachineFunction &MF = *MBB.getParent();
1086 MachineFrameInfo &MFI = *MF.getFrameInfo();
1087 unsigned Align = MFI.getObjectAlignment(FI);
1088 MachineMemOperand *MMO =
1089 MF.getMachineMemOperand(
1090 MachinePointerInfo::getFixedStack(FI),
1091 MachineMemOperand::MOLoad,
1092 MFI.getObjectSize(FI),
1095 switch (RC->getSize()) {
1097 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
1098 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
1099 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
1101 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
1102 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
1103 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
1105 llvm_unreachable("Unknown reg class!");
1108 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
1109 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
1110 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
1111 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
1112 MachineInstrBuilder MIB;
1114 if (Subtarget.hasV5TEOps()) {
1115 MIB = BuildMI(MBB, I, DL, get(ARM::LDRD));
1116 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1117 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1118 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
1120 AddDefaultPred(MIB);
1122 // Fallback to LDM instruction, which has existed since the dawn of
1124 MIB = AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDMIA))
1125 .addFrameIndex(FI).addMemOperand(MMO));
1126 MIB = AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1127 MIB = AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1130 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1131 MIB.addReg(DestReg, RegState::ImplicitDefine);
1133 llvm_unreachable("Unknown reg class!");
1136 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
1137 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1138 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
1139 .addFrameIndex(FI).addImm(16)
1140 .addMemOperand(MMO));
1142 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
1144 .addMemOperand(MMO));
1147 llvm_unreachable("Unknown reg class!");
1150 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
1151 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1152 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
1153 .addFrameIndex(FI).addImm(16)
1154 .addMemOperand(MMO));
1156 MachineInstrBuilder MIB =
1157 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1159 .addMemOperand(MMO));
1160 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1161 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1162 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1163 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1164 MIB.addReg(DestReg, RegState::ImplicitDefine);
1167 llvm_unreachable("Unknown reg class!");
1170 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
1171 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1172 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
1173 .addFrameIndex(FI).addImm(16)
1174 .addMemOperand(MMO));
1176 MachineInstrBuilder MIB =
1177 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1179 .addMemOperand(MMO);
1180 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1181 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1182 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1183 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1184 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1185 MIB.addReg(DestReg, RegState::ImplicitDefine);
1188 llvm_unreachable("Unknown reg class!");
1191 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
1192 MachineInstrBuilder MIB =
1193 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1195 .addMemOperand(MMO);
1196 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1197 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1198 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1199 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1200 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
1201 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
1202 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
1203 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
1204 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1205 MIB.addReg(DestReg, RegState::ImplicitDefine);
1207 llvm_unreachable("Unknown reg class!");
1210 llvm_unreachable("Unknown regclass!");
1215 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1216 int &FrameIndex) const {
1217 switch (MI->getOpcode()) {
1220 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
1221 if (MI->getOperand(1).isFI() &&
1222 MI->getOperand(2).isReg() &&
1223 MI->getOperand(3).isImm() &&
1224 MI->getOperand(2).getReg() == 0 &&
1225 MI->getOperand(3).getImm() == 0) {
1226 FrameIndex = MI->getOperand(1).getIndex();
1227 return MI->getOperand(0).getReg();
1235 if (MI->getOperand(1).isFI() &&
1236 MI->getOperand(2).isImm() &&
1237 MI->getOperand(2).getImm() == 0) {
1238 FrameIndex = MI->getOperand(1).getIndex();
1239 return MI->getOperand(0).getReg();
1243 case ARM::VLD1d64TPseudo:
1244 case ARM::VLD1d64QPseudo:
1245 if (MI->getOperand(1).isFI() &&
1246 MI->getOperand(0).getSubReg() == 0) {
1247 FrameIndex = MI->getOperand(1).getIndex();
1248 return MI->getOperand(0).getReg();
1252 if (MI->getOperand(1).isFI() &&
1253 MI->getOperand(0).getSubReg() == 0) {
1254 FrameIndex = MI->getOperand(1).getIndex();
1255 return MI->getOperand(0).getReg();
1263 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1264 int &FrameIndex) const {
1265 const MachineMemOperand *Dummy;
1266 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1270 ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
1271 MachineFunction &MF = *MI->getParent()->getParent();
1272 Reloc::Model RM = MF.getTarget().getRelocationModel();
1274 if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD) {
1275 assert(getSubtarget().getTargetTriple().getObjectFormat() ==
1277 "LOAD_STACK_GUARD currently supported only for MachO.");
1278 expandLoadStackGuard(MI, RM);
1279 MI->getParent()->erase(MI);
1283 // This hook gets to expand COPY instructions before they become
1284 // copyPhysReg() calls. Look for VMOVS instructions that can legally be
1285 // widened to VMOVD. We prefer the VMOVD when possible because it may be
1286 // changed into a VORR that can go down the NEON pipeline.
1287 if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15() ||
1288 Subtarget.isFPOnlySP())
1291 // Look for a copy between even S-registers. That is where we keep floats
1292 // when using NEON v2f32 instructions for f32 arithmetic.
1293 unsigned DstRegS = MI->getOperand(0).getReg();
1294 unsigned SrcRegS = MI->getOperand(1).getReg();
1295 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
1298 const TargetRegisterInfo *TRI = &getRegisterInfo();
1299 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
1301 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
1303 if (!DstRegD || !SrcRegD)
1306 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
1307 // legal if the COPY already defines the full DstRegD, and it isn't a
1308 // sub-register insertion.
1309 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
1312 // A dead copy shouldn't show up here, but reject it just in case.
1313 if (MI->getOperand(0).isDead())
1316 // All clear, widen the COPY.
1317 DEBUG(dbgs() << "widening: " << *MI);
1318 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1320 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
1321 // or some other super-register.
1322 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
1323 if (ImpDefIdx != -1)
1324 MI->RemoveOperand(ImpDefIdx);
1326 // Change the opcode and operands.
1327 MI->setDesc(get(ARM::VMOVD));
1328 MI->getOperand(0).setReg(DstRegD);
1329 MI->getOperand(1).setReg(SrcRegD);
1330 AddDefaultPred(MIB);
1332 // We are now reading SrcRegD instead of SrcRegS. This may upset the
1333 // register scavenger and machine verifier, so we need to indicate that we
1334 // are reading an undefined value from SrcRegD, but a proper value from
1336 MI->getOperand(1).setIsUndef();
1337 MIB.addReg(SrcRegS, RegState::Implicit);
1339 // SrcRegD may actually contain an unrelated value in the ssub_1
1340 // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
1341 if (MI->getOperand(1).isKill()) {
1342 MI->getOperand(1).setIsKill(false);
1343 MI->addRegisterKilled(SrcRegS, TRI, true);
1346 DEBUG(dbgs() << "replaced by: " << *MI);
1350 /// Create a copy of a const pool value. Update CPI to the new index and return
1352 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1353 MachineConstantPool *MCP = MF.getConstantPool();
1354 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1356 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1357 assert(MCPE.isMachineConstantPoolEntry() &&
1358 "Expecting a machine constantpool entry!");
1359 ARMConstantPoolValue *ACPV =
1360 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1362 unsigned PCLabelId = AFI->createPICLabelUId();
1363 ARMConstantPoolValue *NewCPV = nullptr;
1365 // FIXME: The below assumes PIC relocation model and that the function
1366 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1367 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1368 // instructions, so that's probably OK, but is PIC always correct when
1370 if (ACPV->isGlobalValue())
1371 NewCPV = ARMConstantPoolConstant::
1372 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
1374 else if (ACPV->isExtSymbol())
1375 NewCPV = ARMConstantPoolSymbol::
1376 Create(MF.getFunction()->getContext(),
1377 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
1378 else if (ACPV->isBlockAddress())
1379 NewCPV = ARMConstantPoolConstant::
1380 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
1381 ARMCP::CPBlockAddress, 4);
1382 else if (ACPV->isLSDA())
1383 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
1385 else if (ACPV->isMachineBasicBlock())
1386 NewCPV = ARMConstantPoolMBB::
1387 Create(MF.getFunction()->getContext(),
1388 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
1390 llvm_unreachable("Unexpected ARM constantpool value type!!");
1391 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1395 void ARMBaseInstrInfo::
1396 reMaterialize(MachineBasicBlock &MBB,
1397 MachineBasicBlock::iterator I,
1398 unsigned DestReg, unsigned SubIdx,
1399 const MachineInstr *Orig,
1400 const TargetRegisterInfo &TRI) const {
1401 unsigned Opcode = Orig->getOpcode();
1404 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1405 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1409 case ARM::tLDRpci_pic:
1410 case ARM::t2LDRpci_pic: {
1411 MachineFunction &MF = *MBB.getParent();
1412 unsigned CPI = Orig->getOperand(1).getIndex();
1413 unsigned PCLabelId = duplicateCPV(MF, CPI);
1414 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1416 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1417 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1424 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1425 MachineInstr *MI = TargetInstrInfo::duplicate(Orig, MF);
1426 switch(Orig->getOpcode()) {
1427 case ARM::tLDRpci_pic:
1428 case ARM::t2LDRpci_pic: {
1429 unsigned CPI = Orig->getOperand(1).getIndex();
1430 unsigned PCLabelId = duplicateCPV(MF, CPI);
1431 Orig->getOperand(1).setIndex(CPI);
1432 Orig->getOperand(2).setImm(PCLabelId);
1439 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1440 const MachineInstr *MI1,
1441 const MachineRegisterInfo *MRI) const {
1442 int Opcode = MI0->getOpcode();
1443 if (Opcode == ARM::t2LDRpci ||
1444 Opcode == ARM::t2LDRpci_pic ||
1445 Opcode == ARM::tLDRpci ||
1446 Opcode == ARM::tLDRpci_pic ||
1447 Opcode == ARM::LDRLIT_ga_pcrel ||
1448 Opcode == ARM::LDRLIT_ga_pcrel_ldr ||
1449 Opcode == ARM::tLDRLIT_ga_pcrel ||
1450 Opcode == ARM::MOV_ga_pcrel ||
1451 Opcode == ARM::MOV_ga_pcrel_ldr ||
1452 Opcode == ARM::t2MOV_ga_pcrel) {
1453 if (MI1->getOpcode() != Opcode)
1455 if (MI0->getNumOperands() != MI1->getNumOperands())
1458 const MachineOperand &MO0 = MI0->getOperand(1);
1459 const MachineOperand &MO1 = MI1->getOperand(1);
1460 if (MO0.getOffset() != MO1.getOffset())
1463 if (Opcode == ARM::LDRLIT_ga_pcrel ||
1464 Opcode == ARM::LDRLIT_ga_pcrel_ldr ||
1465 Opcode == ARM::tLDRLIT_ga_pcrel ||
1466 Opcode == ARM::MOV_ga_pcrel ||
1467 Opcode == ARM::MOV_ga_pcrel_ldr ||
1468 Opcode == ARM::t2MOV_ga_pcrel)
1469 // Ignore the PC labels.
1470 return MO0.getGlobal() == MO1.getGlobal();
1472 const MachineFunction *MF = MI0->getParent()->getParent();
1473 const MachineConstantPool *MCP = MF->getConstantPool();
1474 int CPI0 = MO0.getIndex();
1475 int CPI1 = MO1.getIndex();
1476 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1477 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1478 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1479 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1480 if (isARMCP0 && isARMCP1) {
1481 ARMConstantPoolValue *ACPV0 =
1482 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1483 ARMConstantPoolValue *ACPV1 =
1484 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1485 return ACPV0->hasSameValue(ACPV1);
1486 } else if (!isARMCP0 && !isARMCP1) {
1487 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1490 } else if (Opcode == ARM::PICLDR) {
1491 if (MI1->getOpcode() != Opcode)
1493 if (MI0->getNumOperands() != MI1->getNumOperands())
1496 unsigned Addr0 = MI0->getOperand(1).getReg();
1497 unsigned Addr1 = MI1->getOperand(1).getReg();
1498 if (Addr0 != Addr1) {
1500 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1501 !TargetRegisterInfo::isVirtualRegister(Addr1))
1504 // This assumes SSA form.
1505 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1506 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1507 // Check if the loaded value, e.g. a constantpool of a global address, are
1509 if (!produceSameValue(Def0, Def1, MRI))
1513 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1514 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1515 const MachineOperand &MO0 = MI0->getOperand(i);
1516 const MachineOperand &MO1 = MI1->getOperand(i);
1517 if (!MO0.isIdenticalTo(MO1))
1523 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1526 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1527 /// determine if two loads are loading from the same base address. It should
1528 /// only return true if the base pointers are the same and the only differences
1529 /// between the two addresses is the offset. It also returns the offsets by
1532 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1533 /// is permanently disabled.
1534 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1536 int64_t &Offset2) const {
1537 // Don't worry about Thumb: just ARM and Thumb2.
1538 if (Subtarget.isThumb1Only()) return false;
1540 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1543 switch (Load1->getMachineOpcode()) {
1557 case ARM::t2LDRSHi8:
1559 case ARM::t2LDRBi12:
1560 case ARM::t2LDRSHi12:
1564 switch (Load2->getMachineOpcode()) {
1577 case ARM::t2LDRSHi8:
1579 case ARM::t2LDRBi12:
1580 case ARM::t2LDRSHi12:
1584 // Check if base addresses and chain operands match.
1585 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1586 Load1->getOperand(4) != Load2->getOperand(4))
1589 // Index should be Reg0.
1590 if (Load1->getOperand(3) != Load2->getOperand(3))
1593 // Determine the offsets.
1594 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1595 isa<ConstantSDNode>(Load2->getOperand(1))) {
1596 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1597 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1604 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1605 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1606 /// be scheduled togther. On some targets if two loads are loading from
1607 /// addresses in the same cache line, it's better if they are scheduled
1608 /// together. This function takes two integers that represent the load offsets
1609 /// from the common base address. It returns true if it decides it's desirable
1610 /// to schedule the two loads together. "NumLoads" is the number of loads that
1611 /// have already been scheduled after Load1.
1613 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1614 /// is permanently disabled.
1615 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1616 int64_t Offset1, int64_t Offset2,
1617 unsigned NumLoads) const {
1618 // Don't worry about Thumb: just ARM and Thumb2.
1619 if (Subtarget.isThumb1Only()) return false;
1621 assert(Offset2 > Offset1);
1623 if ((Offset2 - Offset1) / 8 > 64)
1626 // Check if the machine opcodes are different. If they are different
1627 // then we consider them to not be of the same base address,
1628 // EXCEPT in the case of Thumb2 byte loads where one is LDRBi8 and the other LDRBi12.
1629 // In this case, they are considered to be the same because they are different
1630 // encoding forms of the same basic instruction.
1631 if ((Load1->getMachineOpcode() != Load2->getMachineOpcode()) &&
1632 !((Load1->getMachineOpcode() == ARM::t2LDRBi8 &&
1633 Load2->getMachineOpcode() == ARM::t2LDRBi12) ||
1634 (Load1->getMachineOpcode() == ARM::t2LDRBi12 &&
1635 Load2->getMachineOpcode() == ARM::t2LDRBi8)))
1636 return false; // FIXME: overly conservative?
1638 // Four loads in a row should be sufficient.
1645 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1646 const MachineBasicBlock *MBB,
1647 const MachineFunction &MF) const {
1648 // Debug info is never a scheduling boundary. It's necessary to be explicit
1649 // due to the special treatment of IT instructions below, otherwise a
1650 // dbg_value followed by an IT will result in the IT instruction being
1651 // considered a scheduling hazard, which is wrong. It should be the actual
1652 // instruction preceding the dbg_value instruction(s), just like it is
1653 // when debug info is not present.
1654 if (MI->isDebugValue())
1657 // Terminators and labels can't be scheduled around.
1658 if (MI->isTerminator() || MI->isPosition())
1661 // Treat the start of the IT block as a scheduling boundary, but schedule
1662 // t2IT along with all instructions following it.
1663 // FIXME: This is a big hammer. But the alternative is to add all potential
1664 // true and anti dependencies to IT block instructions as implicit operands
1665 // to the t2IT instruction. The added compile time and complexity does not
1667 MachineBasicBlock::const_iterator I = MI;
1668 // Make sure to skip any dbg_value instructions
1669 while (++I != MBB->end() && I->isDebugValue())
1671 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1674 // Don't attempt to schedule around any instruction that defines
1675 // a stack-oriented pointer, as it's unlikely to be profitable. This
1676 // saves compile time, because it doesn't require every single
1677 // stack slot reference to depend on the instruction that does the
1679 // Calls don't actually change the stack pointer, even if they have imp-defs.
1680 // No ARM calling conventions change the stack pointer. (X86 calling
1681 // conventions sometimes do).
1682 if (!MI->isCall() && MI->definesRegister(ARM::SP))
1688 bool ARMBaseInstrInfo::
1689 isProfitableToIfCvt(MachineBasicBlock &MBB,
1690 unsigned NumCycles, unsigned ExtraPredCycles,
1691 const BranchProbability &Probability) const {
1695 // Attempt to estimate the relative costs of predication versus branching.
1696 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1697 UnpredCost /= Probability.getDenominator();
1698 UnpredCost += 1; // The branch itself
1699 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1701 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1704 bool ARMBaseInstrInfo::
1705 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1706 unsigned TCycles, unsigned TExtra,
1707 MachineBasicBlock &FMBB,
1708 unsigned FCycles, unsigned FExtra,
1709 const BranchProbability &Probability) const {
1710 if (!TCycles || !FCycles)
1713 // Attempt to estimate the relative costs of predication versus branching.
1714 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1715 TUnpredCost /= Probability.getDenominator();
1717 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1718 unsigned FUnpredCost = Comp * FCycles;
1719 FUnpredCost /= Probability.getDenominator();
1721 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1722 UnpredCost += 1; // The branch itself
1723 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1725 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1729 ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
1730 MachineBasicBlock &FMBB) const {
1731 // Reduce false anti-dependencies to let Swift's out-of-order execution
1732 // engine do its thing.
1733 return Subtarget.isSwift();
1736 /// getInstrPredicate - If instruction is predicated, returns its predicate
1737 /// condition, otherwise returns AL. It also returns the condition code
1738 /// register by reference.
1740 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1741 int PIdx = MI->findFirstPredOperandIdx();
1747 PredReg = MI->getOperand(PIdx+1).getReg();
1748 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1752 int llvm::getMatchingCondBranchOpcode(int Opc) {
1757 if (Opc == ARM::t2B)
1760 llvm_unreachable("Unknown unconditional branch opcode!");
1763 /// commuteInstruction - Handle commutable instructions.
1765 ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1766 switch (MI->getOpcode()) {
1768 case ARM::t2MOVCCr: {
1769 // MOVCC can be commuted by inverting the condition.
1770 unsigned PredReg = 0;
1771 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
1772 // MOVCC AL can't be inverted. Shouldn't happen.
1773 if (CC == ARMCC::AL || PredReg != ARM::CPSR)
1775 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
1778 // After swapping the MOVCC operands, also invert the condition.
1779 MI->getOperand(MI->findFirstPredOperandIdx())
1780 .setImm(ARMCC::getOppositeCondition(CC));
1784 return TargetInstrInfo::commuteInstruction(MI, NewMI);
1787 /// Identify instructions that can be folded into a MOVCC instruction, and
1788 /// return the defining instruction.
1789 static MachineInstr *canFoldIntoMOVCC(unsigned Reg,
1790 const MachineRegisterInfo &MRI,
1791 const TargetInstrInfo *TII) {
1792 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1794 if (!MRI.hasOneNonDBGUse(Reg))
1796 MachineInstr *MI = MRI.getVRegDef(Reg);
1799 // MI is folded into the MOVCC by predicating it.
1800 if (!MI->isPredicable())
1802 // Check if MI has any non-dead defs or physreg uses. This also detects
1803 // predicated instructions which will be reading CPSR.
1804 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
1805 const MachineOperand &MO = MI->getOperand(i);
1806 // Reject frame index operands, PEI can't handle the predicated pseudos.
1807 if (MO.isFI() || MO.isCPI() || MO.isJTI())
1811 // MI can't have any tied operands, that would conflict with predication.
1814 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1816 if (MO.isDef() && !MO.isDead())
1819 bool DontMoveAcrossStores = true;
1820 if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ nullptr,
1821 DontMoveAcrossStores))
1826 bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
1827 SmallVectorImpl<MachineOperand> &Cond,
1828 unsigned &TrueOp, unsigned &FalseOp,
1829 bool &Optimizable) const {
1830 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1831 "Unknown select instruction");
1836 // 3: Condition code.
1840 Cond.push_back(MI->getOperand(3));
1841 Cond.push_back(MI->getOperand(4));
1842 // We can always fold a def.
1847 MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
1848 bool PreferFalse) const {
1849 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1850 "Unknown select instruction");
1851 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1852 MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
1853 bool Invert = !DefMI;
1855 DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this);
1859 // Find new register class to use.
1860 MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
1861 unsigned DestReg = MI->getOperand(0).getReg();
1862 const TargetRegisterClass *PreviousClass = MRI.getRegClass(FalseReg.getReg());
1863 if (!MRI.constrainRegClass(DestReg, PreviousClass))
1866 // Create a new predicated version of DefMI.
1867 // Rfalse is the first use.
1868 MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1869 DefMI->getDesc(), DestReg);
1871 // Copy all the DefMI operands, excluding its (null) predicate.
1872 const MCInstrDesc &DefDesc = DefMI->getDesc();
1873 for (unsigned i = 1, e = DefDesc.getNumOperands();
1874 i != e && !DefDesc.OpInfo[i].isPredicate(); ++i)
1875 NewMI.addOperand(DefMI->getOperand(i));
1877 unsigned CondCode = MI->getOperand(3).getImm();
1879 NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode)));
1881 NewMI.addImm(CondCode);
1882 NewMI.addOperand(MI->getOperand(4));
1884 // DefMI is not the -S version that sets CPSR, so add an optional %noreg.
1885 if (NewMI->hasOptionalDef())
1886 AddDefaultCC(NewMI);
1888 // The output register value when the predicate is false is an implicit
1889 // register operand tied to the first def.
1890 // The tie makes the register allocator ensure the FalseReg is allocated the
1891 // same register as operand 0.
1892 FalseReg.setImplicit();
1893 NewMI.addOperand(FalseReg);
1894 NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
1896 // The caller will erase MI, but not DefMI.
1897 DefMI->eraseFromParent();
1901 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
1902 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
1905 /// This will go away once we can teach tblgen how to set the optional CPSR def
1907 struct AddSubFlagsOpcodePair {
1909 uint16_t MachineOpc;
1912 static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
1913 {ARM::ADDSri, ARM::ADDri},
1914 {ARM::ADDSrr, ARM::ADDrr},
1915 {ARM::ADDSrsi, ARM::ADDrsi},
1916 {ARM::ADDSrsr, ARM::ADDrsr},
1918 {ARM::SUBSri, ARM::SUBri},
1919 {ARM::SUBSrr, ARM::SUBrr},
1920 {ARM::SUBSrsi, ARM::SUBrsi},
1921 {ARM::SUBSrsr, ARM::SUBrsr},
1923 {ARM::RSBSri, ARM::RSBri},
1924 {ARM::RSBSrsi, ARM::RSBrsi},
1925 {ARM::RSBSrsr, ARM::RSBrsr},
1927 {ARM::t2ADDSri, ARM::t2ADDri},
1928 {ARM::t2ADDSrr, ARM::t2ADDrr},
1929 {ARM::t2ADDSrs, ARM::t2ADDrs},
1931 {ARM::t2SUBSri, ARM::t2SUBri},
1932 {ARM::t2SUBSrr, ARM::t2SUBrr},
1933 {ARM::t2SUBSrs, ARM::t2SUBrs},
1935 {ARM::t2RSBSri, ARM::t2RSBri},
1936 {ARM::t2RSBSrs, ARM::t2RSBrs},
1939 unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
1940 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
1941 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
1942 return AddSubFlagsOpcodeMap[i].MachineOpc;
1946 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1947 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1948 unsigned DestReg, unsigned BaseReg, int NumBytes,
1949 ARMCC::CondCodes Pred, unsigned PredReg,
1950 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1951 if (NumBytes == 0 && DestReg != BaseReg) {
1952 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), DestReg)
1953 .addReg(BaseReg, RegState::Kill)
1954 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1955 .setMIFlags(MIFlags);
1959 bool isSub = NumBytes < 0;
1960 if (isSub) NumBytes = -NumBytes;
1963 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1964 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1965 assert(ThisVal && "Didn't extract field correctly");
1967 // We will handle these bits from offset, clear them.
1968 NumBytes &= ~ThisVal;
1970 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1972 // Build the new ADD / SUB.
1973 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1974 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1975 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1976 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1977 .setMIFlags(MIFlags);
1982 static bool isAnySubRegLive(unsigned Reg, const TargetRegisterInfo *TRI,
1984 for (MCSubRegIterator Subreg(Reg, TRI, /* IncludeSelf */ true);
1985 Subreg.isValid(); ++Subreg)
1986 if (MI->getParent()->computeRegisterLiveness(TRI, *Subreg, MI) !=
1987 MachineBasicBlock::LQR_Dead)
1991 bool llvm::tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget,
1992 MachineFunction &MF, MachineInstr *MI,
1993 unsigned NumBytes) {
1994 // This optimisation potentially adds lots of load and store
1995 // micro-operations, it's only really a great benefit to code-size.
1996 if (!MF.getFunction()->getAttributes().hasAttribute(
1997 AttributeSet::FunctionIndex, Attribute::MinSize))
2000 // If only one register is pushed/popped, LLVM can use an LDR/STR
2001 // instead. We can't modify those so make sure we're dealing with an
2002 // instruction we understand.
2003 bool IsPop = isPopOpcode(MI->getOpcode());
2004 bool IsPush = isPushOpcode(MI->getOpcode());
2005 if (!IsPush && !IsPop)
2008 bool IsVFPPushPop = MI->getOpcode() == ARM::VSTMDDB_UPD ||
2009 MI->getOpcode() == ARM::VLDMDIA_UPD;
2010 bool IsT1PushPop = MI->getOpcode() == ARM::tPUSH ||
2011 MI->getOpcode() == ARM::tPOP ||
2012 MI->getOpcode() == ARM::tPOP_RET;
2014 assert((IsT1PushPop || (MI->getOperand(0).getReg() == ARM::SP &&
2015 MI->getOperand(1).getReg() == ARM::SP)) &&
2016 "trying to fold sp update into non-sp-updating push/pop");
2018 // The VFP push & pop act on D-registers, so we can only fold an adjustment
2019 // by a multiple of 8 bytes in correctly. Similarly rN is 4-bytes. Don't try
2020 // if this is violated.
2021 if (NumBytes % (IsVFPPushPop ? 8 : 4) != 0)
2024 // ARM and Thumb2 push/pop insts have explicit "sp, sp" operands (+
2025 // pred) so the list starts at 4. Thumb1 starts after the predicate.
2026 int RegListIdx = IsT1PushPop ? 2 : 4;
2028 // Calculate the space we'll need in terms of registers.
2029 unsigned FirstReg = MI->getOperand(RegListIdx).getReg();
2030 unsigned RD0Reg, RegsNeeded;
2033 RegsNeeded = NumBytes / 8;
2036 RegsNeeded = NumBytes / 4;
2039 // We're going to have to strip all list operands off before
2040 // re-adding them since the order matters, so save the existing ones
2042 SmallVector<MachineOperand, 4> RegList;
2043 for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i)
2044 RegList.push_back(MI->getOperand(i));
2046 const TargetRegisterInfo *TRI = MF.getRegInfo().getTargetRegisterInfo();
2047 const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(&MF);
2049 // Now try to find enough space in the reglist to allocate NumBytes.
2050 for (unsigned CurReg = FirstReg - 1; CurReg >= RD0Reg && RegsNeeded;
2053 // Pushing any register is completely harmless, mark the
2054 // register involved as undef since we don't care about it in
2056 RegList.push_back(MachineOperand::CreateReg(CurReg, false, false,
2057 false, false, true));
2062 // However, we can only pop an extra register if it's not live. For
2063 // registers live within the function we might clobber a return value
2064 // register; the other way a register can be live here is if it's
2066 // TODO: Currently, computeRegisterLiveness() does not report "live" if a
2067 // sub reg is live. When computeRegisterLiveness() works for sub reg, it
2068 // can replace isAnySubRegLive().
2069 if (isCalleeSavedRegister(CurReg, CSRegs) ||
2070 isAnySubRegLive(CurReg, TRI, MI)) {
2071 // VFP pops don't allow holes in the register list, so any skip is fatal
2072 // for our transformation. GPR pops do, so we should just keep looking.
2079 // Mark the unimportant registers as <def,dead> in the POP.
2080 RegList.push_back(MachineOperand::CreateReg(CurReg, true, false, false,
2088 // Finally we know we can profitably perform the optimisation so go
2089 // ahead: strip all existing registers off and add them back again
2090 // in the right order.
2091 for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i)
2092 MI->RemoveOperand(i);
2094 // Add the complete list back in.
2095 MachineInstrBuilder MIB(MF, &*MI);
2096 for (int i = RegList.size() - 1; i >= 0; --i)
2097 MIB.addOperand(RegList[i]);
2102 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
2103 unsigned FrameReg, int &Offset,
2104 const ARMBaseInstrInfo &TII) {
2105 unsigned Opcode = MI.getOpcode();
2106 const MCInstrDesc &Desc = MI.getDesc();
2107 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
2110 // Memory operands in inline assembly always use AddrMode2.
2111 if (Opcode == ARM::INLINEASM)
2112 AddrMode = ARMII::AddrMode2;
2114 if (Opcode == ARM::ADDri) {
2115 Offset += MI.getOperand(FrameRegIdx+1).getImm();
2117 // Turn it into a move.
2118 MI.setDesc(TII.get(ARM::MOVr));
2119 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
2120 MI.RemoveOperand(FrameRegIdx+1);
2123 } else if (Offset < 0) {
2126 MI.setDesc(TII.get(ARM::SUBri));
2129 // Common case: small offset, fits into instruction.
2130 if (ARM_AM::getSOImmVal(Offset) != -1) {
2131 // Replace the FrameIndex with sp / fp
2132 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
2133 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
2138 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
2140 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
2141 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
2143 // We will handle these bits from offset, clear them.
2144 Offset &= ~ThisImmVal;
2146 // Get the properly encoded SOImmVal field.
2147 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
2148 "Bit extraction didn't work?");
2149 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
2151 unsigned ImmIdx = 0;
2153 unsigned NumBits = 0;
2156 case ARMII::AddrMode_i12: {
2157 ImmIdx = FrameRegIdx + 1;
2158 InstrOffs = MI.getOperand(ImmIdx).getImm();
2162 case ARMII::AddrMode2: {
2163 ImmIdx = FrameRegIdx+2;
2164 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
2165 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
2170 case ARMII::AddrMode3: {
2171 ImmIdx = FrameRegIdx+2;
2172 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
2173 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
2178 case ARMII::AddrMode4:
2179 case ARMII::AddrMode6:
2180 // Can't fold any offset even if it's zero.
2182 case ARMII::AddrMode5: {
2183 ImmIdx = FrameRegIdx+1;
2184 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
2185 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
2192 llvm_unreachable("Unsupported addressing mode!");
2195 Offset += InstrOffs * Scale;
2196 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
2202 // Attempt to fold address comp. if opcode has offset bits
2204 // Common case: small offset, fits into instruction.
2205 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
2206 int ImmedOffset = Offset / Scale;
2207 unsigned Mask = (1 << NumBits) - 1;
2208 if ((unsigned)Offset <= Mask * Scale) {
2209 // Replace the FrameIndex with sp
2210 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
2211 // FIXME: When addrmode2 goes away, this will simplify (like the
2212 // T2 version), as the LDR.i12 versions don't need the encoding
2213 // tricks for the offset value.
2215 if (AddrMode == ARMII::AddrMode_i12)
2216 ImmedOffset = -ImmedOffset;
2218 ImmedOffset |= 1 << NumBits;
2220 ImmOp.ChangeToImmediate(ImmedOffset);
2225 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
2226 ImmedOffset = ImmedOffset & Mask;
2228 if (AddrMode == ARMII::AddrMode_i12)
2229 ImmedOffset = -ImmedOffset;
2231 ImmedOffset |= 1 << NumBits;
2233 ImmOp.ChangeToImmediate(ImmedOffset);
2234 Offset &= ~(Mask*Scale);
2238 Offset = (isSub) ? -Offset : Offset;
2242 /// analyzeCompare - For a comparison instruction, return the source registers
2243 /// in SrcReg and SrcReg2 if having two register operands, and the value it
2244 /// compares against in CmpValue. Return true if the comparison instruction
2245 /// can be analyzed.
2246 bool ARMBaseInstrInfo::
2247 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
2248 int &CmpMask, int &CmpValue) const {
2249 switch (MI->getOpcode()) {
2253 SrcReg = MI->getOperand(0).getReg();
2256 CmpValue = MI->getOperand(1).getImm();
2260 SrcReg = MI->getOperand(0).getReg();
2261 SrcReg2 = MI->getOperand(1).getReg();
2267 SrcReg = MI->getOperand(0).getReg();
2269 CmpMask = MI->getOperand(1).getImm();
2277 /// isSuitableForMask - Identify a suitable 'and' instruction that
2278 /// operates on the given source register and applies the same mask
2279 /// as a 'tst' instruction. Provide a limited look-through for copies.
2280 /// When successful, MI will hold the found instruction.
2281 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
2282 int CmpMask, bool CommonUse) {
2283 switch (MI->getOpcode()) {
2286 if (CmpMask != MI->getOperand(2).getImm())
2288 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
2292 // Walk down one instruction which is potentially an 'and'.
2293 const MachineInstr &Copy = *MI;
2294 MachineBasicBlock::iterator AND(
2295 std::next(MachineBasicBlock::iterator(MI)));
2296 if (AND == MI->getParent()->end()) return false;
2298 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
2306 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2307 /// the condition code if we modify the instructions such that flags are
2309 inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) {
2311 default: return ARMCC::AL;
2312 case ARMCC::EQ: return ARMCC::EQ;
2313 case ARMCC::NE: return ARMCC::NE;
2314 case ARMCC::HS: return ARMCC::LS;
2315 case ARMCC::LO: return ARMCC::HI;
2316 case ARMCC::HI: return ARMCC::LO;
2317 case ARMCC::LS: return ARMCC::HS;
2318 case ARMCC::GE: return ARMCC::LE;
2319 case ARMCC::LT: return ARMCC::GT;
2320 case ARMCC::GT: return ARMCC::LT;
2321 case ARMCC::LE: return ARMCC::GE;
2325 /// isRedundantFlagInstr - check whether the first instruction, whose only
2326 /// purpose is to update flags, can be made redundant.
2327 /// CMPrr can be made redundant by SUBrr if the operands are the same.
2328 /// CMPri can be made redundant by SUBri if the operands are the same.
2329 /// This function can be extended later on.
2330 inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg,
2331 unsigned SrcReg2, int ImmValue,
2333 if ((CmpI->getOpcode() == ARM::CMPrr ||
2334 CmpI->getOpcode() == ARM::t2CMPrr) &&
2335 (OI->getOpcode() == ARM::SUBrr ||
2336 OI->getOpcode() == ARM::t2SUBrr) &&
2337 ((OI->getOperand(1).getReg() == SrcReg &&
2338 OI->getOperand(2).getReg() == SrcReg2) ||
2339 (OI->getOperand(1).getReg() == SrcReg2 &&
2340 OI->getOperand(2).getReg() == SrcReg)))
2343 if ((CmpI->getOpcode() == ARM::CMPri ||
2344 CmpI->getOpcode() == ARM::t2CMPri) &&
2345 (OI->getOpcode() == ARM::SUBri ||
2346 OI->getOpcode() == ARM::t2SUBri) &&
2347 OI->getOperand(1).getReg() == SrcReg &&
2348 OI->getOperand(2).getImm() == ImmValue)
2353 /// optimizeCompareInstr - Convert the instruction supplying the argument to the
2354 /// comparison into one that sets the zero bit in the flags register;
2355 /// Remove a redundant Compare instruction if an earlier instruction can set the
2356 /// flags in the same way as Compare.
2357 /// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two
2358 /// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the
2359 /// condition code of instructions which use the flags.
2360 bool ARMBaseInstrInfo::
2361 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
2362 int CmpMask, int CmpValue,
2363 const MachineRegisterInfo *MRI) const {
2364 // Get the unique definition of SrcReg.
2365 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
2366 if (!MI) return false;
2368 // Masked compares sometimes use the same register as the corresponding 'and'.
2369 if (CmpMask != ~0) {
2370 if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) {
2372 for (MachineRegisterInfo::use_instr_iterator
2373 UI = MRI->use_instr_begin(SrcReg), UE = MRI->use_instr_end();
2375 if (UI->getParent() != CmpInstr->getParent()) continue;
2376 MachineInstr *PotentialAND = &*UI;
2377 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) ||
2378 isPredicated(PotentialAND))
2383 if (!MI) return false;
2387 // Get ready to iterate backward from CmpInstr.
2388 MachineBasicBlock::iterator I = CmpInstr, E = MI,
2389 B = CmpInstr->getParent()->begin();
2391 // Early exit if CmpInstr is at the beginning of the BB.
2392 if (I == B) return false;
2394 // There are two possible candidates which can be changed to set CPSR:
2395 // One is MI, the other is a SUB instruction.
2396 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
2397 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
2398 MachineInstr *Sub = nullptr;
2400 // MI is not a candidate for CMPrr.
2402 else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
2403 // Conservatively refuse to convert an instruction which isn't in the same
2404 // BB as the comparison.
2405 // For CMPri, we need to check Sub, thus we can't return here.
2406 if (CmpInstr->getOpcode() == ARM::CMPri ||
2407 CmpInstr->getOpcode() == ARM::t2CMPri)
2413 // Check that CPSR isn't set between the comparison instruction and the one we
2414 // want to change. At the same time, search for Sub.
2415 const TargetRegisterInfo *TRI = &getRegisterInfo();
2417 for (; I != E; --I) {
2418 const MachineInstr &Instr = *I;
2420 if (Instr.modifiesRegister(ARM::CPSR, TRI) ||
2421 Instr.readsRegister(ARM::CPSR, TRI))
2422 // This instruction modifies or uses CPSR after the one we want to
2423 // change. We can't do this transformation.
2426 // Check whether CmpInstr can be made redundant by the current instruction.
2427 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
2433 // The 'and' is below the comparison instruction.
2437 // Return false if no candidates exist.
2441 // The single candidate is called MI.
2444 // We can't use a predicated instruction - it doesn't always write the flags.
2445 if (isPredicated(MI))
2448 switch (MI->getOpcode()) {
2482 case ARM::t2EORri: {
2483 // Scan forward for the use of CPSR
2484 // When checking against MI: if it's a conditional code requires
2485 // checking of V bit, then this is not safe to do.
2486 // It is safe to remove CmpInstr if CPSR is redefined or killed.
2487 // If we are done with the basic block, we need to check whether CPSR is
2489 SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
2491 bool isSafe = false;
2493 E = CmpInstr->getParent()->end();
2494 while (!isSafe && ++I != E) {
2495 const MachineInstr &Instr = *I;
2496 for (unsigned IO = 0, EO = Instr.getNumOperands();
2497 !isSafe && IO != EO; ++IO) {
2498 const MachineOperand &MO = Instr.getOperand(IO);
2499 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
2503 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
2509 // Condition code is after the operand before CPSR except for VSELs.
2510 ARMCC::CondCodes CC;
2511 bool IsInstrVSel = true;
2512 switch (Instr.getOpcode()) {
2514 IsInstrVSel = false;
2515 CC = (ARMCC::CondCodes)Instr.getOperand(IO - 1).getImm();
2536 ARMCC::CondCodes NewCC = getSwappedCondition(CC);
2537 if (NewCC == ARMCC::AL)
2539 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
2540 // on CMP needs to be updated to be based on SUB.
2541 // Push the condition code operands to OperandsToUpdate.
2542 // If it is safe to remove CmpInstr, the condition code of these
2543 // operands will be modified.
2544 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
2545 Sub->getOperand(2).getReg() == SrcReg) {
2546 // VSel doesn't support condition code update.
2549 OperandsToUpdate.push_back(
2550 std::make_pair(&((*I).getOperand(IO - 1)), NewCC));
2555 // CPSR can be used multiple times, we should continue.
2568 // If CPSR is not killed nor re-defined, we should check whether it is
2569 // live-out. If it is live-out, do not optimize.
2571 MachineBasicBlock *MBB = CmpInstr->getParent();
2572 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
2573 SE = MBB->succ_end(); SI != SE; ++SI)
2574 if ((*SI)->isLiveIn(ARM::CPSR))
2578 // Toggle the optional operand to CPSR.
2579 MI->getOperand(5).setReg(ARM::CPSR);
2580 MI->getOperand(5).setIsDef(true);
2581 assert(!isPredicated(MI) && "Can't use flags from predicated instruction");
2582 CmpInstr->eraseFromParent();
2584 // Modify the condition code of operands in OperandsToUpdate.
2585 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2586 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2587 for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++)
2588 OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second);
2596 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
2597 MachineInstr *DefMI, unsigned Reg,
2598 MachineRegisterInfo *MRI) const {
2599 // Fold large immediates into add, sub, or, xor.
2600 unsigned DefOpc = DefMI->getOpcode();
2601 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
2603 if (!DefMI->getOperand(1).isImm())
2604 // Could be t2MOVi32imm <ga:xx>
2607 if (!MRI->hasOneNonDBGUse(Reg))
2610 const MCInstrDesc &DefMCID = DefMI->getDesc();
2611 if (DefMCID.hasOptionalDef()) {
2612 unsigned NumOps = DefMCID.getNumOperands();
2613 const MachineOperand &MO = DefMI->getOperand(NumOps-1);
2614 if (MO.getReg() == ARM::CPSR && !MO.isDead())
2615 // If DefMI defines CPSR and it is not dead, it's obviously not safe
2620 const MCInstrDesc &UseMCID = UseMI->getDesc();
2621 if (UseMCID.hasOptionalDef()) {
2622 unsigned NumOps = UseMCID.getNumOperands();
2623 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
2624 // If the instruction sets the flag, do not attempt this optimization
2625 // since it may change the semantics of the code.
2629 unsigned UseOpc = UseMI->getOpcode();
2630 unsigned NewUseOpc = 0;
2631 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
2632 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
2633 bool Commute = false;
2635 default: return false;
2643 case ARM::t2EORrr: {
2644 Commute = UseMI->getOperand(2).getReg() != Reg;
2651 NewUseOpc = ARM::SUBri;
2657 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
2659 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
2660 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
2663 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
2664 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
2665 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
2669 case ARM::t2SUBrr: {
2673 NewUseOpc = ARM::t2SUBri;
2678 case ARM::t2EORrr: {
2679 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
2681 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
2682 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
2685 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
2686 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
2687 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
2695 unsigned OpIdx = Commute ? 2 : 1;
2696 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
2697 bool isKill = UseMI->getOperand(OpIdx).isKill();
2698 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
2699 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
2700 UseMI, UseMI->getDebugLoc(),
2701 get(NewUseOpc), NewReg)
2702 .addReg(Reg1, getKillRegState(isKill))
2703 .addImm(SOImmValV1)));
2704 UseMI->setDesc(get(NewUseOpc));
2705 UseMI->getOperand(1).setReg(NewReg);
2706 UseMI->getOperand(1).setIsKill();
2707 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
2708 DefMI->eraseFromParent();
2712 static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData,
2713 const MachineInstr *MI) {
2714 switch (MI->getOpcode()) {
2716 const MCInstrDesc &Desc = MI->getDesc();
2717 int UOps = ItinData->getNumMicroOps(Desc.getSchedClass());
2718 assert(UOps >= 0 && "bad # UOps");
2726 unsigned ShOpVal = MI->getOperand(3).getImm();
2727 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2728 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2731 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2732 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2739 if (!MI->getOperand(2).getReg())
2742 unsigned ShOpVal = MI->getOperand(3).getImm();
2743 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2744 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2747 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2748 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2755 return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2;
2757 case ARM::LDRSB_POST:
2758 case ARM::LDRSH_POST: {
2759 unsigned Rt = MI->getOperand(0).getReg();
2760 unsigned Rm = MI->getOperand(3).getReg();
2761 return (Rt == Rm) ? 4 : 3;
2764 case ARM::LDR_PRE_REG:
2765 case ARM::LDRB_PRE_REG: {
2766 unsigned Rt = MI->getOperand(0).getReg();
2767 unsigned Rm = MI->getOperand(3).getReg();
2770 unsigned ShOpVal = MI->getOperand(4).getImm();
2771 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2772 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2775 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2776 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2781 case ARM::STR_PRE_REG:
2782 case ARM::STRB_PRE_REG: {
2783 unsigned ShOpVal = MI->getOperand(4).getImm();
2784 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2785 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2788 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2789 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2795 case ARM::STRH_PRE: {
2796 unsigned Rt = MI->getOperand(0).getReg();
2797 unsigned Rm = MI->getOperand(3).getReg();
2802 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub)
2806 case ARM::LDR_POST_REG:
2807 case ARM::LDRB_POST_REG:
2808 case ARM::LDRH_POST: {
2809 unsigned Rt = MI->getOperand(0).getReg();
2810 unsigned Rm = MI->getOperand(3).getReg();
2811 return (Rt == Rm) ? 3 : 2;
2814 case ARM::LDR_PRE_IMM:
2815 case ARM::LDRB_PRE_IMM:
2816 case ARM::LDR_POST_IMM:
2817 case ARM::LDRB_POST_IMM:
2818 case ARM::STRB_POST_IMM:
2819 case ARM::STRB_POST_REG:
2820 case ARM::STRB_PRE_IMM:
2821 case ARM::STRH_POST:
2822 case ARM::STR_POST_IMM:
2823 case ARM::STR_POST_REG:
2824 case ARM::STR_PRE_IMM:
2827 case ARM::LDRSB_PRE:
2828 case ARM::LDRSH_PRE: {
2829 unsigned Rm = MI->getOperand(3).getReg();
2832 unsigned Rt = MI->getOperand(0).getReg();
2835 unsigned ShOpVal = MI->getOperand(4).getImm();
2836 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2837 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2840 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2841 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2847 unsigned Rt = MI->getOperand(0).getReg();
2848 unsigned Rn = MI->getOperand(2).getReg();
2849 unsigned Rm = MI->getOperand(3).getReg();
2851 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2852 return (Rt == Rn) ? 3 : 2;
2856 unsigned Rm = MI->getOperand(3).getReg();
2858 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2862 case ARM::LDRD_POST:
2863 case ARM::t2LDRD_POST:
2866 case ARM::STRD_POST:
2867 case ARM::t2STRD_POST:
2870 case ARM::LDRD_PRE: {
2871 unsigned Rt = MI->getOperand(0).getReg();
2872 unsigned Rn = MI->getOperand(3).getReg();
2873 unsigned Rm = MI->getOperand(4).getReg();
2875 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2876 return (Rt == Rn) ? 4 : 3;
2879 case ARM::t2LDRD_PRE: {
2880 unsigned Rt = MI->getOperand(0).getReg();
2881 unsigned Rn = MI->getOperand(3).getReg();
2882 return (Rt == Rn) ? 4 : 3;
2885 case ARM::STRD_PRE: {
2886 unsigned Rm = MI->getOperand(4).getReg();
2888 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2892 case ARM::t2STRD_PRE:
2895 case ARM::t2LDR_POST:
2896 case ARM::t2LDRB_POST:
2897 case ARM::t2LDRB_PRE:
2898 case ARM::t2LDRSBi12:
2899 case ARM::t2LDRSBi8:
2900 case ARM::t2LDRSBpci:
2902 case ARM::t2LDRH_POST:
2903 case ARM::t2LDRH_PRE:
2905 case ARM::t2LDRSB_POST:
2906 case ARM::t2LDRSB_PRE:
2907 case ARM::t2LDRSH_POST:
2908 case ARM::t2LDRSH_PRE:
2909 case ARM::t2LDRSHi12:
2910 case ARM::t2LDRSHi8:
2911 case ARM::t2LDRSHpci:
2915 case ARM::t2LDRDi8: {
2916 unsigned Rt = MI->getOperand(0).getReg();
2917 unsigned Rn = MI->getOperand(2).getReg();
2918 return (Rt == Rn) ? 3 : 2;
2921 case ARM::t2STRB_POST:
2922 case ARM::t2STRB_PRE:
2925 case ARM::t2STRH_POST:
2926 case ARM::t2STRH_PRE:
2928 case ARM::t2STR_POST:
2929 case ARM::t2STR_PRE:
2935 // Return the number of 32-bit words loaded by LDM or stored by STM. If this
2936 // can't be easily determined return 0 (missing MachineMemOperand).
2938 // FIXME: The current MachineInstr design does not support relying on machine
2939 // mem operands to determine the width of a memory access. Instead, we expect
2940 // the target to provide this information based on the instruction opcode and
2941 // operands. However, using MachineMemOperand is the best solution now for
2944 // 1) getNumMicroOps tries to infer LDM memory width from the total number of MI
2945 // operands. This is much more dangerous than using the MachineMemOperand
2946 // sizes because CodeGen passes can insert/remove optional machine operands. In
2947 // fact, it's totally incorrect for preRA passes and appears to be wrong for
2948 // postRA passes as well.
2950 // 2) getNumLDMAddresses is only used by the scheduling machine model and any
2951 // machine model that calls this should handle the unknown (zero size) case.
2953 // Long term, we should require a target hook that verifies MachineMemOperand
2954 // sizes during MC lowering. That target hook should be local to MC lowering
2955 // because we can't ensure that it is aware of other MI forms. Doing this will
2956 // ensure that MachineMemOperands are correctly propagated through all passes.
2957 unsigned ARMBaseInstrInfo::getNumLDMAddresses(const MachineInstr *MI) const {
2959 for (MachineInstr::mmo_iterator I = MI->memoperands_begin(),
2960 E = MI->memoperands_end(); I != E; ++I) {
2961 Size += (*I)->getSize();
2967 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
2968 const MachineInstr *MI) const {
2969 if (!ItinData || ItinData->isEmpty())
2972 const MCInstrDesc &Desc = MI->getDesc();
2973 unsigned Class = Desc.getSchedClass();
2974 int ItinUOps = ItinData->getNumMicroOps(Class);
2975 if (ItinUOps >= 0) {
2976 if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore()))
2977 return getNumMicroOpsSwiftLdSt(ItinData, MI);
2982 unsigned Opc = MI->getOpcode();
2985 llvm_unreachable("Unexpected multi-uops instruction!");
2990 // The number of uOps for load / store multiple are determined by the number
2993 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
2994 // same cycle. The scheduling for the first load / store must be done
2995 // separately by assuming the address is not 64-bit aligned.
2997 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
2998 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
2999 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
3001 case ARM::VLDMDIA_UPD:
3002 case ARM::VLDMDDB_UPD:
3004 case ARM::VLDMSIA_UPD:
3005 case ARM::VLDMSDB_UPD:
3007 case ARM::VSTMDIA_UPD:
3008 case ARM::VSTMDDB_UPD:
3010 case ARM::VSTMSIA_UPD:
3011 case ARM::VSTMSDB_UPD: {
3012 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
3013 return (NumRegs / 2) + (NumRegs % 2) + 1;
3016 case ARM::LDMIA_RET:
3021 case ARM::LDMIA_UPD:
3022 case ARM::LDMDA_UPD:
3023 case ARM::LDMDB_UPD:
3024 case ARM::LDMIB_UPD:
3029 case ARM::STMIA_UPD:
3030 case ARM::STMDA_UPD:
3031 case ARM::STMDB_UPD:
3032 case ARM::STMIB_UPD:
3034 case ARM::tLDMIA_UPD:
3035 case ARM::tSTMIA_UPD:
3039 case ARM::t2LDMIA_RET:
3042 case ARM::t2LDMIA_UPD:
3043 case ARM::t2LDMDB_UPD:
3046 case ARM::t2STMIA_UPD:
3047 case ARM::t2STMDB_UPD: {
3048 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
3049 if (Subtarget.isSwift()) {
3050 int UOps = 1 + NumRegs; // One for address computation, one for each ld / st.
3053 case ARM::VLDMDIA_UPD:
3054 case ARM::VLDMDDB_UPD:
3055 case ARM::VLDMSIA_UPD:
3056 case ARM::VLDMSDB_UPD:
3057 case ARM::VSTMDIA_UPD:
3058 case ARM::VSTMDDB_UPD:
3059 case ARM::VSTMSIA_UPD:
3060 case ARM::VSTMSDB_UPD:
3061 case ARM::LDMIA_UPD:
3062 case ARM::LDMDA_UPD:
3063 case ARM::LDMDB_UPD:
3064 case ARM::LDMIB_UPD:
3065 case ARM::STMIA_UPD:
3066 case ARM::STMDA_UPD:
3067 case ARM::STMDB_UPD:
3068 case ARM::STMIB_UPD:
3069 case ARM::tLDMIA_UPD:
3070 case ARM::tSTMIA_UPD:
3071 case ARM::t2LDMIA_UPD:
3072 case ARM::t2LDMDB_UPD:
3073 case ARM::t2STMIA_UPD:
3074 case ARM::t2STMDB_UPD:
3075 ++UOps; // One for base register writeback.
3077 case ARM::LDMIA_RET:
3079 case ARM::t2LDMIA_RET:
3080 UOps += 2; // One for base reg wb, one for write to pc.
3084 } else if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3087 // 4 registers would be issued: 2, 2.
3088 // 5 registers would be issued: 2, 2, 1.
3089 int A8UOps = (NumRegs / 2);
3093 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3094 int A9UOps = (NumRegs / 2);
3095 // If there are odd number of registers or if it's not 64-bit aligned,
3096 // then it takes an extra AGU (Address Generation Unit) cycle.
3097 if ((NumRegs % 2) ||
3098 !MI->hasOneMemOperand() ||
3099 (*MI->memoperands_begin())->getAlignment() < 8)
3103 // Assume the worst.
3111 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
3112 const MCInstrDesc &DefMCID,
3114 unsigned DefIdx, unsigned DefAlign) const {
3115 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
3117 // Def is the address writeback.
3118 return ItinData->getOperandCycle(DefClass, DefIdx);
3121 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3122 // (regno / 2) + (regno % 2) + 1
3123 DefCycle = RegNo / 2 + 1;
3126 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3128 bool isSLoad = false;
3130 switch (DefMCID.getOpcode()) {
3133 case ARM::VLDMSIA_UPD:
3134 case ARM::VLDMSDB_UPD:
3139 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
3140 // then it takes an extra cycle.
3141 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
3144 // Assume the worst.
3145 DefCycle = RegNo + 2;
3152 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
3153 const MCInstrDesc &DefMCID,
3155 unsigned DefIdx, unsigned DefAlign) const {
3156 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
3158 // Def is the address writeback.
3159 return ItinData->getOperandCycle(DefClass, DefIdx);
3162 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3163 // 4 registers would be issued: 1, 2, 1.
3164 // 5 registers would be issued: 1, 2, 2.
3165 DefCycle = RegNo / 2;
3168 // Result latency is issue cycle + 2: E2.
3170 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3171 DefCycle = (RegNo / 2);
3172 // If there are odd number of registers or if it's not 64-bit aligned,
3173 // then it takes an extra AGU (Address Generation Unit) cycle.
3174 if ((RegNo % 2) || DefAlign < 8)
3176 // Result latency is AGU cycles + 2.
3179 // Assume the worst.
3180 DefCycle = RegNo + 2;
3187 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
3188 const MCInstrDesc &UseMCID,
3190 unsigned UseIdx, unsigned UseAlign) const {
3191 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
3193 return ItinData->getOperandCycle(UseClass, UseIdx);
3196 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3197 // (regno / 2) + (regno % 2) + 1
3198 UseCycle = RegNo / 2 + 1;
3201 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3203 bool isSStore = false;
3205 switch (UseMCID.getOpcode()) {
3208 case ARM::VSTMSIA_UPD:
3209 case ARM::VSTMSDB_UPD:
3214 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
3215 // then it takes an extra cycle.
3216 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
3219 // Assume the worst.
3220 UseCycle = RegNo + 2;
3227 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
3228 const MCInstrDesc &UseMCID,
3230 unsigned UseIdx, unsigned UseAlign) const {
3231 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
3233 return ItinData->getOperandCycle(UseClass, UseIdx);
3236 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3237 UseCycle = RegNo / 2;
3242 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3243 UseCycle = (RegNo / 2);
3244 // If there are odd number of registers or if it's not 64-bit aligned,
3245 // then it takes an extra AGU (Address Generation Unit) cycle.
3246 if ((RegNo % 2) || UseAlign < 8)
3249 // Assume the worst.
3256 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3257 const MCInstrDesc &DefMCID,
3258 unsigned DefIdx, unsigned DefAlign,
3259 const MCInstrDesc &UseMCID,
3260 unsigned UseIdx, unsigned UseAlign) const {
3261 unsigned DefClass = DefMCID.getSchedClass();
3262 unsigned UseClass = UseMCID.getSchedClass();
3264 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
3265 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
3267 // This may be a def / use of a variable_ops instruction, the operand
3268 // latency might be determinable dynamically. Let the target try to
3271 bool LdmBypass = false;
3272 switch (DefMCID.getOpcode()) {
3274 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3278 case ARM::VLDMDIA_UPD:
3279 case ARM::VLDMDDB_UPD:
3281 case ARM::VLDMSIA_UPD:
3282 case ARM::VLDMSDB_UPD:
3283 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
3286 case ARM::LDMIA_RET:
3291 case ARM::LDMIA_UPD:
3292 case ARM::LDMDA_UPD:
3293 case ARM::LDMDB_UPD:
3294 case ARM::LDMIB_UPD:
3296 case ARM::tLDMIA_UPD:
3298 case ARM::t2LDMIA_RET:
3301 case ARM::t2LDMIA_UPD:
3302 case ARM::t2LDMDB_UPD:
3304 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
3309 // We can't seem to determine the result latency of the def, assume it's 2.
3313 switch (UseMCID.getOpcode()) {
3315 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
3319 case ARM::VSTMDIA_UPD:
3320 case ARM::VSTMDDB_UPD:
3322 case ARM::VSTMSIA_UPD:
3323 case ARM::VSTMSDB_UPD:
3324 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3331 case ARM::STMIA_UPD:
3332 case ARM::STMDA_UPD:
3333 case ARM::STMDB_UPD:
3334 case ARM::STMIB_UPD:
3335 case ARM::tSTMIA_UPD:
3340 case ARM::t2STMIA_UPD:
3341 case ARM::t2STMDB_UPD:
3342 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3347 // Assume it's read in the first stage.
3350 UseCycle = DefCycle - UseCycle + 1;
3353 // It's a variable_ops instruction so we can't use DefIdx here. Just use
3354 // first def operand.
3355 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
3358 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
3359 UseClass, UseIdx)) {
3367 static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
3368 const MachineInstr *MI, unsigned Reg,
3369 unsigned &DefIdx, unsigned &Dist) {
3372 MachineBasicBlock::const_iterator I = MI; ++I;
3373 MachineBasicBlock::const_instr_iterator II = std::prev(I.getInstrIterator());
3374 assert(II->isInsideBundle() && "Empty bundle?");
3377 while (II->isInsideBundle()) {
3378 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
3385 assert(Idx != -1 && "Cannot find bundled definition!");
3390 static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
3391 const MachineInstr *MI, unsigned Reg,
3392 unsigned &UseIdx, unsigned &Dist) {
3395 MachineBasicBlock::const_instr_iterator II = MI; ++II;
3396 assert(II->isInsideBundle() && "Empty bundle?");
3397 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3399 // FIXME: This doesn't properly handle multiple uses.
3401 while (II != E && II->isInsideBundle()) {
3402 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
3405 if (II->getOpcode() != ARM::t2IT)
3419 /// Return the number of cycles to add to (or subtract from) the static
3420 /// itinerary based on the def opcode and alignment. The caller will ensure that
3421 /// adjusted latency is at least one cycle.
3422 static int adjustDefLatency(const ARMSubtarget &Subtarget,
3423 const MachineInstr *DefMI,
3424 const MCInstrDesc *DefMCID, unsigned DefAlign) {
3426 if (Subtarget.isCortexA8() || Subtarget.isLikeA9() || Subtarget.isCortexA7()) {
3427 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3428 // variants are one cycle cheaper.
3429 switch (DefMCID->getOpcode()) {
3433 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3434 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3436 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3443 case ARM::t2LDRSHs: {
3444 // Thumb2 mode: lsl only.
3445 unsigned ShAmt = DefMI->getOperand(3).getImm();
3446 if (ShAmt == 0 || ShAmt == 2)
3451 } else if (Subtarget.isSwift()) {
3452 // FIXME: Properly handle all of the latency adjustments for address
3454 switch (DefMCID->getOpcode()) {
3458 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3459 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
3460 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3463 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3464 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
3467 ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3474 case ARM::t2LDRSHs: {
3475 // Thumb2 mode: lsl only.
3476 unsigned ShAmt = DefMI->getOperand(3).getImm();
3477 if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3)
3484 if (DefAlign < 8 && Subtarget.isLikeA9()) {
3485 switch (DefMCID->getOpcode()) {
3491 case ARM::VLD1q8wb_fixed:
3492 case ARM::VLD1q16wb_fixed:
3493 case ARM::VLD1q32wb_fixed:
3494 case ARM::VLD1q64wb_fixed:
3495 case ARM::VLD1q8wb_register:
3496 case ARM::VLD1q16wb_register:
3497 case ARM::VLD1q32wb_register:
3498 case ARM::VLD1q64wb_register:
3505 case ARM::VLD2d8wb_fixed:
3506 case ARM::VLD2d16wb_fixed:
3507 case ARM::VLD2d32wb_fixed:
3508 case ARM::VLD2q8wb_fixed:
3509 case ARM::VLD2q16wb_fixed:
3510 case ARM::VLD2q32wb_fixed:
3511 case ARM::VLD2d8wb_register:
3512 case ARM::VLD2d16wb_register:
3513 case ARM::VLD2d32wb_register:
3514 case ARM::VLD2q8wb_register:
3515 case ARM::VLD2q16wb_register:
3516 case ARM::VLD2q32wb_register:
3521 case ARM::VLD3d8_UPD:
3522 case ARM::VLD3d16_UPD:
3523 case ARM::VLD3d32_UPD:
3524 case ARM::VLD1d64Twb_fixed:
3525 case ARM::VLD1d64Twb_register:
3526 case ARM::VLD3q8_UPD:
3527 case ARM::VLD3q16_UPD:
3528 case ARM::VLD3q32_UPD:
3533 case ARM::VLD4d8_UPD:
3534 case ARM::VLD4d16_UPD:
3535 case ARM::VLD4d32_UPD:
3536 case ARM::VLD1d64Qwb_fixed:
3537 case ARM::VLD1d64Qwb_register:
3538 case ARM::VLD4q8_UPD:
3539 case ARM::VLD4q16_UPD:
3540 case ARM::VLD4q32_UPD:
3541 case ARM::VLD1DUPq8:
3542 case ARM::VLD1DUPq16:
3543 case ARM::VLD1DUPq32:
3544 case ARM::VLD1DUPq8wb_fixed:
3545 case ARM::VLD1DUPq16wb_fixed:
3546 case ARM::VLD1DUPq32wb_fixed:
3547 case ARM::VLD1DUPq8wb_register:
3548 case ARM::VLD1DUPq16wb_register:
3549 case ARM::VLD1DUPq32wb_register:
3550 case ARM::VLD2DUPd8:
3551 case ARM::VLD2DUPd16:
3552 case ARM::VLD2DUPd32:
3553 case ARM::VLD2DUPd8wb_fixed:
3554 case ARM::VLD2DUPd16wb_fixed:
3555 case ARM::VLD2DUPd32wb_fixed:
3556 case ARM::VLD2DUPd8wb_register:
3557 case ARM::VLD2DUPd16wb_register:
3558 case ARM::VLD2DUPd32wb_register:
3559 case ARM::VLD4DUPd8:
3560 case ARM::VLD4DUPd16:
3561 case ARM::VLD4DUPd32:
3562 case ARM::VLD4DUPd8_UPD:
3563 case ARM::VLD4DUPd16_UPD:
3564 case ARM::VLD4DUPd32_UPD:
3566 case ARM::VLD1LNd16:
3567 case ARM::VLD1LNd32:
3568 case ARM::VLD1LNd8_UPD:
3569 case ARM::VLD1LNd16_UPD:
3570 case ARM::VLD1LNd32_UPD:
3572 case ARM::VLD2LNd16:
3573 case ARM::VLD2LNd32:
3574 case ARM::VLD2LNq16:
3575 case ARM::VLD2LNq32:
3576 case ARM::VLD2LNd8_UPD:
3577 case ARM::VLD2LNd16_UPD:
3578 case ARM::VLD2LNd32_UPD:
3579 case ARM::VLD2LNq16_UPD:
3580 case ARM::VLD2LNq32_UPD:
3582 case ARM::VLD4LNd16:
3583 case ARM::VLD4LNd32:
3584 case ARM::VLD4LNq16:
3585 case ARM::VLD4LNq32:
3586 case ARM::VLD4LNd8_UPD:
3587 case ARM::VLD4LNd16_UPD:
3588 case ARM::VLD4LNd32_UPD:
3589 case ARM::VLD4LNq16_UPD:
3590 case ARM::VLD4LNq32_UPD:
3591 // If the address is not 64-bit aligned, the latencies of these
3592 // instructions increases by one.
3603 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3604 const MachineInstr *DefMI, unsigned DefIdx,
3605 const MachineInstr *UseMI,
3606 unsigned UseIdx) const {
3607 // No operand latency. The caller may fall back to getInstrLatency.
3608 if (!ItinData || ItinData->isEmpty())
3611 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
3612 unsigned Reg = DefMO.getReg();
3613 const MCInstrDesc *DefMCID = &DefMI->getDesc();
3614 const MCInstrDesc *UseMCID = &UseMI->getDesc();
3616 unsigned DefAdj = 0;
3617 if (DefMI->isBundle()) {
3618 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
3619 DefMCID = &DefMI->getDesc();
3621 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
3622 DefMI->isRegSequence() || DefMI->isImplicitDef()) {
3626 unsigned UseAdj = 0;
3627 if (UseMI->isBundle()) {
3629 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
3630 Reg, NewUseIdx, UseAdj);
3636 UseMCID = &UseMI->getDesc();
3639 if (Reg == ARM::CPSR) {
3640 if (DefMI->getOpcode() == ARM::FMSTAT) {
3641 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
3642 return Subtarget.isLikeA9() ? 1 : 20;
3645 // CPSR set and branch can be paired in the same cycle.
3646 if (UseMI->isBranch())
3649 // Otherwise it takes the instruction latency (generally one).
3650 unsigned Latency = getInstrLatency(ItinData, DefMI);
3652 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
3653 // its uses. Instructions which are otherwise scheduled between them may
3654 // incur a code size penalty (not able to use the CPSR setting 16-bit
3656 if (Latency > 0 && Subtarget.isThumb2()) {
3657 const MachineFunction *MF = DefMI->getParent()->getParent();
3658 if (MF->getFunction()->getAttributes().
3659 hasAttribute(AttributeSet::FunctionIndex,
3660 Attribute::OptimizeForSize))
3666 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
3669 unsigned DefAlign = DefMI->hasOneMemOperand()
3670 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
3671 unsigned UseAlign = UseMI->hasOneMemOperand()
3672 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
3674 // Get the itinerary's latency if possible, and handle variable_ops.
3675 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
3676 *UseMCID, UseIdx, UseAlign);
3677 // Unable to find operand latency. The caller may resort to getInstrLatency.
3681 // Adjust for IT block position.
3682 int Adj = DefAdj + UseAdj;
3684 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3685 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
3686 if (Adj >= 0 || (int)Latency > -Adj) {
3687 return Latency + Adj;
3689 // Return the itinerary latency, which may be zero but not less than zero.
3694 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3695 SDNode *DefNode, unsigned DefIdx,
3696 SDNode *UseNode, unsigned UseIdx) const {
3697 if (!DefNode->isMachineOpcode())
3700 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
3702 if (isZeroCost(DefMCID.Opcode))
3705 if (!ItinData || ItinData->isEmpty())
3706 return DefMCID.mayLoad() ? 3 : 1;
3708 if (!UseNode->isMachineOpcode()) {
3709 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
3710 if (Subtarget.isLikeA9() || Subtarget.isSwift())
3711 return Latency <= 2 ? 1 : Latency - 1;
3713 return Latency <= 3 ? 1 : Latency - 2;
3716 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
3717 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
3718 unsigned DefAlign = !DefMN->memoperands_empty()
3719 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
3720 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
3721 unsigned UseAlign = !UseMN->memoperands_empty()
3722 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
3723 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
3724 UseMCID, UseIdx, UseAlign);
3727 (Subtarget.isCortexA8() || Subtarget.isLikeA9() ||
3728 Subtarget.isCortexA7())) {
3729 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3730 // variants are one cycle cheaper.
3731 switch (DefMCID.getOpcode()) {
3736 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3737 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3739 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3746 case ARM::t2LDRSHs: {
3747 // Thumb2 mode: lsl only.
3749 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3750 if (ShAmt == 0 || ShAmt == 2)
3755 } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) {
3756 // FIXME: Properly handle all of the latency adjustments for address
3758 switch (DefMCID.getOpcode()) {
3763 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3764 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3766 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3767 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3769 else if (ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3776 case ARM::t2LDRSHs: {
3777 // Thumb2 mode: lsl 0-3 only.
3784 if (DefAlign < 8 && Subtarget.isLikeA9())
3785 switch (DefMCID.getOpcode()) {
3791 case ARM::VLD1q8wb_register:
3792 case ARM::VLD1q16wb_register:
3793 case ARM::VLD1q32wb_register:
3794 case ARM::VLD1q64wb_register:
3795 case ARM::VLD1q8wb_fixed:
3796 case ARM::VLD1q16wb_fixed:
3797 case ARM::VLD1q32wb_fixed:
3798 case ARM::VLD1q64wb_fixed:
3802 case ARM::VLD2q8Pseudo:
3803 case ARM::VLD2q16Pseudo:
3804 case ARM::VLD2q32Pseudo:
3805 case ARM::VLD2d8wb_fixed:
3806 case ARM::VLD2d16wb_fixed:
3807 case ARM::VLD2d32wb_fixed:
3808 case ARM::VLD2q8PseudoWB_fixed:
3809 case ARM::VLD2q16PseudoWB_fixed:
3810 case ARM::VLD2q32PseudoWB_fixed:
3811 case ARM::VLD2d8wb_register:
3812 case ARM::VLD2d16wb_register:
3813 case ARM::VLD2d32wb_register:
3814 case ARM::VLD2q8PseudoWB_register:
3815 case ARM::VLD2q16PseudoWB_register:
3816 case ARM::VLD2q32PseudoWB_register:
3817 case ARM::VLD3d8Pseudo:
3818 case ARM::VLD3d16Pseudo:
3819 case ARM::VLD3d32Pseudo:
3820 case ARM::VLD1d64TPseudo:
3821 case ARM::VLD1d64TPseudoWB_fixed:
3822 case ARM::VLD3d8Pseudo_UPD:
3823 case ARM::VLD3d16Pseudo_UPD:
3824 case ARM::VLD3d32Pseudo_UPD:
3825 case ARM::VLD3q8Pseudo_UPD:
3826 case ARM::VLD3q16Pseudo_UPD:
3827 case ARM::VLD3q32Pseudo_UPD:
3828 case ARM::VLD3q8oddPseudo:
3829 case ARM::VLD3q16oddPseudo:
3830 case ARM::VLD3q32oddPseudo:
3831 case ARM::VLD3q8oddPseudo_UPD:
3832 case ARM::VLD3q16oddPseudo_UPD:
3833 case ARM::VLD3q32oddPseudo_UPD:
3834 case ARM::VLD4d8Pseudo:
3835 case ARM::VLD4d16Pseudo:
3836 case ARM::VLD4d32Pseudo:
3837 case ARM::VLD1d64QPseudo:
3838 case ARM::VLD1d64QPseudoWB_fixed:
3839 case ARM::VLD4d8Pseudo_UPD:
3840 case ARM::VLD4d16Pseudo_UPD:
3841 case ARM::VLD4d32Pseudo_UPD:
3842 case ARM::VLD4q8Pseudo_UPD:
3843 case ARM::VLD4q16Pseudo_UPD:
3844 case ARM::VLD4q32Pseudo_UPD:
3845 case ARM::VLD4q8oddPseudo:
3846 case ARM::VLD4q16oddPseudo:
3847 case ARM::VLD4q32oddPseudo:
3848 case ARM::VLD4q8oddPseudo_UPD:
3849 case ARM::VLD4q16oddPseudo_UPD:
3850 case ARM::VLD4q32oddPseudo_UPD:
3851 case ARM::VLD1DUPq8:
3852 case ARM::VLD1DUPq16:
3853 case ARM::VLD1DUPq32:
3854 case ARM::VLD1DUPq8wb_fixed:
3855 case ARM::VLD1DUPq16wb_fixed:
3856 case ARM::VLD1DUPq32wb_fixed:
3857 case ARM::VLD1DUPq8wb_register:
3858 case ARM::VLD1DUPq16wb_register:
3859 case ARM::VLD1DUPq32wb_register:
3860 case ARM::VLD2DUPd8:
3861 case ARM::VLD2DUPd16:
3862 case ARM::VLD2DUPd32:
3863 case ARM::VLD2DUPd8wb_fixed:
3864 case ARM::VLD2DUPd16wb_fixed:
3865 case ARM::VLD2DUPd32wb_fixed:
3866 case ARM::VLD2DUPd8wb_register:
3867 case ARM::VLD2DUPd16wb_register:
3868 case ARM::VLD2DUPd32wb_register:
3869 case ARM::VLD4DUPd8Pseudo:
3870 case ARM::VLD4DUPd16Pseudo:
3871 case ARM::VLD4DUPd32Pseudo:
3872 case ARM::VLD4DUPd8Pseudo_UPD:
3873 case ARM::VLD4DUPd16Pseudo_UPD:
3874 case ARM::VLD4DUPd32Pseudo_UPD:
3875 case ARM::VLD1LNq8Pseudo:
3876 case ARM::VLD1LNq16Pseudo:
3877 case ARM::VLD1LNq32Pseudo:
3878 case ARM::VLD1LNq8Pseudo_UPD:
3879 case ARM::VLD1LNq16Pseudo_UPD:
3880 case ARM::VLD1LNq32Pseudo_UPD:
3881 case ARM::VLD2LNd8Pseudo:
3882 case ARM::VLD2LNd16Pseudo:
3883 case ARM::VLD2LNd32Pseudo:
3884 case ARM::VLD2LNq16Pseudo:
3885 case ARM::VLD2LNq32Pseudo:
3886 case ARM::VLD2LNd8Pseudo_UPD:
3887 case ARM::VLD2LNd16Pseudo_UPD:
3888 case ARM::VLD2LNd32Pseudo_UPD:
3889 case ARM::VLD2LNq16Pseudo_UPD:
3890 case ARM::VLD2LNq32Pseudo_UPD:
3891 case ARM::VLD4LNd8Pseudo:
3892 case ARM::VLD4LNd16Pseudo:
3893 case ARM::VLD4LNd32Pseudo:
3894 case ARM::VLD4LNq16Pseudo:
3895 case ARM::VLD4LNq32Pseudo:
3896 case ARM::VLD4LNd8Pseudo_UPD:
3897 case ARM::VLD4LNd16Pseudo_UPD:
3898 case ARM::VLD4LNd32Pseudo_UPD:
3899 case ARM::VLD4LNq16Pseudo_UPD:
3900 case ARM::VLD4LNq32Pseudo_UPD:
3901 // If the address is not 64-bit aligned, the latencies of these
3902 // instructions increases by one.
3910 unsigned ARMBaseInstrInfo::getPredicationCost(const MachineInstr *MI) const {
3911 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3912 MI->isRegSequence() || MI->isImplicitDef())
3918 const MCInstrDesc &MCID = MI->getDesc();
3920 if (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR)) {
3921 // When predicated, CPSR is an additional source operand for CPSR updating
3922 // instructions, this apparently increases their latencies.
3928 unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3929 const MachineInstr *MI,
3930 unsigned *PredCost) const {
3931 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3932 MI->isRegSequence() || MI->isImplicitDef())
3935 // An instruction scheduler typically runs on unbundled instructions, however
3936 // other passes may query the latency of a bundled instruction.
3937 if (MI->isBundle()) {
3938 unsigned Latency = 0;
3939 MachineBasicBlock::const_instr_iterator I = MI;
3940 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3941 while (++I != E && I->isInsideBundle()) {
3942 if (I->getOpcode() != ARM::t2IT)
3943 Latency += getInstrLatency(ItinData, I, PredCost);
3948 const MCInstrDesc &MCID = MI->getDesc();
3949 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) {
3950 // When predicated, CPSR is an additional source operand for CPSR updating
3951 // instructions, this apparently increases their latencies.
3954 // Be sure to call getStageLatency for an empty itinerary in case it has a
3955 // valid MinLatency property.
3957 return MI->mayLoad() ? 3 : 1;
3959 unsigned Class = MCID.getSchedClass();
3961 // For instructions with variable uops, use uops as latency.
3962 if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0)
3963 return getNumMicroOps(ItinData, MI);
3965 // For the common case, fall back on the itinerary's latency.
3966 unsigned Latency = ItinData->getStageLatency(Class);
3968 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3969 unsigned DefAlign = MI->hasOneMemOperand()
3970 ? (*MI->memoperands_begin())->getAlignment() : 0;
3971 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
3972 if (Adj >= 0 || (int)Latency > -Adj) {
3973 return Latency + Adj;
3978 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3979 SDNode *Node) const {
3980 if (!Node->isMachineOpcode())
3983 if (!ItinData || ItinData->isEmpty())
3986 unsigned Opcode = Node->getMachineOpcode();
3989 return ItinData->getStageLatency(get(Opcode).getSchedClass());
3996 bool ARMBaseInstrInfo::
3997 hasHighOperandLatency(const InstrItineraryData *ItinData,
3998 const MachineRegisterInfo *MRI,
3999 const MachineInstr *DefMI, unsigned DefIdx,
4000 const MachineInstr *UseMI, unsigned UseIdx) const {
4001 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
4002 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
4003 if (Subtarget.isCortexA8() &&
4004 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
4005 // CortexA8 VFP instructions are not pipelined.
4008 // Hoist VFP / NEON instructions with 4 or higher latency.
4009 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
4011 Latency = getInstrLatency(ItinData, DefMI);
4014 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
4015 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
4018 bool ARMBaseInstrInfo::
4019 hasLowDefLatency(const InstrItineraryData *ItinData,
4020 const MachineInstr *DefMI, unsigned DefIdx) const {
4021 if (!ItinData || ItinData->isEmpty())
4024 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
4025 if (DDomain == ARMII::DomainGeneral) {
4026 unsigned DefClass = DefMI->getDesc().getSchedClass();
4027 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
4028 return (DefCycle != -1 && DefCycle <= 2);
4033 bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
4034 StringRef &ErrInfo) const {
4035 if (convertAddSubFlagsOpcode(MI->getOpcode())) {
4036 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
4042 // LoadStackGuard has so far only been implemented for MachO. Different code
4043 // sequence is needed for other targets.
4044 void ARMBaseInstrInfo::expandLoadStackGuardBase(MachineBasicBlock::iterator MI,
4045 unsigned LoadImmOpc,
4047 Reloc::Model RM) const {
4048 MachineBasicBlock &MBB = *MI->getParent();
4049 DebugLoc DL = MI->getDebugLoc();
4050 unsigned Reg = MI->getOperand(0).getReg();
4051 const GlobalValue *GV =
4052 cast<GlobalValue>((*MI->memoperands_begin())->getValue());
4053 MachineInstrBuilder MIB;
4055 BuildMI(MBB, MI, DL, get(LoadImmOpc), Reg)
4056 .addGlobalAddress(GV, 0, ARMII::MO_NONLAZY);
4058 if (Subtarget.GVIsIndirectSymbol(GV, RM)) {
4059 MIB = BuildMI(MBB, MI, DL, get(LoadOpc), Reg);
4060 MIB.addReg(Reg, RegState::Kill).addImm(0);
4061 unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
4062 MachineMemOperand *MMO = MBB.getParent()->
4063 getMachineMemOperand(MachinePointerInfo::getGOT(), Flag, 4, 4);
4064 MIB.addMemOperand(MMO);
4065 AddDefaultPred(MIB);
4068 MIB = BuildMI(MBB, MI, DL, get(LoadOpc), Reg);
4069 MIB.addReg(Reg, RegState::Kill).addImm(0);
4070 MIB.setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
4071 AddDefaultPred(MIB);
4075 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
4076 unsigned &AddSubOpc,
4077 bool &NegAcc, bool &HasLane) const {
4078 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
4079 if (I == MLxEntryMap.end())
4082 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
4083 MulOpc = Entry.MulOpc;
4084 AddSubOpc = Entry.AddSubOpc;
4085 NegAcc = Entry.NegAcc;
4086 HasLane = Entry.HasLane;
4090 //===----------------------------------------------------------------------===//
4091 // Execution domains.
4092 //===----------------------------------------------------------------------===//
4094 // Some instructions go down the NEON pipeline, some go down the VFP pipeline,
4095 // and some can go down both. The vmov instructions go down the VFP pipeline,
4096 // but they can be changed to vorr equivalents that are executed by the NEON
4099 // We use the following execution domain numbering:
4107 // Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
4109 std::pair<uint16_t, uint16_t>
4110 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
4111 // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
4112 // if they are not predicated.
4113 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
4114 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
4116 // CortexA9 is particularly picky about mixing the two and wants these
4118 if (Subtarget.isCortexA9() && !isPredicated(MI) &&
4119 (MI->getOpcode() == ARM::VMOVRS ||
4120 MI->getOpcode() == ARM::VMOVSR ||
4121 MI->getOpcode() == ARM::VMOVS))
4122 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
4124 // No other instructions can be swizzled, so just determine their domain.
4125 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
4127 if (Domain & ARMII::DomainNEON)
4128 return std::make_pair(ExeNEON, 0);
4130 // Certain instructions can go either way on Cortex-A8.
4131 // Treat them as NEON instructions.
4132 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
4133 return std::make_pair(ExeNEON, 0);
4135 if (Domain & ARMII::DomainVFP)
4136 return std::make_pair(ExeVFP, 0);
4138 return std::make_pair(ExeGeneric, 0);
4141 static unsigned getCorrespondingDRegAndLane(const TargetRegisterInfo *TRI,
4142 unsigned SReg, unsigned &Lane) {
4143 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_0, &ARM::DPRRegClass);
4146 if (DReg != ARM::NoRegister)
4150 DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1, &ARM::DPRRegClass);
4152 assert(DReg && "S-register with no D super-register?");
4156 /// getImplicitSPRUseForDPRUse - Given a use of a DPR register and lane,
4157 /// set ImplicitSReg to a register number that must be marked as implicit-use or
4158 /// zero if no register needs to be defined as implicit-use.
4160 /// If the function cannot determine if an SPR should be marked implicit use or
4161 /// not, it returns false.
4163 /// This function handles cases where an instruction is being modified from taking
4164 /// an SPR to a DPR[Lane]. A use of the DPR is being added, which may conflict
4165 /// with an earlier def of an SPR corresponding to DPR[Lane^1] (i.e. the other
4166 /// lane of the DPR).
4168 /// If the other SPR is defined, an implicit-use of it should be added. Else,
4169 /// (including the case where the DPR itself is defined), it should not.
4171 static bool getImplicitSPRUseForDPRUse(const TargetRegisterInfo *TRI,
4173 unsigned DReg, unsigned Lane,
4174 unsigned &ImplicitSReg) {
4175 // If the DPR is defined or used already, the other SPR lane will be chained
4176 // correctly, so there is nothing to be done.
4177 if (MI->definesRegister(DReg, TRI) || MI->readsRegister(DReg, TRI)) {
4182 // Otherwise we need to go searching to see if the SPR is set explicitly.
4183 ImplicitSReg = TRI->getSubReg(DReg,
4184 (Lane & 1) ? ARM::ssub_0 : ARM::ssub_1);
4185 MachineBasicBlock::LivenessQueryResult LQR =
4186 MI->getParent()->computeRegisterLiveness(TRI, ImplicitSReg, MI);
4188 if (LQR == MachineBasicBlock::LQR_Live)
4190 else if (LQR == MachineBasicBlock::LQR_Unknown)
4193 // If the register is known not to be live, there is no need to add an
4200 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
4201 unsigned DstReg, SrcReg, DReg;
4203 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
4204 const TargetRegisterInfo *TRI = &getRegisterInfo();
4205 switch (MI->getOpcode()) {
4207 llvm_unreachable("cannot handle opcode!");
4210 if (Domain != ExeNEON)
4213 // Zap the predicate operands.
4214 assert(!isPredicated(MI) && "Cannot predicate a VORRd");
4216 // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
4217 DstReg = MI->getOperand(0).getReg();
4218 SrcReg = MI->getOperand(1).getReg();
4220 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4221 MI->RemoveOperand(i-1);
4223 // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits)
4224 MI->setDesc(get(ARM::VORRd));
4225 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
4230 if (Domain != ExeNEON)
4232 assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
4234 // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits)
4235 DstReg = MI->getOperand(0).getReg();
4236 SrcReg = MI->getOperand(1).getReg();
4238 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4239 MI->RemoveOperand(i-1);
4241 DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane);
4243 // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps)
4244 // Note that DSrc has been widened and the other lane may be undef, which
4245 // contaminates the entire register.
4246 MI->setDesc(get(ARM::VGETLNi32));
4247 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
4248 .addReg(DReg, RegState::Undef)
4251 // The old source should be an implicit use, otherwise we might think it
4252 // was dead before here.
4253 MIB.addReg(SrcReg, RegState::Implicit);
4256 if (Domain != ExeNEON)
4258 assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
4260 // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits)
4261 DstReg = MI->getOperand(0).getReg();
4262 SrcReg = MI->getOperand(1).getReg();
4264 DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane);
4266 unsigned ImplicitSReg;
4267 if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg))
4270 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4271 MI->RemoveOperand(i-1);
4273 // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps)
4274 // Again DDst may be undefined at the beginning of this instruction.
4275 MI->setDesc(get(ARM::VSETLNi32));
4276 MIB.addReg(DReg, RegState::Define)
4277 .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI)))
4280 AddDefaultPred(MIB);
4282 // The narrower destination must be marked as set to keep previous chains
4284 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4285 if (ImplicitSReg != 0)
4286 MIB.addReg(ImplicitSReg, RegState::Implicit);
4290 if (Domain != ExeNEON)
4293 // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits)
4294 DstReg = MI->getOperand(0).getReg();
4295 SrcReg = MI->getOperand(1).getReg();
4297 unsigned DstLane = 0, SrcLane = 0, DDst, DSrc;
4298 DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane);
4299 DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane);
4301 unsigned ImplicitSReg;
4302 if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg))
4305 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4306 MI->RemoveOperand(i-1);
4309 // Destination can be:
4310 // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits)
4311 MI->setDesc(get(ARM::VDUPLN32d));
4312 MIB.addReg(DDst, RegState::Define)
4313 .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI)))
4315 AddDefaultPred(MIB);
4317 // Neither the source or the destination are naturally represented any
4318 // more, so add them in manually.
4319 MIB.addReg(DstReg, RegState::Implicit | RegState::Define);
4320 MIB.addReg(SrcReg, RegState::Implicit);
4321 if (ImplicitSReg != 0)
4322 MIB.addReg(ImplicitSReg, RegState::Implicit);
4326 // In general there's no single instruction that can perform an S <-> S
4327 // move in NEON space, but a pair of VEXT instructions *can* do the
4328 // job. It turns out that the VEXTs needed will only use DSrc once, with
4329 // the position based purely on the combination of lane-0 and lane-1
4330 // involved. For example
4331 // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1
4332 // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1
4333 // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1
4334 // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1
4336 // Pattern of the MachineInstrs is:
4337 // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits)
4338 MachineInstrBuilder NewMIB;
4339 NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4340 get(ARM::VEXTd32), DDst);
4342 // On the first instruction, both DSrc and DDst may be <undef> if present.
4343 // Specifically when the original instruction didn't have them as an
4345 unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst;
4346 bool CurUndef = !MI->readsRegister(CurReg, TRI);
4347 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
4349 CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst;
4350 CurUndef = !MI->readsRegister(CurReg, TRI);
4351 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
4354 AddDefaultPred(NewMIB);
4356 if (SrcLane == DstLane)
4357 NewMIB.addReg(SrcReg, RegState::Implicit);
4359 MI->setDesc(get(ARM::VEXTd32));
4360 MIB.addReg(DDst, RegState::Define);
4362 // On the second instruction, DDst has definitely been defined above, so
4363 // it is not <undef>. DSrc, if present, can be <undef> as above.
4364 CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst;
4365 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4366 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4368 CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst;
4369 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4370 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4373 AddDefaultPred(MIB);
4375 if (SrcLane != DstLane)
4376 MIB.addReg(SrcReg, RegState::Implicit);
4378 // As before, the original destination is no longer represented, add it
4380 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4381 if (ImplicitSReg != 0)
4382 MIB.addReg(ImplicitSReg, RegState::Implicit);
4389 //===----------------------------------------------------------------------===//
4390 // Partial register updates
4391 //===----------------------------------------------------------------------===//
4393 // Swift renames NEON registers with 64-bit granularity. That means any
4394 // instruction writing an S-reg implicitly reads the containing D-reg. The
4395 // problem is mostly avoided by translating f32 operations to v2f32 operations
4396 // on D-registers, but f32 loads are still a problem.
4398 // These instructions can load an f32 into a NEON register:
4400 // VLDRS - Only writes S, partial D update.
4401 // VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops.
4402 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
4404 // FCONSTD can be used as a dependency-breaking instruction.
4405 unsigned ARMBaseInstrInfo::
4406 getPartialRegUpdateClearance(const MachineInstr *MI,
4408 const TargetRegisterInfo *TRI) const {
4409 if (!SwiftPartialUpdateClearance ||
4410 !(Subtarget.isSwift() || Subtarget.isCortexA15()))
4413 assert(TRI && "Need TRI instance");
4415 const MachineOperand &MO = MI->getOperand(OpNum);
4418 unsigned Reg = MO.getReg();
4421 switch(MI->getOpcode()) {
4422 // Normal instructions writing only an S-register.
4427 case ARM::VMOVv4i16:
4428 case ARM::VMOVv2i32:
4429 case ARM::VMOVv2f32:
4430 case ARM::VMOVv1i64:
4431 UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI);
4434 // Explicitly reads the dependency.
4435 case ARM::VLD1LNd32:
4442 // If this instruction actually reads a value from Reg, there is no unwanted
4444 if (UseOp != -1 && MI->getOperand(UseOp).readsReg())
4447 // We must be able to clobber the whole D-reg.
4448 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4449 // Virtual register must be a foo:ssub_0<def,undef> operand.
4450 if (!MO.getSubReg() || MI->readsVirtualRegister(Reg))
4452 } else if (ARM::SPRRegClass.contains(Reg)) {
4453 // Physical register: MI must define the full D-reg.
4454 unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0,
4456 if (!DReg || !MI->definesRegister(DReg, TRI))
4460 // MI has an unwanted D-register dependency.
4461 // Avoid defs in the previous N instructrions.
4462 return SwiftPartialUpdateClearance;
4465 // Break a partial register dependency after getPartialRegUpdateClearance
4466 // returned non-zero.
4467 void ARMBaseInstrInfo::
4468 breakPartialRegDependency(MachineBasicBlock::iterator MI,
4470 const TargetRegisterInfo *TRI) const {
4471 assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def");
4472 assert(TRI && "Need TRI instance");
4474 const MachineOperand &MO = MI->getOperand(OpNum);
4475 unsigned Reg = MO.getReg();
4476 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
4477 "Can't break virtual register dependencies.");
4478 unsigned DReg = Reg;
4480 // If MI defines an S-reg, find the corresponding D super-register.
4481 if (ARM::SPRRegClass.contains(Reg)) {
4482 DReg = ARM::D0 + (Reg - ARM::S0) / 2;
4483 assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken");
4486 assert(ARM::DPRRegClass.contains(DReg) && "Can only break D-reg deps");
4487 assert(MI->definesRegister(DReg, TRI) && "MI doesn't clobber full D-reg");
4489 // FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines
4490 // the full D-register by loading the same value to both lanes. The
4491 // instruction is micro-coded with 2 uops, so don't do this until we can
4492 // properly schedule micro-coded instructions. The dispatcher stalls cause
4493 // too big regressions.
4495 // Insert the dependency-breaking FCONSTD before MI.
4496 // 96 is the encoding of 0.5, but the actual value doesn't matter here.
4497 AddDefaultPred(BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4498 get(ARM::FCONSTD), DReg).addImm(96));
4499 MI->addRegisterKilled(DReg, TRI, true);
4502 void ARMBaseInstrInfo::getUnconditionalBranch(
4503 MCInst &Branch, const MCSymbolRefExpr *BranchTarget) const {
4504 if (Subtarget.isThumb())
4505 Branch.setOpcode(ARM::tB);
4506 else if (Subtarget.isThumb2())
4507 Branch.setOpcode(ARM::t2B);
4509 Branch.setOpcode(ARM::Bcc);
4511 Branch.addOperand(MCOperand::CreateExpr(BranchTarget));
4512 Branch.addOperand(MCOperand::CreateImm(ARMCC::AL));
4513 Branch.addOperand(MCOperand::CreateReg(0));
4516 void ARMBaseInstrInfo::getTrap(MCInst &MI) const {
4517 if (Subtarget.isThumb())
4518 MI.setOpcode(ARM::tTRAP);
4519 else if (Subtarget.useNaClTrap())
4520 MI.setOpcode(ARM::TRAPNaCl);
4522 MI.setOpcode(ARM::TRAP);
4525 bool ARMBaseInstrInfo::hasNOP() const {
4526 return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
4529 bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
4530 if (MI->getNumOperands() < 4)
4532 unsigned ShOpVal = MI->getOperand(3).getImm();
4533 unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
4534 // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
4535 if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) ||
4536 ((ShImm == 1 || ShImm == 2) &&
4537 ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl))
4543 bool ARMBaseInstrInfo::getRegSequenceLikeInputs(
4544 const MachineInstr &MI, unsigned DefIdx,
4545 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
4546 assert(DefIdx < MI.getDesc().getNumDefs() && "Invalid definition index");
4547 assert(MI.isRegSequenceLike() && "Invalid kind of instruction");
4549 switch (MI.getOpcode()) {
4551 // dX = VMOVDRR rY, rZ
4553 // dX = REG_SEQUENCE rY, ssub_0, rZ, ssub_1
4554 // Populate the InputRegs accordingly.
4556 const MachineOperand *MOReg = &MI.getOperand(1);
4557 InputRegs.push_back(
4558 RegSubRegPairAndIdx(MOReg->getReg(), MOReg->getSubReg(), ARM::ssub_0));
4560 MOReg = &MI.getOperand(2);
4561 InputRegs.push_back(
4562 RegSubRegPairAndIdx(MOReg->getReg(), MOReg->getSubReg(), ARM::ssub_1));
4565 llvm_unreachable("Target dependent opcode missing");
4568 bool ARMBaseInstrInfo::getExtractSubregLikeInputs(
4569 const MachineInstr &MI, unsigned DefIdx,
4570 RegSubRegPairAndIdx &InputReg) const {
4571 assert(DefIdx < MI.getDesc().getNumDefs() && "Invalid definition index");
4572 assert(MI.isExtractSubregLike() && "Invalid kind of instruction");
4574 switch (MI.getOpcode()) {
4576 // rX, rY = VMOVRRD dZ
4578 // rX = EXTRACT_SUBREG dZ, ssub_0
4579 // rY = EXTRACT_SUBREG dZ, ssub_1
4580 const MachineOperand &MOReg = MI.getOperand(2);
4581 InputReg.Reg = MOReg.getReg();
4582 InputReg.SubReg = MOReg.getSubReg();
4583 InputReg.SubIdx = DefIdx == 0 ? ARM::ssub_0 : ARM::ssub_1;
4586 llvm_unreachable("Target dependent opcode missing");
4589 bool ARMBaseInstrInfo::getInsertSubregLikeInputs(
4590 const MachineInstr &MI, unsigned DefIdx, RegSubRegPair &BaseReg,
4591 RegSubRegPairAndIdx &InsertedReg) const {
4592 assert(DefIdx < MI.getDesc().getNumDefs() && "Invalid definition index");
4593 assert(MI.isInsertSubregLike() && "Invalid kind of instruction");
4595 switch (MI.getOpcode()) {
4596 case ARM::VSETLNi32:
4597 // dX = VSETLNi32 dY, rZ, imm
4598 const MachineOperand &MOBaseReg = MI.getOperand(1);
4599 const MachineOperand &MOInsertedReg = MI.getOperand(2);
4600 const MachineOperand &MOIndex = MI.getOperand(3);
4601 BaseReg.Reg = MOBaseReg.getReg();
4602 BaseReg.SubReg = MOBaseReg.getSubReg();
4604 InsertedReg.Reg = MOInsertedReg.getReg();
4605 InsertedReg.SubReg = MOInsertedReg.getSubReg();
4606 InsertedReg.SubIdx = MOIndex.getImm() == 0 ? ARM::ssub_0 : ARM::ssub_1;
4609 llvm_unreachable("Target dependent opcode missing");