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/Support/BranchProbability.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
40 #define GET_INSTRINFO_CTOR
41 #include "ARMGenInstrInfo.inc"
46 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
47 cl::desc("Enable ARM 2-addr to 3-addr conv"));
50 WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true),
51 cl::desc("Widen ARM vmovs to vmovd when possible"));
53 static cl::opt<unsigned>
54 SwiftPartialUpdateClearance("swift-partial-update-clearance",
55 cl::Hidden, cl::init(12),
56 cl::desc("Clearance before partial register updates"));
58 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
60 uint16_t MLxOpc; // MLA / MLS opcode
61 uint16_t MulOpc; // Expanded multiplication opcode
62 uint16_t AddSubOpc; // Expanded add / sub opcode
63 bool NegAcc; // True if the acc is negated before the add / sub.
64 bool HasLane; // True if instruction has an extra "lane" operand.
67 static const ARM_MLxEntry ARM_MLxTable[] = {
68 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
70 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
71 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
72 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
73 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
74 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
75 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
76 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
77 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
80 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
81 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
82 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
83 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
84 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
85 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
86 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
87 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
90 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
91 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
93 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
94 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
95 assert(false && "Duplicated entries?");
96 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
97 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
101 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
102 // currently defaults to no prepass hazard recognizer.
103 ScheduleHazardRecognizer *ARMBaseInstrInfo::
104 CreateTargetHazardRecognizer(const TargetMachine *TM,
105 const ScheduleDAG *DAG) const {
106 if (usePreRAHazardRecognizer()) {
107 const InstrItineraryData *II = TM->getInstrItineraryData();
108 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
110 return TargetInstrInfo::CreateTargetHazardRecognizer(TM, DAG);
113 ScheduleHazardRecognizer *ARMBaseInstrInfo::
114 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
115 const ScheduleDAG *DAG) const {
116 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
117 return (ScheduleHazardRecognizer *)new ARMHazardRecognizer(II, DAG);
118 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
122 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
123 MachineBasicBlock::iterator &MBBI,
124 LiveVariables *LV) const {
125 // FIXME: Thumb2 support.
130 MachineInstr *MI = MBBI;
131 MachineFunction &MF = *MI->getParent()->getParent();
132 uint64_t TSFlags = MI->getDesc().TSFlags;
134 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
135 default: return NULL;
136 case ARMII::IndexModePre:
139 case ARMII::IndexModePost:
143 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
145 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
149 MachineInstr *UpdateMI = NULL;
150 MachineInstr *MemMI = NULL;
151 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
152 const MCInstrDesc &MCID = MI->getDesc();
153 unsigned NumOps = MCID.getNumOperands();
154 bool isLoad = !MI->mayStore();
155 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
156 const MachineOperand &Base = MI->getOperand(2);
157 const MachineOperand &Offset = MI->getOperand(NumOps-3);
158 unsigned WBReg = WB.getReg();
159 unsigned BaseReg = Base.getReg();
160 unsigned OffReg = Offset.getReg();
161 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
162 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
164 default: llvm_unreachable("Unknown indexed op!");
165 case ARMII::AddrMode2: {
166 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
167 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
169 if (ARM_AM::getSOImmVal(Amt) == -1)
170 // Can't encode it in a so_imm operand. This transformation will
171 // add more than 1 instruction. Abandon!
173 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
174 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
175 .addReg(BaseReg).addImm(Amt)
176 .addImm(Pred).addReg(0).addReg(0);
177 } else if (Amt != 0) {
178 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
179 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
180 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
181 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
182 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
183 .addImm(Pred).addReg(0).addReg(0);
185 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
186 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
187 .addReg(BaseReg).addReg(OffReg)
188 .addImm(Pred).addReg(0).addReg(0);
191 case ARMII::AddrMode3 : {
192 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
193 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
195 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
196 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
197 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
198 .addReg(BaseReg).addImm(Amt)
199 .addImm(Pred).addReg(0).addReg(0);
201 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
202 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
203 .addReg(BaseReg).addReg(OffReg)
204 .addImm(Pred).addReg(0).addReg(0);
209 std::vector<MachineInstr*> NewMIs;
212 MemMI = BuildMI(MF, MI->getDebugLoc(),
213 get(MemOpc), MI->getOperand(0).getReg())
214 .addReg(WBReg).addImm(0).addImm(Pred);
216 MemMI = BuildMI(MF, MI->getDebugLoc(),
217 get(MemOpc)).addReg(MI->getOperand(1).getReg())
218 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
219 NewMIs.push_back(MemMI);
220 NewMIs.push_back(UpdateMI);
223 MemMI = BuildMI(MF, MI->getDebugLoc(),
224 get(MemOpc), MI->getOperand(0).getReg())
225 .addReg(BaseReg).addImm(0).addImm(Pred);
227 MemMI = BuildMI(MF, MI->getDebugLoc(),
228 get(MemOpc)).addReg(MI->getOperand(1).getReg())
229 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
231 UpdateMI->getOperand(0).setIsDead();
232 NewMIs.push_back(UpdateMI);
233 NewMIs.push_back(MemMI);
236 // Transfer LiveVariables states, kill / dead info.
238 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
239 MachineOperand &MO = MI->getOperand(i);
240 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
241 unsigned Reg = MO.getReg();
243 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
245 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
247 LV->addVirtualRegisterDead(Reg, NewMI);
249 if (MO.isUse() && MO.isKill()) {
250 for (unsigned j = 0; j < 2; ++j) {
251 // Look at the two new MI's in reverse order.
252 MachineInstr *NewMI = NewMIs[j];
253 if (!NewMI->readsRegister(Reg))
255 LV->addVirtualRegisterKilled(Reg, NewMI);
256 if (VI.removeKill(MI))
257 VI.Kills.push_back(NewMI);
265 MFI->insert(MBBI, NewMIs[1]);
266 MFI->insert(MBBI, NewMIs[0]);
272 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
273 MachineBasicBlock *&FBB,
274 SmallVectorImpl<MachineOperand> &Cond,
275 bool AllowModify) const {
279 MachineBasicBlock::iterator I = MBB.end();
280 if (I == MBB.begin())
281 return false; // Empty blocks are easy.
284 // Walk backwards from the end of the basic block until the branch is
285 // analyzed or we give up.
286 while (isPredicated(I) || I->isTerminator()) {
288 // Flag to be raised on unanalyzeable instructions. This is useful in cases
289 // where we want to clean up on the end of the basic block before we bail
291 bool CantAnalyze = false;
293 // Skip over DEBUG values and predicated nonterminators.
294 while (I->isDebugValue() || !I->isTerminator()) {
295 if (I == MBB.begin())
300 if (isIndirectBranchOpcode(I->getOpcode()) ||
301 isJumpTableBranchOpcode(I->getOpcode())) {
302 // Indirect branches and jump tables can't be analyzed, but we still want
303 // to clean up any instructions at the tail of the basic block.
305 } else if (isUncondBranchOpcode(I->getOpcode())) {
306 TBB = I->getOperand(0).getMBB();
307 } else if (isCondBranchOpcode(I->getOpcode())) {
308 // Bail out if we encounter multiple conditional branches.
312 assert(!FBB && "FBB should have been null.");
314 TBB = I->getOperand(0).getMBB();
315 Cond.push_back(I->getOperand(1));
316 Cond.push_back(I->getOperand(2));
317 } else if (I->isReturn()) {
318 // Returns can't be analyzed, but we should run cleanup.
319 CantAnalyze = !isPredicated(I);
321 // We encountered other unrecognized terminator. Bail out immediately.
325 // Cleanup code - to be run for unpredicated unconditional branches and
327 if (!isPredicated(I) &&
328 (isUncondBranchOpcode(I->getOpcode()) ||
329 isIndirectBranchOpcode(I->getOpcode()) ||
330 isJumpTableBranchOpcode(I->getOpcode()) ||
332 // Forget any previous condition branch information - it no longer applies.
336 // If we can modify the function, delete everything below this
337 // unconditional branch.
339 MachineBasicBlock::iterator DI = llvm::next(I);
340 while (DI != MBB.end()) {
341 MachineInstr *InstToDelete = DI;
343 InstToDelete->eraseFromParent();
351 if (I == MBB.begin())
357 // We made it past the terminators without bailing out - we must have
358 // analyzed this branch successfully.
363 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
364 MachineBasicBlock::iterator I = MBB.end();
365 if (I == MBB.begin()) return 0;
367 while (I->isDebugValue()) {
368 if (I == MBB.begin())
372 if (!isUncondBranchOpcode(I->getOpcode()) &&
373 !isCondBranchOpcode(I->getOpcode()))
376 // Remove the branch.
377 I->eraseFromParent();
381 if (I == MBB.begin()) return 1;
383 if (!isCondBranchOpcode(I->getOpcode()))
386 // Remove the branch.
387 I->eraseFromParent();
392 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
393 MachineBasicBlock *FBB,
394 const SmallVectorImpl<MachineOperand> &Cond,
396 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
397 int BOpc = !AFI->isThumbFunction()
398 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
399 int BccOpc = !AFI->isThumbFunction()
400 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
401 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
403 // Shouldn't be a fall through.
404 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
405 assert((Cond.size() == 2 || Cond.size() == 0) &&
406 "ARM branch conditions have two components!");
409 if (Cond.empty()) { // Unconditional branch?
411 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
413 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
415 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
416 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
420 // Two-way conditional branch.
421 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
422 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
424 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
426 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
430 bool ARMBaseInstrInfo::
431 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
432 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
433 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
437 bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
438 if (MI->isBundle()) {
439 MachineBasicBlock::const_instr_iterator I = MI;
440 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
441 while (++I != E && I->isInsideBundle()) {
442 int PIdx = I->findFirstPredOperandIdx();
443 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
449 int PIdx = MI->findFirstPredOperandIdx();
450 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
453 bool ARMBaseInstrInfo::
454 PredicateInstruction(MachineInstr *MI,
455 const SmallVectorImpl<MachineOperand> &Pred) const {
456 unsigned Opc = MI->getOpcode();
457 if (isUncondBranchOpcode(Opc)) {
458 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
459 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
460 .addImm(Pred[0].getImm())
461 .addReg(Pred[1].getReg());
465 int PIdx = MI->findFirstPredOperandIdx();
467 MachineOperand &PMO = MI->getOperand(PIdx);
468 PMO.setImm(Pred[0].getImm());
469 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
475 bool ARMBaseInstrInfo::
476 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
477 const SmallVectorImpl<MachineOperand> &Pred2) const {
478 if (Pred1.size() > 2 || Pred2.size() > 2)
481 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
482 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
492 return CC2 == ARMCC::HI;
494 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
496 return CC2 == ARMCC::GT;
498 return CC2 == ARMCC::LT;
502 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
503 std::vector<MachineOperand> &Pred) const {
505 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
506 const MachineOperand &MO = MI->getOperand(i);
507 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
508 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) {
517 /// isPredicable - Return true if the specified instruction can be predicated.
518 /// By default, this returns true for every instruction with a
519 /// PredicateOperand.
520 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
521 if (!MI->isPredicable())
524 ARMFunctionInfo *AFI =
525 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
527 if (AFI->isThumb2Function()) {
528 if (getSubtarget().hasV8Ops())
529 return isV8EligibleForIT(MI);
530 } else { // non-Thumb
531 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON)
538 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
539 LLVM_ATTRIBUTE_NOINLINE
540 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
542 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
544 assert(JTI < JT.size());
545 return JT[JTI].MBBs.size();
548 /// GetInstSize - Return the size of the specified MachineInstr.
550 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
551 const MachineBasicBlock &MBB = *MI->getParent();
552 const MachineFunction *MF = MBB.getParent();
553 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
555 const MCInstrDesc &MCID = MI->getDesc();
557 return MCID.getSize();
559 // If this machine instr is an inline asm, measure it.
560 if (MI->getOpcode() == ARM::INLINEASM)
561 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
564 unsigned Opc = MI->getOpcode();
566 case TargetOpcode::IMPLICIT_DEF:
567 case TargetOpcode::KILL:
568 case TargetOpcode::PROLOG_LABEL:
569 case TargetOpcode::EH_LABEL:
570 case TargetOpcode::DBG_VALUE:
572 case TargetOpcode::BUNDLE:
573 return getInstBundleLength(MI);
574 case ARM::MOVi16_ga_pcrel:
575 case ARM::MOVTi16_ga_pcrel:
576 case ARM::t2MOVi16_ga_pcrel:
577 case ARM::t2MOVTi16_ga_pcrel:
580 case ARM::t2MOVi32imm:
582 case ARM::CONSTPOOL_ENTRY:
583 // If this machine instr is a constant pool entry, its size is recorded as
585 return MI->getOperand(2).getImm();
586 case ARM::Int_eh_sjlj_longjmp:
588 case ARM::tInt_eh_sjlj_longjmp:
590 case ARM::Int_eh_sjlj_setjmp:
591 case ARM::Int_eh_sjlj_setjmp_nofp:
593 case ARM::tInt_eh_sjlj_setjmp:
594 case ARM::t2Int_eh_sjlj_setjmp:
595 case ARM::t2Int_eh_sjlj_setjmp_nofp:
603 case ARM::t2TBH_JT: {
604 // These are jumptable branches, i.e. a branch followed by an inlined
605 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
606 // entry is one byte; TBH two byte each.
607 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
608 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
609 unsigned NumOps = MCID.getNumOperands();
610 MachineOperand JTOP =
611 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
612 unsigned JTI = JTOP.getIndex();
613 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
615 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
616 assert(JTI < JT.size());
617 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
618 // 4 aligned. The assembler / linker may add 2 byte padding just before
619 // the JT entries. The size does not include this padding; the
620 // constant islands pass does separate bookkeeping for it.
621 // FIXME: If we know the size of the function is less than (1 << 16) *2
622 // bytes, we can use 16-bit entries instead. Then there won't be an
624 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
625 unsigned NumEntries = getNumJTEntries(JT, JTI);
626 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
627 // Make sure the instruction that follows TBB is 2-byte aligned.
628 // FIXME: Constant island pass should insert an "ALIGN" instruction
631 return NumEntries * EntrySize + InstSize;
634 // Otherwise, pseudo-instruction sizes are zero.
639 unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
641 MachineBasicBlock::const_instr_iterator I = MI;
642 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
643 while (++I != E && I->isInsideBundle()) {
644 assert(!I->isBundle() && "No nested bundle!");
645 Size += GetInstSizeInBytes(&*I);
650 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
651 MachineBasicBlock::iterator I, DebugLoc DL,
652 unsigned DestReg, unsigned SrcReg,
653 bool KillSrc) const {
654 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
655 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
657 if (GPRDest && GPRSrc) {
658 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
659 .addReg(SrcReg, getKillRegState(KillSrc))));
663 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
664 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
667 if (SPRDest && SPRSrc)
669 else if (GPRDest && SPRSrc)
671 else if (SPRDest && GPRSrc)
673 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
675 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
679 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
680 MIB.addReg(SrcReg, getKillRegState(KillSrc));
681 if (Opc == ARM::VORRq)
682 MIB.addReg(SrcReg, getKillRegState(KillSrc));
687 // Handle register classes that require multiple instructions.
688 unsigned BeginIdx = 0;
689 unsigned SubRegs = 0;
692 // Use VORRq when possible.
693 if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
694 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 2;
695 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
696 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 4;
697 // Fall back to VMOVD.
698 else if (ARM::DPairRegClass.contains(DestReg, SrcReg))
699 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2;
700 else if (ARM::DTripleRegClass.contains(DestReg, SrcReg))
701 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3;
702 else if (ARM::DQuadRegClass.contains(DestReg, SrcReg))
703 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4;
704 else if (ARM::GPRPairRegClass.contains(DestReg, SrcReg))
705 Opc = ARM::MOVr, BeginIdx = ARM::gsub_0, SubRegs = 2;
707 else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg))
708 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2, Spacing = 2;
709 else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg))
710 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3, Spacing = 2;
711 else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg))
712 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4, Spacing = 2;
714 assert(Opc && "Impossible reg-to-reg copy");
716 const TargetRegisterInfo *TRI = &getRegisterInfo();
717 MachineInstrBuilder Mov;
719 // Copy register tuples backward when the first Dest reg overlaps with SrcReg.
720 if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) {
721 BeginIdx = BeginIdx + ((SubRegs-1)*Spacing);
725 SmallSet<unsigned, 4> DstRegs;
727 for (unsigned i = 0; i != SubRegs; ++i) {
728 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i*Spacing);
729 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i*Spacing);
730 assert(Dst && Src && "Bad sub-register");
732 assert(!DstRegs.count(Src) && "destructive vector copy");
735 Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst)
737 // VORR takes two source operands.
738 if (Opc == ARM::VORRq)
740 Mov = AddDefaultPred(Mov);
742 if (Opc == ARM::MOVr)
743 Mov = AddDefaultCC(Mov);
745 // Add implicit super-register defs and kills to the last instruction.
746 Mov->addRegisterDefined(DestReg, TRI);
748 Mov->addRegisterKilled(SrcReg, TRI);
751 const MachineInstrBuilder &
752 ARMBaseInstrInfo::AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
753 unsigned SubIdx, unsigned State,
754 const TargetRegisterInfo *TRI) const {
756 return MIB.addReg(Reg, State);
758 if (TargetRegisterInfo::isPhysicalRegister(Reg))
759 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
760 return MIB.addReg(Reg, State, SubIdx);
763 void ARMBaseInstrInfo::
764 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
765 unsigned SrcReg, bool isKill, int FI,
766 const TargetRegisterClass *RC,
767 const TargetRegisterInfo *TRI) const {
769 if (I != MBB.end()) DL = I->getDebugLoc();
770 MachineFunction &MF = *MBB.getParent();
771 MachineFrameInfo &MFI = *MF.getFrameInfo();
772 unsigned Align = MFI.getObjectAlignment(FI);
774 MachineMemOperand *MMO =
775 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
776 MachineMemOperand::MOStore,
777 MFI.getObjectSize(FI),
780 switch (RC->getSize()) {
782 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
783 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
784 .addReg(SrcReg, getKillRegState(isKill))
785 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
786 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
787 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
788 .addReg(SrcReg, getKillRegState(isKill))
789 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
791 llvm_unreachable("Unknown reg class!");
794 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
795 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
796 .addReg(SrcReg, getKillRegState(isKill))
797 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
798 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
799 if (Subtarget.hasV5TEOps()) {
800 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::STRD));
801 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
802 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
803 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
807 // Fallback to STM instruction, which has existed since the dawn of
809 MachineInstrBuilder MIB =
810 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STMIA))
811 .addFrameIndex(FI).addMemOperand(MMO));
812 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
813 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
816 llvm_unreachable("Unknown reg class!");
819 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
820 // Use aligned spills if the stack can be realigned.
821 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
822 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
823 .addFrameIndex(FI).addImm(16)
824 .addReg(SrcReg, getKillRegState(isKill))
825 .addMemOperand(MMO));
827 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
828 .addReg(SrcReg, getKillRegState(isKill))
830 .addMemOperand(MMO));
833 llvm_unreachable("Unknown reg class!");
836 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
837 // Use aligned spills if the stack can be realigned.
838 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
839 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
840 .addFrameIndex(FI).addImm(16)
841 .addReg(SrcReg, getKillRegState(isKill))
842 .addMemOperand(MMO));
844 MachineInstrBuilder MIB =
845 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
848 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
849 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
850 AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
853 llvm_unreachable("Unknown reg class!");
856 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
857 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
858 // FIXME: It's possible to only store part of the QQ register if the
859 // spilled def has a sub-register index.
860 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
861 .addFrameIndex(FI).addImm(16)
862 .addReg(SrcReg, getKillRegState(isKill))
863 .addMemOperand(MMO));
865 MachineInstrBuilder MIB =
866 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
869 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
870 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
871 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
872 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
875 llvm_unreachable("Unknown reg class!");
878 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
879 MachineInstrBuilder MIB =
880 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
883 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
884 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
885 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
886 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
887 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
888 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
889 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
890 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
892 llvm_unreachable("Unknown reg class!");
895 llvm_unreachable("Unknown reg class!");
900 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
901 int &FrameIndex) const {
902 switch (MI->getOpcode()) {
905 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
906 if (MI->getOperand(1).isFI() &&
907 MI->getOperand(2).isReg() &&
908 MI->getOperand(3).isImm() &&
909 MI->getOperand(2).getReg() == 0 &&
910 MI->getOperand(3).getImm() == 0) {
911 FrameIndex = MI->getOperand(1).getIndex();
912 return MI->getOperand(0).getReg();
920 if (MI->getOperand(1).isFI() &&
921 MI->getOperand(2).isImm() &&
922 MI->getOperand(2).getImm() == 0) {
923 FrameIndex = MI->getOperand(1).getIndex();
924 return MI->getOperand(0).getReg();
928 case ARM::VST1d64TPseudo:
929 case ARM::VST1d64QPseudo:
930 if (MI->getOperand(0).isFI() &&
931 MI->getOperand(2).getSubReg() == 0) {
932 FrameIndex = MI->getOperand(0).getIndex();
933 return MI->getOperand(2).getReg();
937 if (MI->getOperand(1).isFI() &&
938 MI->getOperand(0).getSubReg() == 0) {
939 FrameIndex = MI->getOperand(1).getIndex();
940 return MI->getOperand(0).getReg();
948 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
949 int &FrameIndex) const {
950 const MachineMemOperand *Dummy;
951 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
954 void ARMBaseInstrInfo::
955 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
956 unsigned DestReg, int FI,
957 const TargetRegisterClass *RC,
958 const TargetRegisterInfo *TRI) const {
960 if (I != MBB.end()) DL = I->getDebugLoc();
961 MachineFunction &MF = *MBB.getParent();
962 MachineFrameInfo &MFI = *MF.getFrameInfo();
963 unsigned Align = MFI.getObjectAlignment(FI);
964 MachineMemOperand *MMO =
965 MF.getMachineMemOperand(
966 MachinePointerInfo::getFixedStack(FI),
967 MachineMemOperand::MOLoad,
968 MFI.getObjectSize(FI),
971 switch (RC->getSize()) {
973 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
974 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
975 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
977 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
978 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
979 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
981 llvm_unreachable("Unknown reg class!");
984 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
985 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
986 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
987 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
988 MachineInstrBuilder MIB;
990 if (Subtarget.hasV5TEOps()) {
991 MIB = BuildMI(MBB, I, DL, get(ARM::LDRD));
992 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
993 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
994 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
998 // Fallback to LDM instruction, which has existed since the dawn of
1000 MIB = AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDMIA))
1001 .addFrameIndex(FI).addMemOperand(MMO));
1002 MIB = AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1003 MIB = AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1006 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1007 MIB.addReg(DestReg, RegState::ImplicitDefine);
1009 llvm_unreachable("Unknown reg class!");
1012 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
1013 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1014 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
1015 .addFrameIndex(FI).addImm(16)
1016 .addMemOperand(MMO));
1018 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
1020 .addMemOperand(MMO));
1023 llvm_unreachable("Unknown reg class!");
1026 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
1027 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1028 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
1029 .addFrameIndex(FI).addImm(16)
1030 .addMemOperand(MMO));
1032 MachineInstrBuilder MIB =
1033 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1035 .addMemOperand(MMO));
1036 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1037 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1038 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1039 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1040 MIB.addReg(DestReg, RegState::ImplicitDefine);
1043 llvm_unreachable("Unknown reg class!");
1046 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
1047 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1048 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
1049 .addFrameIndex(FI).addImm(16)
1050 .addMemOperand(MMO));
1052 MachineInstrBuilder MIB =
1053 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1055 .addMemOperand(MMO);
1056 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1057 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1058 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1059 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1060 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1061 MIB.addReg(DestReg, RegState::ImplicitDefine);
1064 llvm_unreachable("Unknown reg class!");
1067 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
1068 MachineInstrBuilder MIB =
1069 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1071 .addMemOperand(MMO);
1072 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1073 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1074 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1075 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1076 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
1077 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
1078 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
1079 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
1080 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1081 MIB.addReg(DestReg, RegState::ImplicitDefine);
1083 llvm_unreachable("Unknown reg class!");
1086 llvm_unreachable("Unknown regclass!");
1091 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1092 int &FrameIndex) const {
1093 switch (MI->getOpcode()) {
1096 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
1097 if (MI->getOperand(1).isFI() &&
1098 MI->getOperand(2).isReg() &&
1099 MI->getOperand(3).isImm() &&
1100 MI->getOperand(2).getReg() == 0 &&
1101 MI->getOperand(3).getImm() == 0) {
1102 FrameIndex = MI->getOperand(1).getIndex();
1103 return MI->getOperand(0).getReg();
1111 if (MI->getOperand(1).isFI() &&
1112 MI->getOperand(2).isImm() &&
1113 MI->getOperand(2).getImm() == 0) {
1114 FrameIndex = MI->getOperand(1).getIndex();
1115 return MI->getOperand(0).getReg();
1119 case ARM::VLD1d64TPseudo:
1120 case ARM::VLD1d64QPseudo:
1121 if (MI->getOperand(1).isFI() &&
1122 MI->getOperand(0).getSubReg() == 0) {
1123 FrameIndex = MI->getOperand(1).getIndex();
1124 return MI->getOperand(0).getReg();
1128 if (MI->getOperand(1).isFI() &&
1129 MI->getOperand(0).getSubReg() == 0) {
1130 FrameIndex = MI->getOperand(1).getIndex();
1131 return MI->getOperand(0).getReg();
1139 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1140 int &FrameIndex) const {
1141 const MachineMemOperand *Dummy;
1142 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1145 bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
1146 // This hook gets to expand COPY instructions before they become
1147 // copyPhysReg() calls. Look for VMOVS instructions that can legally be
1148 // widened to VMOVD. We prefer the VMOVD when possible because it may be
1149 // changed into a VORR that can go down the NEON pipeline.
1150 if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15())
1153 // Look for a copy between even S-registers. That is where we keep floats
1154 // when using NEON v2f32 instructions for f32 arithmetic.
1155 unsigned DstRegS = MI->getOperand(0).getReg();
1156 unsigned SrcRegS = MI->getOperand(1).getReg();
1157 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
1160 const TargetRegisterInfo *TRI = &getRegisterInfo();
1161 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
1163 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
1165 if (!DstRegD || !SrcRegD)
1168 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
1169 // legal if the COPY already defines the full DstRegD, and it isn't a
1170 // sub-register insertion.
1171 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
1174 // A dead copy shouldn't show up here, but reject it just in case.
1175 if (MI->getOperand(0).isDead())
1178 // All clear, widen the COPY.
1179 DEBUG(dbgs() << "widening: " << *MI);
1180 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1182 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
1183 // or some other super-register.
1184 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
1185 if (ImpDefIdx != -1)
1186 MI->RemoveOperand(ImpDefIdx);
1188 // Change the opcode and operands.
1189 MI->setDesc(get(ARM::VMOVD));
1190 MI->getOperand(0).setReg(DstRegD);
1191 MI->getOperand(1).setReg(SrcRegD);
1192 AddDefaultPred(MIB);
1194 // We are now reading SrcRegD instead of SrcRegS. This may upset the
1195 // register scavenger and machine verifier, so we need to indicate that we
1196 // are reading an undefined value from SrcRegD, but a proper value from
1198 MI->getOperand(1).setIsUndef();
1199 MIB.addReg(SrcRegS, RegState::Implicit);
1201 // SrcRegD may actually contain an unrelated value in the ssub_1
1202 // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
1203 if (MI->getOperand(1).isKill()) {
1204 MI->getOperand(1).setIsKill(false);
1205 MI->addRegisterKilled(SrcRegS, TRI, true);
1208 DEBUG(dbgs() << "replaced by: " << *MI);
1212 /// Create a copy of a const pool value. Update CPI to the new index and return
1214 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1215 MachineConstantPool *MCP = MF.getConstantPool();
1216 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1218 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1219 assert(MCPE.isMachineConstantPoolEntry() &&
1220 "Expecting a machine constantpool entry!");
1221 ARMConstantPoolValue *ACPV =
1222 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1224 unsigned PCLabelId = AFI->createPICLabelUId();
1225 ARMConstantPoolValue *NewCPV = 0;
1226 // FIXME: The below assumes PIC relocation model and that the function
1227 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1228 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1229 // instructions, so that's probably OK, but is PIC always correct when
1231 if (ACPV->isGlobalValue())
1232 NewCPV = ARMConstantPoolConstant::
1233 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
1235 else if (ACPV->isExtSymbol())
1236 NewCPV = ARMConstantPoolSymbol::
1237 Create(MF.getFunction()->getContext(),
1238 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
1239 else if (ACPV->isBlockAddress())
1240 NewCPV = ARMConstantPoolConstant::
1241 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
1242 ARMCP::CPBlockAddress, 4);
1243 else if (ACPV->isLSDA())
1244 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
1246 else if (ACPV->isMachineBasicBlock())
1247 NewCPV = ARMConstantPoolMBB::
1248 Create(MF.getFunction()->getContext(),
1249 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
1251 llvm_unreachable("Unexpected ARM constantpool value type!!");
1252 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1256 void ARMBaseInstrInfo::
1257 reMaterialize(MachineBasicBlock &MBB,
1258 MachineBasicBlock::iterator I,
1259 unsigned DestReg, unsigned SubIdx,
1260 const MachineInstr *Orig,
1261 const TargetRegisterInfo &TRI) const {
1262 unsigned Opcode = Orig->getOpcode();
1265 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1266 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1270 case ARM::tLDRpci_pic:
1271 case ARM::t2LDRpci_pic: {
1272 MachineFunction &MF = *MBB.getParent();
1273 unsigned CPI = Orig->getOperand(1).getIndex();
1274 unsigned PCLabelId = duplicateCPV(MF, CPI);
1275 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1277 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1278 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1285 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1286 MachineInstr *MI = TargetInstrInfo::duplicate(Orig, MF);
1287 switch(Orig->getOpcode()) {
1288 case ARM::tLDRpci_pic:
1289 case ARM::t2LDRpci_pic: {
1290 unsigned CPI = Orig->getOperand(1).getIndex();
1291 unsigned PCLabelId = duplicateCPV(MF, CPI);
1292 Orig->getOperand(1).setIndex(CPI);
1293 Orig->getOperand(2).setImm(PCLabelId);
1300 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1301 const MachineInstr *MI1,
1302 const MachineRegisterInfo *MRI) const {
1303 int Opcode = MI0->getOpcode();
1304 if (Opcode == ARM::t2LDRpci ||
1305 Opcode == ARM::t2LDRpci_pic ||
1306 Opcode == ARM::tLDRpci ||
1307 Opcode == ARM::tLDRpci_pic ||
1308 Opcode == ARM::MOV_ga_dyn ||
1309 Opcode == ARM::MOV_ga_pcrel ||
1310 Opcode == ARM::MOV_ga_pcrel_ldr ||
1311 Opcode == ARM::t2MOV_ga_dyn ||
1312 Opcode == ARM::t2MOV_ga_pcrel) {
1313 if (MI1->getOpcode() != Opcode)
1315 if (MI0->getNumOperands() != MI1->getNumOperands())
1318 const MachineOperand &MO0 = MI0->getOperand(1);
1319 const MachineOperand &MO1 = MI1->getOperand(1);
1320 if (MO0.getOffset() != MO1.getOffset())
1323 if (Opcode == ARM::MOV_ga_dyn ||
1324 Opcode == ARM::MOV_ga_pcrel ||
1325 Opcode == ARM::MOV_ga_pcrel_ldr ||
1326 Opcode == ARM::t2MOV_ga_dyn ||
1327 Opcode == ARM::t2MOV_ga_pcrel)
1328 // Ignore the PC labels.
1329 return MO0.getGlobal() == MO1.getGlobal();
1331 const MachineFunction *MF = MI0->getParent()->getParent();
1332 const MachineConstantPool *MCP = MF->getConstantPool();
1333 int CPI0 = MO0.getIndex();
1334 int CPI1 = MO1.getIndex();
1335 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1336 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1337 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1338 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1339 if (isARMCP0 && isARMCP1) {
1340 ARMConstantPoolValue *ACPV0 =
1341 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1342 ARMConstantPoolValue *ACPV1 =
1343 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1344 return ACPV0->hasSameValue(ACPV1);
1345 } else if (!isARMCP0 && !isARMCP1) {
1346 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1349 } else if (Opcode == ARM::PICLDR) {
1350 if (MI1->getOpcode() != Opcode)
1352 if (MI0->getNumOperands() != MI1->getNumOperands())
1355 unsigned Addr0 = MI0->getOperand(1).getReg();
1356 unsigned Addr1 = MI1->getOperand(1).getReg();
1357 if (Addr0 != Addr1) {
1359 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1360 !TargetRegisterInfo::isVirtualRegister(Addr1))
1363 // This assumes SSA form.
1364 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1365 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1366 // Check if the loaded value, e.g. a constantpool of a global address, are
1368 if (!produceSameValue(Def0, Def1, MRI))
1372 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1373 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1374 const MachineOperand &MO0 = MI0->getOperand(i);
1375 const MachineOperand &MO1 = MI1->getOperand(i);
1376 if (!MO0.isIdenticalTo(MO1))
1382 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1385 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1386 /// determine if two loads are loading from the same base address. It should
1387 /// only return true if the base pointers are the same and the only differences
1388 /// between the two addresses is the offset. It also returns the offsets by
1391 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1392 /// is permanently disabled.
1393 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1395 int64_t &Offset2) const {
1396 // Don't worry about Thumb: just ARM and Thumb2.
1397 if (Subtarget.isThumb1Only()) return false;
1399 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1402 switch (Load1->getMachineOpcode()) {
1416 case ARM::t2LDRSHi8:
1418 case ARM::t2LDRBi12:
1419 case ARM::t2LDRSHi12:
1423 switch (Load2->getMachineOpcode()) {
1436 case ARM::t2LDRSHi8:
1438 case ARM::t2LDRBi12:
1439 case ARM::t2LDRSHi12:
1443 // Check if base addresses and chain operands match.
1444 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1445 Load1->getOperand(4) != Load2->getOperand(4))
1448 // Index should be Reg0.
1449 if (Load1->getOperand(3) != Load2->getOperand(3))
1452 // Determine the offsets.
1453 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1454 isa<ConstantSDNode>(Load2->getOperand(1))) {
1455 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1456 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1463 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1464 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1465 /// be scheduled togther. On some targets if two loads are loading from
1466 /// addresses in the same cache line, it's better if they are scheduled
1467 /// together. This function takes two integers that represent the load offsets
1468 /// from the common base address. It returns true if it decides it's desirable
1469 /// to schedule the two loads together. "NumLoads" is the number of loads that
1470 /// have already been scheduled after Load1.
1472 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1473 /// is permanently disabled.
1474 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1475 int64_t Offset1, int64_t Offset2,
1476 unsigned NumLoads) const {
1477 // Don't worry about Thumb: just ARM and Thumb2.
1478 if (Subtarget.isThumb1Only()) return false;
1480 assert(Offset2 > Offset1);
1482 if ((Offset2 - Offset1) / 8 > 64)
1485 // Check if the machine opcodes are different. If they are different
1486 // then we consider them to not be of the same base address,
1487 // EXCEPT in the case of Thumb2 byte loads where one is LDRBi8 and the other LDRBi12.
1488 // In this case, they are considered to be the same because they are different
1489 // encoding forms of the same basic instruction.
1490 if ((Load1->getMachineOpcode() != Load2->getMachineOpcode()) &&
1491 !((Load1->getMachineOpcode() == ARM::t2LDRBi8 &&
1492 Load2->getMachineOpcode() == ARM::t2LDRBi12) ||
1493 (Load1->getMachineOpcode() == ARM::t2LDRBi12 &&
1494 Load2->getMachineOpcode() == ARM::t2LDRBi8)))
1495 return false; // FIXME: overly conservative?
1497 // Four loads in a row should be sufficient.
1504 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1505 const MachineBasicBlock *MBB,
1506 const MachineFunction &MF) const {
1507 // Debug info is never a scheduling boundary. It's necessary to be explicit
1508 // due to the special treatment of IT instructions below, otherwise a
1509 // dbg_value followed by an IT will result in the IT instruction being
1510 // considered a scheduling hazard, which is wrong. It should be the actual
1511 // instruction preceding the dbg_value instruction(s), just like it is
1512 // when debug info is not present.
1513 if (MI->isDebugValue())
1516 // Terminators and labels can't be scheduled around.
1517 if (MI->isTerminator() || MI->isLabel())
1520 // Treat the start of the IT block as a scheduling boundary, but schedule
1521 // t2IT along with all instructions following it.
1522 // FIXME: This is a big hammer. But the alternative is to add all potential
1523 // true and anti dependencies to IT block instructions as implicit operands
1524 // to the t2IT instruction. The added compile time and complexity does not
1526 MachineBasicBlock::const_iterator I = MI;
1527 // Make sure to skip any dbg_value instructions
1528 while (++I != MBB->end() && I->isDebugValue())
1530 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1533 // Don't attempt to schedule around any instruction that defines
1534 // a stack-oriented pointer, as it's unlikely to be profitable. This
1535 // saves compile time, because it doesn't require every single
1536 // stack slot reference to depend on the instruction that does the
1538 // Calls don't actually change the stack pointer, even if they have imp-defs.
1539 // No ARM calling conventions change the stack pointer. (X86 calling
1540 // conventions sometimes do).
1541 if (!MI->isCall() && MI->definesRegister(ARM::SP))
1547 bool ARMBaseInstrInfo::
1548 isProfitableToIfCvt(MachineBasicBlock &MBB,
1549 unsigned NumCycles, unsigned ExtraPredCycles,
1550 const BranchProbability &Probability) const {
1554 // Attempt to estimate the relative costs of predication versus branching.
1555 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1556 UnpredCost /= Probability.getDenominator();
1557 UnpredCost += 1; // The branch itself
1558 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1560 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1563 bool ARMBaseInstrInfo::
1564 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1565 unsigned TCycles, unsigned TExtra,
1566 MachineBasicBlock &FMBB,
1567 unsigned FCycles, unsigned FExtra,
1568 const BranchProbability &Probability) const {
1569 if (!TCycles || !FCycles)
1572 // Attempt to estimate the relative costs of predication versus branching.
1573 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1574 TUnpredCost /= Probability.getDenominator();
1576 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1577 unsigned FUnpredCost = Comp * FCycles;
1578 FUnpredCost /= Probability.getDenominator();
1580 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1581 UnpredCost += 1; // The branch itself
1582 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1584 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1588 ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
1589 MachineBasicBlock &FMBB) const {
1590 // Reduce false anti-dependencies to let Swift's out-of-order execution
1591 // engine do its thing.
1592 return Subtarget.isSwift();
1595 /// getInstrPredicate - If instruction is predicated, returns its predicate
1596 /// condition, otherwise returns AL. It also returns the condition code
1597 /// register by reference.
1599 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1600 int PIdx = MI->findFirstPredOperandIdx();
1606 PredReg = MI->getOperand(PIdx+1).getReg();
1607 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1611 int llvm::getMatchingCondBranchOpcode(int Opc) {
1616 if (Opc == ARM::t2B)
1619 llvm_unreachable("Unknown unconditional branch opcode!");
1622 /// commuteInstruction - Handle commutable instructions.
1624 ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1625 switch (MI->getOpcode()) {
1627 case ARM::t2MOVCCr: {
1628 // MOVCC can be commuted by inverting the condition.
1629 unsigned PredReg = 0;
1630 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
1631 // MOVCC AL can't be inverted. Shouldn't happen.
1632 if (CC == ARMCC::AL || PredReg != ARM::CPSR)
1634 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
1637 // After swapping the MOVCC operands, also invert the condition.
1638 MI->getOperand(MI->findFirstPredOperandIdx())
1639 .setImm(ARMCC::getOppositeCondition(CC));
1643 return TargetInstrInfo::commuteInstruction(MI, NewMI);
1646 /// Identify instructions that can be folded into a MOVCC instruction, and
1647 /// return the defining instruction.
1648 static MachineInstr *canFoldIntoMOVCC(unsigned Reg,
1649 const MachineRegisterInfo &MRI,
1650 const TargetInstrInfo *TII) {
1651 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1653 if (!MRI.hasOneNonDBGUse(Reg))
1655 MachineInstr *MI = MRI.getVRegDef(Reg);
1658 // MI is folded into the MOVCC by predicating it.
1659 if (!MI->isPredicable())
1661 // Check if MI has any non-dead defs or physreg uses. This also detects
1662 // predicated instructions which will be reading CPSR.
1663 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
1664 const MachineOperand &MO = MI->getOperand(i);
1665 // Reject frame index operands, PEI can't handle the predicated pseudos.
1666 if (MO.isFI() || MO.isCPI() || MO.isJTI())
1670 // MI can't have any tied operands, that would conflict with predication.
1673 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1675 if (MO.isDef() && !MO.isDead())
1678 bool DontMoveAcrossStores = true;
1679 if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ 0, DontMoveAcrossStores))
1684 bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
1685 SmallVectorImpl<MachineOperand> &Cond,
1686 unsigned &TrueOp, unsigned &FalseOp,
1687 bool &Optimizable) const {
1688 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1689 "Unknown select instruction");
1694 // 3: Condition code.
1698 Cond.push_back(MI->getOperand(3));
1699 Cond.push_back(MI->getOperand(4));
1700 // We can always fold a def.
1705 MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
1706 bool PreferFalse) const {
1707 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1708 "Unknown select instruction");
1709 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1710 MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
1711 bool Invert = !DefMI;
1713 DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this);
1717 // Create a new predicated version of DefMI.
1718 // Rfalse is the first use.
1719 MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1721 MI->getOperand(0).getReg());
1723 // Copy all the DefMI operands, excluding its (null) predicate.
1724 const MCInstrDesc &DefDesc = DefMI->getDesc();
1725 for (unsigned i = 1, e = DefDesc.getNumOperands();
1726 i != e && !DefDesc.OpInfo[i].isPredicate(); ++i)
1727 NewMI.addOperand(DefMI->getOperand(i));
1729 unsigned CondCode = MI->getOperand(3).getImm();
1731 NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode)));
1733 NewMI.addImm(CondCode);
1734 NewMI.addOperand(MI->getOperand(4));
1736 // DefMI is not the -S version that sets CPSR, so add an optional %noreg.
1737 if (NewMI->hasOptionalDef())
1738 AddDefaultCC(NewMI);
1740 // The output register value when the predicate is false is an implicit
1741 // register operand tied to the first def.
1742 // The tie makes the register allocator ensure the FalseReg is allocated the
1743 // same register as operand 0.
1744 MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
1745 FalseReg.setImplicit();
1746 NewMI.addOperand(FalseReg);
1747 NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
1749 // The caller will erase MI, but not DefMI.
1750 DefMI->eraseFromParent();
1754 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
1755 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
1758 /// This will go away once we can teach tblgen how to set the optional CPSR def
1760 struct AddSubFlagsOpcodePair {
1762 uint16_t MachineOpc;
1765 static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
1766 {ARM::ADDSri, ARM::ADDri},
1767 {ARM::ADDSrr, ARM::ADDrr},
1768 {ARM::ADDSrsi, ARM::ADDrsi},
1769 {ARM::ADDSrsr, ARM::ADDrsr},
1771 {ARM::SUBSri, ARM::SUBri},
1772 {ARM::SUBSrr, ARM::SUBrr},
1773 {ARM::SUBSrsi, ARM::SUBrsi},
1774 {ARM::SUBSrsr, ARM::SUBrsr},
1776 {ARM::RSBSri, ARM::RSBri},
1777 {ARM::RSBSrsi, ARM::RSBrsi},
1778 {ARM::RSBSrsr, ARM::RSBrsr},
1780 {ARM::t2ADDSri, ARM::t2ADDri},
1781 {ARM::t2ADDSrr, ARM::t2ADDrr},
1782 {ARM::t2ADDSrs, ARM::t2ADDrs},
1784 {ARM::t2SUBSri, ARM::t2SUBri},
1785 {ARM::t2SUBSrr, ARM::t2SUBrr},
1786 {ARM::t2SUBSrs, ARM::t2SUBrs},
1788 {ARM::t2RSBSri, ARM::t2RSBri},
1789 {ARM::t2RSBSrs, ARM::t2RSBrs},
1792 unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
1793 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
1794 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
1795 return AddSubFlagsOpcodeMap[i].MachineOpc;
1799 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1800 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1801 unsigned DestReg, unsigned BaseReg, int NumBytes,
1802 ARMCC::CondCodes Pred, unsigned PredReg,
1803 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1804 bool isSub = NumBytes < 0;
1805 if (isSub) NumBytes = -NumBytes;
1808 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1809 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1810 assert(ThisVal && "Didn't extract field correctly");
1812 // We will handle these bits from offset, clear them.
1813 NumBytes &= ~ThisVal;
1815 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1817 // Build the new ADD / SUB.
1818 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1819 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1820 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1821 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1822 .setMIFlags(MIFlags);
1827 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1828 unsigned FrameReg, int &Offset,
1829 const ARMBaseInstrInfo &TII) {
1830 unsigned Opcode = MI.getOpcode();
1831 const MCInstrDesc &Desc = MI.getDesc();
1832 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1835 // Memory operands in inline assembly always use AddrMode2.
1836 if (Opcode == ARM::INLINEASM)
1837 AddrMode = ARMII::AddrMode2;
1839 if (Opcode == ARM::ADDri) {
1840 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1842 // Turn it into a move.
1843 MI.setDesc(TII.get(ARM::MOVr));
1844 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1845 MI.RemoveOperand(FrameRegIdx+1);
1848 } else if (Offset < 0) {
1851 MI.setDesc(TII.get(ARM::SUBri));
1854 // Common case: small offset, fits into instruction.
1855 if (ARM_AM::getSOImmVal(Offset) != -1) {
1856 // Replace the FrameIndex with sp / fp
1857 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1858 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1863 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1865 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1866 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1868 // We will handle these bits from offset, clear them.
1869 Offset &= ~ThisImmVal;
1871 // Get the properly encoded SOImmVal field.
1872 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1873 "Bit extraction didn't work?");
1874 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1876 unsigned ImmIdx = 0;
1878 unsigned NumBits = 0;
1881 case ARMII::AddrMode_i12: {
1882 ImmIdx = FrameRegIdx + 1;
1883 InstrOffs = MI.getOperand(ImmIdx).getImm();
1887 case ARMII::AddrMode2: {
1888 ImmIdx = FrameRegIdx+2;
1889 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1890 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1895 case ARMII::AddrMode3: {
1896 ImmIdx = FrameRegIdx+2;
1897 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1898 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1903 case ARMII::AddrMode4:
1904 case ARMII::AddrMode6:
1905 // Can't fold any offset even if it's zero.
1907 case ARMII::AddrMode5: {
1908 ImmIdx = FrameRegIdx+1;
1909 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1910 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1917 llvm_unreachable("Unsupported addressing mode!");
1920 Offset += InstrOffs * Scale;
1921 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1927 // Attempt to fold address comp. if opcode has offset bits
1929 // Common case: small offset, fits into instruction.
1930 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1931 int ImmedOffset = Offset / Scale;
1932 unsigned Mask = (1 << NumBits) - 1;
1933 if ((unsigned)Offset <= Mask * Scale) {
1934 // Replace the FrameIndex with sp
1935 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1936 // FIXME: When addrmode2 goes away, this will simplify (like the
1937 // T2 version), as the LDR.i12 versions don't need the encoding
1938 // tricks for the offset value.
1940 if (AddrMode == ARMII::AddrMode_i12)
1941 ImmedOffset = -ImmedOffset;
1943 ImmedOffset |= 1 << NumBits;
1945 ImmOp.ChangeToImmediate(ImmedOffset);
1950 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1951 ImmedOffset = ImmedOffset & Mask;
1953 if (AddrMode == ARMII::AddrMode_i12)
1954 ImmedOffset = -ImmedOffset;
1956 ImmedOffset |= 1 << NumBits;
1958 ImmOp.ChangeToImmediate(ImmedOffset);
1959 Offset &= ~(Mask*Scale);
1963 Offset = (isSub) ? -Offset : Offset;
1967 /// analyzeCompare - For a comparison instruction, return the source registers
1968 /// in SrcReg and SrcReg2 if having two register operands, and the value it
1969 /// compares against in CmpValue. Return true if the comparison instruction
1970 /// can be analyzed.
1971 bool ARMBaseInstrInfo::
1972 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
1973 int &CmpMask, int &CmpValue) const {
1974 switch (MI->getOpcode()) {
1978 SrcReg = MI->getOperand(0).getReg();
1981 CmpValue = MI->getOperand(1).getImm();
1985 SrcReg = MI->getOperand(0).getReg();
1986 SrcReg2 = MI->getOperand(1).getReg();
1992 SrcReg = MI->getOperand(0).getReg();
1994 CmpMask = MI->getOperand(1).getImm();
2002 /// isSuitableForMask - Identify a suitable 'and' instruction that
2003 /// operates on the given source register and applies the same mask
2004 /// as a 'tst' instruction. Provide a limited look-through for copies.
2005 /// When successful, MI will hold the found instruction.
2006 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
2007 int CmpMask, bool CommonUse) {
2008 switch (MI->getOpcode()) {
2011 if (CmpMask != MI->getOperand(2).getImm())
2013 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
2017 // Walk down one instruction which is potentially an 'and'.
2018 const MachineInstr &Copy = *MI;
2019 MachineBasicBlock::iterator AND(
2020 llvm::next(MachineBasicBlock::iterator(MI)));
2021 if (AND == MI->getParent()->end()) return false;
2023 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
2031 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2032 /// the condition code if we modify the instructions such that flags are
2034 inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) {
2036 default: return ARMCC::AL;
2037 case ARMCC::EQ: return ARMCC::EQ;
2038 case ARMCC::NE: return ARMCC::NE;
2039 case ARMCC::HS: return ARMCC::LS;
2040 case ARMCC::LO: return ARMCC::HI;
2041 case ARMCC::HI: return ARMCC::LO;
2042 case ARMCC::LS: return ARMCC::HS;
2043 case ARMCC::GE: return ARMCC::LE;
2044 case ARMCC::LT: return ARMCC::GT;
2045 case ARMCC::GT: return ARMCC::LT;
2046 case ARMCC::LE: return ARMCC::GE;
2050 /// isRedundantFlagInstr - check whether the first instruction, whose only
2051 /// purpose is to update flags, can be made redundant.
2052 /// CMPrr can be made redundant by SUBrr if the operands are the same.
2053 /// CMPri can be made redundant by SUBri if the operands are the same.
2054 /// This function can be extended later on.
2055 inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg,
2056 unsigned SrcReg2, int ImmValue,
2058 if ((CmpI->getOpcode() == ARM::CMPrr ||
2059 CmpI->getOpcode() == ARM::t2CMPrr) &&
2060 (OI->getOpcode() == ARM::SUBrr ||
2061 OI->getOpcode() == ARM::t2SUBrr) &&
2062 ((OI->getOperand(1).getReg() == SrcReg &&
2063 OI->getOperand(2).getReg() == SrcReg2) ||
2064 (OI->getOperand(1).getReg() == SrcReg2 &&
2065 OI->getOperand(2).getReg() == SrcReg)))
2068 if ((CmpI->getOpcode() == ARM::CMPri ||
2069 CmpI->getOpcode() == ARM::t2CMPri) &&
2070 (OI->getOpcode() == ARM::SUBri ||
2071 OI->getOpcode() == ARM::t2SUBri) &&
2072 OI->getOperand(1).getReg() == SrcReg &&
2073 OI->getOperand(2).getImm() == ImmValue)
2078 /// optimizeCompareInstr - Convert the instruction supplying the argument to the
2079 /// comparison into one that sets the zero bit in the flags register;
2080 /// Remove a redundant Compare instruction if an earlier instruction can set the
2081 /// flags in the same way as Compare.
2082 /// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two
2083 /// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the
2084 /// condition code of instructions which use the flags.
2085 bool ARMBaseInstrInfo::
2086 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
2087 int CmpMask, int CmpValue,
2088 const MachineRegisterInfo *MRI) const {
2089 // Get the unique definition of SrcReg.
2090 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
2091 if (!MI) return false;
2093 // Masked compares sometimes use the same register as the corresponding 'and'.
2094 if (CmpMask != ~0) {
2095 if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) {
2097 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
2098 UE = MRI->use_end(); UI != UE; ++UI) {
2099 if (UI->getParent() != CmpInstr->getParent()) continue;
2100 MachineInstr *PotentialAND = &*UI;
2101 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) ||
2102 isPredicated(PotentialAND))
2107 if (!MI) return false;
2111 // Get ready to iterate backward from CmpInstr.
2112 MachineBasicBlock::iterator I = CmpInstr, E = MI,
2113 B = CmpInstr->getParent()->begin();
2115 // Early exit if CmpInstr is at the beginning of the BB.
2116 if (I == B) return false;
2118 // There are two possible candidates which can be changed to set CPSR:
2119 // One is MI, the other is a SUB instruction.
2120 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
2121 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
2122 MachineInstr *Sub = NULL;
2124 // MI is not a candidate for CMPrr.
2126 else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
2127 // Conservatively refuse to convert an instruction which isn't in the same
2128 // BB as the comparison.
2129 // For CMPri, we need to check Sub, thus we can't return here.
2130 if (CmpInstr->getOpcode() == ARM::CMPri ||
2131 CmpInstr->getOpcode() == ARM::t2CMPri)
2137 // Check that CPSR isn't set between the comparison instruction and the one we
2138 // want to change. At the same time, search for Sub.
2139 const TargetRegisterInfo *TRI = &getRegisterInfo();
2141 for (; I != E; --I) {
2142 const MachineInstr &Instr = *I;
2144 if (Instr.modifiesRegister(ARM::CPSR, TRI) ||
2145 Instr.readsRegister(ARM::CPSR, TRI))
2146 // This instruction modifies or uses CPSR after the one we want to
2147 // change. We can't do this transformation.
2150 // Check whether CmpInstr can be made redundant by the current instruction.
2151 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
2157 // The 'and' is below the comparison instruction.
2161 // Return false if no candidates exist.
2165 // The single candidate is called MI.
2168 // We can't use a predicated instruction - it doesn't always write the flags.
2169 if (isPredicated(MI))
2172 switch (MI->getOpcode()) {
2206 case ARM::t2EORri: {
2207 // Scan forward for the use of CPSR
2208 // When checking against MI: if it's a conditional code requires
2209 // checking of V bit, then this is not safe to do.
2210 // It is safe to remove CmpInstr if CPSR is redefined or killed.
2211 // If we are done with the basic block, we need to check whether CPSR is
2213 SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
2215 bool isSafe = false;
2217 E = CmpInstr->getParent()->end();
2218 while (!isSafe && ++I != E) {
2219 const MachineInstr &Instr = *I;
2220 for (unsigned IO = 0, EO = Instr.getNumOperands();
2221 !isSafe && IO != EO; ++IO) {
2222 const MachineOperand &MO = Instr.getOperand(IO);
2223 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
2227 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
2233 // Condition code is after the operand before CPSR.
2234 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
2236 ARMCC::CondCodes NewCC = getSwappedCondition(CC);
2237 if (NewCC == ARMCC::AL)
2239 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
2240 // on CMP needs to be updated to be based on SUB.
2241 // Push the condition code operands to OperandsToUpdate.
2242 // If it is safe to remove CmpInstr, the condition code of these
2243 // operands will be modified.
2244 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
2245 Sub->getOperand(2).getReg() == SrcReg)
2246 OperandsToUpdate.push_back(std::make_pair(&((*I).getOperand(IO-1)),
2252 // CPSR can be used multiple times, we should continue.
2265 // If CPSR is not killed nor re-defined, we should check whether it is
2266 // live-out. If it is live-out, do not optimize.
2268 MachineBasicBlock *MBB = CmpInstr->getParent();
2269 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
2270 SE = MBB->succ_end(); SI != SE; ++SI)
2271 if ((*SI)->isLiveIn(ARM::CPSR))
2275 // Toggle the optional operand to CPSR.
2276 MI->getOperand(5).setReg(ARM::CPSR);
2277 MI->getOperand(5).setIsDef(true);
2278 assert(!isPredicated(MI) && "Can't use flags from predicated instruction");
2279 CmpInstr->eraseFromParent();
2281 // Modify the condition code of operands in OperandsToUpdate.
2282 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2283 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2284 for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++)
2285 OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second);
2293 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
2294 MachineInstr *DefMI, unsigned Reg,
2295 MachineRegisterInfo *MRI) const {
2296 // Fold large immediates into add, sub, or, xor.
2297 unsigned DefOpc = DefMI->getOpcode();
2298 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
2300 if (!DefMI->getOperand(1).isImm())
2301 // Could be t2MOVi32imm <ga:xx>
2304 if (!MRI->hasOneNonDBGUse(Reg))
2307 const MCInstrDesc &DefMCID = DefMI->getDesc();
2308 if (DefMCID.hasOptionalDef()) {
2309 unsigned NumOps = DefMCID.getNumOperands();
2310 const MachineOperand &MO = DefMI->getOperand(NumOps-1);
2311 if (MO.getReg() == ARM::CPSR && !MO.isDead())
2312 // If DefMI defines CPSR and it is not dead, it's obviously not safe
2317 const MCInstrDesc &UseMCID = UseMI->getDesc();
2318 if (UseMCID.hasOptionalDef()) {
2319 unsigned NumOps = UseMCID.getNumOperands();
2320 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
2321 // If the instruction sets the flag, do not attempt this optimization
2322 // since it may change the semantics of the code.
2326 unsigned UseOpc = UseMI->getOpcode();
2327 unsigned NewUseOpc = 0;
2328 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
2329 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
2330 bool Commute = false;
2332 default: return false;
2340 case ARM::t2EORrr: {
2341 Commute = UseMI->getOperand(2).getReg() != Reg;
2348 NewUseOpc = ARM::SUBri;
2354 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
2356 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
2357 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
2360 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
2361 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
2362 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
2366 case ARM::t2SUBrr: {
2370 NewUseOpc = ARM::t2SUBri;
2375 case ARM::t2EORrr: {
2376 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
2378 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
2379 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
2382 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
2383 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
2384 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
2392 unsigned OpIdx = Commute ? 2 : 1;
2393 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
2394 bool isKill = UseMI->getOperand(OpIdx).isKill();
2395 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
2396 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
2397 UseMI, UseMI->getDebugLoc(),
2398 get(NewUseOpc), NewReg)
2399 .addReg(Reg1, getKillRegState(isKill))
2400 .addImm(SOImmValV1)));
2401 UseMI->setDesc(get(NewUseOpc));
2402 UseMI->getOperand(1).setReg(NewReg);
2403 UseMI->getOperand(1).setIsKill();
2404 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
2405 DefMI->eraseFromParent();
2409 static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData,
2410 const MachineInstr *MI) {
2411 switch (MI->getOpcode()) {
2413 const MCInstrDesc &Desc = MI->getDesc();
2414 int UOps = ItinData->getNumMicroOps(Desc.getSchedClass());
2415 assert(UOps >= 0 && "bad # UOps");
2423 unsigned ShOpVal = MI->getOperand(3).getImm();
2424 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2425 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2428 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2429 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2436 if (!MI->getOperand(2).getReg())
2439 unsigned ShOpVal = MI->getOperand(3).getImm();
2440 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2441 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2444 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2445 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2452 return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2;
2454 case ARM::LDRSB_POST:
2455 case ARM::LDRSH_POST: {
2456 unsigned Rt = MI->getOperand(0).getReg();
2457 unsigned Rm = MI->getOperand(3).getReg();
2458 return (Rt == Rm) ? 4 : 3;
2461 case ARM::LDR_PRE_REG:
2462 case ARM::LDRB_PRE_REG: {
2463 unsigned Rt = MI->getOperand(0).getReg();
2464 unsigned Rm = MI->getOperand(3).getReg();
2467 unsigned ShOpVal = MI->getOperand(4).getImm();
2468 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2469 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2472 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2473 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2478 case ARM::STR_PRE_REG:
2479 case ARM::STRB_PRE_REG: {
2480 unsigned ShOpVal = MI->getOperand(4).getImm();
2481 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2482 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2485 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2486 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2492 case ARM::STRH_PRE: {
2493 unsigned Rt = MI->getOperand(0).getReg();
2494 unsigned Rm = MI->getOperand(3).getReg();
2499 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub)
2503 case ARM::LDR_POST_REG:
2504 case ARM::LDRB_POST_REG:
2505 case ARM::LDRH_POST: {
2506 unsigned Rt = MI->getOperand(0).getReg();
2507 unsigned Rm = MI->getOperand(3).getReg();
2508 return (Rt == Rm) ? 3 : 2;
2511 case ARM::LDR_PRE_IMM:
2512 case ARM::LDRB_PRE_IMM:
2513 case ARM::LDR_POST_IMM:
2514 case ARM::LDRB_POST_IMM:
2515 case ARM::STRB_POST_IMM:
2516 case ARM::STRB_POST_REG:
2517 case ARM::STRB_PRE_IMM:
2518 case ARM::STRH_POST:
2519 case ARM::STR_POST_IMM:
2520 case ARM::STR_POST_REG:
2521 case ARM::STR_PRE_IMM:
2524 case ARM::LDRSB_PRE:
2525 case ARM::LDRSH_PRE: {
2526 unsigned Rm = MI->getOperand(3).getReg();
2529 unsigned Rt = MI->getOperand(0).getReg();
2532 unsigned ShOpVal = MI->getOperand(4).getImm();
2533 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2534 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2537 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2538 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2544 unsigned Rt = MI->getOperand(0).getReg();
2545 unsigned Rn = MI->getOperand(2).getReg();
2546 unsigned Rm = MI->getOperand(3).getReg();
2548 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2549 return (Rt == Rn) ? 3 : 2;
2553 unsigned Rm = MI->getOperand(3).getReg();
2555 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2559 case ARM::LDRD_POST:
2560 case ARM::t2LDRD_POST:
2563 case ARM::STRD_POST:
2564 case ARM::t2STRD_POST:
2567 case ARM::LDRD_PRE: {
2568 unsigned Rt = MI->getOperand(0).getReg();
2569 unsigned Rn = MI->getOperand(3).getReg();
2570 unsigned Rm = MI->getOperand(4).getReg();
2572 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2573 return (Rt == Rn) ? 4 : 3;
2576 case ARM::t2LDRD_PRE: {
2577 unsigned Rt = MI->getOperand(0).getReg();
2578 unsigned Rn = MI->getOperand(3).getReg();
2579 return (Rt == Rn) ? 4 : 3;
2582 case ARM::STRD_PRE: {
2583 unsigned Rm = MI->getOperand(4).getReg();
2585 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2589 case ARM::t2STRD_PRE:
2592 case ARM::t2LDR_POST:
2593 case ARM::t2LDRB_POST:
2594 case ARM::t2LDRB_PRE:
2595 case ARM::t2LDRSBi12:
2596 case ARM::t2LDRSBi8:
2597 case ARM::t2LDRSBpci:
2599 case ARM::t2LDRH_POST:
2600 case ARM::t2LDRH_PRE:
2602 case ARM::t2LDRSB_POST:
2603 case ARM::t2LDRSB_PRE:
2604 case ARM::t2LDRSH_POST:
2605 case ARM::t2LDRSH_PRE:
2606 case ARM::t2LDRSHi12:
2607 case ARM::t2LDRSHi8:
2608 case ARM::t2LDRSHpci:
2612 case ARM::t2LDRDi8: {
2613 unsigned Rt = MI->getOperand(0).getReg();
2614 unsigned Rn = MI->getOperand(2).getReg();
2615 return (Rt == Rn) ? 3 : 2;
2618 case ARM::t2STRB_POST:
2619 case ARM::t2STRB_PRE:
2622 case ARM::t2STRH_POST:
2623 case ARM::t2STRH_PRE:
2625 case ARM::t2STR_POST:
2626 case ARM::t2STR_PRE:
2632 // Return the number of 32-bit words loaded by LDM or stored by STM. If this
2633 // can't be easily determined return 0 (missing MachineMemOperand).
2635 // FIXME: The current MachineInstr design does not support relying on machine
2636 // mem operands to determine the width of a memory access. Instead, we expect
2637 // the target to provide this information based on the instruction opcode and
2638 // operands. However, using MachineMemOperand is a the best solution now for
2641 // 1) getNumMicroOps tries to infer LDM memory width from the total number of MI
2642 // operands. This is much more dangerous than using the MachineMemOperand
2643 // sizes because CodeGen passes can insert/remove optional machine operands. In
2644 // fact, it's totally incorrect for preRA passes and appears to be wrong for
2645 // postRA passes as well.
2647 // 2) getNumLDMAddresses is only used by the scheduling machine model and any
2648 // machine model that calls this should handle the unknown (zero size) case.
2650 // Long term, we should require a target hook that verifies MachineMemOperand
2651 // sizes during MC lowering. That target hook should be local to MC lowering
2652 // because we can't ensure that it is aware of other MI forms. Doing this will
2653 // ensure that MachineMemOperands are correctly propagated through all passes.
2654 unsigned ARMBaseInstrInfo::getNumLDMAddresses(const MachineInstr *MI) const {
2656 for (MachineInstr::mmo_iterator I = MI->memoperands_begin(),
2657 E = MI->memoperands_end(); I != E; ++I) {
2658 Size += (*I)->getSize();
2664 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
2665 const MachineInstr *MI) const {
2666 if (!ItinData || ItinData->isEmpty())
2669 const MCInstrDesc &Desc = MI->getDesc();
2670 unsigned Class = Desc.getSchedClass();
2671 int ItinUOps = ItinData->getNumMicroOps(Class);
2672 if (ItinUOps >= 0) {
2673 if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore()))
2674 return getNumMicroOpsSwiftLdSt(ItinData, MI);
2679 unsigned Opc = MI->getOpcode();
2682 llvm_unreachable("Unexpected multi-uops instruction!");
2687 // The number of uOps for load / store multiple are determined by the number
2690 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
2691 // same cycle. The scheduling for the first load / store must be done
2692 // separately by assuming the address is not 64-bit aligned.
2694 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
2695 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
2696 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
2698 case ARM::VLDMDIA_UPD:
2699 case ARM::VLDMDDB_UPD:
2701 case ARM::VLDMSIA_UPD:
2702 case ARM::VLDMSDB_UPD:
2704 case ARM::VSTMDIA_UPD:
2705 case ARM::VSTMDDB_UPD:
2707 case ARM::VSTMSIA_UPD:
2708 case ARM::VSTMSDB_UPD: {
2709 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
2710 return (NumRegs / 2) + (NumRegs % 2) + 1;
2713 case ARM::LDMIA_RET:
2718 case ARM::LDMIA_UPD:
2719 case ARM::LDMDA_UPD:
2720 case ARM::LDMDB_UPD:
2721 case ARM::LDMIB_UPD:
2726 case ARM::STMIA_UPD:
2727 case ARM::STMDA_UPD:
2728 case ARM::STMDB_UPD:
2729 case ARM::STMIB_UPD:
2731 case ARM::tLDMIA_UPD:
2732 case ARM::tSTMIA_UPD:
2736 case ARM::t2LDMIA_RET:
2739 case ARM::t2LDMIA_UPD:
2740 case ARM::t2LDMDB_UPD:
2743 case ARM::t2STMIA_UPD:
2744 case ARM::t2STMDB_UPD: {
2745 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
2746 if (Subtarget.isSwift()) {
2747 int UOps = 1 + NumRegs; // One for address computation, one for each ld / st.
2750 case ARM::VLDMDIA_UPD:
2751 case ARM::VLDMDDB_UPD:
2752 case ARM::VLDMSIA_UPD:
2753 case ARM::VLDMSDB_UPD:
2754 case ARM::VSTMDIA_UPD:
2755 case ARM::VSTMDDB_UPD:
2756 case ARM::VSTMSIA_UPD:
2757 case ARM::VSTMSDB_UPD:
2758 case ARM::LDMIA_UPD:
2759 case ARM::LDMDA_UPD:
2760 case ARM::LDMDB_UPD:
2761 case ARM::LDMIB_UPD:
2762 case ARM::STMIA_UPD:
2763 case ARM::STMDA_UPD:
2764 case ARM::STMDB_UPD:
2765 case ARM::STMIB_UPD:
2766 case ARM::tLDMIA_UPD:
2767 case ARM::tSTMIA_UPD:
2768 case ARM::t2LDMIA_UPD:
2769 case ARM::t2LDMDB_UPD:
2770 case ARM::t2STMIA_UPD:
2771 case ARM::t2STMDB_UPD:
2772 ++UOps; // One for base register writeback.
2774 case ARM::LDMIA_RET:
2776 case ARM::t2LDMIA_RET:
2777 UOps += 2; // One for base reg wb, one for write to pc.
2781 } else if (Subtarget.isCortexA8()) {
2784 // 4 registers would be issued: 2, 2.
2785 // 5 registers would be issued: 2, 2, 1.
2786 int A8UOps = (NumRegs / 2);
2790 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2791 int A9UOps = (NumRegs / 2);
2792 // If there are odd number of registers or if it's not 64-bit aligned,
2793 // then it takes an extra AGU (Address Generation Unit) cycle.
2794 if ((NumRegs % 2) ||
2795 !MI->hasOneMemOperand() ||
2796 (*MI->memoperands_begin())->getAlignment() < 8)
2800 // Assume the worst.
2808 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
2809 const MCInstrDesc &DefMCID,
2811 unsigned DefIdx, unsigned DefAlign) const {
2812 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2814 // Def is the address writeback.
2815 return ItinData->getOperandCycle(DefClass, DefIdx);
2818 if (Subtarget.isCortexA8()) {
2819 // (regno / 2) + (regno % 2) + 1
2820 DefCycle = RegNo / 2 + 1;
2823 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2825 bool isSLoad = false;
2827 switch (DefMCID.getOpcode()) {
2830 case ARM::VLDMSIA_UPD:
2831 case ARM::VLDMSDB_UPD:
2836 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2837 // then it takes an extra cycle.
2838 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
2841 // Assume the worst.
2842 DefCycle = RegNo + 2;
2849 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
2850 const MCInstrDesc &DefMCID,
2852 unsigned DefIdx, unsigned DefAlign) const {
2853 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2855 // Def is the address writeback.
2856 return ItinData->getOperandCycle(DefClass, DefIdx);
2859 if (Subtarget.isCortexA8()) {
2860 // 4 registers would be issued: 1, 2, 1.
2861 // 5 registers would be issued: 1, 2, 2.
2862 DefCycle = RegNo / 2;
2865 // Result latency is issue cycle + 2: E2.
2867 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2868 DefCycle = (RegNo / 2);
2869 // If there are odd number of registers or if it's not 64-bit aligned,
2870 // then it takes an extra AGU (Address Generation Unit) cycle.
2871 if ((RegNo % 2) || DefAlign < 8)
2873 // Result latency is AGU cycles + 2.
2876 // Assume the worst.
2877 DefCycle = RegNo + 2;
2884 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
2885 const MCInstrDesc &UseMCID,
2887 unsigned UseIdx, unsigned UseAlign) const {
2888 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2890 return ItinData->getOperandCycle(UseClass, UseIdx);
2893 if (Subtarget.isCortexA8()) {
2894 // (regno / 2) + (regno % 2) + 1
2895 UseCycle = RegNo / 2 + 1;
2898 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2900 bool isSStore = false;
2902 switch (UseMCID.getOpcode()) {
2905 case ARM::VSTMSIA_UPD:
2906 case ARM::VSTMSDB_UPD:
2911 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2912 // then it takes an extra cycle.
2913 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2916 // Assume the worst.
2917 UseCycle = RegNo + 2;
2924 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2925 const MCInstrDesc &UseMCID,
2927 unsigned UseIdx, unsigned UseAlign) const {
2928 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2930 return ItinData->getOperandCycle(UseClass, UseIdx);
2933 if (Subtarget.isCortexA8()) {
2934 UseCycle = RegNo / 2;
2939 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2940 UseCycle = (RegNo / 2);
2941 // If there are odd number of registers or if it's not 64-bit aligned,
2942 // then it takes an extra AGU (Address Generation Unit) cycle.
2943 if ((RegNo % 2) || UseAlign < 8)
2946 // Assume the worst.
2953 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2954 const MCInstrDesc &DefMCID,
2955 unsigned DefIdx, unsigned DefAlign,
2956 const MCInstrDesc &UseMCID,
2957 unsigned UseIdx, unsigned UseAlign) const {
2958 unsigned DefClass = DefMCID.getSchedClass();
2959 unsigned UseClass = UseMCID.getSchedClass();
2961 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2962 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2964 // This may be a def / use of a variable_ops instruction, the operand
2965 // latency might be determinable dynamically. Let the target try to
2968 bool LdmBypass = false;
2969 switch (DefMCID.getOpcode()) {
2971 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2975 case ARM::VLDMDIA_UPD:
2976 case ARM::VLDMDDB_UPD:
2978 case ARM::VLDMSIA_UPD:
2979 case ARM::VLDMSDB_UPD:
2980 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2983 case ARM::LDMIA_RET:
2988 case ARM::LDMIA_UPD:
2989 case ARM::LDMDA_UPD:
2990 case ARM::LDMDB_UPD:
2991 case ARM::LDMIB_UPD:
2993 case ARM::tLDMIA_UPD:
2995 case ARM::t2LDMIA_RET:
2998 case ARM::t2LDMIA_UPD:
2999 case ARM::t2LDMDB_UPD:
3001 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
3006 // We can't seem to determine the result latency of the def, assume it's 2.
3010 switch (UseMCID.getOpcode()) {
3012 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
3016 case ARM::VSTMDIA_UPD:
3017 case ARM::VSTMDDB_UPD:
3019 case ARM::VSTMSIA_UPD:
3020 case ARM::VSTMSDB_UPD:
3021 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3028 case ARM::STMIA_UPD:
3029 case ARM::STMDA_UPD:
3030 case ARM::STMDB_UPD:
3031 case ARM::STMIB_UPD:
3032 case ARM::tSTMIA_UPD:
3037 case ARM::t2STMIA_UPD:
3038 case ARM::t2STMDB_UPD:
3039 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3044 // Assume it's read in the first stage.
3047 UseCycle = DefCycle - UseCycle + 1;
3050 // It's a variable_ops instruction so we can't use DefIdx here. Just use
3051 // first def operand.
3052 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
3055 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
3056 UseClass, UseIdx)) {
3064 static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
3065 const MachineInstr *MI, unsigned Reg,
3066 unsigned &DefIdx, unsigned &Dist) {
3069 MachineBasicBlock::const_iterator I = MI; ++I;
3070 MachineBasicBlock::const_instr_iterator II =
3071 llvm::prior(I.getInstrIterator());
3072 assert(II->isInsideBundle() && "Empty bundle?");
3075 while (II->isInsideBundle()) {
3076 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
3083 assert(Idx != -1 && "Cannot find bundled definition!");
3088 static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
3089 const MachineInstr *MI, unsigned Reg,
3090 unsigned &UseIdx, unsigned &Dist) {
3093 MachineBasicBlock::const_instr_iterator II = MI; ++II;
3094 assert(II->isInsideBundle() && "Empty bundle?");
3095 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3097 // FIXME: This doesn't properly handle multiple uses.
3099 while (II != E && II->isInsideBundle()) {
3100 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
3103 if (II->getOpcode() != ARM::t2IT)
3117 /// Return the number of cycles to add to (or subtract from) the static
3118 /// itinerary based on the def opcode and alignment. The caller will ensure that
3119 /// adjusted latency is at least one cycle.
3120 static int adjustDefLatency(const ARMSubtarget &Subtarget,
3121 const MachineInstr *DefMI,
3122 const MCInstrDesc *DefMCID, unsigned DefAlign) {
3124 if (Subtarget.isCortexA8() || Subtarget.isLikeA9()) {
3125 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3126 // variants are one cycle cheaper.
3127 switch (DefMCID->getOpcode()) {
3131 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3132 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3134 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3141 case ARM::t2LDRSHs: {
3142 // Thumb2 mode: lsl only.
3143 unsigned ShAmt = DefMI->getOperand(3).getImm();
3144 if (ShAmt == 0 || ShAmt == 2)
3149 } else if (Subtarget.isSwift()) {
3150 // FIXME: Properly handle all of the latency adjustments for address
3152 switch (DefMCID->getOpcode()) {
3156 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3157 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
3158 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3161 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3162 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
3165 ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3172 case ARM::t2LDRSHs: {
3173 // Thumb2 mode: lsl only.
3174 unsigned ShAmt = DefMI->getOperand(3).getImm();
3175 if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3)
3182 if (DefAlign < 8 && Subtarget.isLikeA9()) {
3183 switch (DefMCID->getOpcode()) {
3189 case ARM::VLD1q8wb_fixed:
3190 case ARM::VLD1q16wb_fixed:
3191 case ARM::VLD1q32wb_fixed:
3192 case ARM::VLD1q64wb_fixed:
3193 case ARM::VLD1q8wb_register:
3194 case ARM::VLD1q16wb_register:
3195 case ARM::VLD1q32wb_register:
3196 case ARM::VLD1q64wb_register:
3203 case ARM::VLD2d8wb_fixed:
3204 case ARM::VLD2d16wb_fixed:
3205 case ARM::VLD2d32wb_fixed:
3206 case ARM::VLD2q8wb_fixed:
3207 case ARM::VLD2q16wb_fixed:
3208 case ARM::VLD2q32wb_fixed:
3209 case ARM::VLD2d8wb_register:
3210 case ARM::VLD2d16wb_register:
3211 case ARM::VLD2d32wb_register:
3212 case ARM::VLD2q8wb_register:
3213 case ARM::VLD2q16wb_register:
3214 case ARM::VLD2q32wb_register:
3219 case ARM::VLD3d8_UPD:
3220 case ARM::VLD3d16_UPD:
3221 case ARM::VLD3d32_UPD:
3222 case ARM::VLD1d64Twb_fixed:
3223 case ARM::VLD1d64Twb_register:
3224 case ARM::VLD3q8_UPD:
3225 case ARM::VLD3q16_UPD:
3226 case ARM::VLD3q32_UPD:
3231 case ARM::VLD4d8_UPD:
3232 case ARM::VLD4d16_UPD:
3233 case ARM::VLD4d32_UPD:
3234 case ARM::VLD1d64Qwb_fixed:
3235 case ARM::VLD1d64Qwb_register:
3236 case ARM::VLD4q8_UPD:
3237 case ARM::VLD4q16_UPD:
3238 case ARM::VLD4q32_UPD:
3239 case ARM::VLD1DUPq8:
3240 case ARM::VLD1DUPq16:
3241 case ARM::VLD1DUPq32:
3242 case ARM::VLD1DUPq8wb_fixed:
3243 case ARM::VLD1DUPq16wb_fixed:
3244 case ARM::VLD1DUPq32wb_fixed:
3245 case ARM::VLD1DUPq8wb_register:
3246 case ARM::VLD1DUPq16wb_register:
3247 case ARM::VLD1DUPq32wb_register:
3248 case ARM::VLD2DUPd8:
3249 case ARM::VLD2DUPd16:
3250 case ARM::VLD2DUPd32:
3251 case ARM::VLD2DUPd8wb_fixed:
3252 case ARM::VLD2DUPd16wb_fixed:
3253 case ARM::VLD2DUPd32wb_fixed:
3254 case ARM::VLD2DUPd8wb_register:
3255 case ARM::VLD2DUPd16wb_register:
3256 case ARM::VLD2DUPd32wb_register:
3257 case ARM::VLD4DUPd8:
3258 case ARM::VLD4DUPd16:
3259 case ARM::VLD4DUPd32:
3260 case ARM::VLD4DUPd8_UPD:
3261 case ARM::VLD4DUPd16_UPD:
3262 case ARM::VLD4DUPd32_UPD:
3264 case ARM::VLD1LNd16:
3265 case ARM::VLD1LNd32:
3266 case ARM::VLD1LNd8_UPD:
3267 case ARM::VLD1LNd16_UPD:
3268 case ARM::VLD1LNd32_UPD:
3270 case ARM::VLD2LNd16:
3271 case ARM::VLD2LNd32:
3272 case ARM::VLD2LNq16:
3273 case ARM::VLD2LNq32:
3274 case ARM::VLD2LNd8_UPD:
3275 case ARM::VLD2LNd16_UPD:
3276 case ARM::VLD2LNd32_UPD:
3277 case ARM::VLD2LNq16_UPD:
3278 case ARM::VLD2LNq32_UPD:
3280 case ARM::VLD4LNd16:
3281 case ARM::VLD4LNd32:
3282 case ARM::VLD4LNq16:
3283 case ARM::VLD4LNq32:
3284 case ARM::VLD4LNd8_UPD:
3285 case ARM::VLD4LNd16_UPD:
3286 case ARM::VLD4LNd32_UPD:
3287 case ARM::VLD4LNq16_UPD:
3288 case ARM::VLD4LNq32_UPD:
3289 // If the address is not 64-bit aligned, the latencies of these
3290 // instructions increases by one.
3301 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3302 const MachineInstr *DefMI, unsigned DefIdx,
3303 const MachineInstr *UseMI,
3304 unsigned UseIdx) const {
3305 // No operand latency. The caller may fall back to getInstrLatency.
3306 if (!ItinData || ItinData->isEmpty())
3309 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
3310 unsigned Reg = DefMO.getReg();
3311 const MCInstrDesc *DefMCID = &DefMI->getDesc();
3312 const MCInstrDesc *UseMCID = &UseMI->getDesc();
3314 unsigned DefAdj = 0;
3315 if (DefMI->isBundle()) {
3316 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
3317 DefMCID = &DefMI->getDesc();
3319 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
3320 DefMI->isRegSequence() || DefMI->isImplicitDef()) {
3324 unsigned UseAdj = 0;
3325 if (UseMI->isBundle()) {
3327 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
3328 Reg, NewUseIdx, UseAdj);
3334 UseMCID = &UseMI->getDesc();
3337 if (Reg == ARM::CPSR) {
3338 if (DefMI->getOpcode() == ARM::FMSTAT) {
3339 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
3340 return Subtarget.isLikeA9() ? 1 : 20;
3343 // CPSR set and branch can be paired in the same cycle.
3344 if (UseMI->isBranch())
3347 // Otherwise it takes the instruction latency (generally one).
3348 unsigned Latency = getInstrLatency(ItinData, DefMI);
3350 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
3351 // its uses. Instructions which are otherwise scheduled between them may
3352 // incur a code size penalty (not able to use the CPSR setting 16-bit
3354 if (Latency > 0 && Subtarget.isThumb2()) {
3355 const MachineFunction *MF = DefMI->getParent()->getParent();
3356 if (MF->getFunction()->getAttributes().
3357 hasAttribute(AttributeSet::FunctionIndex,
3358 Attribute::OptimizeForSize))
3364 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
3367 unsigned DefAlign = DefMI->hasOneMemOperand()
3368 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
3369 unsigned UseAlign = UseMI->hasOneMemOperand()
3370 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
3372 // Get the itinerary's latency if possible, and handle variable_ops.
3373 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
3374 *UseMCID, UseIdx, UseAlign);
3375 // Unable to find operand latency. The caller may resort to getInstrLatency.
3379 // Adjust for IT block position.
3380 int Adj = DefAdj + UseAdj;
3382 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3383 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
3384 if (Adj >= 0 || (int)Latency > -Adj) {
3385 return Latency + Adj;
3387 // Return the itinerary latency, which may be zero but not less than zero.
3392 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3393 SDNode *DefNode, unsigned DefIdx,
3394 SDNode *UseNode, unsigned UseIdx) const {
3395 if (!DefNode->isMachineOpcode())
3398 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
3400 if (isZeroCost(DefMCID.Opcode))
3403 if (!ItinData || ItinData->isEmpty())
3404 return DefMCID.mayLoad() ? 3 : 1;
3406 if (!UseNode->isMachineOpcode()) {
3407 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
3408 if (Subtarget.isLikeA9() || Subtarget.isSwift())
3409 return Latency <= 2 ? 1 : Latency - 1;
3411 return Latency <= 3 ? 1 : Latency - 2;
3414 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
3415 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
3416 unsigned DefAlign = !DefMN->memoperands_empty()
3417 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
3418 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
3419 unsigned UseAlign = !UseMN->memoperands_empty()
3420 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
3421 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
3422 UseMCID, UseIdx, UseAlign);
3425 (Subtarget.isCortexA8() || Subtarget.isLikeA9())) {
3426 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3427 // variants are one cycle cheaper.
3428 switch (DefMCID.getOpcode()) {
3433 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
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.
3446 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3447 if (ShAmt == 0 || ShAmt == 2)
3452 } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) {
3453 // FIXME: Properly handle all of the latency adjustments for address
3455 switch (DefMCID.getOpcode()) {
3460 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3461 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3463 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3464 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3466 else if (ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3473 case ARM::t2LDRSHs: {
3474 // Thumb2 mode: lsl 0-3 only.
3481 if (DefAlign < 8 && Subtarget.isLikeA9())
3482 switch (DefMCID.getOpcode()) {
3488 case ARM::VLD1q8wb_register:
3489 case ARM::VLD1q16wb_register:
3490 case ARM::VLD1q32wb_register:
3491 case ARM::VLD1q64wb_register:
3492 case ARM::VLD1q8wb_fixed:
3493 case ARM::VLD1q16wb_fixed:
3494 case ARM::VLD1q32wb_fixed:
3495 case ARM::VLD1q64wb_fixed:
3499 case ARM::VLD2q8Pseudo:
3500 case ARM::VLD2q16Pseudo:
3501 case ARM::VLD2q32Pseudo:
3502 case ARM::VLD2d8wb_fixed:
3503 case ARM::VLD2d16wb_fixed:
3504 case ARM::VLD2d32wb_fixed:
3505 case ARM::VLD2q8PseudoWB_fixed:
3506 case ARM::VLD2q16PseudoWB_fixed:
3507 case ARM::VLD2q32PseudoWB_fixed:
3508 case ARM::VLD2d8wb_register:
3509 case ARM::VLD2d16wb_register:
3510 case ARM::VLD2d32wb_register:
3511 case ARM::VLD2q8PseudoWB_register:
3512 case ARM::VLD2q16PseudoWB_register:
3513 case ARM::VLD2q32PseudoWB_register:
3514 case ARM::VLD3d8Pseudo:
3515 case ARM::VLD3d16Pseudo:
3516 case ARM::VLD3d32Pseudo:
3517 case ARM::VLD1d64TPseudo:
3518 case ARM::VLD3d8Pseudo_UPD:
3519 case ARM::VLD3d16Pseudo_UPD:
3520 case ARM::VLD3d32Pseudo_UPD:
3521 case ARM::VLD3q8Pseudo_UPD:
3522 case ARM::VLD3q16Pseudo_UPD:
3523 case ARM::VLD3q32Pseudo_UPD:
3524 case ARM::VLD3q8oddPseudo:
3525 case ARM::VLD3q16oddPseudo:
3526 case ARM::VLD3q32oddPseudo:
3527 case ARM::VLD3q8oddPseudo_UPD:
3528 case ARM::VLD3q16oddPseudo_UPD:
3529 case ARM::VLD3q32oddPseudo_UPD:
3530 case ARM::VLD4d8Pseudo:
3531 case ARM::VLD4d16Pseudo:
3532 case ARM::VLD4d32Pseudo:
3533 case ARM::VLD1d64QPseudo:
3534 case ARM::VLD4d8Pseudo_UPD:
3535 case ARM::VLD4d16Pseudo_UPD:
3536 case ARM::VLD4d32Pseudo_UPD:
3537 case ARM::VLD4q8Pseudo_UPD:
3538 case ARM::VLD4q16Pseudo_UPD:
3539 case ARM::VLD4q32Pseudo_UPD:
3540 case ARM::VLD4q8oddPseudo:
3541 case ARM::VLD4q16oddPseudo:
3542 case ARM::VLD4q32oddPseudo:
3543 case ARM::VLD4q8oddPseudo_UPD:
3544 case ARM::VLD4q16oddPseudo_UPD:
3545 case ARM::VLD4q32oddPseudo_UPD:
3546 case ARM::VLD1DUPq8:
3547 case ARM::VLD1DUPq16:
3548 case ARM::VLD1DUPq32:
3549 case ARM::VLD1DUPq8wb_fixed:
3550 case ARM::VLD1DUPq16wb_fixed:
3551 case ARM::VLD1DUPq32wb_fixed:
3552 case ARM::VLD1DUPq8wb_register:
3553 case ARM::VLD1DUPq16wb_register:
3554 case ARM::VLD1DUPq32wb_register:
3555 case ARM::VLD2DUPd8:
3556 case ARM::VLD2DUPd16:
3557 case ARM::VLD2DUPd32:
3558 case ARM::VLD2DUPd8wb_fixed:
3559 case ARM::VLD2DUPd16wb_fixed:
3560 case ARM::VLD2DUPd32wb_fixed:
3561 case ARM::VLD2DUPd8wb_register:
3562 case ARM::VLD2DUPd16wb_register:
3563 case ARM::VLD2DUPd32wb_register:
3564 case ARM::VLD4DUPd8Pseudo:
3565 case ARM::VLD4DUPd16Pseudo:
3566 case ARM::VLD4DUPd32Pseudo:
3567 case ARM::VLD4DUPd8Pseudo_UPD:
3568 case ARM::VLD4DUPd16Pseudo_UPD:
3569 case ARM::VLD4DUPd32Pseudo_UPD:
3570 case ARM::VLD1LNq8Pseudo:
3571 case ARM::VLD1LNq16Pseudo:
3572 case ARM::VLD1LNq32Pseudo:
3573 case ARM::VLD1LNq8Pseudo_UPD:
3574 case ARM::VLD1LNq16Pseudo_UPD:
3575 case ARM::VLD1LNq32Pseudo_UPD:
3576 case ARM::VLD2LNd8Pseudo:
3577 case ARM::VLD2LNd16Pseudo:
3578 case ARM::VLD2LNd32Pseudo:
3579 case ARM::VLD2LNq16Pseudo:
3580 case ARM::VLD2LNq32Pseudo:
3581 case ARM::VLD2LNd8Pseudo_UPD:
3582 case ARM::VLD2LNd16Pseudo_UPD:
3583 case ARM::VLD2LNd32Pseudo_UPD:
3584 case ARM::VLD2LNq16Pseudo_UPD:
3585 case ARM::VLD2LNq32Pseudo_UPD:
3586 case ARM::VLD4LNd8Pseudo:
3587 case ARM::VLD4LNd16Pseudo:
3588 case ARM::VLD4LNd32Pseudo:
3589 case ARM::VLD4LNq16Pseudo:
3590 case ARM::VLD4LNq32Pseudo:
3591 case ARM::VLD4LNd8Pseudo_UPD:
3592 case ARM::VLD4LNd16Pseudo_UPD:
3593 case ARM::VLD4LNd32Pseudo_UPD:
3594 case ARM::VLD4LNq16Pseudo_UPD:
3595 case ARM::VLD4LNq32Pseudo_UPD:
3596 // If the address is not 64-bit aligned, the latencies of these
3597 // instructions increases by one.
3605 unsigned ARMBaseInstrInfo::getPredicationCost(const MachineInstr *MI) const {
3606 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3607 MI->isRegSequence() || MI->isImplicitDef())
3613 const MCInstrDesc &MCID = MI->getDesc();
3615 if (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR)) {
3616 // When predicated, CPSR is an additional source operand for CPSR updating
3617 // instructions, this apparently increases their latencies.
3623 unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3624 const MachineInstr *MI,
3625 unsigned *PredCost) const {
3626 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3627 MI->isRegSequence() || MI->isImplicitDef())
3630 // An instruction scheduler typically runs on unbundled instructions, however
3631 // other passes may query the latency of a bundled instruction.
3632 if (MI->isBundle()) {
3633 unsigned Latency = 0;
3634 MachineBasicBlock::const_instr_iterator I = MI;
3635 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3636 while (++I != E && I->isInsideBundle()) {
3637 if (I->getOpcode() != ARM::t2IT)
3638 Latency += getInstrLatency(ItinData, I, PredCost);
3643 const MCInstrDesc &MCID = MI->getDesc();
3644 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) {
3645 // When predicated, CPSR is an additional source operand for CPSR updating
3646 // instructions, this apparently increases their latencies.
3649 // Be sure to call getStageLatency for an empty itinerary in case it has a
3650 // valid MinLatency property.
3652 return MI->mayLoad() ? 3 : 1;
3654 unsigned Class = MCID.getSchedClass();
3656 // For instructions with variable uops, use uops as latency.
3657 if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0)
3658 return getNumMicroOps(ItinData, MI);
3660 // For the common case, fall back on the itinerary's latency.
3661 unsigned Latency = ItinData->getStageLatency(Class);
3663 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3664 unsigned DefAlign = MI->hasOneMemOperand()
3665 ? (*MI->memoperands_begin())->getAlignment() : 0;
3666 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
3667 if (Adj >= 0 || (int)Latency > -Adj) {
3668 return Latency + Adj;
3673 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3674 SDNode *Node) const {
3675 if (!Node->isMachineOpcode())
3678 if (!ItinData || ItinData->isEmpty())
3681 unsigned Opcode = Node->getMachineOpcode();
3684 return ItinData->getStageLatency(get(Opcode).getSchedClass());
3691 bool ARMBaseInstrInfo::
3692 hasHighOperandLatency(const InstrItineraryData *ItinData,
3693 const MachineRegisterInfo *MRI,
3694 const MachineInstr *DefMI, unsigned DefIdx,
3695 const MachineInstr *UseMI, unsigned UseIdx) const {
3696 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3697 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
3698 if (Subtarget.isCortexA8() &&
3699 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
3700 // CortexA8 VFP instructions are not pipelined.
3703 // Hoist VFP / NEON instructions with 4 or higher latency.
3704 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
3706 Latency = getInstrLatency(ItinData, DefMI);
3709 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
3710 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
3713 bool ARMBaseInstrInfo::
3714 hasLowDefLatency(const InstrItineraryData *ItinData,
3715 const MachineInstr *DefMI, unsigned DefIdx) const {
3716 if (!ItinData || ItinData->isEmpty())
3719 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3720 if (DDomain == ARMII::DomainGeneral) {
3721 unsigned DefClass = DefMI->getDesc().getSchedClass();
3722 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3723 return (DefCycle != -1 && DefCycle <= 2);
3728 bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
3729 StringRef &ErrInfo) const {
3730 if (convertAddSubFlagsOpcode(MI->getOpcode())) {
3731 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
3738 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
3739 unsigned &AddSubOpc,
3740 bool &NegAcc, bool &HasLane) const {
3741 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
3742 if (I == MLxEntryMap.end())
3745 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
3746 MulOpc = Entry.MulOpc;
3747 AddSubOpc = Entry.AddSubOpc;
3748 NegAcc = Entry.NegAcc;
3749 HasLane = Entry.HasLane;
3753 //===----------------------------------------------------------------------===//
3754 // Execution domains.
3755 //===----------------------------------------------------------------------===//
3757 // Some instructions go down the NEON pipeline, some go down the VFP pipeline,
3758 // and some can go down both. The vmov instructions go down the VFP pipeline,
3759 // but they can be changed to vorr equivalents that are executed by the NEON
3762 // We use the following execution domain numbering:
3770 // Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
3772 std::pair<uint16_t, uint16_t>
3773 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
3774 // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
3775 // if they are not predicated.
3776 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
3777 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3779 // CortexA9 is particularly picky about mixing the two and wants these
3781 if (Subtarget.isCortexA9() && !isPredicated(MI) &&
3782 (MI->getOpcode() == ARM::VMOVRS ||
3783 MI->getOpcode() == ARM::VMOVSR ||
3784 MI->getOpcode() == ARM::VMOVS))
3785 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3787 // No other instructions can be swizzled, so just determine their domain.
3788 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
3790 if (Domain & ARMII::DomainNEON)
3791 return std::make_pair(ExeNEON, 0);
3793 // Certain instructions can go either way on Cortex-A8.
3794 // Treat them as NEON instructions.
3795 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
3796 return std::make_pair(ExeNEON, 0);
3798 if (Domain & ARMII::DomainVFP)
3799 return std::make_pair(ExeVFP, 0);
3801 return std::make_pair(ExeGeneric, 0);
3804 static unsigned getCorrespondingDRegAndLane(const TargetRegisterInfo *TRI,
3805 unsigned SReg, unsigned &Lane) {
3806 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_0, &ARM::DPRRegClass);
3809 if (DReg != ARM::NoRegister)
3813 DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1, &ARM::DPRRegClass);
3815 assert(DReg && "S-register with no D super-register?");
3819 /// getImplicitSPRUseForDPRUse - Given a use of a DPR register and lane,
3820 /// set ImplicitSReg to a register number that must be marked as implicit-use or
3821 /// zero if no register needs to be defined as implicit-use.
3823 /// If the function cannot determine if an SPR should be marked implicit use or
3824 /// not, it returns false.
3826 /// This function handles cases where an instruction is being modified from taking
3827 /// an SPR to a DPR[Lane]. A use of the DPR is being added, which may conflict
3828 /// with an earlier def of an SPR corresponding to DPR[Lane^1] (i.e. the other
3829 /// lane of the DPR).
3831 /// If the other SPR is defined, an implicit-use of it should be added. Else,
3832 /// (including the case where the DPR itself is defined), it should not.
3834 static bool getImplicitSPRUseForDPRUse(const TargetRegisterInfo *TRI,
3836 unsigned DReg, unsigned Lane,
3837 unsigned &ImplicitSReg) {
3838 // If the DPR is defined or used already, the other SPR lane will be chained
3839 // correctly, so there is nothing to be done.
3840 if (MI->definesRegister(DReg, TRI) || MI->readsRegister(DReg, TRI)) {
3845 // Otherwise we need to go searching to see if the SPR is set explicitly.
3846 ImplicitSReg = TRI->getSubReg(DReg,
3847 (Lane & 1) ? ARM::ssub_0 : ARM::ssub_1);
3848 MachineBasicBlock::LivenessQueryResult LQR =
3849 MI->getParent()->computeRegisterLiveness(TRI, ImplicitSReg, MI);
3851 if (LQR == MachineBasicBlock::LQR_Live)
3853 else if (LQR == MachineBasicBlock::LQR_Unknown)
3856 // If the register is known not to be live, there is no need to add an
3863 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
3864 unsigned DstReg, SrcReg, DReg;
3866 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3867 const TargetRegisterInfo *TRI = &getRegisterInfo();
3868 switch (MI->getOpcode()) {
3870 llvm_unreachable("cannot handle opcode!");
3873 if (Domain != ExeNEON)
3876 // Zap the predicate operands.
3877 assert(!isPredicated(MI) && "Cannot predicate a VORRd");
3879 // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
3880 DstReg = MI->getOperand(0).getReg();
3881 SrcReg = MI->getOperand(1).getReg();
3883 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3884 MI->RemoveOperand(i-1);
3886 // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits)
3887 MI->setDesc(get(ARM::VORRd));
3888 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3893 if (Domain != ExeNEON)
3895 assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
3897 // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits)
3898 DstReg = MI->getOperand(0).getReg();
3899 SrcReg = MI->getOperand(1).getReg();
3901 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3902 MI->RemoveOperand(i-1);
3904 DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane);
3906 // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps)
3907 // Note that DSrc has been widened and the other lane may be undef, which
3908 // contaminates the entire register.
3909 MI->setDesc(get(ARM::VGETLNi32));
3910 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3911 .addReg(DReg, RegState::Undef)
3914 // The old source should be an implicit use, otherwise we might think it
3915 // was dead before here.
3916 MIB.addReg(SrcReg, RegState::Implicit);
3919 if (Domain != ExeNEON)
3921 assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
3923 // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits)
3924 DstReg = MI->getOperand(0).getReg();
3925 SrcReg = MI->getOperand(1).getReg();
3927 DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane);
3929 unsigned ImplicitSReg;
3930 if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg))
3933 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3934 MI->RemoveOperand(i-1);
3936 // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps)
3937 // Again DDst may be undefined at the beginning of this instruction.
3938 MI->setDesc(get(ARM::VSETLNi32));
3939 MIB.addReg(DReg, RegState::Define)
3940 .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI)))
3943 AddDefaultPred(MIB);
3945 // The narrower destination must be marked as set to keep previous chains
3947 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
3948 if (ImplicitSReg != 0)
3949 MIB.addReg(ImplicitSReg, RegState::Implicit);
3953 if (Domain != ExeNEON)
3956 // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits)
3957 DstReg = MI->getOperand(0).getReg();
3958 SrcReg = MI->getOperand(1).getReg();
3960 unsigned DstLane = 0, SrcLane = 0, DDst, DSrc;
3961 DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane);
3962 DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane);
3964 unsigned ImplicitSReg;
3965 if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg))
3968 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3969 MI->RemoveOperand(i-1);
3972 // Destination can be:
3973 // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits)
3974 MI->setDesc(get(ARM::VDUPLN32d));
3975 MIB.addReg(DDst, RegState::Define)
3976 .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI)))
3978 AddDefaultPred(MIB);
3980 // Neither the source or the destination are naturally represented any
3981 // more, so add them in manually.
3982 MIB.addReg(DstReg, RegState::Implicit | RegState::Define);
3983 MIB.addReg(SrcReg, RegState::Implicit);
3984 if (ImplicitSReg != 0)
3985 MIB.addReg(ImplicitSReg, RegState::Implicit);
3989 // In general there's no single instruction that can perform an S <-> S
3990 // move in NEON space, but a pair of VEXT instructions *can* do the
3991 // job. It turns out that the VEXTs needed will only use DSrc once, with
3992 // the position based purely on the combination of lane-0 and lane-1
3993 // involved. For example
3994 // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1
3995 // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1
3996 // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1
3997 // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1
3999 // Pattern of the MachineInstrs is:
4000 // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits)
4001 MachineInstrBuilder NewMIB;
4002 NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4003 get(ARM::VEXTd32), DDst);
4005 // On the first instruction, both DSrc and DDst may be <undef> if present.
4006 // Specifically when the original instruction didn't have them as an
4008 unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst;
4009 bool CurUndef = !MI->readsRegister(CurReg, TRI);
4010 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
4012 CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst;
4013 CurUndef = !MI->readsRegister(CurReg, TRI);
4014 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
4017 AddDefaultPred(NewMIB);
4019 if (SrcLane == DstLane)
4020 NewMIB.addReg(SrcReg, RegState::Implicit);
4022 MI->setDesc(get(ARM::VEXTd32));
4023 MIB.addReg(DDst, RegState::Define);
4025 // On the second instruction, DDst has definitely been defined above, so
4026 // it is not <undef>. DSrc, if present, can be <undef> as above.
4027 CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst;
4028 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4029 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4031 CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst;
4032 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4033 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4036 AddDefaultPred(MIB);
4038 if (SrcLane != DstLane)
4039 MIB.addReg(SrcReg, RegState::Implicit);
4041 // As before, the original destination is no longer represented, add it
4043 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4044 if (ImplicitSReg != 0)
4045 MIB.addReg(ImplicitSReg, RegState::Implicit);
4052 //===----------------------------------------------------------------------===//
4053 // Partial register updates
4054 //===----------------------------------------------------------------------===//
4056 // Swift renames NEON registers with 64-bit granularity. That means any
4057 // instruction writing an S-reg implicitly reads the containing D-reg. The
4058 // problem is mostly avoided by translating f32 operations to v2f32 operations
4059 // on D-registers, but f32 loads are still a problem.
4061 // These instructions can load an f32 into a NEON register:
4063 // VLDRS - Only writes S, partial D update.
4064 // VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops.
4065 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
4067 // FCONSTD can be used as a dependency-breaking instruction.
4068 unsigned ARMBaseInstrInfo::
4069 getPartialRegUpdateClearance(const MachineInstr *MI,
4071 const TargetRegisterInfo *TRI) const {
4072 if (!SwiftPartialUpdateClearance ||
4073 !(Subtarget.isSwift() || Subtarget.isCortexA15()))
4076 assert(TRI && "Need TRI instance");
4078 const MachineOperand &MO = MI->getOperand(OpNum);
4081 unsigned Reg = MO.getReg();
4084 switch(MI->getOpcode()) {
4085 // Normal instructions writing only an S-register.
4090 case ARM::VMOVv4i16:
4091 case ARM::VMOVv2i32:
4092 case ARM::VMOVv2f32:
4093 case ARM::VMOVv1i64:
4094 UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI);
4097 // Explicitly reads the dependency.
4098 case ARM::VLD1LNd32:
4105 // If this instruction actually reads a value from Reg, there is no unwanted
4107 if (UseOp != -1 && MI->getOperand(UseOp).readsReg())
4110 // We must be able to clobber the whole D-reg.
4111 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4112 // Virtual register must be a foo:ssub_0<def,undef> operand.
4113 if (!MO.getSubReg() || MI->readsVirtualRegister(Reg))
4115 } else if (ARM::SPRRegClass.contains(Reg)) {
4116 // Physical register: MI must define the full D-reg.
4117 unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0,
4119 if (!DReg || !MI->definesRegister(DReg, TRI))
4123 // MI has an unwanted D-register dependency.
4124 // Avoid defs in the previous N instructrions.
4125 return SwiftPartialUpdateClearance;
4128 // Break a partial register dependency after getPartialRegUpdateClearance
4129 // returned non-zero.
4130 void ARMBaseInstrInfo::
4131 breakPartialRegDependency(MachineBasicBlock::iterator MI,
4133 const TargetRegisterInfo *TRI) const {
4134 assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def");
4135 assert(TRI && "Need TRI instance");
4137 const MachineOperand &MO = MI->getOperand(OpNum);
4138 unsigned Reg = MO.getReg();
4139 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
4140 "Can't break virtual register dependencies.");
4141 unsigned DReg = Reg;
4143 // If MI defines an S-reg, find the corresponding D super-register.
4144 if (ARM::SPRRegClass.contains(Reg)) {
4145 DReg = ARM::D0 + (Reg - ARM::S0) / 2;
4146 assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken");
4149 assert(ARM::DPRRegClass.contains(DReg) && "Can only break D-reg deps");
4150 assert(MI->definesRegister(DReg, TRI) && "MI doesn't clobber full D-reg");
4152 // FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines
4153 // the full D-register by loading the same value to both lanes. The
4154 // instruction is micro-coded with 2 uops, so don't do this until we can
4155 // properly schedule micro-coded instructions. The dispatcher stalls cause
4156 // too big regressions.
4158 // Insert the dependency-breaking FCONSTD before MI.
4159 // 96 is the encoding of 0.5, but the actual value doesn't matter here.
4160 AddDefaultPred(BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4161 get(ARM::FCONSTD), DReg).addImm(96));
4162 MI->addRegisterKilled(DReg, TRI, true);
4165 bool ARMBaseInstrInfo::hasNOP() const {
4166 return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
4169 bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
4170 if (MI->getNumOperands() < 4)
4172 unsigned ShOpVal = MI->getOperand(3).getImm();
4173 unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
4174 // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
4175 if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) ||
4176 ((ShImm == 1 || ShImm == 2) &&
4177 ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl))