1 //===-- ARMBaseInstrInfo.cpp - ARM Instruction Information ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the Base ARM implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "ARMBaseInstrInfo.h"
16 #include "ARMBaseRegisterInfo.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMHazardRecognizer.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "MCTargetDesc/ARMAddressingModes.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/CodeGen/LiveVariables.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineJumpTableInfo.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/SelectionDAGNodes.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/GlobalValue.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/Support/BranchProbability.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
39 #define GET_INSTRINFO_CTOR
40 #include "ARMGenInstrInfo.inc"
45 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
46 cl::desc("Enable ARM 2-addr to 3-addr conv"));
49 WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true),
50 cl::desc("Widen ARM vmovs to vmovd when possible"));
52 static cl::opt<unsigned>
53 SwiftPartialUpdateClearance("swift-partial-update-clearance",
54 cl::Hidden, cl::init(12),
55 cl::desc("Clearance before partial register updates"));
57 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
59 uint16_t MLxOpc; // MLA / MLS opcode
60 uint16_t MulOpc; // Expanded multiplication opcode
61 uint16_t AddSubOpc; // Expanded add / sub opcode
62 bool NegAcc; // True if the acc is negated before the add / sub.
63 bool HasLane; // True if instruction has an extra "lane" operand.
66 static const ARM_MLxEntry ARM_MLxTable[] = {
67 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
69 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
70 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
71 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
72 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
73 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
74 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
75 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
76 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
79 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
80 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
81 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
82 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
83 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
84 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
85 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
86 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
89 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
90 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
92 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
93 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
94 assert(false && "Duplicated entries?");
95 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
96 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
100 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
101 // currently defaults to no prepass hazard recognizer.
102 ScheduleHazardRecognizer *ARMBaseInstrInfo::
103 CreateTargetHazardRecognizer(const TargetMachine *TM,
104 const ScheduleDAG *DAG) const {
105 if (usePreRAHazardRecognizer()) {
106 const InstrItineraryData *II = TM->getInstrItineraryData();
107 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
109 return TargetInstrInfo::CreateTargetHazardRecognizer(TM, DAG);
112 ScheduleHazardRecognizer *ARMBaseInstrInfo::
113 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
114 const ScheduleDAG *DAG) const {
115 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
116 return (ScheduleHazardRecognizer *)new ARMHazardRecognizer(II, DAG);
117 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
121 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
122 MachineBasicBlock::iterator &MBBI,
123 LiveVariables *LV) const {
124 // FIXME: Thumb2 support.
129 MachineInstr *MI = MBBI;
130 MachineFunction &MF = *MI->getParent()->getParent();
131 uint64_t TSFlags = MI->getDesc().TSFlags;
133 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
134 default: return NULL;
135 case ARMII::IndexModePre:
138 case ARMII::IndexModePost:
142 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
144 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
148 MachineInstr *UpdateMI = NULL;
149 MachineInstr *MemMI = NULL;
150 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
151 const MCInstrDesc &MCID = MI->getDesc();
152 unsigned NumOps = MCID.getNumOperands();
153 bool isLoad = !MI->mayStore();
154 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
155 const MachineOperand &Base = MI->getOperand(2);
156 const MachineOperand &Offset = MI->getOperand(NumOps-3);
157 unsigned WBReg = WB.getReg();
158 unsigned BaseReg = Base.getReg();
159 unsigned OffReg = Offset.getReg();
160 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
161 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
163 default: llvm_unreachable("Unknown indexed op!");
164 case ARMII::AddrMode2: {
165 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
166 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
168 if (ARM_AM::getSOImmVal(Amt) == -1)
169 // Can't encode it in a so_imm operand. This transformation will
170 // add more than 1 instruction. Abandon!
172 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
173 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
174 .addReg(BaseReg).addImm(Amt)
175 .addImm(Pred).addReg(0).addReg(0);
176 } else if (Amt != 0) {
177 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
178 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
179 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
180 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
181 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
182 .addImm(Pred).addReg(0).addReg(0);
184 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
185 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
186 .addReg(BaseReg).addReg(OffReg)
187 .addImm(Pred).addReg(0).addReg(0);
190 case ARMII::AddrMode3 : {
191 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
192 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
194 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
195 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
196 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
197 .addReg(BaseReg).addImm(Amt)
198 .addImm(Pred).addReg(0).addReg(0);
200 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
201 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
202 .addReg(BaseReg).addReg(OffReg)
203 .addImm(Pred).addReg(0).addReg(0);
208 std::vector<MachineInstr*> NewMIs;
211 MemMI = BuildMI(MF, MI->getDebugLoc(),
212 get(MemOpc), MI->getOperand(0).getReg())
213 .addReg(WBReg).addImm(0).addImm(Pred);
215 MemMI = BuildMI(MF, MI->getDebugLoc(),
216 get(MemOpc)).addReg(MI->getOperand(1).getReg())
217 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
218 NewMIs.push_back(MemMI);
219 NewMIs.push_back(UpdateMI);
222 MemMI = BuildMI(MF, MI->getDebugLoc(),
223 get(MemOpc), MI->getOperand(0).getReg())
224 .addReg(BaseReg).addImm(0).addImm(Pred);
226 MemMI = BuildMI(MF, MI->getDebugLoc(),
227 get(MemOpc)).addReg(MI->getOperand(1).getReg())
228 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
230 UpdateMI->getOperand(0).setIsDead();
231 NewMIs.push_back(UpdateMI);
232 NewMIs.push_back(MemMI);
235 // Transfer LiveVariables states, kill / dead info.
237 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
238 MachineOperand &MO = MI->getOperand(i);
239 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
240 unsigned Reg = MO.getReg();
242 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
244 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
246 LV->addVirtualRegisterDead(Reg, NewMI);
248 if (MO.isUse() && MO.isKill()) {
249 for (unsigned j = 0; j < 2; ++j) {
250 // Look at the two new MI's in reverse order.
251 MachineInstr *NewMI = NewMIs[j];
252 if (!NewMI->readsRegister(Reg))
254 LV->addVirtualRegisterKilled(Reg, NewMI);
255 if (VI.removeKill(MI))
256 VI.Kills.push_back(NewMI);
264 MFI->insert(MBBI, NewMIs[1]);
265 MFI->insert(MBBI, NewMIs[0]);
271 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
272 MachineBasicBlock *&FBB,
273 SmallVectorImpl<MachineOperand> &Cond,
274 bool AllowModify) const {
278 MachineBasicBlock::iterator I = MBB.end();
279 if (I == MBB.begin())
280 return false; // Empty blocks are easy.
283 // Walk backwards from the end of the basic block until the branch is
284 // analyzed or we give up.
285 while (isPredicated(I) || I->isTerminator()) {
287 // Flag to be raised on unanalyzeable instructions. This is useful in cases
288 // where we want to clean up on the end of the basic block before we bail
290 bool CantAnalyze = false;
292 // Skip over DEBUG values and predicated nonterminators.
293 while (I->isDebugValue() || !I->isTerminator()) {
294 if (I == MBB.begin())
299 if (isIndirectBranchOpcode(I->getOpcode()) ||
300 isJumpTableBranchOpcode(I->getOpcode())) {
301 // Indirect branches and jump tables can't be analyzed, but we still want
302 // to clean up any instructions at the tail of the basic block.
304 } else if (isUncondBranchOpcode(I->getOpcode())) {
305 TBB = I->getOperand(0).getMBB();
306 } else if (isCondBranchOpcode(I->getOpcode())) {
307 // Bail out if we encounter multiple conditional branches.
311 assert(!FBB && "FBB should have been null.");
313 TBB = I->getOperand(0).getMBB();
314 Cond.push_back(I->getOperand(1));
315 Cond.push_back(I->getOperand(2));
316 } else if (I->isReturn()) {
317 // Returns can't be analyzed, but we should run cleanup.
318 CantAnalyze = !isPredicated(I);
320 // We encountered other unrecognized terminator. Bail out immediately.
324 // Cleanup code - to be run for unpredicated unconditional branches and
326 if (!isPredicated(I) &&
327 (isUncondBranchOpcode(I->getOpcode()) ||
328 isIndirectBranchOpcode(I->getOpcode()) ||
329 isJumpTableBranchOpcode(I->getOpcode()) ||
331 // Forget any previous condition branch information - it no longer applies.
335 // If we can modify the function, delete everything below this
336 // unconditional branch.
338 MachineBasicBlock::iterator DI = llvm::next(I);
339 while (DI != MBB.end()) {
340 MachineInstr *InstToDelete = DI;
342 InstToDelete->eraseFromParent();
350 if (I == MBB.begin())
356 // We made it past the terminators without bailing out - we must have
357 // analyzed this branch successfully.
362 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
363 MachineBasicBlock::iterator I = MBB.end();
364 if (I == MBB.begin()) return 0;
366 while (I->isDebugValue()) {
367 if (I == MBB.begin())
371 if (!isUncondBranchOpcode(I->getOpcode()) &&
372 !isCondBranchOpcode(I->getOpcode()))
375 // Remove the branch.
376 I->eraseFromParent();
380 if (I == MBB.begin()) return 1;
382 if (!isCondBranchOpcode(I->getOpcode()))
385 // Remove the branch.
386 I->eraseFromParent();
391 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
392 MachineBasicBlock *FBB,
393 const SmallVectorImpl<MachineOperand> &Cond,
395 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
396 int BOpc = !AFI->isThumbFunction()
397 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
398 int BccOpc = !AFI->isThumbFunction()
399 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
400 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
402 // Shouldn't be a fall through.
403 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
404 assert((Cond.size() == 2 || Cond.size() == 0) &&
405 "ARM branch conditions have two components!");
408 if (Cond.empty()) { // Unconditional branch?
410 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
412 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
414 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
415 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
419 // Two-way conditional branch.
420 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
421 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
423 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
425 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
429 bool ARMBaseInstrInfo::
430 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
431 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
432 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
436 bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
437 if (MI->isBundle()) {
438 MachineBasicBlock::const_instr_iterator I = MI;
439 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
440 while (++I != E && I->isInsideBundle()) {
441 int PIdx = I->findFirstPredOperandIdx();
442 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
448 int PIdx = MI->findFirstPredOperandIdx();
449 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
452 bool ARMBaseInstrInfo::
453 PredicateInstruction(MachineInstr *MI,
454 const SmallVectorImpl<MachineOperand> &Pred) const {
455 unsigned Opc = MI->getOpcode();
456 if (isUncondBranchOpcode(Opc)) {
457 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
458 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
459 .addImm(Pred[0].getImm())
460 .addReg(Pred[1].getReg());
464 int PIdx = MI->findFirstPredOperandIdx();
466 MachineOperand &PMO = MI->getOperand(PIdx);
467 PMO.setImm(Pred[0].getImm());
468 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
474 bool ARMBaseInstrInfo::
475 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
476 const SmallVectorImpl<MachineOperand> &Pred2) const {
477 if (Pred1.size() > 2 || Pred2.size() > 2)
480 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
481 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
491 return CC2 == ARMCC::HI;
493 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
495 return CC2 == ARMCC::GT;
497 return CC2 == ARMCC::LT;
501 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
502 std::vector<MachineOperand> &Pred) const {
504 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
505 const MachineOperand &MO = MI->getOperand(i);
506 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
507 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) {
516 /// isPredicable - Return true if the specified instruction can be predicated.
517 /// By default, this returns true for every instruction with a
518 /// PredicateOperand.
519 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
520 if (!MI->isPredicable())
523 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
524 ARMFunctionInfo *AFI =
525 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
526 return AFI->isThumb2Function();
531 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
532 LLVM_ATTRIBUTE_NOINLINE
533 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
535 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
537 assert(JTI < JT.size());
538 return JT[JTI].MBBs.size();
541 /// GetInstSize - Return the size of the specified MachineInstr.
543 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
544 const MachineBasicBlock &MBB = *MI->getParent();
545 const MachineFunction *MF = MBB.getParent();
546 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
548 const MCInstrDesc &MCID = MI->getDesc();
550 return MCID.getSize();
552 // If this machine instr is an inline asm, measure it.
553 if (MI->getOpcode() == ARM::INLINEASM)
554 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
557 unsigned Opc = MI->getOpcode();
559 case TargetOpcode::IMPLICIT_DEF:
560 case TargetOpcode::KILL:
561 case TargetOpcode::PROLOG_LABEL:
562 case TargetOpcode::EH_LABEL:
563 case TargetOpcode::DBG_VALUE:
565 case TargetOpcode::BUNDLE:
566 return getInstBundleLength(MI);
567 case ARM::MOVi16_ga_pcrel:
568 case ARM::MOVTi16_ga_pcrel:
569 case ARM::t2MOVi16_ga_pcrel:
570 case ARM::t2MOVTi16_ga_pcrel:
573 case ARM::t2MOVi32imm:
575 case ARM::CONSTPOOL_ENTRY:
576 // If this machine instr is a constant pool entry, its size is recorded as
578 return MI->getOperand(2).getImm();
579 case ARM::Int_eh_sjlj_longjmp:
581 case ARM::tInt_eh_sjlj_longjmp:
583 case ARM::Int_eh_sjlj_setjmp:
584 case ARM::Int_eh_sjlj_setjmp_nofp:
586 case ARM::tInt_eh_sjlj_setjmp:
587 case ARM::t2Int_eh_sjlj_setjmp:
588 case ARM::t2Int_eh_sjlj_setjmp_nofp:
596 case ARM::t2TBH_JT: {
597 // These are jumptable branches, i.e. a branch followed by an inlined
598 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
599 // entry is one byte; TBH two byte each.
600 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
601 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
602 unsigned NumOps = MCID.getNumOperands();
603 MachineOperand JTOP =
604 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
605 unsigned JTI = JTOP.getIndex();
606 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
608 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
609 assert(JTI < JT.size());
610 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
611 // 4 aligned. The assembler / linker may add 2 byte padding just before
612 // the JT entries. The size does not include this padding; the
613 // constant islands pass does separate bookkeeping for it.
614 // FIXME: If we know the size of the function is less than (1 << 16) *2
615 // bytes, we can use 16-bit entries instead. Then there won't be an
617 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
618 unsigned NumEntries = getNumJTEntries(JT, JTI);
619 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
620 // Make sure the instruction that follows TBB is 2-byte aligned.
621 // FIXME: Constant island pass should insert an "ALIGN" instruction
624 return NumEntries * EntrySize + InstSize;
627 // Otherwise, pseudo-instruction sizes are zero.
632 unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
634 MachineBasicBlock::const_instr_iterator I = MI;
635 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
636 while (++I != E && I->isInsideBundle()) {
637 assert(!I->isBundle() && "No nested bundle!");
638 Size += GetInstSizeInBytes(&*I);
643 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
644 MachineBasicBlock::iterator I, DebugLoc DL,
645 unsigned DestReg, unsigned SrcReg,
646 bool KillSrc) const {
647 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
648 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
650 if (GPRDest && GPRSrc) {
651 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
652 .addReg(SrcReg, getKillRegState(KillSrc))));
656 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
657 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
660 if (SPRDest && SPRSrc)
662 else if (GPRDest && SPRSrc)
664 else if (SPRDest && GPRSrc)
666 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
668 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
672 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
673 MIB.addReg(SrcReg, getKillRegState(KillSrc));
674 if (Opc == ARM::VORRq)
675 MIB.addReg(SrcReg, getKillRegState(KillSrc));
680 // Handle register classes that require multiple instructions.
681 unsigned BeginIdx = 0;
682 unsigned SubRegs = 0;
685 // Use VORRq when possible.
686 if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
687 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 2;
688 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
689 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 4;
690 // Fall back to VMOVD.
691 else if (ARM::DPairRegClass.contains(DestReg, SrcReg))
692 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2;
693 else if (ARM::DTripleRegClass.contains(DestReg, SrcReg))
694 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3;
695 else if (ARM::DQuadRegClass.contains(DestReg, SrcReg))
696 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4;
697 else if (ARM::GPRPairRegClass.contains(DestReg, SrcReg))
698 Opc = ARM::MOVr, BeginIdx = ARM::gsub_0, SubRegs = 2;
700 else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg))
701 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2, Spacing = 2;
702 else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg))
703 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3, Spacing = 2;
704 else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg))
705 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4, Spacing = 2;
707 assert(Opc && "Impossible reg-to-reg copy");
709 const TargetRegisterInfo *TRI = &getRegisterInfo();
710 MachineInstrBuilder Mov;
712 // Copy register tuples backward when the first Dest reg overlaps with SrcReg.
713 if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) {
714 BeginIdx = BeginIdx + ((SubRegs-1)*Spacing);
718 SmallSet<unsigned, 4> DstRegs;
720 for (unsigned i = 0; i != SubRegs; ++i) {
721 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i*Spacing);
722 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i*Spacing);
723 assert(Dst && Src && "Bad sub-register");
725 assert(!DstRegs.count(Src) && "destructive vector copy");
728 Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst)
730 // VORR takes two source operands.
731 if (Opc == ARM::VORRq)
733 Mov = AddDefaultPred(Mov);
735 if (Opc == ARM::MOVr)
736 Mov = AddDefaultCC(Mov);
738 // Add implicit super-register defs and kills to the last instruction.
739 Mov->addRegisterDefined(DestReg, TRI);
741 Mov->addRegisterKilled(SrcReg, TRI);
744 const MachineInstrBuilder &
745 ARMBaseInstrInfo::AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
746 unsigned SubIdx, unsigned State,
747 const TargetRegisterInfo *TRI) const {
749 return MIB.addReg(Reg, State);
751 if (TargetRegisterInfo::isPhysicalRegister(Reg))
752 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
753 return MIB.addReg(Reg, State, SubIdx);
756 void ARMBaseInstrInfo::
757 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
758 unsigned SrcReg, bool isKill, int FI,
759 const TargetRegisterClass *RC,
760 const TargetRegisterInfo *TRI) const {
762 if (I != MBB.end()) DL = I->getDebugLoc();
763 MachineFunction &MF = *MBB.getParent();
764 MachineFrameInfo &MFI = *MF.getFrameInfo();
765 unsigned Align = MFI.getObjectAlignment(FI);
767 MachineMemOperand *MMO =
768 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
769 MachineMemOperand::MOStore,
770 MFI.getObjectSize(FI),
773 switch (RC->getSize()) {
775 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
776 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
777 .addReg(SrcReg, getKillRegState(isKill))
778 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
779 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
780 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
781 .addReg(SrcReg, getKillRegState(isKill))
782 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
784 llvm_unreachable("Unknown reg class!");
787 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
788 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
789 .addReg(SrcReg, getKillRegState(isKill))
790 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
791 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
792 if (Subtarget.hasV5TEOps()) {
793 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::STRD));
794 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
795 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
796 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
800 // Fallback to STM instruction, which has existed since the dawn of
802 MachineInstrBuilder MIB =
803 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STMIA))
804 .addFrameIndex(FI).addMemOperand(MMO));
805 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
806 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
809 llvm_unreachable("Unknown reg class!");
812 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
813 // Use aligned spills if the stack can be realigned.
814 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
815 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
816 .addFrameIndex(FI).addImm(16)
817 .addReg(SrcReg, getKillRegState(isKill))
818 .addMemOperand(MMO));
820 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
821 .addReg(SrcReg, getKillRegState(isKill))
823 .addMemOperand(MMO));
826 llvm_unreachable("Unknown reg class!");
829 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
830 // Use aligned spills if the stack can be realigned.
831 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
832 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
833 .addFrameIndex(FI).addImm(16)
834 .addReg(SrcReg, getKillRegState(isKill))
835 .addMemOperand(MMO));
837 MachineInstrBuilder MIB =
838 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
841 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
842 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
843 AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
846 llvm_unreachable("Unknown reg class!");
849 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
850 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
851 // FIXME: It's possible to only store part of the QQ register if the
852 // spilled def has a sub-register index.
853 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
854 .addFrameIndex(FI).addImm(16)
855 .addReg(SrcReg, getKillRegState(isKill))
856 .addMemOperand(MMO));
858 MachineInstrBuilder MIB =
859 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
862 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
863 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
864 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
865 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
868 llvm_unreachable("Unknown reg class!");
871 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
872 MachineInstrBuilder MIB =
873 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
876 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
877 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
878 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
879 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
880 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
881 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
882 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
883 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
885 llvm_unreachable("Unknown reg class!");
888 llvm_unreachable("Unknown reg class!");
893 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
894 int &FrameIndex) const {
895 switch (MI->getOpcode()) {
898 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
899 if (MI->getOperand(1).isFI() &&
900 MI->getOperand(2).isReg() &&
901 MI->getOperand(3).isImm() &&
902 MI->getOperand(2).getReg() == 0 &&
903 MI->getOperand(3).getImm() == 0) {
904 FrameIndex = MI->getOperand(1).getIndex();
905 return MI->getOperand(0).getReg();
913 if (MI->getOperand(1).isFI() &&
914 MI->getOperand(2).isImm() &&
915 MI->getOperand(2).getImm() == 0) {
916 FrameIndex = MI->getOperand(1).getIndex();
917 return MI->getOperand(0).getReg();
921 case ARM::VST1d64TPseudo:
922 case ARM::VST1d64QPseudo:
923 if (MI->getOperand(0).isFI() &&
924 MI->getOperand(2).getSubReg() == 0) {
925 FrameIndex = MI->getOperand(0).getIndex();
926 return MI->getOperand(2).getReg();
930 if (MI->getOperand(1).isFI() &&
931 MI->getOperand(0).getSubReg() == 0) {
932 FrameIndex = MI->getOperand(1).getIndex();
933 return MI->getOperand(0).getReg();
941 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
942 int &FrameIndex) const {
943 const MachineMemOperand *Dummy;
944 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
947 void ARMBaseInstrInfo::
948 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
949 unsigned DestReg, int FI,
950 const TargetRegisterClass *RC,
951 const TargetRegisterInfo *TRI) const {
953 if (I != MBB.end()) DL = I->getDebugLoc();
954 MachineFunction &MF = *MBB.getParent();
955 MachineFrameInfo &MFI = *MF.getFrameInfo();
956 unsigned Align = MFI.getObjectAlignment(FI);
957 MachineMemOperand *MMO =
958 MF.getMachineMemOperand(
959 MachinePointerInfo::getFixedStack(FI),
960 MachineMemOperand::MOLoad,
961 MFI.getObjectSize(FI),
964 switch (RC->getSize()) {
966 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
967 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
968 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
970 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
971 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
972 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
974 llvm_unreachable("Unknown reg class!");
977 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
978 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
979 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
980 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
981 MachineInstrBuilder MIB;
983 if (Subtarget.hasV5TEOps()) {
984 MIB = BuildMI(MBB, I, DL, get(ARM::LDRD));
985 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
986 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
987 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
991 // Fallback to LDM instruction, which has existed since the dawn of
993 MIB = AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDMIA))
994 .addFrameIndex(FI).addMemOperand(MMO));
995 MIB = AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
996 MIB = AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
999 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1000 MIB.addReg(DestReg, RegState::ImplicitDefine);
1002 llvm_unreachable("Unknown reg class!");
1005 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
1006 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1007 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
1008 .addFrameIndex(FI).addImm(16)
1009 .addMemOperand(MMO));
1011 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
1013 .addMemOperand(MMO));
1016 llvm_unreachable("Unknown reg class!");
1019 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
1020 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1021 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
1022 .addFrameIndex(FI).addImm(16)
1023 .addMemOperand(MMO));
1025 MachineInstrBuilder MIB =
1026 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1028 .addMemOperand(MMO));
1029 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1030 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1031 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1032 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1033 MIB.addReg(DestReg, RegState::ImplicitDefine);
1036 llvm_unreachable("Unknown reg class!");
1039 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
1040 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1041 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
1042 .addFrameIndex(FI).addImm(16)
1043 .addMemOperand(MMO));
1045 MachineInstrBuilder MIB =
1046 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1048 .addMemOperand(MMO);
1049 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1050 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1051 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1052 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1053 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1054 MIB.addReg(DestReg, RegState::ImplicitDefine);
1057 llvm_unreachable("Unknown reg class!");
1060 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
1061 MachineInstrBuilder MIB =
1062 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1064 .addMemOperand(MMO);
1065 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1066 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1067 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1068 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1069 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
1070 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
1071 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
1072 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
1073 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1074 MIB.addReg(DestReg, RegState::ImplicitDefine);
1076 llvm_unreachable("Unknown reg class!");
1079 llvm_unreachable("Unknown regclass!");
1084 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1085 int &FrameIndex) const {
1086 switch (MI->getOpcode()) {
1089 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
1090 if (MI->getOperand(1).isFI() &&
1091 MI->getOperand(2).isReg() &&
1092 MI->getOperand(3).isImm() &&
1093 MI->getOperand(2).getReg() == 0 &&
1094 MI->getOperand(3).getImm() == 0) {
1095 FrameIndex = MI->getOperand(1).getIndex();
1096 return MI->getOperand(0).getReg();
1104 if (MI->getOperand(1).isFI() &&
1105 MI->getOperand(2).isImm() &&
1106 MI->getOperand(2).getImm() == 0) {
1107 FrameIndex = MI->getOperand(1).getIndex();
1108 return MI->getOperand(0).getReg();
1112 case ARM::VLD1d64TPseudo:
1113 case ARM::VLD1d64QPseudo:
1114 if (MI->getOperand(1).isFI() &&
1115 MI->getOperand(0).getSubReg() == 0) {
1116 FrameIndex = MI->getOperand(1).getIndex();
1117 return MI->getOperand(0).getReg();
1121 if (MI->getOperand(1).isFI() &&
1122 MI->getOperand(0).getSubReg() == 0) {
1123 FrameIndex = MI->getOperand(1).getIndex();
1124 return MI->getOperand(0).getReg();
1132 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1133 int &FrameIndex) const {
1134 const MachineMemOperand *Dummy;
1135 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1138 bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
1139 // This hook gets to expand COPY instructions before they become
1140 // copyPhysReg() calls. Look for VMOVS instructions that can legally be
1141 // widened to VMOVD. We prefer the VMOVD when possible because it may be
1142 // changed into a VORR that can go down the NEON pipeline.
1143 if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15())
1146 // Look for a copy between even S-registers. That is where we keep floats
1147 // when using NEON v2f32 instructions for f32 arithmetic.
1148 unsigned DstRegS = MI->getOperand(0).getReg();
1149 unsigned SrcRegS = MI->getOperand(1).getReg();
1150 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
1153 const TargetRegisterInfo *TRI = &getRegisterInfo();
1154 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
1156 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
1158 if (!DstRegD || !SrcRegD)
1161 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
1162 // legal if the COPY already defines the full DstRegD, and it isn't a
1163 // sub-register insertion.
1164 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
1167 // A dead copy shouldn't show up here, but reject it just in case.
1168 if (MI->getOperand(0).isDead())
1171 // All clear, widen the COPY.
1172 DEBUG(dbgs() << "widening: " << *MI);
1173 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1175 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
1176 // or some other super-register.
1177 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
1178 if (ImpDefIdx != -1)
1179 MI->RemoveOperand(ImpDefIdx);
1181 // Change the opcode and operands.
1182 MI->setDesc(get(ARM::VMOVD));
1183 MI->getOperand(0).setReg(DstRegD);
1184 MI->getOperand(1).setReg(SrcRegD);
1185 AddDefaultPred(MIB);
1187 // We are now reading SrcRegD instead of SrcRegS. This may upset the
1188 // register scavenger and machine verifier, so we need to indicate that we
1189 // are reading an undefined value from SrcRegD, but a proper value from
1191 MI->getOperand(1).setIsUndef();
1192 MIB.addReg(SrcRegS, RegState::Implicit);
1194 // SrcRegD may actually contain an unrelated value in the ssub_1
1195 // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
1196 if (MI->getOperand(1).isKill()) {
1197 MI->getOperand(1).setIsKill(false);
1198 MI->addRegisterKilled(SrcRegS, TRI, true);
1201 DEBUG(dbgs() << "replaced by: " << *MI);
1205 /// Create a copy of a const pool value. Update CPI to the new index and return
1207 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1208 MachineConstantPool *MCP = MF.getConstantPool();
1209 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1211 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1212 assert(MCPE.isMachineConstantPoolEntry() &&
1213 "Expecting a machine constantpool entry!");
1214 ARMConstantPoolValue *ACPV =
1215 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1217 unsigned PCLabelId = AFI->createPICLabelUId();
1218 ARMConstantPoolValue *NewCPV = 0;
1219 // FIXME: The below assumes PIC relocation model and that the function
1220 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1221 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1222 // instructions, so that's probably OK, but is PIC always correct when
1224 if (ACPV->isGlobalValue())
1225 NewCPV = ARMConstantPoolConstant::
1226 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
1228 else if (ACPV->isExtSymbol())
1229 NewCPV = ARMConstantPoolSymbol::
1230 Create(MF.getFunction()->getContext(),
1231 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
1232 else if (ACPV->isBlockAddress())
1233 NewCPV = ARMConstantPoolConstant::
1234 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
1235 ARMCP::CPBlockAddress, 4);
1236 else if (ACPV->isLSDA())
1237 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
1239 else if (ACPV->isMachineBasicBlock())
1240 NewCPV = ARMConstantPoolMBB::
1241 Create(MF.getFunction()->getContext(),
1242 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
1244 llvm_unreachable("Unexpected ARM constantpool value type!!");
1245 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1249 void ARMBaseInstrInfo::
1250 reMaterialize(MachineBasicBlock &MBB,
1251 MachineBasicBlock::iterator I,
1252 unsigned DestReg, unsigned SubIdx,
1253 const MachineInstr *Orig,
1254 const TargetRegisterInfo &TRI) const {
1255 unsigned Opcode = Orig->getOpcode();
1258 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1259 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1263 case ARM::tLDRpci_pic:
1264 case ARM::t2LDRpci_pic: {
1265 MachineFunction &MF = *MBB.getParent();
1266 unsigned CPI = Orig->getOperand(1).getIndex();
1267 unsigned PCLabelId = duplicateCPV(MF, CPI);
1268 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1270 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1271 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1278 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1279 MachineInstr *MI = TargetInstrInfo::duplicate(Orig, MF);
1280 switch(Orig->getOpcode()) {
1281 case ARM::tLDRpci_pic:
1282 case ARM::t2LDRpci_pic: {
1283 unsigned CPI = Orig->getOperand(1).getIndex();
1284 unsigned PCLabelId = duplicateCPV(MF, CPI);
1285 Orig->getOperand(1).setIndex(CPI);
1286 Orig->getOperand(2).setImm(PCLabelId);
1293 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1294 const MachineInstr *MI1,
1295 const MachineRegisterInfo *MRI) const {
1296 int Opcode = MI0->getOpcode();
1297 if (Opcode == ARM::t2LDRpci ||
1298 Opcode == ARM::t2LDRpci_pic ||
1299 Opcode == ARM::tLDRpci ||
1300 Opcode == ARM::tLDRpci_pic ||
1301 Opcode == ARM::MOV_ga_dyn ||
1302 Opcode == ARM::MOV_ga_pcrel ||
1303 Opcode == ARM::MOV_ga_pcrel_ldr ||
1304 Opcode == ARM::t2MOV_ga_dyn ||
1305 Opcode == ARM::t2MOV_ga_pcrel) {
1306 if (MI1->getOpcode() != Opcode)
1308 if (MI0->getNumOperands() != MI1->getNumOperands())
1311 const MachineOperand &MO0 = MI0->getOperand(1);
1312 const MachineOperand &MO1 = MI1->getOperand(1);
1313 if (MO0.getOffset() != MO1.getOffset())
1316 if (Opcode == ARM::MOV_ga_dyn ||
1317 Opcode == ARM::MOV_ga_pcrel ||
1318 Opcode == ARM::MOV_ga_pcrel_ldr ||
1319 Opcode == ARM::t2MOV_ga_dyn ||
1320 Opcode == ARM::t2MOV_ga_pcrel)
1321 // Ignore the PC labels.
1322 return MO0.getGlobal() == MO1.getGlobal();
1324 const MachineFunction *MF = MI0->getParent()->getParent();
1325 const MachineConstantPool *MCP = MF->getConstantPool();
1326 int CPI0 = MO0.getIndex();
1327 int CPI1 = MO1.getIndex();
1328 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1329 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1330 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1331 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1332 if (isARMCP0 && isARMCP1) {
1333 ARMConstantPoolValue *ACPV0 =
1334 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1335 ARMConstantPoolValue *ACPV1 =
1336 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1337 return ACPV0->hasSameValue(ACPV1);
1338 } else if (!isARMCP0 && !isARMCP1) {
1339 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1342 } else if (Opcode == ARM::PICLDR) {
1343 if (MI1->getOpcode() != Opcode)
1345 if (MI0->getNumOperands() != MI1->getNumOperands())
1348 unsigned Addr0 = MI0->getOperand(1).getReg();
1349 unsigned Addr1 = MI1->getOperand(1).getReg();
1350 if (Addr0 != Addr1) {
1352 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1353 !TargetRegisterInfo::isVirtualRegister(Addr1))
1356 // This assumes SSA form.
1357 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1358 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1359 // Check if the loaded value, e.g. a constantpool of a global address, are
1361 if (!produceSameValue(Def0, Def1, MRI))
1365 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1366 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1367 const MachineOperand &MO0 = MI0->getOperand(i);
1368 const MachineOperand &MO1 = MI1->getOperand(i);
1369 if (!MO0.isIdenticalTo(MO1))
1375 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1378 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1379 /// determine if two loads are loading from the same base address. It should
1380 /// only return true if the base pointers are the same and the only differences
1381 /// between the two addresses is the offset. It also returns the offsets by
1384 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1385 /// is permanently disabled.
1386 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1388 int64_t &Offset2) const {
1389 // Don't worry about Thumb: just ARM and Thumb2.
1390 if (Subtarget.isThumb1Only()) return false;
1392 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1395 switch (Load1->getMachineOpcode()) {
1409 case ARM::t2LDRSHi8:
1411 case ARM::t2LDRBi12:
1412 case ARM::t2LDRSHi12:
1416 switch (Load2->getMachineOpcode()) {
1429 case ARM::t2LDRSHi8:
1431 case ARM::t2LDRBi12:
1432 case ARM::t2LDRSHi12:
1436 // Check if base addresses and chain operands match.
1437 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1438 Load1->getOperand(4) != Load2->getOperand(4))
1441 // Index should be Reg0.
1442 if (Load1->getOperand(3) != Load2->getOperand(3))
1445 // Determine the offsets.
1446 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1447 isa<ConstantSDNode>(Load2->getOperand(1))) {
1448 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1449 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1456 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1457 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1458 /// be scheduled togther. On some targets if two loads are loading from
1459 /// addresses in the same cache line, it's better if they are scheduled
1460 /// together. This function takes two integers that represent the load offsets
1461 /// from the common base address. It returns true if it decides it's desirable
1462 /// to schedule the two loads together. "NumLoads" is the number of loads that
1463 /// have already been scheduled after Load1.
1465 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1466 /// is permanently disabled.
1467 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1468 int64_t Offset1, int64_t Offset2,
1469 unsigned NumLoads) const {
1470 // Don't worry about Thumb: just ARM and Thumb2.
1471 if (Subtarget.isThumb1Only()) return false;
1473 assert(Offset2 > Offset1);
1475 if ((Offset2 - Offset1) / 8 > 64)
1478 // Check if the machine opcodes are different. If they are different
1479 // then we consider them to not be of the same base address,
1480 // EXCEPT in the case of Thumb2 byte loads where one is LDRBi8 and the other LDRBi12.
1481 // In this case, they are considered to be the same because they are different
1482 // encoding forms of the same basic instruction.
1483 if ((Load1->getMachineOpcode() != Load2->getMachineOpcode()) &&
1484 !((Load1->getMachineOpcode() == ARM::t2LDRBi8 &&
1485 Load2->getMachineOpcode() == ARM::t2LDRBi12) ||
1486 (Load1->getMachineOpcode() == ARM::t2LDRBi12 &&
1487 Load2->getMachineOpcode() == ARM::t2LDRBi8)))
1488 return false; // FIXME: overly conservative?
1490 // Four loads in a row should be sufficient.
1497 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1498 const MachineBasicBlock *MBB,
1499 const MachineFunction &MF) const {
1500 // Debug info is never a scheduling boundary. It's necessary to be explicit
1501 // due to the special treatment of IT instructions below, otherwise a
1502 // dbg_value followed by an IT will result in the IT instruction being
1503 // considered a scheduling hazard, which is wrong. It should be the actual
1504 // instruction preceding the dbg_value instruction(s), just like it is
1505 // when debug info is not present.
1506 if (MI->isDebugValue())
1509 // Terminators and labels can't be scheduled around.
1510 if (MI->isTerminator() || MI->isLabel())
1513 // Treat the start of the IT block as a scheduling boundary, but schedule
1514 // t2IT along with all instructions following it.
1515 // FIXME: This is a big hammer. But the alternative is to add all potential
1516 // true and anti dependencies to IT block instructions as implicit operands
1517 // to the t2IT instruction. The added compile time and complexity does not
1519 MachineBasicBlock::const_iterator I = MI;
1520 // Make sure to skip any dbg_value instructions
1521 while (++I != MBB->end() && I->isDebugValue())
1523 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1526 // Don't attempt to schedule around any instruction that defines
1527 // a stack-oriented pointer, as it's unlikely to be profitable. This
1528 // saves compile time, because it doesn't require every single
1529 // stack slot reference to depend on the instruction that does the
1531 // Calls don't actually change the stack pointer, even if they have imp-defs.
1532 // No ARM calling conventions change the stack pointer. (X86 calling
1533 // conventions sometimes do).
1534 if (!MI->isCall() && MI->definesRegister(ARM::SP))
1540 bool ARMBaseInstrInfo::
1541 isProfitableToIfCvt(MachineBasicBlock &MBB,
1542 unsigned NumCycles, unsigned ExtraPredCycles,
1543 const BranchProbability &Probability) const {
1547 // Attempt to estimate the relative costs of predication versus branching.
1548 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1549 UnpredCost /= Probability.getDenominator();
1550 UnpredCost += 1; // The branch itself
1551 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1553 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1556 bool ARMBaseInstrInfo::
1557 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1558 unsigned TCycles, unsigned TExtra,
1559 MachineBasicBlock &FMBB,
1560 unsigned FCycles, unsigned FExtra,
1561 const BranchProbability &Probability) const {
1562 if (!TCycles || !FCycles)
1565 // Attempt to estimate the relative costs of predication versus branching.
1566 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1567 TUnpredCost /= Probability.getDenominator();
1569 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1570 unsigned FUnpredCost = Comp * FCycles;
1571 FUnpredCost /= Probability.getDenominator();
1573 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1574 UnpredCost += 1; // The branch itself
1575 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1577 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1581 ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
1582 MachineBasicBlock &FMBB) const {
1583 // Reduce false anti-dependencies to let Swift's out-of-order execution
1584 // engine do its thing.
1585 return Subtarget.isSwift();
1588 /// getInstrPredicate - If instruction is predicated, returns its predicate
1589 /// condition, otherwise returns AL. It also returns the condition code
1590 /// register by reference.
1592 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1593 int PIdx = MI->findFirstPredOperandIdx();
1599 PredReg = MI->getOperand(PIdx+1).getReg();
1600 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1604 int llvm::getMatchingCondBranchOpcode(int Opc) {
1609 if (Opc == ARM::t2B)
1612 llvm_unreachable("Unknown unconditional branch opcode!");
1615 /// commuteInstruction - Handle commutable instructions.
1617 ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1618 switch (MI->getOpcode()) {
1620 case ARM::t2MOVCCr: {
1621 // MOVCC can be commuted by inverting the condition.
1622 unsigned PredReg = 0;
1623 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
1624 // MOVCC AL can't be inverted. Shouldn't happen.
1625 if (CC == ARMCC::AL || PredReg != ARM::CPSR)
1627 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
1630 // After swapping the MOVCC operands, also invert the condition.
1631 MI->getOperand(MI->findFirstPredOperandIdx())
1632 .setImm(ARMCC::getOppositeCondition(CC));
1636 return TargetInstrInfo::commuteInstruction(MI, NewMI);
1639 /// Identify instructions that can be folded into a MOVCC instruction, and
1640 /// return the defining instruction.
1641 static MachineInstr *canFoldIntoMOVCC(unsigned Reg,
1642 const MachineRegisterInfo &MRI,
1643 const TargetInstrInfo *TII) {
1644 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1646 if (!MRI.hasOneNonDBGUse(Reg))
1648 MachineInstr *MI = MRI.getVRegDef(Reg);
1651 // MI is folded into the MOVCC by predicating it.
1652 if (!MI->isPredicable())
1654 // Check if MI has any non-dead defs or physreg uses. This also detects
1655 // predicated instructions which will be reading CPSR.
1656 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
1657 const MachineOperand &MO = MI->getOperand(i);
1658 // Reject frame index operands, PEI can't handle the predicated pseudos.
1659 if (MO.isFI() || MO.isCPI() || MO.isJTI())
1663 // MI can't have any tied operands, that would conflict with predication.
1666 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1668 if (MO.isDef() && !MO.isDead())
1671 bool DontMoveAcrossStores = true;
1672 if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ 0, DontMoveAcrossStores))
1677 bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
1678 SmallVectorImpl<MachineOperand> &Cond,
1679 unsigned &TrueOp, unsigned &FalseOp,
1680 bool &Optimizable) const {
1681 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1682 "Unknown select instruction");
1687 // 3: Condition code.
1691 Cond.push_back(MI->getOperand(3));
1692 Cond.push_back(MI->getOperand(4));
1693 // We can always fold a def.
1698 MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
1699 bool PreferFalse) const {
1700 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1701 "Unknown select instruction");
1702 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1703 MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
1704 bool Invert = !DefMI;
1706 DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this);
1710 // Create a new predicated version of DefMI.
1711 // Rfalse is the first use.
1712 MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1714 MI->getOperand(0).getReg());
1716 // Copy all the DefMI operands, excluding its (null) predicate.
1717 const MCInstrDesc &DefDesc = DefMI->getDesc();
1718 for (unsigned i = 1, e = DefDesc.getNumOperands();
1719 i != e && !DefDesc.OpInfo[i].isPredicate(); ++i)
1720 NewMI.addOperand(DefMI->getOperand(i));
1722 unsigned CondCode = MI->getOperand(3).getImm();
1724 NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode)));
1726 NewMI.addImm(CondCode);
1727 NewMI.addOperand(MI->getOperand(4));
1729 // DefMI is not the -S version that sets CPSR, so add an optional %noreg.
1730 if (NewMI->hasOptionalDef())
1731 AddDefaultCC(NewMI);
1733 // The output register value when the predicate is false is an implicit
1734 // register operand tied to the first def.
1735 // The tie makes the register allocator ensure the FalseReg is allocated the
1736 // same register as operand 0.
1737 MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
1738 FalseReg.setImplicit();
1739 NewMI.addOperand(FalseReg);
1740 NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
1742 // The caller will erase MI, but not DefMI.
1743 DefMI->eraseFromParent();
1747 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
1748 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
1751 /// This will go away once we can teach tblgen how to set the optional CPSR def
1753 struct AddSubFlagsOpcodePair {
1755 uint16_t MachineOpc;
1758 static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
1759 {ARM::ADDSri, ARM::ADDri},
1760 {ARM::ADDSrr, ARM::ADDrr},
1761 {ARM::ADDSrsi, ARM::ADDrsi},
1762 {ARM::ADDSrsr, ARM::ADDrsr},
1764 {ARM::SUBSri, ARM::SUBri},
1765 {ARM::SUBSrr, ARM::SUBrr},
1766 {ARM::SUBSrsi, ARM::SUBrsi},
1767 {ARM::SUBSrsr, ARM::SUBrsr},
1769 {ARM::RSBSri, ARM::RSBri},
1770 {ARM::RSBSrsi, ARM::RSBrsi},
1771 {ARM::RSBSrsr, ARM::RSBrsr},
1773 {ARM::t2ADDSri, ARM::t2ADDri},
1774 {ARM::t2ADDSrr, ARM::t2ADDrr},
1775 {ARM::t2ADDSrs, ARM::t2ADDrs},
1777 {ARM::t2SUBSri, ARM::t2SUBri},
1778 {ARM::t2SUBSrr, ARM::t2SUBrr},
1779 {ARM::t2SUBSrs, ARM::t2SUBrs},
1781 {ARM::t2RSBSri, ARM::t2RSBri},
1782 {ARM::t2RSBSrs, ARM::t2RSBrs},
1785 unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
1786 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
1787 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
1788 return AddSubFlagsOpcodeMap[i].MachineOpc;
1792 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1793 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1794 unsigned DestReg, unsigned BaseReg, int NumBytes,
1795 ARMCC::CondCodes Pred, unsigned PredReg,
1796 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1797 bool isSub = NumBytes < 0;
1798 if (isSub) NumBytes = -NumBytes;
1801 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1802 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1803 assert(ThisVal && "Didn't extract field correctly");
1805 // We will handle these bits from offset, clear them.
1806 NumBytes &= ~ThisVal;
1808 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1810 // Build the new ADD / SUB.
1811 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1812 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1813 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1814 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1815 .setMIFlags(MIFlags);
1820 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1821 unsigned FrameReg, int &Offset,
1822 const ARMBaseInstrInfo &TII) {
1823 unsigned Opcode = MI.getOpcode();
1824 const MCInstrDesc &Desc = MI.getDesc();
1825 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1828 // Memory operands in inline assembly always use AddrMode2.
1829 if (Opcode == ARM::INLINEASM)
1830 AddrMode = ARMII::AddrMode2;
1832 if (Opcode == ARM::ADDri) {
1833 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1835 // Turn it into a move.
1836 MI.setDesc(TII.get(ARM::MOVr));
1837 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1838 MI.RemoveOperand(FrameRegIdx+1);
1841 } else if (Offset < 0) {
1844 MI.setDesc(TII.get(ARM::SUBri));
1847 // Common case: small offset, fits into instruction.
1848 if (ARM_AM::getSOImmVal(Offset) != -1) {
1849 // Replace the FrameIndex with sp / fp
1850 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1851 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1856 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1858 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1859 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1861 // We will handle these bits from offset, clear them.
1862 Offset &= ~ThisImmVal;
1864 // Get the properly encoded SOImmVal field.
1865 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1866 "Bit extraction didn't work?");
1867 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1869 unsigned ImmIdx = 0;
1871 unsigned NumBits = 0;
1874 case ARMII::AddrMode_i12: {
1875 ImmIdx = FrameRegIdx + 1;
1876 InstrOffs = MI.getOperand(ImmIdx).getImm();
1880 case ARMII::AddrMode2: {
1881 ImmIdx = FrameRegIdx+2;
1882 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1883 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1888 case ARMII::AddrMode3: {
1889 ImmIdx = FrameRegIdx+2;
1890 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1891 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1896 case ARMII::AddrMode4:
1897 case ARMII::AddrMode6:
1898 // Can't fold any offset even if it's zero.
1900 case ARMII::AddrMode5: {
1901 ImmIdx = FrameRegIdx+1;
1902 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1903 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1910 llvm_unreachable("Unsupported addressing mode!");
1913 Offset += InstrOffs * Scale;
1914 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1920 // Attempt to fold address comp. if opcode has offset bits
1922 // Common case: small offset, fits into instruction.
1923 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1924 int ImmedOffset = Offset / Scale;
1925 unsigned Mask = (1 << NumBits) - 1;
1926 if ((unsigned)Offset <= Mask * Scale) {
1927 // Replace the FrameIndex with sp
1928 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1929 // FIXME: When addrmode2 goes away, this will simplify (like the
1930 // T2 version), as the LDR.i12 versions don't need the encoding
1931 // tricks for the offset value.
1933 if (AddrMode == ARMII::AddrMode_i12)
1934 ImmedOffset = -ImmedOffset;
1936 ImmedOffset |= 1 << NumBits;
1938 ImmOp.ChangeToImmediate(ImmedOffset);
1943 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1944 ImmedOffset = ImmedOffset & Mask;
1946 if (AddrMode == ARMII::AddrMode_i12)
1947 ImmedOffset = -ImmedOffset;
1949 ImmedOffset |= 1 << NumBits;
1951 ImmOp.ChangeToImmediate(ImmedOffset);
1952 Offset &= ~(Mask*Scale);
1956 Offset = (isSub) ? -Offset : Offset;
1960 /// analyzeCompare - For a comparison instruction, return the source registers
1961 /// in SrcReg and SrcReg2 if having two register operands, and the value it
1962 /// compares against in CmpValue. Return true if the comparison instruction
1963 /// can be analyzed.
1964 bool ARMBaseInstrInfo::
1965 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
1966 int &CmpMask, int &CmpValue) const {
1967 switch (MI->getOpcode()) {
1971 SrcReg = MI->getOperand(0).getReg();
1974 CmpValue = MI->getOperand(1).getImm();
1978 SrcReg = MI->getOperand(0).getReg();
1979 SrcReg2 = MI->getOperand(1).getReg();
1985 SrcReg = MI->getOperand(0).getReg();
1987 CmpMask = MI->getOperand(1).getImm();
1995 /// isSuitableForMask - Identify a suitable 'and' instruction that
1996 /// operates on the given source register and applies the same mask
1997 /// as a 'tst' instruction. Provide a limited look-through for copies.
1998 /// When successful, MI will hold the found instruction.
1999 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
2000 int CmpMask, bool CommonUse) {
2001 switch (MI->getOpcode()) {
2004 if (CmpMask != MI->getOperand(2).getImm())
2006 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
2010 // Walk down one instruction which is potentially an 'and'.
2011 const MachineInstr &Copy = *MI;
2012 MachineBasicBlock::iterator AND(
2013 llvm::next(MachineBasicBlock::iterator(MI)));
2014 if (AND == MI->getParent()->end()) return false;
2016 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
2024 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2025 /// the condition code if we modify the instructions such that flags are
2027 inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) {
2029 default: return ARMCC::AL;
2030 case ARMCC::EQ: return ARMCC::EQ;
2031 case ARMCC::NE: return ARMCC::NE;
2032 case ARMCC::HS: return ARMCC::LS;
2033 case ARMCC::LO: return ARMCC::HI;
2034 case ARMCC::HI: return ARMCC::LO;
2035 case ARMCC::LS: return ARMCC::HS;
2036 case ARMCC::GE: return ARMCC::LE;
2037 case ARMCC::LT: return ARMCC::GT;
2038 case ARMCC::GT: return ARMCC::LT;
2039 case ARMCC::LE: return ARMCC::GE;
2043 /// isRedundantFlagInstr - check whether the first instruction, whose only
2044 /// purpose is to update flags, can be made redundant.
2045 /// CMPrr can be made redundant by SUBrr if the operands are the same.
2046 /// CMPri can be made redundant by SUBri if the operands are the same.
2047 /// This function can be extended later on.
2048 inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg,
2049 unsigned SrcReg2, int ImmValue,
2051 if ((CmpI->getOpcode() == ARM::CMPrr ||
2052 CmpI->getOpcode() == ARM::t2CMPrr) &&
2053 (OI->getOpcode() == ARM::SUBrr ||
2054 OI->getOpcode() == ARM::t2SUBrr) &&
2055 ((OI->getOperand(1).getReg() == SrcReg &&
2056 OI->getOperand(2).getReg() == SrcReg2) ||
2057 (OI->getOperand(1).getReg() == SrcReg2 &&
2058 OI->getOperand(2).getReg() == SrcReg)))
2061 if ((CmpI->getOpcode() == ARM::CMPri ||
2062 CmpI->getOpcode() == ARM::t2CMPri) &&
2063 (OI->getOpcode() == ARM::SUBri ||
2064 OI->getOpcode() == ARM::t2SUBri) &&
2065 OI->getOperand(1).getReg() == SrcReg &&
2066 OI->getOperand(2).getImm() == ImmValue)
2071 /// optimizeCompareInstr - Convert the instruction supplying the argument to the
2072 /// comparison into one that sets the zero bit in the flags register;
2073 /// Remove a redundant Compare instruction if an earlier instruction can set the
2074 /// flags in the same way as Compare.
2075 /// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two
2076 /// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the
2077 /// condition code of instructions which use the flags.
2078 bool ARMBaseInstrInfo::
2079 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
2080 int CmpMask, int CmpValue,
2081 const MachineRegisterInfo *MRI) const {
2082 // Get the unique definition of SrcReg.
2083 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
2084 if (!MI) return false;
2086 // Masked compares sometimes use the same register as the corresponding 'and'.
2087 if (CmpMask != ~0) {
2088 if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) {
2090 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
2091 UE = MRI->use_end(); UI != UE; ++UI) {
2092 if (UI->getParent() != CmpInstr->getParent()) continue;
2093 MachineInstr *PotentialAND = &*UI;
2094 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) ||
2095 isPredicated(PotentialAND))
2100 if (!MI) return false;
2104 // Get ready to iterate backward from CmpInstr.
2105 MachineBasicBlock::iterator I = CmpInstr, E = MI,
2106 B = CmpInstr->getParent()->begin();
2108 // Early exit if CmpInstr is at the beginning of the BB.
2109 if (I == B) return false;
2111 // There are two possible candidates which can be changed to set CPSR:
2112 // One is MI, the other is a SUB instruction.
2113 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
2114 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
2115 MachineInstr *Sub = NULL;
2117 // MI is not a candidate for CMPrr.
2119 else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
2120 // Conservatively refuse to convert an instruction which isn't in the same
2121 // BB as the comparison.
2122 // For CMPri, we need to check Sub, thus we can't return here.
2123 if (CmpInstr->getOpcode() == ARM::CMPri ||
2124 CmpInstr->getOpcode() == ARM::t2CMPri)
2130 // Check that CPSR isn't set between the comparison instruction and the one we
2131 // want to change. At the same time, search for Sub.
2132 const TargetRegisterInfo *TRI = &getRegisterInfo();
2134 for (; I != E; --I) {
2135 const MachineInstr &Instr = *I;
2137 if (Instr.modifiesRegister(ARM::CPSR, TRI) ||
2138 Instr.readsRegister(ARM::CPSR, TRI))
2139 // This instruction modifies or uses CPSR after the one we want to
2140 // change. We can't do this transformation.
2143 // Check whether CmpInstr can be made redundant by the current instruction.
2144 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
2150 // The 'and' is below the comparison instruction.
2154 // Return false if no candidates exist.
2158 // The single candidate is called MI.
2161 // We can't use a predicated instruction - it doesn't always write the flags.
2162 if (isPredicated(MI))
2165 switch (MI->getOpcode()) {
2199 case ARM::t2EORri: {
2200 // Scan forward for the use of CPSR
2201 // When checking against MI: if it's a conditional code requires
2202 // checking of V bit, then this is not safe to do.
2203 // It is safe to remove CmpInstr if CPSR is redefined or killed.
2204 // If we are done with the basic block, we need to check whether CPSR is
2206 SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
2208 bool isSafe = false;
2210 E = CmpInstr->getParent()->end();
2211 while (!isSafe && ++I != E) {
2212 const MachineInstr &Instr = *I;
2213 for (unsigned IO = 0, EO = Instr.getNumOperands();
2214 !isSafe && IO != EO; ++IO) {
2215 const MachineOperand &MO = Instr.getOperand(IO);
2216 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
2220 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
2226 // Condition code is after the operand before CPSR.
2227 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
2229 ARMCC::CondCodes NewCC = getSwappedCondition(CC);
2230 if (NewCC == ARMCC::AL)
2232 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
2233 // on CMP needs to be updated to be based on SUB.
2234 // Push the condition code operands to OperandsToUpdate.
2235 // If it is safe to remove CmpInstr, the condition code of these
2236 // operands will be modified.
2237 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
2238 Sub->getOperand(2).getReg() == SrcReg)
2239 OperandsToUpdate.push_back(std::make_pair(&((*I).getOperand(IO-1)),
2245 // CPSR can be used multiple times, we should continue.
2258 // If CPSR is not killed nor re-defined, we should check whether it is
2259 // live-out. If it is live-out, do not optimize.
2261 MachineBasicBlock *MBB = CmpInstr->getParent();
2262 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
2263 SE = MBB->succ_end(); SI != SE; ++SI)
2264 if ((*SI)->isLiveIn(ARM::CPSR))
2268 // Toggle the optional operand to CPSR.
2269 MI->getOperand(5).setReg(ARM::CPSR);
2270 MI->getOperand(5).setIsDef(true);
2271 assert(!isPredicated(MI) && "Can't use flags from predicated instruction");
2272 CmpInstr->eraseFromParent();
2274 // Modify the condition code of operands in OperandsToUpdate.
2275 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2276 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2277 for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++)
2278 OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second);
2286 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
2287 MachineInstr *DefMI, unsigned Reg,
2288 MachineRegisterInfo *MRI) const {
2289 // Fold large immediates into add, sub, or, xor.
2290 unsigned DefOpc = DefMI->getOpcode();
2291 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
2293 if (!DefMI->getOperand(1).isImm())
2294 // Could be t2MOVi32imm <ga:xx>
2297 if (!MRI->hasOneNonDBGUse(Reg))
2300 const MCInstrDesc &DefMCID = DefMI->getDesc();
2301 if (DefMCID.hasOptionalDef()) {
2302 unsigned NumOps = DefMCID.getNumOperands();
2303 const MachineOperand &MO = DefMI->getOperand(NumOps-1);
2304 if (MO.getReg() == ARM::CPSR && !MO.isDead())
2305 // If DefMI defines CPSR and it is not dead, it's obviously not safe
2310 const MCInstrDesc &UseMCID = UseMI->getDesc();
2311 if (UseMCID.hasOptionalDef()) {
2312 unsigned NumOps = UseMCID.getNumOperands();
2313 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
2314 // If the instruction sets the flag, do not attempt this optimization
2315 // since it may change the semantics of the code.
2319 unsigned UseOpc = UseMI->getOpcode();
2320 unsigned NewUseOpc = 0;
2321 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
2322 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
2323 bool Commute = false;
2325 default: return false;
2333 case ARM::t2EORrr: {
2334 Commute = UseMI->getOperand(2).getReg() != Reg;
2341 NewUseOpc = ARM::SUBri;
2347 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
2349 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
2350 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
2353 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
2354 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
2355 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
2359 case ARM::t2SUBrr: {
2363 NewUseOpc = ARM::t2SUBri;
2368 case ARM::t2EORrr: {
2369 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
2371 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
2372 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
2375 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
2376 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
2377 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
2385 unsigned OpIdx = Commute ? 2 : 1;
2386 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
2387 bool isKill = UseMI->getOperand(OpIdx).isKill();
2388 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
2389 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
2390 UseMI, UseMI->getDebugLoc(),
2391 get(NewUseOpc), NewReg)
2392 .addReg(Reg1, getKillRegState(isKill))
2393 .addImm(SOImmValV1)));
2394 UseMI->setDesc(get(NewUseOpc));
2395 UseMI->getOperand(1).setReg(NewReg);
2396 UseMI->getOperand(1).setIsKill();
2397 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
2398 DefMI->eraseFromParent();
2402 static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData,
2403 const MachineInstr *MI) {
2404 switch (MI->getOpcode()) {
2406 const MCInstrDesc &Desc = MI->getDesc();
2407 int UOps = ItinData->getNumMicroOps(Desc.getSchedClass());
2408 assert(UOps >= 0 && "bad # UOps");
2416 unsigned ShOpVal = MI->getOperand(3).getImm();
2417 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2418 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2421 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2422 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2429 if (!MI->getOperand(2).getReg())
2432 unsigned ShOpVal = MI->getOperand(3).getImm();
2433 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2434 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2437 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2438 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2445 return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2;
2447 case ARM::LDRSB_POST:
2448 case ARM::LDRSH_POST: {
2449 unsigned Rt = MI->getOperand(0).getReg();
2450 unsigned Rm = MI->getOperand(3).getReg();
2451 return (Rt == Rm) ? 4 : 3;
2454 case ARM::LDR_PRE_REG:
2455 case ARM::LDRB_PRE_REG: {
2456 unsigned Rt = MI->getOperand(0).getReg();
2457 unsigned Rm = MI->getOperand(3).getReg();
2460 unsigned ShOpVal = MI->getOperand(4).getImm();
2461 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2462 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2465 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2466 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2471 case ARM::STR_PRE_REG:
2472 case ARM::STRB_PRE_REG: {
2473 unsigned ShOpVal = MI->getOperand(4).getImm();
2474 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2475 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2478 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2479 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2485 case ARM::STRH_PRE: {
2486 unsigned Rt = MI->getOperand(0).getReg();
2487 unsigned Rm = MI->getOperand(3).getReg();
2492 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub)
2496 case ARM::LDR_POST_REG:
2497 case ARM::LDRB_POST_REG:
2498 case ARM::LDRH_POST: {
2499 unsigned Rt = MI->getOperand(0).getReg();
2500 unsigned Rm = MI->getOperand(3).getReg();
2501 return (Rt == Rm) ? 3 : 2;
2504 case ARM::LDR_PRE_IMM:
2505 case ARM::LDRB_PRE_IMM:
2506 case ARM::LDR_POST_IMM:
2507 case ARM::LDRB_POST_IMM:
2508 case ARM::STRB_POST_IMM:
2509 case ARM::STRB_POST_REG:
2510 case ARM::STRB_PRE_IMM:
2511 case ARM::STRH_POST:
2512 case ARM::STR_POST_IMM:
2513 case ARM::STR_POST_REG:
2514 case ARM::STR_PRE_IMM:
2517 case ARM::LDRSB_PRE:
2518 case ARM::LDRSH_PRE: {
2519 unsigned Rm = MI->getOperand(3).getReg();
2522 unsigned Rt = MI->getOperand(0).getReg();
2525 unsigned ShOpVal = MI->getOperand(4).getImm();
2526 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2527 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2530 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2531 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2537 unsigned Rt = MI->getOperand(0).getReg();
2538 unsigned Rn = MI->getOperand(2).getReg();
2539 unsigned Rm = MI->getOperand(3).getReg();
2541 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2542 return (Rt == Rn) ? 3 : 2;
2546 unsigned Rm = MI->getOperand(3).getReg();
2548 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2552 case ARM::LDRD_POST:
2553 case ARM::t2LDRD_POST:
2556 case ARM::STRD_POST:
2557 case ARM::t2STRD_POST:
2560 case ARM::LDRD_PRE: {
2561 unsigned Rt = MI->getOperand(0).getReg();
2562 unsigned Rn = MI->getOperand(3).getReg();
2563 unsigned Rm = MI->getOperand(4).getReg();
2565 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2566 return (Rt == Rn) ? 4 : 3;
2569 case ARM::t2LDRD_PRE: {
2570 unsigned Rt = MI->getOperand(0).getReg();
2571 unsigned Rn = MI->getOperand(3).getReg();
2572 return (Rt == Rn) ? 4 : 3;
2575 case ARM::STRD_PRE: {
2576 unsigned Rm = MI->getOperand(4).getReg();
2578 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2582 case ARM::t2STRD_PRE:
2585 case ARM::t2LDR_POST:
2586 case ARM::t2LDRB_POST:
2587 case ARM::t2LDRB_PRE:
2588 case ARM::t2LDRSBi12:
2589 case ARM::t2LDRSBi8:
2590 case ARM::t2LDRSBpci:
2592 case ARM::t2LDRH_POST:
2593 case ARM::t2LDRH_PRE:
2595 case ARM::t2LDRSB_POST:
2596 case ARM::t2LDRSB_PRE:
2597 case ARM::t2LDRSH_POST:
2598 case ARM::t2LDRSH_PRE:
2599 case ARM::t2LDRSHi12:
2600 case ARM::t2LDRSHi8:
2601 case ARM::t2LDRSHpci:
2605 case ARM::t2LDRDi8: {
2606 unsigned Rt = MI->getOperand(0).getReg();
2607 unsigned Rn = MI->getOperand(2).getReg();
2608 return (Rt == Rn) ? 3 : 2;
2611 case ARM::t2STRB_POST:
2612 case ARM::t2STRB_PRE:
2615 case ARM::t2STRH_POST:
2616 case ARM::t2STRH_PRE:
2618 case ARM::t2STR_POST:
2619 case ARM::t2STR_PRE:
2625 // Return the number of 32-bit words loaded by LDM or stored by STM. If this
2626 // can't be easily determined return 0 (missing MachineMemOperand).
2628 // FIXME: The current MachineInstr design does not support relying on machine
2629 // mem operands to determine the width of a memory access. Instead, we expect
2630 // the target to provide this information based on the instruction opcode and
2631 // operands. However, using MachineMemOperand is a the best solution now for
2634 // 1) getNumMicroOps tries to infer LDM memory width from the total number of MI
2635 // operands. This is much more dangerous than using the MachineMemOperand
2636 // sizes because CodeGen passes can insert/remove optional machine operands. In
2637 // fact, it's totally incorrect for preRA passes and appears to be wrong for
2638 // postRA passes as well.
2640 // 2) getNumLDMAddresses is only used by the scheduling machine model and any
2641 // machine model that calls this should handle the unknown (zero size) case.
2643 // Long term, we should require a target hook that verifies MachineMemOperand
2644 // sizes during MC lowering. That target hook should be local to MC lowering
2645 // because we can't ensure that it is aware of other MI forms. Doing this will
2646 // ensure that MachineMemOperands are correctly propagated through all passes.
2647 unsigned ARMBaseInstrInfo::getNumLDMAddresses(const MachineInstr *MI) const {
2649 for (MachineInstr::mmo_iterator I = MI->memoperands_begin(),
2650 E = MI->memoperands_end(); I != E; ++I) {
2651 Size += (*I)->getSize();
2657 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
2658 const MachineInstr *MI) const {
2659 if (!ItinData || ItinData->isEmpty())
2662 const MCInstrDesc &Desc = MI->getDesc();
2663 unsigned Class = Desc.getSchedClass();
2664 int ItinUOps = ItinData->getNumMicroOps(Class);
2665 if (ItinUOps >= 0) {
2666 if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore()))
2667 return getNumMicroOpsSwiftLdSt(ItinData, MI);
2672 unsigned Opc = MI->getOpcode();
2675 llvm_unreachable("Unexpected multi-uops instruction!");
2680 // The number of uOps for load / store multiple are determined by the number
2683 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
2684 // same cycle. The scheduling for the first load / store must be done
2685 // separately by assuming the address is not 64-bit aligned.
2687 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
2688 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
2689 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
2691 case ARM::VLDMDIA_UPD:
2692 case ARM::VLDMDDB_UPD:
2694 case ARM::VLDMSIA_UPD:
2695 case ARM::VLDMSDB_UPD:
2697 case ARM::VSTMDIA_UPD:
2698 case ARM::VSTMDDB_UPD:
2700 case ARM::VSTMSIA_UPD:
2701 case ARM::VSTMSDB_UPD: {
2702 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
2703 return (NumRegs / 2) + (NumRegs % 2) + 1;
2706 case ARM::LDMIA_RET:
2711 case ARM::LDMIA_UPD:
2712 case ARM::LDMDA_UPD:
2713 case ARM::LDMDB_UPD:
2714 case ARM::LDMIB_UPD:
2719 case ARM::STMIA_UPD:
2720 case ARM::STMDA_UPD:
2721 case ARM::STMDB_UPD:
2722 case ARM::STMIB_UPD:
2724 case ARM::tLDMIA_UPD:
2725 case ARM::tSTMIA_UPD:
2729 case ARM::t2LDMIA_RET:
2732 case ARM::t2LDMIA_UPD:
2733 case ARM::t2LDMDB_UPD:
2736 case ARM::t2STMIA_UPD:
2737 case ARM::t2STMDB_UPD: {
2738 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
2739 if (Subtarget.isSwift()) {
2740 int UOps = 1 + NumRegs; // One for address computation, one for each ld / st.
2743 case ARM::VLDMDIA_UPD:
2744 case ARM::VLDMDDB_UPD:
2745 case ARM::VLDMSIA_UPD:
2746 case ARM::VLDMSDB_UPD:
2747 case ARM::VSTMDIA_UPD:
2748 case ARM::VSTMDDB_UPD:
2749 case ARM::VSTMSIA_UPD:
2750 case ARM::VSTMSDB_UPD:
2751 case ARM::LDMIA_UPD:
2752 case ARM::LDMDA_UPD:
2753 case ARM::LDMDB_UPD:
2754 case ARM::LDMIB_UPD:
2755 case ARM::STMIA_UPD:
2756 case ARM::STMDA_UPD:
2757 case ARM::STMDB_UPD:
2758 case ARM::STMIB_UPD:
2759 case ARM::tLDMIA_UPD:
2760 case ARM::tSTMIA_UPD:
2761 case ARM::t2LDMIA_UPD:
2762 case ARM::t2LDMDB_UPD:
2763 case ARM::t2STMIA_UPD:
2764 case ARM::t2STMDB_UPD:
2765 ++UOps; // One for base register writeback.
2767 case ARM::LDMIA_RET:
2769 case ARM::t2LDMIA_RET:
2770 UOps += 2; // One for base reg wb, one for write to pc.
2774 } else if (Subtarget.isCortexA8()) {
2777 // 4 registers would be issued: 2, 2.
2778 // 5 registers would be issued: 2, 2, 1.
2779 int A8UOps = (NumRegs / 2);
2783 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2784 int A9UOps = (NumRegs / 2);
2785 // If there are odd number of registers or if it's not 64-bit aligned,
2786 // then it takes an extra AGU (Address Generation Unit) cycle.
2787 if ((NumRegs % 2) ||
2788 !MI->hasOneMemOperand() ||
2789 (*MI->memoperands_begin())->getAlignment() < 8)
2793 // Assume the worst.
2801 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
2802 const MCInstrDesc &DefMCID,
2804 unsigned DefIdx, unsigned DefAlign) const {
2805 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2807 // Def is the address writeback.
2808 return ItinData->getOperandCycle(DefClass, DefIdx);
2811 if (Subtarget.isCortexA8()) {
2812 // (regno / 2) + (regno % 2) + 1
2813 DefCycle = RegNo / 2 + 1;
2816 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2818 bool isSLoad = false;
2820 switch (DefMCID.getOpcode()) {
2823 case ARM::VLDMSIA_UPD:
2824 case ARM::VLDMSDB_UPD:
2829 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2830 // then it takes an extra cycle.
2831 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
2834 // Assume the worst.
2835 DefCycle = RegNo + 2;
2842 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
2843 const MCInstrDesc &DefMCID,
2845 unsigned DefIdx, unsigned DefAlign) const {
2846 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2848 // Def is the address writeback.
2849 return ItinData->getOperandCycle(DefClass, DefIdx);
2852 if (Subtarget.isCortexA8()) {
2853 // 4 registers would be issued: 1, 2, 1.
2854 // 5 registers would be issued: 1, 2, 2.
2855 DefCycle = RegNo / 2;
2858 // Result latency is issue cycle + 2: E2.
2860 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2861 DefCycle = (RegNo / 2);
2862 // If there are odd number of registers or if it's not 64-bit aligned,
2863 // then it takes an extra AGU (Address Generation Unit) cycle.
2864 if ((RegNo % 2) || DefAlign < 8)
2866 // Result latency is AGU cycles + 2.
2869 // Assume the worst.
2870 DefCycle = RegNo + 2;
2877 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
2878 const MCInstrDesc &UseMCID,
2880 unsigned UseIdx, unsigned UseAlign) const {
2881 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2883 return ItinData->getOperandCycle(UseClass, UseIdx);
2886 if (Subtarget.isCortexA8()) {
2887 // (regno / 2) + (regno % 2) + 1
2888 UseCycle = RegNo / 2 + 1;
2891 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2893 bool isSStore = false;
2895 switch (UseMCID.getOpcode()) {
2898 case ARM::VSTMSIA_UPD:
2899 case ARM::VSTMSDB_UPD:
2904 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2905 // then it takes an extra cycle.
2906 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2909 // Assume the worst.
2910 UseCycle = RegNo + 2;
2917 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2918 const MCInstrDesc &UseMCID,
2920 unsigned UseIdx, unsigned UseAlign) const {
2921 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2923 return ItinData->getOperandCycle(UseClass, UseIdx);
2926 if (Subtarget.isCortexA8()) {
2927 UseCycle = RegNo / 2;
2932 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2933 UseCycle = (RegNo / 2);
2934 // If there are odd number of registers or if it's not 64-bit aligned,
2935 // then it takes an extra AGU (Address Generation Unit) cycle.
2936 if ((RegNo % 2) || UseAlign < 8)
2939 // Assume the worst.
2946 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2947 const MCInstrDesc &DefMCID,
2948 unsigned DefIdx, unsigned DefAlign,
2949 const MCInstrDesc &UseMCID,
2950 unsigned UseIdx, unsigned UseAlign) const {
2951 unsigned DefClass = DefMCID.getSchedClass();
2952 unsigned UseClass = UseMCID.getSchedClass();
2954 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2955 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2957 // This may be a def / use of a variable_ops instruction, the operand
2958 // latency might be determinable dynamically. Let the target try to
2961 bool LdmBypass = false;
2962 switch (DefMCID.getOpcode()) {
2964 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2968 case ARM::VLDMDIA_UPD:
2969 case ARM::VLDMDDB_UPD:
2971 case ARM::VLDMSIA_UPD:
2972 case ARM::VLDMSDB_UPD:
2973 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2976 case ARM::LDMIA_RET:
2981 case ARM::LDMIA_UPD:
2982 case ARM::LDMDA_UPD:
2983 case ARM::LDMDB_UPD:
2984 case ARM::LDMIB_UPD:
2986 case ARM::tLDMIA_UPD:
2988 case ARM::t2LDMIA_RET:
2991 case ARM::t2LDMIA_UPD:
2992 case ARM::t2LDMDB_UPD:
2994 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2999 // We can't seem to determine the result latency of the def, assume it's 2.
3003 switch (UseMCID.getOpcode()) {
3005 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
3009 case ARM::VSTMDIA_UPD:
3010 case ARM::VSTMDDB_UPD:
3012 case ARM::VSTMSIA_UPD:
3013 case ARM::VSTMSDB_UPD:
3014 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3021 case ARM::STMIA_UPD:
3022 case ARM::STMDA_UPD:
3023 case ARM::STMDB_UPD:
3024 case ARM::STMIB_UPD:
3025 case ARM::tSTMIA_UPD:
3030 case ARM::t2STMIA_UPD:
3031 case ARM::t2STMDB_UPD:
3032 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3037 // Assume it's read in the first stage.
3040 UseCycle = DefCycle - UseCycle + 1;
3043 // It's a variable_ops instruction so we can't use DefIdx here. Just use
3044 // first def operand.
3045 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
3048 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
3049 UseClass, UseIdx)) {
3057 static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
3058 const MachineInstr *MI, unsigned Reg,
3059 unsigned &DefIdx, unsigned &Dist) {
3062 MachineBasicBlock::const_iterator I = MI; ++I;
3063 MachineBasicBlock::const_instr_iterator II =
3064 llvm::prior(I.getInstrIterator());
3065 assert(II->isInsideBundle() && "Empty bundle?");
3068 while (II->isInsideBundle()) {
3069 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
3076 assert(Idx != -1 && "Cannot find bundled definition!");
3081 static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
3082 const MachineInstr *MI, unsigned Reg,
3083 unsigned &UseIdx, unsigned &Dist) {
3086 MachineBasicBlock::const_instr_iterator II = MI; ++II;
3087 assert(II->isInsideBundle() && "Empty bundle?");
3088 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3090 // FIXME: This doesn't properly handle multiple uses.
3092 while (II != E && II->isInsideBundle()) {
3093 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
3096 if (II->getOpcode() != ARM::t2IT)
3110 /// Return the number of cycles to add to (or subtract from) the static
3111 /// itinerary based on the def opcode and alignment. The caller will ensure that
3112 /// adjusted latency is at least one cycle.
3113 static int adjustDefLatency(const ARMSubtarget &Subtarget,
3114 const MachineInstr *DefMI,
3115 const MCInstrDesc *DefMCID, unsigned DefAlign) {
3117 if (Subtarget.isCortexA8() || Subtarget.isLikeA9()) {
3118 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3119 // variants are one cycle cheaper.
3120 switch (DefMCID->getOpcode()) {
3124 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3125 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3127 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3134 case ARM::t2LDRSHs: {
3135 // Thumb2 mode: lsl only.
3136 unsigned ShAmt = DefMI->getOperand(3).getImm();
3137 if (ShAmt == 0 || ShAmt == 2)
3142 } else if (Subtarget.isSwift()) {
3143 // FIXME: Properly handle all of the latency adjustments for address
3145 switch (DefMCID->getOpcode()) {
3149 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3150 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
3151 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3154 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3155 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
3158 ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3165 case ARM::t2LDRSHs: {
3166 // Thumb2 mode: lsl only.
3167 unsigned ShAmt = DefMI->getOperand(3).getImm();
3168 if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3)
3175 if (DefAlign < 8 && Subtarget.isLikeA9()) {
3176 switch (DefMCID->getOpcode()) {
3182 case ARM::VLD1q8wb_fixed:
3183 case ARM::VLD1q16wb_fixed:
3184 case ARM::VLD1q32wb_fixed:
3185 case ARM::VLD1q64wb_fixed:
3186 case ARM::VLD1q8wb_register:
3187 case ARM::VLD1q16wb_register:
3188 case ARM::VLD1q32wb_register:
3189 case ARM::VLD1q64wb_register:
3196 case ARM::VLD2d8wb_fixed:
3197 case ARM::VLD2d16wb_fixed:
3198 case ARM::VLD2d32wb_fixed:
3199 case ARM::VLD2q8wb_fixed:
3200 case ARM::VLD2q16wb_fixed:
3201 case ARM::VLD2q32wb_fixed:
3202 case ARM::VLD2d8wb_register:
3203 case ARM::VLD2d16wb_register:
3204 case ARM::VLD2d32wb_register:
3205 case ARM::VLD2q8wb_register:
3206 case ARM::VLD2q16wb_register:
3207 case ARM::VLD2q32wb_register:
3212 case ARM::VLD3d8_UPD:
3213 case ARM::VLD3d16_UPD:
3214 case ARM::VLD3d32_UPD:
3215 case ARM::VLD1d64Twb_fixed:
3216 case ARM::VLD1d64Twb_register:
3217 case ARM::VLD3q8_UPD:
3218 case ARM::VLD3q16_UPD:
3219 case ARM::VLD3q32_UPD:
3224 case ARM::VLD4d8_UPD:
3225 case ARM::VLD4d16_UPD:
3226 case ARM::VLD4d32_UPD:
3227 case ARM::VLD1d64Qwb_fixed:
3228 case ARM::VLD1d64Qwb_register:
3229 case ARM::VLD4q8_UPD:
3230 case ARM::VLD4q16_UPD:
3231 case ARM::VLD4q32_UPD:
3232 case ARM::VLD1DUPq8:
3233 case ARM::VLD1DUPq16:
3234 case ARM::VLD1DUPq32:
3235 case ARM::VLD1DUPq8wb_fixed:
3236 case ARM::VLD1DUPq16wb_fixed:
3237 case ARM::VLD1DUPq32wb_fixed:
3238 case ARM::VLD1DUPq8wb_register:
3239 case ARM::VLD1DUPq16wb_register:
3240 case ARM::VLD1DUPq32wb_register:
3241 case ARM::VLD2DUPd8:
3242 case ARM::VLD2DUPd16:
3243 case ARM::VLD2DUPd32:
3244 case ARM::VLD2DUPd8wb_fixed:
3245 case ARM::VLD2DUPd16wb_fixed:
3246 case ARM::VLD2DUPd32wb_fixed:
3247 case ARM::VLD2DUPd8wb_register:
3248 case ARM::VLD2DUPd16wb_register:
3249 case ARM::VLD2DUPd32wb_register:
3250 case ARM::VLD4DUPd8:
3251 case ARM::VLD4DUPd16:
3252 case ARM::VLD4DUPd32:
3253 case ARM::VLD4DUPd8_UPD:
3254 case ARM::VLD4DUPd16_UPD:
3255 case ARM::VLD4DUPd32_UPD:
3257 case ARM::VLD1LNd16:
3258 case ARM::VLD1LNd32:
3259 case ARM::VLD1LNd8_UPD:
3260 case ARM::VLD1LNd16_UPD:
3261 case ARM::VLD1LNd32_UPD:
3263 case ARM::VLD2LNd16:
3264 case ARM::VLD2LNd32:
3265 case ARM::VLD2LNq16:
3266 case ARM::VLD2LNq32:
3267 case ARM::VLD2LNd8_UPD:
3268 case ARM::VLD2LNd16_UPD:
3269 case ARM::VLD2LNd32_UPD:
3270 case ARM::VLD2LNq16_UPD:
3271 case ARM::VLD2LNq32_UPD:
3273 case ARM::VLD4LNd16:
3274 case ARM::VLD4LNd32:
3275 case ARM::VLD4LNq16:
3276 case ARM::VLD4LNq32:
3277 case ARM::VLD4LNd8_UPD:
3278 case ARM::VLD4LNd16_UPD:
3279 case ARM::VLD4LNd32_UPD:
3280 case ARM::VLD4LNq16_UPD:
3281 case ARM::VLD4LNq32_UPD:
3282 // If the address is not 64-bit aligned, the latencies of these
3283 // instructions increases by one.
3294 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3295 const MachineInstr *DefMI, unsigned DefIdx,
3296 const MachineInstr *UseMI,
3297 unsigned UseIdx) const {
3298 // No operand latency. The caller may fall back to getInstrLatency.
3299 if (!ItinData || ItinData->isEmpty())
3302 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
3303 unsigned Reg = DefMO.getReg();
3304 const MCInstrDesc *DefMCID = &DefMI->getDesc();
3305 const MCInstrDesc *UseMCID = &UseMI->getDesc();
3307 unsigned DefAdj = 0;
3308 if (DefMI->isBundle()) {
3309 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
3310 DefMCID = &DefMI->getDesc();
3312 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
3313 DefMI->isRegSequence() || DefMI->isImplicitDef()) {
3317 unsigned UseAdj = 0;
3318 if (UseMI->isBundle()) {
3320 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
3321 Reg, NewUseIdx, UseAdj);
3327 UseMCID = &UseMI->getDesc();
3330 if (Reg == ARM::CPSR) {
3331 if (DefMI->getOpcode() == ARM::FMSTAT) {
3332 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
3333 return Subtarget.isLikeA9() ? 1 : 20;
3336 // CPSR set and branch can be paired in the same cycle.
3337 if (UseMI->isBranch())
3340 // Otherwise it takes the instruction latency (generally one).
3341 unsigned Latency = getInstrLatency(ItinData, DefMI);
3343 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
3344 // its uses. Instructions which are otherwise scheduled between them may
3345 // incur a code size penalty (not able to use the CPSR setting 16-bit
3347 if (Latency > 0 && Subtarget.isThumb2()) {
3348 const MachineFunction *MF = DefMI->getParent()->getParent();
3349 if (MF->getFunction()->getAttributes().
3350 hasAttribute(AttributeSet::FunctionIndex,
3351 Attribute::OptimizeForSize))
3357 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
3360 unsigned DefAlign = DefMI->hasOneMemOperand()
3361 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
3362 unsigned UseAlign = UseMI->hasOneMemOperand()
3363 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
3365 // Get the itinerary's latency if possible, and handle variable_ops.
3366 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
3367 *UseMCID, UseIdx, UseAlign);
3368 // Unable to find operand latency. The caller may resort to getInstrLatency.
3372 // Adjust for IT block position.
3373 int Adj = DefAdj + UseAdj;
3375 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3376 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
3377 if (Adj >= 0 || (int)Latency > -Adj) {
3378 return Latency + Adj;
3380 // Return the itinerary latency, which may be zero but not less than zero.
3385 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3386 SDNode *DefNode, unsigned DefIdx,
3387 SDNode *UseNode, unsigned UseIdx) const {
3388 if (!DefNode->isMachineOpcode())
3391 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
3393 if (isZeroCost(DefMCID.Opcode))
3396 if (!ItinData || ItinData->isEmpty())
3397 return DefMCID.mayLoad() ? 3 : 1;
3399 if (!UseNode->isMachineOpcode()) {
3400 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
3401 if (Subtarget.isLikeA9() || Subtarget.isSwift())
3402 return Latency <= 2 ? 1 : Latency - 1;
3404 return Latency <= 3 ? 1 : Latency - 2;
3407 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
3408 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
3409 unsigned DefAlign = !DefMN->memoperands_empty()
3410 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
3411 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
3412 unsigned UseAlign = !UseMN->memoperands_empty()
3413 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
3414 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
3415 UseMCID, UseIdx, UseAlign);
3418 (Subtarget.isCortexA8() || Subtarget.isLikeA9())) {
3419 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3420 // variants are one cycle cheaper.
3421 switch (DefMCID.getOpcode()) {
3426 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3427 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3429 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3436 case ARM::t2LDRSHs: {
3437 // Thumb2 mode: lsl only.
3439 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3440 if (ShAmt == 0 || ShAmt == 2)
3445 } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) {
3446 // FIXME: Properly handle all of the latency adjustments for address
3448 switch (DefMCID.getOpcode()) {
3453 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3454 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3456 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3457 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3459 else if (ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3466 case ARM::t2LDRSHs: {
3467 // Thumb2 mode: lsl 0-3 only.
3474 if (DefAlign < 8 && Subtarget.isLikeA9())
3475 switch (DefMCID.getOpcode()) {
3481 case ARM::VLD1q8wb_register:
3482 case ARM::VLD1q16wb_register:
3483 case ARM::VLD1q32wb_register:
3484 case ARM::VLD1q64wb_register:
3485 case ARM::VLD1q8wb_fixed:
3486 case ARM::VLD1q16wb_fixed:
3487 case ARM::VLD1q32wb_fixed:
3488 case ARM::VLD1q64wb_fixed:
3492 case ARM::VLD2q8Pseudo:
3493 case ARM::VLD2q16Pseudo:
3494 case ARM::VLD2q32Pseudo:
3495 case ARM::VLD2d8wb_fixed:
3496 case ARM::VLD2d16wb_fixed:
3497 case ARM::VLD2d32wb_fixed:
3498 case ARM::VLD2q8PseudoWB_fixed:
3499 case ARM::VLD2q16PseudoWB_fixed:
3500 case ARM::VLD2q32PseudoWB_fixed:
3501 case ARM::VLD2d8wb_register:
3502 case ARM::VLD2d16wb_register:
3503 case ARM::VLD2d32wb_register:
3504 case ARM::VLD2q8PseudoWB_register:
3505 case ARM::VLD2q16PseudoWB_register:
3506 case ARM::VLD2q32PseudoWB_register:
3507 case ARM::VLD3d8Pseudo:
3508 case ARM::VLD3d16Pseudo:
3509 case ARM::VLD3d32Pseudo:
3510 case ARM::VLD1d64TPseudo:
3511 case ARM::VLD3d8Pseudo_UPD:
3512 case ARM::VLD3d16Pseudo_UPD:
3513 case ARM::VLD3d32Pseudo_UPD:
3514 case ARM::VLD3q8Pseudo_UPD:
3515 case ARM::VLD3q16Pseudo_UPD:
3516 case ARM::VLD3q32Pseudo_UPD:
3517 case ARM::VLD3q8oddPseudo:
3518 case ARM::VLD3q16oddPseudo:
3519 case ARM::VLD3q32oddPseudo:
3520 case ARM::VLD3q8oddPseudo_UPD:
3521 case ARM::VLD3q16oddPseudo_UPD:
3522 case ARM::VLD3q32oddPseudo_UPD:
3523 case ARM::VLD4d8Pseudo:
3524 case ARM::VLD4d16Pseudo:
3525 case ARM::VLD4d32Pseudo:
3526 case ARM::VLD1d64QPseudo:
3527 case ARM::VLD4d8Pseudo_UPD:
3528 case ARM::VLD4d16Pseudo_UPD:
3529 case ARM::VLD4d32Pseudo_UPD:
3530 case ARM::VLD4q8Pseudo_UPD:
3531 case ARM::VLD4q16Pseudo_UPD:
3532 case ARM::VLD4q32Pseudo_UPD:
3533 case ARM::VLD4q8oddPseudo:
3534 case ARM::VLD4q16oddPseudo:
3535 case ARM::VLD4q32oddPseudo:
3536 case ARM::VLD4q8oddPseudo_UPD:
3537 case ARM::VLD4q16oddPseudo_UPD:
3538 case ARM::VLD4q32oddPseudo_UPD:
3539 case ARM::VLD1DUPq8:
3540 case ARM::VLD1DUPq16:
3541 case ARM::VLD1DUPq32:
3542 case ARM::VLD1DUPq8wb_fixed:
3543 case ARM::VLD1DUPq16wb_fixed:
3544 case ARM::VLD1DUPq32wb_fixed:
3545 case ARM::VLD1DUPq8wb_register:
3546 case ARM::VLD1DUPq16wb_register:
3547 case ARM::VLD1DUPq32wb_register:
3548 case ARM::VLD2DUPd8:
3549 case ARM::VLD2DUPd16:
3550 case ARM::VLD2DUPd32:
3551 case ARM::VLD2DUPd8wb_fixed:
3552 case ARM::VLD2DUPd16wb_fixed:
3553 case ARM::VLD2DUPd32wb_fixed:
3554 case ARM::VLD2DUPd8wb_register:
3555 case ARM::VLD2DUPd16wb_register:
3556 case ARM::VLD2DUPd32wb_register:
3557 case ARM::VLD4DUPd8Pseudo:
3558 case ARM::VLD4DUPd16Pseudo:
3559 case ARM::VLD4DUPd32Pseudo:
3560 case ARM::VLD4DUPd8Pseudo_UPD:
3561 case ARM::VLD4DUPd16Pseudo_UPD:
3562 case ARM::VLD4DUPd32Pseudo_UPD:
3563 case ARM::VLD1LNq8Pseudo:
3564 case ARM::VLD1LNq16Pseudo:
3565 case ARM::VLD1LNq32Pseudo:
3566 case ARM::VLD1LNq8Pseudo_UPD:
3567 case ARM::VLD1LNq16Pseudo_UPD:
3568 case ARM::VLD1LNq32Pseudo_UPD:
3569 case ARM::VLD2LNd8Pseudo:
3570 case ARM::VLD2LNd16Pseudo:
3571 case ARM::VLD2LNd32Pseudo:
3572 case ARM::VLD2LNq16Pseudo:
3573 case ARM::VLD2LNq32Pseudo:
3574 case ARM::VLD2LNd8Pseudo_UPD:
3575 case ARM::VLD2LNd16Pseudo_UPD:
3576 case ARM::VLD2LNd32Pseudo_UPD:
3577 case ARM::VLD2LNq16Pseudo_UPD:
3578 case ARM::VLD2LNq32Pseudo_UPD:
3579 case ARM::VLD4LNd8Pseudo:
3580 case ARM::VLD4LNd16Pseudo:
3581 case ARM::VLD4LNd32Pseudo:
3582 case ARM::VLD4LNq16Pseudo:
3583 case ARM::VLD4LNq32Pseudo:
3584 case ARM::VLD4LNd8Pseudo_UPD:
3585 case ARM::VLD4LNd16Pseudo_UPD:
3586 case ARM::VLD4LNd32Pseudo_UPD:
3587 case ARM::VLD4LNq16Pseudo_UPD:
3588 case ARM::VLD4LNq32Pseudo_UPD:
3589 // If the address is not 64-bit aligned, the latencies of these
3590 // instructions increases by one.
3598 unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3599 const MachineInstr *MI,
3600 unsigned *PredCost) const {
3601 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3602 MI->isRegSequence() || MI->isImplicitDef())
3605 // An instruction scheduler typically runs on unbundled instructions, however
3606 // other passes may query the latency of a bundled instruction.
3607 if (MI->isBundle()) {
3608 unsigned Latency = 0;
3609 MachineBasicBlock::const_instr_iterator I = MI;
3610 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3611 while (++I != E && I->isInsideBundle()) {
3612 if (I->getOpcode() != ARM::t2IT)
3613 Latency += getInstrLatency(ItinData, I, PredCost);
3618 const MCInstrDesc &MCID = MI->getDesc();
3619 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) {
3620 // When predicated, CPSR is an additional source operand for CPSR updating
3621 // instructions, this apparently increases their latencies.
3624 // Be sure to call getStageLatency for an empty itinerary in case it has a
3625 // valid MinLatency property.
3627 return MI->mayLoad() ? 3 : 1;
3629 unsigned Class = MCID.getSchedClass();
3631 // For instructions with variable uops, use uops as latency.
3632 if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0)
3633 return getNumMicroOps(ItinData, MI);
3635 // For the common case, fall back on the itinerary's latency.
3636 unsigned Latency = ItinData->getStageLatency(Class);
3638 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3639 unsigned DefAlign = MI->hasOneMemOperand()
3640 ? (*MI->memoperands_begin())->getAlignment() : 0;
3641 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
3642 if (Adj >= 0 || (int)Latency > -Adj) {
3643 return Latency + Adj;
3648 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3649 SDNode *Node) const {
3650 if (!Node->isMachineOpcode())
3653 if (!ItinData || ItinData->isEmpty())
3656 unsigned Opcode = Node->getMachineOpcode();
3659 return ItinData->getStageLatency(get(Opcode).getSchedClass());
3666 bool ARMBaseInstrInfo::
3667 hasHighOperandLatency(const InstrItineraryData *ItinData,
3668 const MachineRegisterInfo *MRI,
3669 const MachineInstr *DefMI, unsigned DefIdx,
3670 const MachineInstr *UseMI, unsigned UseIdx) const {
3671 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3672 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
3673 if (Subtarget.isCortexA8() &&
3674 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
3675 // CortexA8 VFP instructions are not pipelined.
3678 // Hoist VFP / NEON instructions with 4 or higher latency.
3679 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
3681 Latency = getInstrLatency(ItinData, DefMI);
3684 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
3685 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
3688 bool ARMBaseInstrInfo::
3689 hasLowDefLatency(const InstrItineraryData *ItinData,
3690 const MachineInstr *DefMI, unsigned DefIdx) const {
3691 if (!ItinData || ItinData->isEmpty())
3694 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3695 if (DDomain == ARMII::DomainGeneral) {
3696 unsigned DefClass = DefMI->getDesc().getSchedClass();
3697 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3698 return (DefCycle != -1 && DefCycle <= 2);
3703 bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
3704 StringRef &ErrInfo) const {
3705 if (convertAddSubFlagsOpcode(MI->getOpcode())) {
3706 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
3713 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
3714 unsigned &AddSubOpc,
3715 bool &NegAcc, bool &HasLane) const {
3716 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
3717 if (I == MLxEntryMap.end())
3720 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
3721 MulOpc = Entry.MulOpc;
3722 AddSubOpc = Entry.AddSubOpc;
3723 NegAcc = Entry.NegAcc;
3724 HasLane = Entry.HasLane;
3728 //===----------------------------------------------------------------------===//
3729 // Execution domains.
3730 //===----------------------------------------------------------------------===//
3732 // Some instructions go down the NEON pipeline, some go down the VFP pipeline,
3733 // and some can go down both. The vmov instructions go down the VFP pipeline,
3734 // but they can be changed to vorr equivalents that are executed by the NEON
3737 // We use the following execution domain numbering:
3745 // Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
3747 std::pair<uint16_t, uint16_t>
3748 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
3749 // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
3750 // if they are not predicated.
3751 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
3752 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3754 // CortexA9 is particularly picky about mixing the two and wants these
3756 if (Subtarget.isCortexA9() && !isPredicated(MI) &&
3757 (MI->getOpcode() == ARM::VMOVRS ||
3758 MI->getOpcode() == ARM::VMOVSR ||
3759 MI->getOpcode() == ARM::VMOVS))
3760 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3762 // No other instructions can be swizzled, so just determine their domain.
3763 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
3765 if (Domain & ARMII::DomainNEON)
3766 return std::make_pair(ExeNEON, 0);
3768 // Certain instructions can go either way on Cortex-A8.
3769 // Treat them as NEON instructions.
3770 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
3771 return std::make_pair(ExeNEON, 0);
3773 if (Domain & ARMII::DomainVFP)
3774 return std::make_pair(ExeVFP, 0);
3776 return std::make_pair(ExeGeneric, 0);
3779 static unsigned getCorrespondingDRegAndLane(const TargetRegisterInfo *TRI,
3780 unsigned SReg, unsigned &Lane) {
3781 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_0, &ARM::DPRRegClass);
3784 if (DReg != ARM::NoRegister)
3788 DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1, &ARM::DPRRegClass);
3790 assert(DReg && "S-register with no D super-register?");
3794 /// getImplicitSPRUseForDPRUse - Given a use of a DPR register and lane,
3795 /// set ImplicitSReg to a register number that must be marked as implicit-use or
3796 /// zero if no register needs to be defined as implicit-use.
3798 /// If the function cannot determine if an SPR should be marked implicit use or
3799 /// not, it returns false.
3801 /// This function handles cases where an instruction is being modified from taking
3802 /// an SPR to a DPR[Lane]. A use of the DPR is being added, which may conflict
3803 /// with an earlier def of an SPR corresponding to DPR[Lane^1] (i.e. the other
3804 /// lane of the DPR).
3806 /// If the other SPR is defined, an implicit-use of it should be added. Else,
3807 /// (including the case where the DPR itself is defined), it should not.
3809 static bool getImplicitSPRUseForDPRUse(const TargetRegisterInfo *TRI,
3811 unsigned DReg, unsigned Lane,
3812 unsigned &ImplicitSReg) {
3813 // If the DPR is defined or used already, the other SPR lane will be chained
3814 // correctly, so there is nothing to be done.
3815 if (MI->definesRegister(DReg, TRI) || MI->readsRegister(DReg, TRI)) {
3820 // Otherwise we need to go searching to see if the SPR is set explicitly.
3821 ImplicitSReg = TRI->getSubReg(DReg,
3822 (Lane & 1) ? ARM::ssub_0 : ARM::ssub_1);
3823 MachineBasicBlock::LivenessQueryResult LQR =
3824 MI->getParent()->computeRegisterLiveness(TRI, ImplicitSReg, MI);
3826 if (LQR == MachineBasicBlock::LQR_Live)
3828 else if (LQR == MachineBasicBlock::LQR_Unknown)
3831 // If the register is known not to be live, there is no need to add an
3838 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
3839 unsigned DstReg, SrcReg, DReg;
3841 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3842 const TargetRegisterInfo *TRI = &getRegisterInfo();
3843 switch (MI->getOpcode()) {
3845 llvm_unreachable("cannot handle opcode!");
3848 if (Domain != ExeNEON)
3851 // Zap the predicate operands.
3852 assert(!isPredicated(MI) && "Cannot predicate a VORRd");
3854 // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
3855 DstReg = MI->getOperand(0).getReg();
3856 SrcReg = MI->getOperand(1).getReg();
3858 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3859 MI->RemoveOperand(i-1);
3861 // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits)
3862 MI->setDesc(get(ARM::VORRd));
3863 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3868 if (Domain != ExeNEON)
3870 assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
3872 // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits)
3873 DstReg = MI->getOperand(0).getReg();
3874 SrcReg = MI->getOperand(1).getReg();
3876 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3877 MI->RemoveOperand(i-1);
3879 DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane);
3881 // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps)
3882 // Note that DSrc has been widened and the other lane may be undef, which
3883 // contaminates the entire register.
3884 MI->setDesc(get(ARM::VGETLNi32));
3885 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3886 .addReg(DReg, RegState::Undef)
3889 // The old source should be an implicit use, otherwise we might think it
3890 // was dead before here.
3891 MIB.addReg(SrcReg, RegState::Implicit);
3894 if (Domain != ExeNEON)
3896 assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
3898 // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits)
3899 DstReg = MI->getOperand(0).getReg();
3900 SrcReg = MI->getOperand(1).getReg();
3902 DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane);
3904 unsigned ImplicitSReg;
3905 if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg))
3908 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3909 MI->RemoveOperand(i-1);
3911 // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps)
3912 // Again DDst may be undefined at the beginning of this instruction.
3913 MI->setDesc(get(ARM::VSETLNi32));
3914 MIB.addReg(DReg, RegState::Define)
3915 .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI)))
3918 AddDefaultPred(MIB);
3920 // The narrower destination must be marked as set to keep previous chains
3922 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
3923 if (ImplicitSReg != 0)
3924 MIB.addReg(ImplicitSReg, RegState::Implicit);
3928 if (Domain != ExeNEON)
3931 // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits)
3932 DstReg = MI->getOperand(0).getReg();
3933 SrcReg = MI->getOperand(1).getReg();
3935 unsigned DstLane = 0, SrcLane = 0, DDst, DSrc;
3936 DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane);
3937 DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane);
3939 unsigned ImplicitSReg;
3940 if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg))
3943 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3944 MI->RemoveOperand(i-1);
3947 // Destination can be:
3948 // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits)
3949 MI->setDesc(get(ARM::VDUPLN32d));
3950 MIB.addReg(DDst, RegState::Define)
3951 .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI)))
3953 AddDefaultPred(MIB);
3955 // Neither the source or the destination are naturally represented any
3956 // more, so add them in manually.
3957 MIB.addReg(DstReg, RegState::Implicit | RegState::Define);
3958 MIB.addReg(SrcReg, RegState::Implicit);
3959 if (ImplicitSReg != 0)
3960 MIB.addReg(ImplicitSReg, RegState::Implicit);
3964 // In general there's no single instruction that can perform an S <-> S
3965 // move in NEON space, but a pair of VEXT instructions *can* do the
3966 // job. It turns out that the VEXTs needed will only use DSrc once, with
3967 // the position based purely on the combination of lane-0 and lane-1
3968 // involved. For example
3969 // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1
3970 // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1
3971 // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1
3972 // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1
3974 // Pattern of the MachineInstrs is:
3975 // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits)
3976 MachineInstrBuilder NewMIB;
3977 NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
3978 get(ARM::VEXTd32), DDst);
3980 // On the first instruction, both DSrc and DDst may be <undef> if present.
3981 // Specifically when the original instruction didn't have them as an
3983 unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst;
3984 bool CurUndef = !MI->readsRegister(CurReg, TRI);
3985 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
3987 CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst;
3988 CurUndef = !MI->readsRegister(CurReg, TRI);
3989 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
3992 AddDefaultPred(NewMIB);
3994 if (SrcLane == DstLane)
3995 NewMIB.addReg(SrcReg, RegState::Implicit);
3997 MI->setDesc(get(ARM::VEXTd32));
3998 MIB.addReg(DDst, RegState::Define);
4000 // On the second instruction, DDst has definitely been defined above, so
4001 // it is not <undef>. DSrc, if present, can be <undef> as above.
4002 CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst;
4003 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4004 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4006 CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst;
4007 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4008 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4011 AddDefaultPred(MIB);
4013 if (SrcLane != DstLane)
4014 MIB.addReg(SrcReg, RegState::Implicit);
4016 // As before, the original destination is no longer represented, add it
4018 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4019 if (ImplicitSReg != 0)
4020 MIB.addReg(ImplicitSReg, RegState::Implicit);
4027 //===----------------------------------------------------------------------===//
4028 // Partial register updates
4029 //===----------------------------------------------------------------------===//
4031 // Swift renames NEON registers with 64-bit granularity. That means any
4032 // instruction writing an S-reg implicitly reads the containing D-reg. The
4033 // problem is mostly avoided by translating f32 operations to v2f32 operations
4034 // on D-registers, but f32 loads are still a problem.
4036 // These instructions can load an f32 into a NEON register:
4038 // VLDRS - Only writes S, partial D update.
4039 // VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops.
4040 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
4042 // FCONSTD can be used as a dependency-breaking instruction.
4043 unsigned ARMBaseInstrInfo::
4044 getPartialRegUpdateClearance(const MachineInstr *MI,
4046 const TargetRegisterInfo *TRI) const {
4047 if (!SwiftPartialUpdateClearance ||
4048 !(Subtarget.isSwift() || Subtarget.isCortexA15()))
4051 assert(TRI && "Need TRI instance");
4053 const MachineOperand &MO = MI->getOperand(OpNum);
4056 unsigned Reg = MO.getReg();
4059 switch(MI->getOpcode()) {
4060 // Normal instructions writing only an S-register.
4065 case ARM::VMOVv4i16:
4066 case ARM::VMOVv2i32:
4067 case ARM::VMOVv2f32:
4068 case ARM::VMOVv1i64:
4069 UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI);
4072 // Explicitly reads the dependency.
4073 case ARM::VLD1LNd32:
4080 // If this instruction actually reads a value from Reg, there is no unwanted
4082 if (UseOp != -1 && MI->getOperand(UseOp).readsReg())
4085 // We must be able to clobber the whole D-reg.
4086 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4087 // Virtual register must be a foo:ssub_0<def,undef> operand.
4088 if (!MO.getSubReg() || MI->readsVirtualRegister(Reg))
4090 } else if (ARM::SPRRegClass.contains(Reg)) {
4091 // Physical register: MI must define the full D-reg.
4092 unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0,
4094 if (!DReg || !MI->definesRegister(DReg, TRI))
4098 // MI has an unwanted D-register dependency.
4099 // Avoid defs in the previous N instructrions.
4100 return SwiftPartialUpdateClearance;
4103 // Break a partial register dependency after getPartialRegUpdateClearance
4104 // returned non-zero.
4105 void ARMBaseInstrInfo::
4106 breakPartialRegDependency(MachineBasicBlock::iterator MI,
4108 const TargetRegisterInfo *TRI) const {
4109 assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def");
4110 assert(TRI && "Need TRI instance");
4112 const MachineOperand &MO = MI->getOperand(OpNum);
4113 unsigned Reg = MO.getReg();
4114 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
4115 "Can't break virtual register dependencies.");
4116 unsigned DReg = Reg;
4118 // If MI defines an S-reg, find the corresponding D super-register.
4119 if (ARM::SPRRegClass.contains(Reg)) {
4120 DReg = ARM::D0 + (Reg - ARM::S0) / 2;
4121 assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken");
4124 assert(ARM::DPRRegClass.contains(DReg) && "Can only break D-reg deps");
4125 assert(MI->definesRegister(DReg, TRI) && "MI doesn't clobber full D-reg");
4127 // FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines
4128 // the full D-register by loading the same value to both lanes. The
4129 // instruction is micro-coded with 2 uops, so don't do this until we can
4130 // properly schedule micro-coded instuctions. The dispatcher stalls cause
4131 // too big regressions.
4133 // Insert the dependency-breaking FCONSTD before MI.
4134 // 96 is the encoding of 0.5, but the actual value doesn't matter here.
4135 AddDefaultPred(BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4136 get(ARM::FCONSTD), DReg).addImm(96));
4137 MI->addRegisterKilled(DReg, TRI, true);
4140 bool ARMBaseInstrInfo::hasNOP() const {
4141 return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
4144 bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
4145 if (MI->getNumOperands() < 4)
4147 unsigned ShOpVal = MI->getOperand(3).getImm();
4148 unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
4149 // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
4150 if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) ||
4151 ((ShImm == 1 || ShImm == 2) &&
4152 ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl))