1 //===-- ARMBaseInstrInfo.cpp - ARM Instruction Information ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the Base ARM implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
15 #include "ARMBaseInstrInfo.h"
16 #include "ARMBaseRegisterInfo.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMFeatures.h"
19 #include "ARMHazardRecognizer.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "MCTargetDesc/ARMAddressingModes.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/CodeGen/LiveVariables.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineJumpTableInfo.h"
28 #include "llvm/CodeGen/MachineMemOperand.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAGNodes.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/GlobalValue.h"
34 #include "llvm/MC/MCAsmInfo.h"
35 #include "llvm/Support/BranchProbability.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
40 #define GET_INSTRINFO_CTOR_DTOR
41 #include "ARMGenInstrInfo.inc"
46 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
47 cl::desc("Enable ARM 2-addr to 3-addr conv"));
50 WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true),
51 cl::desc("Widen ARM vmovs to vmovd when possible"));
53 static cl::opt<unsigned>
54 SwiftPartialUpdateClearance("swift-partial-update-clearance",
55 cl::Hidden, cl::init(12),
56 cl::desc("Clearance before partial register updates"));
58 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
60 uint16_t MLxOpc; // MLA / MLS opcode
61 uint16_t MulOpc; // Expanded multiplication opcode
62 uint16_t AddSubOpc; // Expanded add / sub opcode
63 bool NegAcc; // True if the acc is negated before the add / sub.
64 bool HasLane; // True if instruction has an extra "lane" operand.
67 static const ARM_MLxEntry ARM_MLxTable[] = {
68 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
70 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
71 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
72 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
73 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
74 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
75 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
76 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
77 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
80 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
81 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
82 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
83 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
84 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
85 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
86 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
87 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
90 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
91 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
93 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
94 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
95 assert(false && "Duplicated entries?");
96 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
97 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
101 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
102 // currently defaults to no prepass hazard recognizer.
103 ScheduleHazardRecognizer *ARMBaseInstrInfo::
104 CreateTargetHazardRecognizer(const TargetMachine *TM,
105 const ScheduleDAG *DAG) const {
106 if (usePreRAHazardRecognizer()) {
107 const InstrItineraryData *II = TM->getInstrItineraryData();
108 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
110 return TargetInstrInfo::CreateTargetHazardRecognizer(TM, DAG);
113 ScheduleHazardRecognizer *ARMBaseInstrInfo::
114 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
115 const ScheduleDAG *DAG) const {
116 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
117 return (ScheduleHazardRecognizer *)new ARMHazardRecognizer(II, DAG);
118 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
122 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
123 MachineBasicBlock::iterator &MBBI,
124 LiveVariables *LV) const {
125 // FIXME: Thumb2 support.
130 MachineInstr *MI = MBBI;
131 MachineFunction &MF = *MI->getParent()->getParent();
132 uint64_t TSFlags = MI->getDesc().TSFlags;
134 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
135 default: return NULL;
136 case ARMII::IndexModePre:
139 case ARMII::IndexModePost:
143 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
145 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
149 MachineInstr *UpdateMI = NULL;
150 MachineInstr *MemMI = NULL;
151 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
152 const MCInstrDesc &MCID = MI->getDesc();
153 unsigned NumOps = MCID.getNumOperands();
154 bool isLoad = !MI->mayStore();
155 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
156 const MachineOperand &Base = MI->getOperand(2);
157 const MachineOperand &Offset = MI->getOperand(NumOps-3);
158 unsigned WBReg = WB.getReg();
159 unsigned BaseReg = Base.getReg();
160 unsigned OffReg = Offset.getReg();
161 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
162 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
164 default: llvm_unreachable("Unknown indexed op!");
165 case ARMII::AddrMode2: {
166 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
167 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
169 if (ARM_AM::getSOImmVal(Amt) == -1)
170 // Can't encode it in a so_imm operand. This transformation will
171 // add more than 1 instruction. Abandon!
173 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
174 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
175 .addReg(BaseReg).addImm(Amt)
176 .addImm(Pred).addReg(0).addReg(0);
177 } else if (Amt != 0) {
178 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
179 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
180 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
181 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
182 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
183 .addImm(Pred).addReg(0).addReg(0);
185 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
186 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
187 .addReg(BaseReg).addReg(OffReg)
188 .addImm(Pred).addReg(0).addReg(0);
191 case ARMII::AddrMode3 : {
192 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
193 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
195 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
196 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
197 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
198 .addReg(BaseReg).addImm(Amt)
199 .addImm(Pred).addReg(0).addReg(0);
201 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
202 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
203 .addReg(BaseReg).addReg(OffReg)
204 .addImm(Pred).addReg(0).addReg(0);
209 std::vector<MachineInstr*> NewMIs;
212 MemMI = BuildMI(MF, MI->getDebugLoc(),
213 get(MemOpc), MI->getOperand(0).getReg())
214 .addReg(WBReg).addImm(0).addImm(Pred);
216 MemMI = BuildMI(MF, MI->getDebugLoc(),
217 get(MemOpc)).addReg(MI->getOperand(1).getReg())
218 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
219 NewMIs.push_back(MemMI);
220 NewMIs.push_back(UpdateMI);
223 MemMI = BuildMI(MF, MI->getDebugLoc(),
224 get(MemOpc), MI->getOperand(0).getReg())
225 .addReg(BaseReg).addImm(0).addImm(Pred);
227 MemMI = BuildMI(MF, MI->getDebugLoc(),
228 get(MemOpc)).addReg(MI->getOperand(1).getReg())
229 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
231 UpdateMI->getOperand(0).setIsDead();
232 NewMIs.push_back(UpdateMI);
233 NewMIs.push_back(MemMI);
236 // Transfer LiveVariables states, kill / dead info.
238 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
239 MachineOperand &MO = MI->getOperand(i);
240 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
241 unsigned Reg = MO.getReg();
243 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
245 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
247 LV->addVirtualRegisterDead(Reg, NewMI);
249 if (MO.isUse() && MO.isKill()) {
250 for (unsigned j = 0; j < 2; ++j) {
251 // Look at the two new MI's in reverse order.
252 MachineInstr *NewMI = NewMIs[j];
253 if (!NewMI->readsRegister(Reg))
255 LV->addVirtualRegisterKilled(Reg, NewMI);
256 if (VI.removeKill(MI))
257 VI.Kills.push_back(NewMI);
265 MFI->insert(MBBI, NewMIs[1]);
266 MFI->insert(MBBI, NewMIs[0]);
272 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
273 MachineBasicBlock *&FBB,
274 SmallVectorImpl<MachineOperand> &Cond,
275 bool AllowModify) const {
279 MachineBasicBlock::iterator I = MBB.end();
280 if (I == MBB.begin())
281 return false; // Empty blocks are easy.
284 // Walk backwards from the end of the basic block until the branch is
285 // analyzed or we give up.
286 while (isPredicated(I) || I->isTerminator() || I->isDebugValue()) {
288 // Flag to be raised on unanalyzeable instructions. This is useful in cases
289 // where we want to clean up on the end of the basic block before we bail
291 bool CantAnalyze = false;
293 // Skip over DEBUG values and predicated nonterminators.
294 while (I->isDebugValue() || !I->isTerminator()) {
295 if (I == MBB.begin())
300 if (isIndirectBranchOpcode(I->getOpcode()) ||
301 isJumpTableBranchOpcode(I->getOpcode())) {
302 // Indirect branches and jump tables can't be analyzed, but we still want
303 // to clean up any instructions at the tail of the basic block.
305 } else if (isUncondBranchOpcode(I->getOpcode())) {
306 TBB = I->getOperand(0).getMBB();
307 } else if (isCondBranchOpcode(I->getOpcode())) {
308 // Bail out if we encounter multiple conditional branches.
312 assert(!FBB && "FBB should have been null.");
314 TBB = I->getOperand(0).getMBB();
315 Cond.push_back(I->getOperand(1));
316 Cond.push_back(I->getOperand(2));
317 } else if (I->isReturn()) {
318 // Returns can't be analyzed, but we should run cleanup.
319 CantAnalyze = !isPredicated(I);
321 // We encountered other unrecognized terminator. Bail out immediately.
325 // Cleanup code - to be run for unpredicated unconditional branches and
327 if (!isPredicated(I) &&
328 (isUncondBranchOpcode(I->getOpcode()) ||
329 isIndirectBranchOpcode(I->getOpcode()) ||
330 isJumpTableBranchOpcode(I->getOpcode()) ||
332 // Forget any previous condition branch information - it no longer applies.
336 // If we can modify the function, delete everything below this
337 // unconditional branch.
339 MachineBasicBlock::iterator DI = std::next(I);
340 while (DI != MBB.end()) {
341 MachineInstr *InstToDelete = DI;
343 InstToDelete->eraseFromParent();
351 if (I == MBB.begin())
357 // We made it past the terminators without bailing out - we must have
358 // analyzed this branch successfully.
363 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
364 MachineBasicBlock::iterator I = MBB.end();
365 if (I == MBB.begin()) return 0;
367 while (I->isDebugValue()) {
368 if (I == MBB.begin())
372 if (!isUncondBranchOpcode(I->getOpcode()) &&
373 !isCondBranchOpcode(I->getOpcode()))
376 // Remove the branch.
377 I->eraseFromParent();
381 if (I == MBB.begin()) return 1;
383 if (!isCondBranchOpcode(I->getOpcode()))
386 // Remove the branch.
387 I->eraseFromParent();
392 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
393 MachineBasicBlock *FBB,
394 const SmallVectorImpl<MachineOperand> &Cond,
396 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
397 int BOpc = !AFI->isThumbFunction()
398 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
399 int BccOpc = !AFI->isThumbFunction()
400 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
401 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
403 // Shouldn't be a fall through.
404 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
405 assert((Cond.size() == 2 || Cond.size() == 0) &&
406 "ARM branch conditions have two components!");
409 if (Cond.empty()) { // Unconditional branch?
411 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
413 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
415 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
416 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
420 // Two-way conditional branch.
421 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
422 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
424 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
426 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
430 bool ARMBaseInstrInfo::
431 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
432 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
433 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
437 bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
438 if (MI->isBundle()) {
439 MachineBasicBlock::const_instr_iterator I = MI;
440 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
441 while (++I != E && I->isInsideBundle()) {
442 int PIdx = I->findFirstPredOperandIdx();
443 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
449 int PIdx = MI->findFirstPredOperandIdx();
450 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
453 bool ARMBaseInstrInfo::
454 PredicateInstruction(MachineInstr *MI,
455 const SmallVectorImpl<MachineOperand> &Pred) const {
456 unsigned Opc = MI->getOpcode();
457 if (isUncondBranchOpcode(Opc)) {
458 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
459 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
460 .addImm(Pred[0].getImm())
461 .addReg(Pred[1].getReg());
465 int PIdx = MI->findFirstPredOperandIdx();
467 MachineOperand &PMO = MI->getOperand(PIdx);
468 PMO.setImm(Pred[0].getImm());
469 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
475 bool ARMBaseInstrInfo::
476 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
477 const SmallVectorImpl<MachineOperand> &Pred2) const {
478 if (Pred1.size() > 2 || Pred2.size() > 2)
481 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
482 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
492 return CC2 == ARMCC::HI;
494 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
496 return CC2 == ARMCC::GT;
498 return CC2 == ARMCC::LT;
502 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
503 std::vector<MachineOperand> &Pred) const {
505 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
506 const MachineOperand &MO = MI->getOperand(i);
507 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
508 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) {
517 /// isPredicable - Return true if the specified instruction can be predicated.
518 /// By default, this returns true for every instruction with a
519 /// PredicateOperand.
520 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
521 if (!MI->isPredicable())
524 ARMFunctionInfo *AFI =
525 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
527 if (AFI->isThumb2Function()) {
528 if (getSubtarget().restrictIT())
529 return isV8EligibleForIT(MI);
530 } else { // non-Thumb
531 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON)
539 template <> bool IsCPSRDead<MachineInstr>(MachineInstr *MI) {
540 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
541 const MachineOperand &MO = MI->getOperand(i);
542 if (!MO.isReg() || MO.isUndef() || MO.isUse())
544 if (MO.getReg() != ARM::CPSR)
549 // all definitions of CPSR are dead
554 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
555 LLVM_ATTRIBUTE_NOINLINE
556 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
558 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
560 assert(JTI < JT.size());
561 return JT[JTI].MBBs.size();
564 /// GetInstSize - Return the size of the specified MachineInstr.
566 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
567 const MachineBasicBlock &MBB = *MI->getParent();
568 const MachineFunction *MF = MBB.getParent();
569 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
571 const MCInstrDesc &MCID = MI->getDesc();
573 return MCID.getSize();
575 // If this machine instr is an inline asm, measure it.
576 if (MI->getOpcode() == ARM::INLINEASM)
577 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
578 unsigned Opc = MI->getOpcode();
581 // pseudo-instruction sizes are zero.
583 case TargetOpcode::BUNDLE:
584 return getInstBundleLength(MI);
585 case ARM::MOVi16_ga_pcrel:
586 case ARM::MOVTi16_ga_pcrel:
587 case ARM::t2MOVi16_ga_pcrel:
588 case ARM::t2MOVTi16_ga_pcrel:
591 case ARM::t2MOVi32imm:
593 case ARM::CONSTPOOL_ENTRY:
594 // If this machine instr is a constant pool entry, its size is recorded as
596 return MI->getOperand(2).getImm();
597 case ARM::Int_eh_sjlj_longjmp:
599 case ARM::tInt_eh_sjlj_longjmp:
601 case ARM::Int_eh_sjlj_setjmp:
602 case ARM::Int_eh_sjlj_setjmp_nofp:
604 case ARM::tInt_eh_sjlj_setjmp:
605 case ARM::t2Int_eh_sjlj_setjmp:
606 case ARM::t2Int_eh_sjlj_setjmp_nofp:
614 case ARM::t2TBH_JT: {
615 // These are jumptable branches, i.e. a branch followed by an inlined
616 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
617 // entry is one byte; TBH two byte each.
618 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
619 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
620 unsigned NumOps = MCID.getNumOperands();
621 MachineOperand JTOP =
622 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
623 unsigned JTI = JTOP.getIndex();
624 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
626 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
627 assert(JTI < JT.size());
628 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
629 // 4 aligned. The assembler / linker may add 2 byte padding just before
630 // the JT entries. The size does not include this padding; the
631 // constant islands pass does separate bookkeeping for it.
632 // FIXME: If we know the size of the function is less than (1 << 16) *2
633 // bytes, we can use 16-bit entries instead. Then there won't be an
635 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
636 unsigned NumEntries = getNumJTEntries(JT, JTI);
637 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
638 // Make sure the instruction that follows TBB is 2-byte aligned.
639 // FIXME: Constant island pass should insert an "ALIGN" instruction
642 return NumEntries * EntrySize + InstSize;
647 unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
649 MachineBasicBlock::const_instr_iterator I = MI;
650 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
651 while (++I != E && I->isInsideBundle()) {
652 assert(!I->isBundle() && "No nested bundle!");
653 Size += GetInstSizeInBytes(&*I);
658 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
659 MachineBasicBlock::iterator I, DebugLoc DL,
660 unsigned DestReg, unsigned SrcReg,
661 bool KillSrc) const {
662 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
663 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
665 if (GPRDest && GPRSrc) {
666 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
667 .addReg(SrcReg, getKillRegState(KillSrc))));
671 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
672 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
675 if (SPRDest && SPRSrc)
677 else if (GPRDest && SPRSrc)
679 else if (SPRDest && GPRSrc)
681 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
683 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
687 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
688 MIB.addReg(SrcReg, getKillRegState(KillSrc));
689 if (Opc == ARM::VORRq)
690 MIB.addReg(SrcReg, getKillRegState(KillSrc));
695 // Handle register classes that require multiple instructions.
696 unsigned BeginIdx = 0;
697 unsigned SubRegs = 0;
700 // Use VORRq when possible.
701 if (ARM::QQPRRegClass.contains(DestReg, SrcReg)) {
703 BeginIdx = ARM::qsub_0;
705 } else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg)) {
707 BeginIdx = ARM::qsub_0;
709 // Fall back to VMOVD.
710 } else if (ARM::DPairRegClass.contains(DestReg, SrcReg)) {
712 BeginIdx = ARM::dsub_0;
714 } else if (ARM::DTripleRegClass.contains(DestReg, SrcReg)) {
716 BeginIdx = ARM::dsub_0;
718 } else if (ARM::DQuadRegClass.contains(DestReg, SrcReg)) {
720 BeginIdx = ARM::dsub_0;
722 } else if (ARM::GPRPairRegClass.contains(DestReg, SrcReg)) {
723 Opc = Subtarget.isThumb2() ? ARM::tMOVr : ARM::MOVr;
724 BeginIdx = ARM::gsub_0;
726 } else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg)) {
728 BeginIdx = ARM::dsub_0;
731 } else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg)) {
733 BeginIdx = ARM::dsub_0;
736 } else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg)) {
738 BeginIdx = ARM::dsub_0;
743 assert(Opc && "Impossible reg-to-reg copy");
745 const TargetRegisterInfo *TRI = &getRegisterInfo();
746 MachineInstrBuilder Mov;
748 // Copy register tuples backward when the first Dest reg overlaps with SrcReg.
749 if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) {
750 BeginIdx = BeginIdx + ((SubRegs - 1) * Spacing);
754 SmallSet<unsigned, 4> DstRegs;
756 for (unsigned i = 0; i != SubRegs; ++i) {
757 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i * Spacing);
758 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i * Spacing);
759 assert(Dst && Src && "Bad sub-register");
761 assert(!DstRegs.count(Src) && "destructive vector copy");
764 Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst).addReg(Src);
765 // VORR takes two source operands.
766 if (Opc == ARM::VORRq)
768 Mov = AddDefaultPred(Mov);
770 if (Opc == ARM::MOVr)
771 Mov = AddDefaultCC(Mov);
773 // Add implicit super-register defs and kills to the last instruction.
774 Mov->addRegisterDefined(DestReg, TRI);
776 Mov->addRegisterKilled(SrcReg, TRI);
779 const MachineInstrBuilder &
780 ARMBaseInstrInfo::AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
781 unsigned SubIdx, unsigned State,
782 const TargetRegisterInfo *TRI) const {
784 return MIB.addReg(Reg, State);
786 if (TargetRegisterInfo::isPhysicalRegister(Reg))
787 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
788 return MIB.addReg(Reg, State, SubIdx);
791 void ARMBaseInstrInfo::
792 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
793 unsigned SrcReg, bool isKill, int FI,
794 const TargetRegisterClass *RC,
795 const TargetRegisterInfo *TRI) const {
797 if (I != MBB.end()) DL = I->getDebugLoc();
798 MachineFunction &MF = *MBB.getParent();
799 MachineFrameInfo &MFI = *MF.getFrameInfo();
800 unsigned Align = MFI.getObjectAlignment(FI);
802 MachineMemOperand *MMO =
803 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
804 MachineMemOperand::MOStore,
805 MFI.getObjectSize(FI),
808 switch (RC->getSize()) {
810 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
811 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
812 .addReg(SrcReg, getKillRegState(isKill))
813 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
814 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
815 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
816 .addReg(SrcReg, getKillRegState(isKill))
817 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
819 llvm_unreachable("Unknown reg class!");
822 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
823 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
824 .addReg(SrcReg, getKillRegState(isKill))
825 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
826 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
827 if (Subtarget.hasV5TEOps()) {
828 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::STRD));
829 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
830 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
831 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
835 // Fallback to STM instruction, which has existed since the dawn of
837 MachineInstrBuilder MIB =
838 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STMIA))
839 .addFrameIndex(FI).addMemOperand(MMO));
840 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
841 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
844 llvm_unreachable("Unknown reg class!");
847 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
848 // Use aligned spills if the stack can be realigned.
849 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
850 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
851 .addFrameIndex(FI).addImm(16)
852 .addReg(SrcReg, getKillRegState(isKill))
853 .addMemOperand(MMO));
855 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
856 .addReg(SrcReg, getKillRegState(isKill))
858 .addMemOperand(MMO));
861 llvm_unreachable("Unknown reg class!");
864 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
865 // Use aligned spills if the stack can be realigned.
866 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
867 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
868 .addFrameIndex(FI).addImm(16)
869 .addReg(SrcReg, getKillRegState(isKill))
870 .addMemOperand(MMO));
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 AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
881 llvm_unreachable("Unknown reg class!");
884 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
885 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
886 // FIXME: It's possible to only store part of the QQ register if the
887 // spilled def has a sub-register index.
888 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
889 .addFrameIndex(FI).addImm(16)
890 .addReg(SrcReg, getKillRegState(isKill))
891 .addMemOperand(MMO));
893 MachineInstrBuilder MIB =
894 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
897 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
898 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
899 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
900 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
903 llvm_unreachable("Unknown reg class!");
906 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
907 MachineInstrBuilder MIB =
908 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
911 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
912 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
913 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
914 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
915 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
916 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
917 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
918 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
920 llvm_unreachable("Unknown reg class!");
923 llvm_unreachable("Unknown reg class!");
928 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
929 int &FrameIndex) const {
930 switch (MI->getOpcode()) {
933 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
934 if (MI->getOperand(1).isFI() &&
935 MI->getOperand(2).isReg() &&
936 MI->getOperand(3).isImm() &&
937 MI->getOperand(2).getReg() == 0 &&
938 MI->getOperand(3).getImm() == 0) {
939 FrameIndex = MI->getOperand(1).getIndex();
940 return MI->getOperand(0).getReg();
948 if (MI->getOperand(1).isFI() &&
949 MI->getOperand(2).isImm() &&
950 MI->getOperand(2).getImm() == 0) {
951 FrameIndex = MI->getOperand(1).getIndex();
952 return MI->getOperand(0).getReg();
956 case ARM::VST1d64TPseudo:
957 case ARM::VST1d64QPseudo:
958 if (MI->getOperand(0).isFI() &&
959 MI->getOperand(2).getSubReg() == 0) {
960 FrameIndex = MI->getOperand(0).getIndex();
961 return MI->getOperand(2).getReg();
965 if (MI->getOperand(1).isFI() &&
966 MI->getOperand(0).getSubReg() == 0) {
967 FrameIndex = MI->getOperand(1).getIndex();
968 return MI->getOperand(0).getReg();
976 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
977 int &FrameIndex) const {
978 const MachineMemOperand *Dummy;
979 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
982 void ARMBaseInstrInfo::
983 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
984 unsigned DestReg, int FI,
985 const TargetRegisterClass *RC,
986 const TargetRegisterInfo *TRI) const {
988 if (I != MBB.end()) DL = I->getDebugLoc();
989 MachineFunction &MF = *MBB.getParent();
990 MachineFrameInfo &MFI = *MF.getFrameInfo();
991 unsigned Align = MFI.getObjectAlignment(FI);
992 MachineMemOperand *MMO =
993 MF.getMachineMemOperand(
994 MachinePointerInfo::getFixedStack(FI),
995 MachineMemOperand::MOLoad,
996 MFI.getObjectSize(FI),
999 switch (RC->getSize()) {
1001 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
1002 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
1003 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
1005 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
1006 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
1007 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
1009 llvm_unreachable("Unknown reg class!");
1012 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
1013 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
1014 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
1015 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
1016 MachineInstrBuilder MIB;
1018 if (Subtarget.hasV5TEOps()) {
1019 MIB = BuildMI(MBB, I, DL, get(ARM::LDRD));
1020 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1021 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1022 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
1024 AddDefaultPred(MIB);
1026 // Fallback to LDM instruction, which has existed since the dawn of
1028 MIB = AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDMIA))
1029 .addFrameIndex(FI).addMemOperand(MMO));
1030 MIB = AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1031 MIB = AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1034 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1035 MIB.addReg(DestReg, RegState::ImplicitDefine);
1037 llvm_unreachable("Unknown reg class!");
1040 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
1041 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1042 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
1043 .addFrameIndex(FI).addImm(16)
1044 .addMemOperand(MMO));
1046 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
1048 .addMemOperand(MMO));
1051 llvm_unreachable("Unknown reg class!");
1054 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
1055 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1056 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
1057 .addFrameIndex(FI).addImm(16)
1058 .addMemOperand(MMO));
1060 MachineInstrBuilder MIB =
1061 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1063 .addMemOperand(MMO));
1064 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1065 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1066 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1067 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1068 MIB.addReg(DestReg, RegState::ImplicitDefine);
1071 llvm_unreachable("Unknown reg class!");
1074 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
1075 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1076 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
1077 .addFrameIndex(FI).addImm(16)
1078 .addMemOperand(MMO));
1080 MachineInstrBuilder MIB =
1081 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1083 .addMemOperand(MMO);
1084 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1085 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1086 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1087 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1088 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1089 MIB.addReg(DestReg, RegState::ImplicitDefine);
1092 llvm_unreachable("Unknown reg class!");
1095 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
1096 MachineInstrBuilder MIB =
1097 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1099 .addMemOperand(MMO);
1100 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1101 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1102 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1103 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1104 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
1105 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
1106 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
1107 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
1108 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1109 MIB.addReg(DestReg, RegState::ImplicitDefine);
1111 llvm_unreachable("Unknown reg class!");
1114 llvm_unreachable("Unknown regclass!");
1119 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1120 int &FrameIndex) const {
1121 switch (MI->getOpcode()) {
1124 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
1125 if (MI->getOperand(1).isFI() &&
1126 MI->getOperand(2).isReg() &&
1127 MI->getOperand(3).isImm() &&
1128 MI->getOperand(2).getReg() == 0 &&
1129 MI->getOperand(3).getImm() == 0) {
1130 FrameIndex = MI->getOperand(1).getIndex();
1131 return MI->getOperand(0).getReg();
1139 if (MI->getOperand(1).isFI() &&
1140 MI->getOperand(2).isImm() &&
1141 MI->getOperand(2).getImm() == 0) {
1142 FrameIndex = MI->getOperand(1).getIndex();
1143 return MI->getOperand(0).getReg();
1147 case ARM::VLD1d64TPseudo:
1148 case ARM::VLD1d64QPseudo:
1149 if (MI->getOperand(1).isFI() &&
1150 MI->getOperand(0).getSubReg() == 0) {
1151 FrameIndex = MI->getOperand(1).getIndex();
1152 return MI->getOperand(0).getReg();
1156 if (MI->getOperand(1).isFI() &&
1157 MI->getOperand(0).getSubReg() == 0) {
1158 FrameIndex = MI->getOperand(1).getIndex();
1159 return MI->getOperand(0).getReg();
1167 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1168 int &FrameIndex) const {
1169 const MachineMemOperand *Dummy;
1170 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1173 bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
1174 // This hook gets to expand COPY instructions before they become
1175 // copyPhysReg() calls. Look for VMOVS instructions that can legally be
1176 // widened to VMOVD. We prefer the VMOVD when possible because it may be
1177 // changed into a VORR that can go down the NEON pipeline.
1178 if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15())
1181 // Look for a copy between even S-registers. That is where we keep floats
1182 // when using NEON v2f32 instructions for f32 arithmetic.
1183 unsigned DstRegS = MI->getOperand(0).getReg();
1184 unsigned SrcRegS = MI->getOperand(1).getReg();
1185 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
1188 const TargetRegisterInfo *TRI = &getRegisterInfo();
1189 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
1191 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
1193 if (!DstRegD || !SrcRegD)
1196 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
1197 // legal if the COPY already defines the full DstRegD, and it isn't a
1198 // sub-register insertion.
1199 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
1202 // A dead copy shouldn't show up here, but reject it just in case.
1203 if (MI->getOperand(0).isDead())
1206 // All clear, widen the COPY.
1207 DEBUG(dbgs() << "widening: " << *MI);
1208 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1210 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
1211 // or some other super-register.
1212 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
1213 if (ImpDefIdx != -1)
1214 MI->RemoveOperand(ImpDefIdx);
1216 // Change the opcode and operands.
1217 MI->setDesc(get(ARM::VMOVD));
1218 MI->getOperand(0).setReg(DstRegD);
1219 MI->getOperand(1).setReg(SrcRegD);
1220 AddDefaultPred(MIB);
1222 // We are now reading SrcRegD instead of SrcRegS. This may upset the
1223 // register scavenger and machine verifier, so we need to indicate that we
1224 // are reading an undefined value from SrcRegD, but a proper value from
1226 MI->getOperand(1).setIsUndef();
1227 MIB.addReg(SrcRegS, RegState::Implicit);
1229 // SrcRegD may actually contain an unrelated value in the ssub_1
1230 // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
1231 if (MI->getOperand(1).isKill()) {
1232 MI->getOperand(1).setIsKill(false);
1233 MI->addRegisterKilled(SrcRegS, TRI, true);
1236 DEBUG(dbgs() << "replaced by: " << *MI);
1240 /// Create a copy of a const pool value. Update CPI to the new index and return
1242 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1243 MachineConstantPool *MCP = MF.getConstantPool();
1244 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1246 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1247 assert(MCPE.isMachineConstantPoolEntry() &&
1248 "Expecting a machine constantpool entry!");
1249 ARMConstantPoolValue *ACPV =
1250 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1252 unsigned PCLabelId = AFI->createPICLabelUId();
1253 ARMConstantPoolValue *NewCPV = 0;
1255 // FIXME: The below assumes PIC relocation model and that the function
1256 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1257 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1258 // instructions, so that's probably OK, but is PIC always correct when
1260 if (ACPV->isGlobalValue())
1261 NewCPV = ARMConstantPoolConstant::
1262 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
1264 else if (ACPV->isExtSymbol())
1265 NewCPV = ARMConstantPoolSymbol::
1266 Create(MF.getFunction()->getContext(),
1267 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
1268 else if (ACPV->isBlockAddress())
1269 NewCPV = ARMConstantPoolConstant::
1270 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
1271 ARMCP::CPBlockAddress, 4);
1272 else if (ACPV->isLSDA())
1273 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
1275 else if (ACPV->isMachineBasicBlock())
1276 NewCPV = ARMConstantPoolMBB::
1277 Create(MF.getFunction()->getContext(),
1278 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
1280 llvm_unreachable("Unexpected ARM constantpool value type!!");
1281 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1285 void ARMBaseInstrInfo::
1286 reMaterialize(MachineBasicBlock &MBB,
1287 MachineBasicBlock::iterator I,
1288 unsigned DestReg, unsigned SubIdx,
1289 const MachineInstr *Orig,
1290 const TargetRegisterInfo &TRI) const {
1291 unsigned Opcode = Orig->getOpcode();
1294 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1295 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1299 case ARM::tLDRpci_pic:
1300 case ARM::t2LDRpci_pic: {
1301 MachineFunction &MF = *MBB.getParent();
1302 unsigned CPI = Orig->getOperand(1).getIndex();
1303 unsigned PCLabelId = duplicateCPV(MF, CPI);
1304 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1306 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1307 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1314 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1315 MachineInstr *MI = TargetInstrInfo::duplicate(Orig, MF);
1316 switch(Orig->getOpcode()) {
1317 case ARM::tLDRpci_pic:
1318 case ARM::t2LDRpci_pic: {
1319 unsigned CPI = Orig->getOperand(1).getIndex();
1320 unsigned PCLabelId = duplicateCPV(MF, CPI);
1321 Orig->getOperand(1).setIndex(CPI);
1322 Orig->getOperand(2).setImm(PCLabelId);
1329 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1330 const MachineInstr *MI1,
1331 const MachineRegisterInfo *MRI) const {
1332 int Opcode = MI0->getOpcode();
1333 if (Opcode == ARM::t2LDRpci ||
1334 Opcode == ARM::t2LDRpci_pic ||
1335 Opcode == ARM::tLDRpci ||
1336 Opcode == ARM::tLDRpci_pic ||
1337 Opcode == ARM::LDRLIT_ga_pcrel ||
1338 Opcode == ARM::LDRLIT_ga_pcrel_ldr ||
1339 Opcode == ARM::tLDRLIT_ga_pcrel ||
1340 Opcode == ARM::MOV_ga_pcrel ||
1341 Opcode == ARM::MOV_ga_pcrel_ldr ||
1342 Opcode == ARM::t2MOV_ga_pcrel) {
1343 if (MI1->getOpcode() != Opcode)
1345 if (MI0->getNumOperands() != MI1->getNumOperands())
1348 const MachineOperand &MO0 = MI0->getOperand(1);
1349 const MachineOperand &MO1 = MI1->getOperand(1);
1350 if (MO0.getOffset() != MO1.getOffset())
1353 if (Opcode == ARM::LDRLIT_ga_pcrel ||
1354 Opcode == ARM::LDRLIT_ga_pcrel_ldr ||
1355 Opcode == ARM::tLDRLIT_ga_pcrel ||
1356 Opcode == ARM::MOV_ga_pcrel ||
1357 Opcode == ARM::MOV_ga_pcrel_ldr ||
1358 Opcode == ARM::t2MOV_ga_pcrel)
1359 // Ignore the PC labels.
1360 return MO0.getGlobal() == MO1.getGlobal();
1362 const MachineFunction *MF = MI0->getParent()->getParent();
1363 const MachineConstantPool *MCP = MF->getConstantPool();
1364 int CPI0 = MO0.getIndex();
1365 int CPI1 = MO1.getIndex();
1366 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1367 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1368 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1369 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1370 if (isARMCP0 && isARMCP1) {
1371 ARMConstantPoolValue *ACPV0 =
1372 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1373 ARMConstantPoolValue *ACPV1 =
1374 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1375 return ACPV0->hasSameValue(ACPV1);
1376 } else if (!isARMCP0 && !isARMCP1) {
1377 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1380 } else if (Opcode == ARM::PICLDR) {
1381 if (MI1->getOpcode() != Opcode)
1383 if (MI0->getNumOperands() != MI1->getNumOperands())
1386 unsigned Addr0 = MI0->getOperand(1).getReg();
1387 unsigned Addr1 = MI1->getOperand(1).getReg();
1388 if (Addr0 != Addr1) {
1390 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1391 !TargetRegisterInfo::isVirtualRegister(Addr1))
1394 // This assumes SSA form.
1395 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1396 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1397 // Check if the loaded value, e.g. a constantpool of a global address, are
1399 if (!produceSameValue(Def0, Def1, MRI))
1403 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1404 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1405 const MachineOperand &MO0 = MI0->getOperand(i);
1406 const MachineOperand &MO1 = MI1->getOperand(i);
1407 if (!MO0.isIdenticalTo(MO1))
1413 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1416 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1417 /// determine if two loads are loading from the same base address. It should
1418 /// only return true if the base pointers are the same and the only differences
1419 /// between the two addresses is the offset. It also returns the offsets by
1422 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1423 /// is permanently disabled.
1424 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1426 int64_t &Offset2) const {
1427 // Don't worry about Thumb: just ARM and Thumb2.
1428 if (Subtarget.isThumb1Only()) return false;
1430 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1433 switch (Load1->getMachineOpcode()) {
1447 case ARM::t2LDRSHi8:
1449 case ARM::t2LDRBi12:
1450 case ARM::t2LDRSHi12:
1454 switch (Load2->getMachineOpcode()) {
1467 case ARM::t2LDRSHi8:
1469 case ARM::t2LDRBi12:
1470 case ARM::t2LDRSHi12:
1474 // Check if base addresses and chain operands match.
1475 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1476 Load1->getOperand(4) != Load2->getOperand(4))
1479 // Index should be Reg0.
1480 if (Load1->getOperand(3) != Load2->getOperand(3))
1483 // Determine the offsets.
1484 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1485 isa<ConstantSDNode>(Load2->getOperand(1))) {
1486 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1487 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1494 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1495 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1496 /// be scheduled togther. On some targets if two loads are loading from
1497 /// addresses in the same cache line, it's better if they are scheduled
1498 /// together. This function takes two integers that represent the load offsets
1499 /// from the common base address. It returns true if it decides it's desirable
1500 /// to schedule the two loads together. "NumLoads" is the number of loads that
1501 /// have already been scheduled after Load1.
1503 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1504 /// is permanently disabled.
1505 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1506 int64_t Offset1, int64_t Offset2,
1507 unsigned NumLoads) const {
1508 // Don't worry about Thumb: just ARM and Thumb2.
1509 if (Subtarget.isThumb1Only()) return false;
1511 assert(Offset2 > Offset1);
1513 if ((Offset2 - Offset1) / 8 > 64)
1516 // Check if the machine opcodes are different. If they are different
1517 // then we consider them to not be of the same base address,
1518 // EXCEPT in the case of Thumb2 byte loads where one is LDRBi8 and the other LDRBi12.
1519 // In this case, they are considered to be the same because they are different
1520 // encoding forms of the same basic instruction.
1521 if ((Load1->getMachineOpcode() != Load2->getMachineOpcode()) &&
1522 !((Load1->getMachineOpcode() == ARM::t2LDRBi8 &&
1523 Load2->getMachineOpcode() == ARM::t2LDRBi12) ||
1524 (Load1->getMachineOpcode() == ARM::t2LDRBi12 &&
1525 Load2->getMachineOpcode() == ARM::t2LDRBi8)))
1526 return false; // FIXME: overly conservative?
1528 // Four loads in a row should be sufficient.
1535 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1536 const MachineBasicBlock *MBB,
1537 const MachineFunction &MF) const {
1538 // Debug info is never a scheduling boundary. It's necessary to be explicit
1539 // due to the special treatment of IT instructions below, otherwise a
1540 // dbg_value followed by an IT will result in the IT instruction being
1541 // considered a scheduling hazard, which is wrong. It should be the actual
1542 // instruction preceding the dbg_value instruction(s), just like it is
1543 // when debug info is not present.
1544 if (MI->isDebugValue())
1547 // Terminators and labels can't be scheduled around.
1548 if (MI->isTerminator() || MI->isPosition())
1551 // Treat the start of the IT block as a scheduling boundary, but schedule
1552 // t2IT along with all instructions following it.
1553 // FIXME: This is a big hammer. But the alternative is to add all potential
1554 // true and anti dependencies to IT block instructions as implicit operands
1555 // to the t2IT instruction. The added compile time and complexity does not
1557 MachineBasicBlock::const_iterator I = MI;
1558 // Make sure to skip any dbg_value instructions
1559 while (++I != MBB->end() && I->isDebugValue())
1561 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1564 // Don't attempt to schedule around any instruction that defines
1565 // a stack-oriented pointer, as it's unlikely to be profitable. This
1566 // saves compile time, because it doesn't require every single
1567 // stack slot reference to depend on the instruction that does the
1569 // Calls don't actually change the stack pointer, even if they have imp-defs.
1570 // No ARM calling conventions change the stack pointer. (X86 calling
1571 // conventions sometimes do).
1572 if (!MI->isCall() && MI->definesRegister(ARM::SP))
1578 bool ARMBaseInstrInfo::
1579 isProfitableToIfCvt(MachineBasicBlock &MBB,
1580 unsigned NumCycles, unsigned ExtraPredCycles,
1581 const BranchProbability &Probability) const {
1585 // Attempt to estimate the relative costs of predication versus branching.
1586 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1587 UnpredCost /= Probability.getDenominator();
1588 UnpredCost += 1; // The branch itself
1589 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1591 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1594 bool ARMBaseInstrInfo::
1595 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1596 unsigned TCycles, unsigned TExtra,
1597 MachineBasicBlock &FMBB,
1598 unsigned FCycles, unsigned FExtra,
1599 const BranchProbability &Probability) const {
1600 if (!TCycles || !FCycles)
1603 // Attempt to estimate the relative costs of predication versus branching.
1604 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1605 TUnpredCost /= Probability.getDenominator();
1607 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1608 unsigned FUnpredCost = Comp * FCycles;
1609 FUnpredCost /= Probability.getDenominator();
1611 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1612 UnpredCost += 1; // The branch itself
1613 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1615 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1619 ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
1620 MachineBasicBlock &FMBB) const {
1621 // Reduce false anti-dependencies to let Swift's out-of-order execution
1622 // engine do its thing.
1623 return Subtarget.isSwift();
1626 /// getInstrPredicate - If instruction is predicated, returns its predicate
1627 /// condition, otherwise returns AL. It also returns the condition code
1628 /// register by reference.
1630 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1631 int PIdx = MI->findFirstPredOperandIdx();
1637 PredReg = MI->getOperand(PIdx+1).getReg();
1638 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1642 int llvm::getMatchingCondBranchOpcode(int Opc) {
1647 if (Opc == ARM::t2B)
1650 llvm_unreachable("Unknown unconditional branch opcode!");
1653 /// commuteInstruction - Handle commutable instructions.
1655 ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1656 switch (MI->getOpcode()) {
1658 case ARM::t2MOVCCr: {
1659 // MOVCC can be commuted by inverting the condition.
1660 unsigned PredReg = 0;
1661 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
1662 // MOVCC AL can't be inverted. Shouldn't happen.
1663 if (CC == ARMCC::AL || PredReg != ARM::CPSR)
1665 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
1668 // After swapping the MOVCC operands, also invert the condition.
1669 MI->getOperand(MI->findFirstPredOperandIdx())
1670 .setImm(ARMCC::getOppositeCondition(CC));
1674 return TargetInstrInfo::commuteInstruction(MI, NewMI);
1677 /// Identify instructions that can be folded into a MOVCC instruction, and
1678 /// return the defining instruction.
1679 static MachineInstr *canFoldIntoMOVCC(unsigned Reg,
1680 const MachineRegisterInfo &MRI,
1681 const TargetInstrInfo *TII) {
1682 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1684 if (!MRI.hasOneNonDBGUse(Reg))
1686 MachineInstr *MI = MRI.getVRegDef(Reg);
1689 // MI is folded into the MOVCC by predicating it.
1690 if (!MI->isPredicable())
1692 // Check if MI has any non-dead defs or physreg uses. This also detects
1693 // predicated instructions which will be reading CPSR.
1694 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
1695 const MachineOperand &MO = MI->getOperand(i);
1696 // Reject frame index operands, PEI can't handle the predicated pseudos.
1697 if (MO.isFI() || MO.isCPI() || MO.isJTI())
1701 // MI can't have any tied operands, that would conflict with predication.
1704 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1706 if (MO.isDef() && !MO.isDead())
1709 bool DontMoveAcrossStores = true;
1710 if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ 0, DontMoveAcrossStores))
1715 bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
1716 SmallVectorImpl<MachineOperand> &Cond,
1717 unsigned &TrueOp, unsigned &FalseOp,
1718 bool &Optimizable) const {
1719 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1720 "Unknown select instruction");
1725 // 3: Condition code.
1729 Cond.push_back(MI->getOperand(3));
1730 Cond.push_back(MI->getOperand(4));
1731 // We can always fold a def.
1736 MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
1737 bool PreferFalse) const {
1738 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1739 "Unknown select instruction");
1740 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1741 MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
1742 bool Invert = !DefMI;
1744 DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this);
1748 // Find new register class to use.
1749 MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
1750 unsigned DestReg = MI->getOperand(0).getReg();
1751 const TargetRegisterClass *PreviousClass = MRI.getRegClass(FalseReg.getReg());
1752 if (!MRI.constrainRegClass(DestReg, PreviousClass))
1755 // Create a new predicated version of DefMI.
1756 // Rfalse is the first use.
1757 MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1758 DefMI->getDesc(), DestReg);
1760 // Copy all the DefMI operands, excluding its (null) predicate.
1761 const MCInstrDesc &DefDesc = DefMI->getDesc();
1762 for (unsigned i = 1, e = DefDesc.getNumOperands();
1763 i != e && !DefDesc.OpInfo[i].isPredicate(); ++i)
1764 NewMI.addOperand(DefMI->getOperand(i));
1766 unsigned CondCode = MI->getOperand(3).getImm();
1768 NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode)));
1770 NewMI.addImm(CondCode);
1771 NewMI.addOperand(MI->getOperand(4));
1773 // DefMI is not the -S version that sets CPSR, so add an optional %noreg.
1774 if (NewMI->hasOptionalDef())
1775 AddDefaultCC(NewMI);
1777 // The output register value when the predicate is false is an implicit
1778 // register operand tied to the first def.
1779 // The tie makes the register allocator ensure the FalseReg is allocated the
1780 // same register as operand 0.
1781 FalseReg.setImplicit();
1782 NewMI.addOperand(FalseReg);
1783 NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
1785 // The caller will erase MI, but not DefMI.
1786 DefMI->eraseFromParent();
1790 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
1791 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
1794 /// This will go away once we can teach tblgen how to set the optional CPSR def
1796 struct AddSubFlagsOpcodePair {
1798 uint16_t MachineOpc;
1801 static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
1802 {ARM::ADDSri, ARM::ADDri},
1803 {ARM::ADDSrr, ARM::ADDrr},
1804 {ARM::ADDSrsi, ARM::ADDrsi},
1805 {ARM::ADDSrsr, ARM::ADDrsr},
1807 {ARM::SUBSri, ARM::SUBri},
1808 {ARM::SUBSrr, ARM::SUBrr},
1809 {ARM::SUBSrsi, ARM::SUBrsi},
1810 {ARM::SUBSrsr, ARM::SUBrsr},
1812 {ARM::RSBSri, ARM::RSBri},
1813 {ARM::RSBSrsi, ARM::RSBrsi},
1814 {ARM::RSBSrsr, ARM::RSBrsr},
1816 {ARM::t2ADDSri, ARM::t2ADDri},
1817 {ARM::t2ADDSrr, ARM::t2ADDrr},
1818 {ARM::t2ADDSrs, ARM::t2ADDrs},
1820 {ARM::t2SUBSri, ARM::t2SUBri},
1821 {ARM::t2SUBSrr, ARM::t2SUBrr},
1822 {ARM::t2SUBSrs, ARM::t2SUBrs},
1824 {ARM::t2RSBSri, ARM::t2RSBri},
1825 {ARM::t2RSBSrs, ARM::t2RSBrs},
1828 unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
1829 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
1830 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
1831 return AddSubFlagsOpcodeMap[i].MachineOpc;
1835 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1836 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1837 unsigned DestReg, unsigned BaseReg, int NumBytes,
1838 ARMCC::CondCodes Pred, unsigned PredReg,
1839 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1840 if (NumBytes == 0 && DestReg != BaseReg) {
1841 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), DestReg)
1842 .addReg(BaseReg, RegState::Kill)
1843 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1844 .setMIFlags(MIFlags);
1848 bool isSub = NumBytes < 0;
1849 if (isSub) NumBytes = -NumBytes;
1852 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1853 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1854 assert(ThisVal && "Didn't extract field correctly");
1856 // We will handle these bits from offset, clear them.
1857 NumBytes &= ~ThisVal;
1859 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1861 // Build the new ADD / SUB.
1862 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1863 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1864 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1865 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1866 .setMIFlags(MIFlags);
1871 static bool isAnySubRegLive(unsigned Reg, const TargetRegisterInfo *TRI,
1873 for (MCSubRegIterator Subreg(Reg, TRI, /* IncludeSelf */ true);
1874 Subreg.isValid(); ++Subreg)
1875 if (MI->getParent()->computeRegisterLiveness(TRI, *Subreg, MI) !=
1876 MachineBasicBlock::LQR_Dead)
1880 bool llvm::tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget,
1881 MachineFunction &MF, MachineInstr *MI,
1882 unsigned NumBytes) {
1883 // This optimisation potentially adds lots of load and store
1884 // micro-operations, it's only really a great benefit to code-size.
1885 if (!Subtarget.isMinSize())
1888 // If only one register is pushed/popped, LLVM can use an LDR/STR
1889 // instead. We can't modify those so make sure we're dealing with an
1890 // instruction we understand.
1891 bool IsPop = isPopOpcode(MI->getOpcode());
1892 bool IsPush = isPushOpcode(MI->getOpcode());
1893 if (!IsPush && !IsPop)
1896 bool IsVFPPushPop = MI->getOpcode() == ARM::VSTMDDB_UPD ||
1897 MI->getOpcode() == ARM::VLDMDIA_UPD;
1898 bool IsT1PushPop = MI->getOpcode() == ARM::tPUSH ||
1899 MI->getOpcode() == ARM::tPOP ||
1900 MI->getOpcode() == ARM::tPOP_RET;
1902 assert((IsT1PushPop || (MI->getOperand(0).getReg() == ARM::SP &&
1903 MI->getOperand(1).getReg() == ARM::SP)) &&
1904 "trying to fold sp update into non-sp-updating push/pop");
1906 // The VFP push & pop act on D-registers, so we can only fold an adjustment
1907 // by a multiple of 8 bytes in correctly. Similarly rN is 4-bytes. Don't try
1908 // if this is violated.
1909 if (NumBytes % (IsVFPPushPop ? 8 : 4) != 0)
1912 // ARM and Thumb2 push/pop insts have explicit "sp, sp" operands (+
1913 // pred) so the list starts at 4. Thumb1 starts after the predicate.
1914 int RegListIdx = IsT1PushPop ? 2 : 4;
1916 // Calculate the space we'll need in terms of registers.
1917 unsigned FirstReg = MI->getOperand(RegListIdx).getReg();
1918 unsigned RD0Reg, RegsNeeded;
1921 RegsNeeded = NumBytes / 8;
1924 RegsNeeded = NumBytes / 4;
1927 // We're going to have to strip all list operands off before
1928 // re-adding them since the order matters, so save the existing ones
1930 SmallVector<MachineOperand, 4> RegList;
1931 for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i)
1932 RegList.push_back(MI->getOperand(i));
1934 const TargetRegisterInfo *TRI = MF.getRegInfo().getTargetRegisterInfo();
1935 const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(&MF);
1937 // Now try to find enough space in the reglist to allocate NumBytes.
1938 for (unsigned CurReg = FirstReg - 1; CurReg >= RD0Reg && RegsNeeded;
1941 // Pushing any register is completely harmless, mark the
1942 // register involved as undef since we don't care about it in
1944 RegList.push_back(MachineOperand::CreateReg(CurReg, false, false,
1945 false, false, true));
1950 // However, we can only pop an extra register if it's not live. For
1951 // registers live within the function we might clobber a return value
1952 // register; the other way a register can be live here is if it's
1954 // TODO: Currently, computeRegisterLiveness() does not report "live" if a
1955 // sub reg is live. When computeRegisterLiveness() works for sub reg, it
1956 // can replace isAnySubRegLive().
1957 if (isCalleeSavedRegister(CurReg, CSRegs) ||
1958 isAnySubRegLive(CurReg, TRI, MI)) {
1959 // VFP pops don't allow holes in the register list, so any skip is fatal
1960 // for our transformation. GPR pops do, so we should just keep looking.
1967 // Mark the unimportant registers as <def,dead> in the POP.
1968 RegList.push_back(MachineOperand::CreateReg(CurReg, true, false, false,
1976 // Finally we know we can profitably perform the optimisation so go
1977 // ahead: strip all existing registers off and add them back again
1978 // in the right order.
1979 for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i)
1980 MI->RemoveOperand(i);
1982 // Add the complete list back in.
1983 MachineInstrBuilder MIB(MF, &*MI);
1984 for (int i = RegList.size() - 1; i >= 0; --i)
1985 MIB.addOperand(RegList[i]);
1990 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1991 unsigned FrameReg, int &Offset,
1992 const ARMBaseInstrInfo &TII) {
1993 unsigned Opcode = MI.getOpcode();
1994 const MCInstrDesc &Desc = MI.getDesc();
1995 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1998 // Memory operands in inline assembly always use AddrMode2.
1999 if (Opcode == ARM::INLINEASM)
2000 AddrMode = ARMII::AddrMode2;
2002 if (Opcode == ARM::ADDri) {
2003 Offset += MI.getOperand(FrameRegIdx+1).getImm();
2005 // Turn it into a move.
2006 MI.setDesc(TII.get(ARM::MOVr));
2007 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
2008 MI.RemoveOperand(FrameRegIdx+1);
2011 } else if (Offset < 0) {
2014 MI.setDesc(TII.get(ARM::SUBri));
2017 // Common case: small offset, fits into instruction.
2018 if (ARM_AM::getSOImmVal(Offset) != -1) {
2019 // Replace the FrameIndex with sp / fp
2020 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
2021 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
2026 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
2028 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
2029 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
2031 // We will handle these bits from offset, clear them.
2032 Offset &= ~ThisImmVal;
2034 // Get the properly encoded SOImmVal field.
2035 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
2036 "Bit extraction didn't work?");
2037 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
2039 unsigned ImmIdx = 0;
2041 unsigned NumBits = 0;
2044 case ARMII::AddrMode_i12: {
2045 ImmIdx = FrameRegIdx + 1;
2046 InstrOffs = MI.getOperand(ImmIdx).getImm();
2050 case ARMII::AddrMode2: {
2051 ImmIdx = FrameRegIdx+2;
2052 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
2053 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
2058 case ARMII::AddrMode3: {
2059 ImmIdx = FrameRegIdx+2;
2060 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
2061 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
2066 case ARMII::AddrMode4:
2067 case ARMII::AddrMode6:
2068 // Can't fold any offset even if it's zero.
2070 case ARMII::AddrMode5: {
2071 ImmIdx = FrameRegIdx+1;
2072 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
2073 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
2080 llvm_unreachable("Unsupported addressing mode!");
2083 Offset += InstrOffs * Scale;
2084 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
2090 // Attempt to fold address comp. if opcode has offset bits
2092 // Common case: small offset, fits into instruction.
2093 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
2094 int ImmedOffset = Offset / Scale;
2095 unsigned Mask = (1 << NumBits) - 1;
2096 if ((unsigned)Offset <= Mask * Scale) {
2097 // Replace the FrameIndex with sp
2098 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
2099 // FIXME: When addrmode2 goes away, this will simplify (like the
2100 // T2 version), as the LDR.i12 versions don't need the encoding
2101 // tricks for the offset value.
2103 if (AddrMode == ARMII::AddrMode_i12)
2104 ImmedOffset = -ImmedOffset;
2106 ImmedOffset |= 1 << NumBits;
2108 ImmOp.ChangeToImmediate(ImmedOffset);
2113 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
2114 ImmedOffset = ImmedOffset & Mask;
2116 if (AddrMode == ARMII::AddrMode_i12)
2117 ImmedOffset = -ImmedOffset;
2119 ImmedOffset |= 1 << NumBits;
2121 ImmOp.ChangeToImmediate(ImmedOffset);
2122 Offset &= ~(Mask*Scale);
2126 Offset = (isSub) ? -Offset : Offset;
2130 /// analyzeCompare - For a comparison instruction, return the source registers
2131 /// in SrcReg and SrcReg2 if having two register operands, and the value it
2132 /// compares against in CmpValue. Return true if the comparison instruction
2133 /// can be analyzed.
2134 bool ARMBaseInstrInfo::
2135 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
2136 int &CmpMask, int &CmpValue) const {
2137 switch (MI->getOpcode()) {
2141 SrcReg = MI->getOperand(0).getReg();
2144 CmpValue = MI->getOperand(1).getImm();
2148 SrcReg = MI->getOperand(0).getReg();
2149 SrcReg2 = MI->getOperand(1).getReg();
2155 SrcReg = MI->getOperand(0).getReg();
2157 CmpMask = MI->getOperand(1).getImm();
2165 /// isSuitableForMask - Identify a suitable 'and' instruction that
2166 /// operates on the given source register and applies the same mask
2167 /// as a 'tst' instruction. Provide a limited look-through for copies.
2168 /// When successful, MI will hold the found instruction.
2169 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
2170 int CmpMask, bool CommonUse) {
2171 switch (MI->getOpcode()) {
2174 if (CmpMask != MI->getOperand(2).getImm())
2176 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
2180 // Walk down one instruction which is potentially an 'and'.
2181 const MachineInstr &Copy = *MI;
2182 MachineBasicBlock::iterator AND(
2183 std::next(MachineBasicBlock::iterator(MI)));
2184 if (AND == MI->getParent()->end()) return false;
2186 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
2194 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2195 /// the condition code if we modify the instructions such that flags are
2197 inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) {
2199 default: return ARMCC::AL;
2200 case ARMCC::EQ: return ARMCC::EQ;
2201 case ARMCC::NE: return ARMCC::NE;
2202 case ARMCC::HS: return ARMCC::LS;
2203 case ARMCC::LO: return ARMCC::HI;
2204 case ARMCC::HI: return ARMCC::LO;
2205 case ARMCC::LS: return ARMCC::HS;
2206 case ARMCC::GE: return ARMCC::LE;
2207 case ARMCC::LT: return ARMCC::GT;
2208 case ARMCC::GT: return ARMCC::LT;
2209 case ARMCC::LE: return ARMCC::GE;
2213 /// isRedundantFlagInstr - check whether the first instruction, whose only
2214 /// purpose is to update flags, can be made redundant.
2215 /// CMPrr can be made redundant by SUBrr if the operands are the same.
2216 /// CMPri can be made redundant by SUBri if the operands are the same.
2217 /// This function can be extended later on.
2218 inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg,
2219 unsigned SrcReg2, int ImmValue,
2221 if ((CmpI->getOpcode() == ARM::CMPrr ||
2222 CmpI->getOpcode() == ARM::t2CMPrr) &&
2223 (OI->getOpcode() == ARM::SUBrr ||
2224 OI->getOpcode() == ARM::t2SUBrr) &&
2225 ((OI->getOperand(1).getReg() == SrcReg &&
2226 OI->getOperand(2).getReg() == SrcReg2) ||
2227 (OI->getOperand(1).getReg() == SrcReg2 &&
2228 OI->getOperand(2).getReg() == SrcReg)))
2231 if ((CmpI->getOpcode() == ARM::CMPri ||
2232 CmpI->getOpcode() == ARM::t2CMPri) &&
2233 (OI->getOpcode() == ARM::SUBri ||
2234 OI->getOpcode() == ARM::t2SUBri) &&
2235 OI->getOperand(1).getReg() == SrcReg &&
2236 OI->getOperand(2).getImm() == ImmValue)
2241 /// optimizeCompareInstr - Convert the instruction supplying the argument to the
2242 /// comparison into one that sets the zero bit in the flags register;
2243 /// Remove a redundant Compare instruction if an earlier instruction can set the
2244 /// flags in the same way as Compare.
2245 /// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two
2246 /// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the
2247 /// condition code of instructions which use the flags.
2248 bool ARMBaseInstrInfo::
2249 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
2250 int CmpMask, int CmpValue,
2251 const MachineRegisterInfo *MRI) const {
2252 // Get the unique definition of SrcReg.
2253 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
2254 if (!MI) return false;
2256 // Masked compares sometimes use the same register as the corresponding 'and'.
2257 if (CmpMask != ~0) {
2258 if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) {
2260 for (MachineRegisterInfo::use_instr_iterator
2261 UI = MRI->use_instr_begin(SrcReg), UE = MRI->use_instr_end();
2263 if (UI->getParent() != CmpInstr->getParent()) continue;
2264 MachineInstr *PotentialAND = &*UI;
2265 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) ||
2266 isPredicated(PotentialAND))
2271 if (!MI) return false;
2275 // Get ready to iterate backward from CmpInstr.
2276 MachineBasicBlock::iterator I = CmpInstr, E = MI,
2277 B = CmpInstr->getParent()->begin();
2279 // Early exit if CmpInstr is at the beginning of the BB.
2280 if (I == B) return false;
2282 // There are two possible candidates which can be changed to set CPSR:
2283 // One is MI, the other is a SUB instruction.
2284 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
2285 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
2286 MachineInstr *Sub = NULL;
2288 // MI is not a candidate for CMPrr.
2290 else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
2291 // Conservatively refuse to convert an instruction which isn't in the same
2292 // BB as the comparison.
2293 // For CMPri, we need to check Sub, thus we can't return here.
2294 if (CmpInstr->getOpcode() == ARM::CMPri ||
2295 CmpInstr->getOpcode() == ARM::t2CMPri)
2301 // Check that CPSR isn't set between the comparison instruction and the one we
2302 // want to change. At the same time, search for Sub.
2303 const TargetRegisterInfo *TRI = &getRegisterInfo();
2305 for (; I != E; --I) {
2306 const MachineInstr &Instr = *I;
2308 if (Instr.modifiesRegister(ARM::CPSR, TRI) ||
2309 Instr.readsRegister(ARM::CPSR, TRI))
2310 // This instruction modifies or uses CPSR after the one we want to
2311 // change. We can't do this transformation.
2314 // Check whether CmpInstr can be made redundant by the current instruction.
2315 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
2321 // The 'and' is below the comparison instruction.
2325 // Return false if no candidates exist.
2329 // The single candidate is called MI.
2332 // We can't use a predicated instruction - it doesn't always write the flags.
2333 if (isPredicated(MI))
2336 switch (MI->getOpcode()) {
2370 case ARM::t2EORri: {
2371 // Scan forward for the use of CPSR
2372 // When checking against MI: if it's a conditional code requires
2373 // checking of V bit, then this is not safe to do.
2374 // It is safe to remove CmpInstr if CPSR is redefined or killed.
2375 // If we are done with the basic block, we need to check whether CPSR is
2377 SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
2379 bool isSafe = false;
2381 E = CmpInstr->getParent()->end();
2382 while (!isSafe && ++I != E) {
2383 const MachineInstr &Instr = *I;
2384 for (unsigned IO = 0, EO = Instr.getNumOperands();
2385 !isSafe && IO != EO; ++IO) {
2386 const MachineOperand &MO = Instr.getOperand(IO);
2387 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
2391 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
2397 // Condition code is after the operand before CPSR except for VSELs.
2398 ARMCC::CondCodes CC;
2399 bool IsInstrVSel = true;
2400 switch (Instr.getOpcode()) {
2402 IsInstrVSel = false;
2403 CC = (ARMCC::CondCodes)Instr.getOperand(IO - 1).getImm();
2424 ARMCC::CondCodes NewCC = getSwappedCondition(CC);
2425 if (NewCC == ARMCC::AL)
2427 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
2428 // on CMP needs to be updated to be based on SUB.
2429 // Push the condition code operands to OperandsToUpdate.
2430 // If it is safe to remove CmpInstr, the condition code of these
2431 // operands will be modified.
2432 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
2433 Sub->getOperand(2).getReg() == SrcReg) {
2434 // VSel doesn't support condition code update.
2437 OperandsToUpdate.push_back(
2438 std::make_pair(&((*I).getOperand(IO - 1)), NewCC));
2443 // CPSR can be used multiple times, we should continue.
2456 // If CPSR is not killed nor re-defined, we should check whether it is
2457 // live-out. If it is live-out, do not optimize.
2459 MachineBasicBlock *MBB = CmpInstr->getParent();
2460 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
2461 SE = MBB->succ_end(); SI != SE; ++SI)
2462 if ((*SI)->isLiveIn(ARM::CPSR))
2466 // Toggle the optional operand to CPSR.
2467 MI->getOperand(5).setReg(ARM::CPSR);
2468 MI->getOperand(5).setIsDef(true);
2469 assert(!isPredicated(MI) && "Can't use flags from predicated instruction");
2470 CmpInstr->eraseFromParent();
2472 // Modify the condition code of operands in OperandsToUpdate.
2473 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2474 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2475 for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++)
2476 OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second);
2484 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
2485 MachineInstr *DefMI, unsigned Reg,
2486 MachineRegisterInfo *MRI) const {
2487 // Fold large immediates into add, sub, or, xor.
2488 unsigned DefOpc = DefMI->getOpcode();
2489 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
2491 if (!DefMI->getOperand(1).isImm())
2492 // Could be t2MOVi32imm <ga:xx>
2495 if (!MRI->hasOneNonDBGUse(Reg))
2498 const MCInstrDesc &DefMCID = DefMI->getDesc();
2499 if (DefMCID.hasOptionalDef()) {
2500 unsigned NumOps = DefMCID.getNumOperands();
2501 const MachineOperand &MO = DefMI->getOperand(NumOps-1);
2502 if (MO.getReg() == ARM::CPSR && !MO.isDead())
2503 // If DefMI defines CPSR and it is not dead, it's obviously not safe
2508 const MCInstrDesc &UseMCID = UseMI->getDesc();
2509 if (UseMCID.hasOptionalDef()) {
2510 unsigned NumOps = UseMCID.getNumOperands();
2511 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
2512 // If the instruction sets the flag, do not attempt this optimization
2513 // since it may change the semantics of the code.
2517 unsigned UseOpc = UseMI->getOpcode();
2518 unsigned NewUseOpc = 0;
2519 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
2520 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
2521 bool Commute = false;
2523 default: return false;
2531 case ARM::t2EORrr: {
2532 Commute = UseMI->getOperand(2).getReg() != Reg;
2539 NewUseOpc = ARM::SUBri;
2545 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
2547 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
2548 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
2551 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
2552 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
2553 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
2557 case ARM::t2SUBrr: {
2561 NewUseOpc = ARM::t2SUBri;
2566 case ARM::t2EORrr: {
2567 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
2569 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
2570 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
2573 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
2574 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
2575 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
2583 unsigned OpIdx = Commute ? 2 : 1;
2584 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
2585 bool isKill = UseMI->getOperand(OpIdx).isKill();
2586 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
2587 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
2588 UseMI, UseMI->getDebugLoc(),
2589 get(NewUseOpc), NewReg)
2590 .addReg(Reg1, getKillRegState(isKill))
2591 .addImm(SOImmValV1)));
2592 UseMI->setDesc(get(NewUseOpc));
2593 UseMI->getOperand(1).setReg(NewReg);
2594 UseMI->getOperand(1).setIsKill();
2595 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
2596 DefMI->eraseFromParent();
2600 static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData,
2601 const MachineInstr *MI) {
2602 switch (MI->getOpcode()) {
2604 const MCInstrDesc &Desc = MI->getDesc();
2605 int UOps = ItinData->getNumMicroOps(Desc.getSchedClass());
2606 assert(UOps >= 0 && "bad # UOps");
2614 unsigned ShOpVal = MI->getOperand(3).getImm();
2615 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2616 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2619 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2620 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2627 if (!MI->getOperand(2).getReg())
2630 unsigned ShOpVal = MI->getOperand(3).getImm();
2631 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2632 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2635 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2636 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2643 return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2;
2645 case ARM::LDRSB_POST:
2646 case ARM::LDRSH_POST: {
2647 unsigned Rt = MI->getOperand(0).getReg();
2648 unsigned Rm = MI->getOperand(3).getReg();
2649 return (Rt == Rm) ? 4 : 3;
2652 case ARM::LDR_PRE_REG:
2653 case ARM::LDRB_PRE_REG: {
2654 unsigned Rt = MI->getOperand(0).getReg();
2655 unsigned Rm = MI->getOperand(3).getReg();
2658 unsigned ShOpVal = MI->getOperand(4).getImm();
2659 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2660 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2663 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2664 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2669 case ARM::STR_PRE_REG:
2670 case ARM::STRB_PRE_REG: {
2671 unsigned ShOpVal = MI->getOperand(4).getImm();
2672 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2673 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2676 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2677 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2683 case ARM::STRH_PRE: {
2684 unsigned Rt = MI->getOperand(0).getReg();
2685 unsigned Rm = MI->getOperand(3).getReg();
2690 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub)
2694 case ARM::LDR_POST_REG:
2695 case ARM::LDRB_POST_REG:
2696 case ARM::LDRH_POST: {
2697 unsigned Rt = MI->getOperand(0).getReg();
2698 unsigned Rm = MI->getOperand(3).getReg();
2699 return (Rt == Rm) ? 3 : 2;
2702 case ARM::LDR_PRE_IMM:
2703 case ARM::LDRB_PRE_IMM:
2704 case ARM::LDR_POST_IMM:
2705 case ARM::LDRB_POST_IMM:
2706 case ARM::STRB_POST_IMM:
2707 case ARM::STRB_POST_REG:
2708 case ARM::STRB_PRE_IMM:
2709 case ARM::STRH_POST:
2710 case ARM::STR_POST_IMM:
2711 case ARM::STR_POST_REG:
2712 case ARM::STR_PRE_IMM:
2715 case ARM::LDRSB_PRE:
2716 case ARM::LDRSH_PRE: {
2717 unsigned Rm = MI->getOperand(3).getReg();
2720 unsigned Rt = MI->getOperand(0).getReg();
2723 unsigned ShOpVal = MI->getOperand(4).getImm();
2724 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2725 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2728 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2729 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2735 unsigned Rt = MI->getOperand(0).getReg();
2736 unsigned Rn = MI->getOperand(2).getReg();
2737 unsigned Rm = MI->getOperand(3).getReg();
2739 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2740 return (Rt == Rn) ? 3 : 2;
2744 unsigned Rm = MI->getOperand(3).getReg();
2746 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2750 case ARM::LDRD_POST:
2751 case ARM::t2LDRD_POST:
2754 case ARM::STRD_POST:
2755 case ARM::t2STRD_POST:
2758 case ARM::LDRD_PRE: {
2759 unsigned Rt = MI->getOperand(0).getReg();
2760 unsigned Rn = MI->getOperand(3).getReg();
2761 unsigned Rm = MI->getOperand(4).getReg();
2763 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2764 return (Rt == Rn) ? 4 : 3;
2767 case ARM::t2LDRD_PRE: {
2768 unsigned Rt = MI->getOperand(0).getReg();
2769 unsigned Rn = MI->getOperand(3).getReg();
2770 return (Rt == Rn) ? 4 : 3;
2773 case ARM::STRD_PRE: {
2774 unsigned Rm = MI->getOperand(4).getReg();
2776 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2780 case ARM::t2STRD_PRE:
2783 case ARM::t2LDR_POST:
2784 case ARM::t2LDRB_POST:
2785 case ARM::t2LDRB_PRE:
2786 case ARM::t2LDRSBi12:
2787 case ARM::t2LDRSBi8:
2788 case ARM::t2LDRSBpci:
2790 case ARM::t2LDRH_POST:
2791 case ARM::t2LDRH_PRE:
2793 case ARM::t2LDRSB_POST:
2794 case ARM::t2LDRSB_PRE:
2795 case ARM::t2LDRSH_POST:
2796 case ARM::t2LDRSH_PRE:
2797 case ARM::t2LDRSHi12:
2798 case ARM::t2LDRSHi8:
2799 case ARM::t2LDRSHpci:
2803 case ARM::t2LDRDi8: {
2804 unsigned Rt = MI->getOperand(0).getReg();
2805 unsigned Rn = MI->getOperand(2).getReg();
2806 return (Rt == Rn) ? 3 : 2;
2809 case ARM::t2STRB_POST:
2810 case ARM::t2STRB_PRE:
2813 case ARM::t2STRH_POST:
2814 case ARM::t2STRH_PRE:
2816 case ARM::t2STR_POST:
2817 case ARM::t2STR_PRE:
2823 // Return the number of 32-bit words loaded by LDM or stored by STM. If this
2824 // can't be easily determined return 0 (missing MachineMemOperand).
2826 // FIXME: The current MachineInstr design does not support relying on machine
2827 // mem operands to determine the width of a memory access. Instead, we expect
2828 // the target to provide this information based on the instruction opcode and
2829 // operands. However, using MachineMemOperand is a the best solution now for
2832 // 1) getNumMicroOps tries to infer LDM memory width from the total number of MI
2833 // operands. This is much more dangerous than using the MachineMemOperand
2834 // sizes because CodeGen passes can insert/remove optional machine operands. In
2835 // fact, it's totally incorrect for preRA passes and appears to be wrong for
2836 // postRA passes as well.
2838 // 2) getNumLDMAddresses is only used by the scheduling machine model and any
2839 // machine model that calls this should handle the unknown (zero size) case.
2841 // Long term, we should require a target hook that verifies MachineMemOperand
2842 // sizes during MC lowering. That target hook should be local to MC lowering
2843 // because we can't ensure that it is aware of other MI forms. Doing this will
2844 // ensure that MachineMemOperands are correctly propagated through all passes.
2845 unsigned ARMBaseInstrInfo::getNumLDMAddresses(const MachineInstr *MI) const {
2847 for (MachineInstr::mmo_iterator I = MI->memoperands_begin(),
2848 E = MI->memoperands_end(); I != E; ++I) {
2849 Size += (*I)->getSize();
2855 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
2856 const MachineInstr *MI) const {
2857 if (!ItinData || ItinData->isEmpty())
2860 const MCInstrDesc &Desc = MI->getDesc();
2861 unsigned Class = Desc.getSchedClass();
2862 int ItinUOps = ItinData->getNumMicroOps(Class);
2863 if (ItinUOps >= 0) {
2864 if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore()))
2865 return getNumMicroOpsSwiftLdSt(ItinData, MI);
2870 unsigned Opc = MI->getOpcode();
2873 llvm_unreachable("Unexpected multi-uops instruction!");
2878 // The number of uOps for load / store multiple are determined by the number
2881 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
2882 // same cycle. The scheduling for the first load / store must be done
2883 // separately by assuming the address is not 64-bit aligned.
2885 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
2886 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
2887 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
2889 case ARM::VLDMDIA_UPD:
2890 case ARM::VLDMDDB_UPD:
2892 case ARM::VLDMSIA_UPD:
2893 case ARM::VLDMSDB_UPD:
2895 case ARM::VSTMDIA_UPD:
2896 case ARM::VSTMDDB_UPD:
2898 case ARM::VSTMSIA_UPD:
2899 case ARM::VSTMSDB_UPD: {
2900 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
2901 return (NumRegs / 2) + (NumRegs % 2) + 1;
2904 case ARM::LDMIA_RET:
2909 case ARM::LDMIA_UPD:
2910 case ARM::LDMDA_UPD:
2911 case ARM::LDMDB_UPD:
2912 case ARM::LDMIB_UPD:
2917 case ARM::STMIA_UPD:
2918 case ARM::STMDA_UPD:
2919 case ARM::STMDB_UPD:
2920 case ARM::STMIB_UPD:
2922 case ARM::tLDMIA_UPD:
2923 case ARM::tSTMIA_UPD:
2927 case ARM::t2LDMIA_RET:
2930 case ARM::t2LDMIA_UPD:
2931 case ARM::t2LDMDB_UPD:
2934 case ARM::t2STMIA_UPD:
2935 case ARM::t2STMDB_UPD: {
2936 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
2937 if (Subtarget.isSwift()) {
2938 int UOps = 1 + NumRegs; // One for address computation, one for each ld / st.
2941 case ARM::VLDMDIA_UPD:
2942 case ARM::VLDMDDB_UPD:
2943 case ARM::VLDMSIA_UPD:
2944 case ARM::VLDMSDB_UPD:
2945 case ARM::VSTMDIA_UPD:
2946 case ARM::VSTMDDB_UPD:
2947 case ARM::VSTMSIA_UPD:
2948 case ARM::VSTMSDB_UPD:
2949 case ARM::LDMIA_UPD:
2950 case ARM::LDMDA_UPD:
2951 case ARM::LDMDB_UPD:
2952 case ARM::LDMIB_UPD:
2953 case ARM::STMIA_UPD:
2954 case ARM::STMDA_UPD:
2955 case ARM::STMDB_UPD:
2956 case ARM::STMIB_UPD:
2957 case ARM::tLDMIA_UPD:
2958 case ARM::tSTMIA_UPD:
2959 case ARM::t2LDMIA_UPD:
2960 case ARM::t2LDMDB_UPD:
2961 case ARM::t2STMIA_UPD:
2962 case ARM::t2STMDB_UPD:
2963 ++UOps; // One for base register writeback.
2965 case ARM::LDMIA_RET:
2967 case ARM::t2LDMIA_RET:
2968 UOps += 2; // One for base reg wb, one for write to pc.
2972 } else if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
2975 // 4 registers would be issued: 2, 2.
2976 // 5 registers would be issued: 2, 2, 1.
2977 int A8UOps = (NumRegs / 2);
2981 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2982 int A9UOps = (NumRegs / 2);
2983 // If there are odd number of registers or if it's not 64-bit aligned,
2984 // then it takes an extra AGU (Address Generation Unit) cycle.
2985 if ((NumRegs % 2) ||
2986 !MI->hasOneMemOperand() ||
2987 (*MI->memoperands_begin())->getAlignment() < 8)
2991 // Assume the worst.
2999 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
3000 const MCInstrDesc &DefMCID,
3002 unsigned DefIdx, unsigned DefAlign) const {
3003 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
3005 // Def is the address writeback.
3006 return ItinData->getOperandCycle(DefClass, DefIdx);
3009 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3010 // (regno / 2) + (regno % 2) + 1
3011 DefCycle = RegNo / 2 + 1;
3014 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3016 bool isSLoad = false;
3018 switch (DefMCID.getOpcode()) {
3021 case ARM::VLDMSIA_UPD:
3022 case ARM::VLDMSDB_UPD:
3027 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
3028 // then it takes an extra cycle.
3029 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
3032 // Assume the worst.
3033 DefCycle = RegNo + 2;
3040 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
3041 const MCInstrDesc &DefMCID,
3043 unsigned DefIdx, unsigned DefAlign) const {
3044 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
3046 // Def is the address writeback.
3047 return ItinData->getOperandCycle(DefClass, DefIdx);
3050 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3051 // 4 registers would be issued: 1, 2, 1.
3052 // 5 registers would be issued: 1, 2, 2.
3053 DefCycle = RegNo / 2;
3056 // Result latency is issue cycle + 2: E2.
3058 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3059 DefCycle = (RegNo / 2);
3060 // If there are odd number of registers or if it's not 64-bit aligned,
3061 // then it takes an extra AGU (Address Generation Unit) cycle.
3062 if ((RegNo % 2) || DefAlign < 8)
3064 // Result latency is AGU cycles + 2.
3067 // Assume the worst.
3068 DefCycle = RegNo + 2;
3075 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
3076 const MCInstrDesc &UseMCID,
3078 unsigned UseIdx, unsigned UseAlign) const {
3079 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
3081 return ItinData->getOperandCycle(UseClass, UseIdx);
3084 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3085 // (regno / 2) + (regno % 2) + 1
3086 UseCycle = RegNo / 2 + 1;
3089 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3091 bool isSStore = false;
3093 switch (UseMCID.getOpcode()) {
3096 case ARM::VSTMSIA_UPD:
3097 case ARM::VSTMSDB_UPD:
3102 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
3103 // then it takes an extra cycle.
3104 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
3107 // Assume the worst.
3108 UseCycle = RegNo + 2;
3115 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
3116 const MCInstrDesc &UseMCID,
3118 unsigned UseIdx, unsigned UseAlign) const {
3119 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
3121 return ItinData->getOperandCycle(UseClass, UseIdx);
3124 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) {
3125 UseCycle = RegNo / 2;
3130 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
3131 UseCycle = (RegNo / 2);
3132 // If there are odd number of registers or if it's not 64-bit aligned,
3133 // then it takes an extra AGU (Address Generation Unit) cycle.
3134 if ((RegNo % 2) || UseAlign < 8)
3137 // Assume the worst.
3144 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3145 const MCInstrDesc &DefMCID,
3146 unsigned DefIdx, unsigned DefAlign,
3147 const MCInstrDesc &UseMCID,
3148 unsigned UseIdx, unsigned UseAlign) const {
3149 unsigned DefClass = DefMCID.getSchedClass();
3150 unsigned UseClass = UseMCID.getSchedClass();
3152 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
3153 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
3155 // This may be a def / use of a variable_ops instruction, the operand
3156 // latency might be determinable dynamically. Let the target try to
3159 bool LdmBypass = false;
3160 switch (DefMCID.getOpcode()) {
3162 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3166 case ARM::VLDMDIA_UPD:
3167 case ARM::VLDMDDB_UPD:
3169 case ARM::VLDMSIA_UPD:
3170 case ARM::VLDMSDB_UPD:
3171 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
3174 case ARM::LDMIA_RET:
3179 case ARM::LDMIA_UPD:
3180 case ARM::LDMDA_UPD:
3181 case ARM::LDMDB_UPD:
3182 case ARM::LDMIB_UPD:
3184 case ARM::tLDMIA_UPD:
3186 case ARM::t2LDMIA_RET:
3189 case ARM::t2LDMIA_UPD:
3190 case ARM::t2LDMDB_UPD:
3192 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
3197 // We can't seem to determine the result latency of the def, assume it's 2.
3201 switch (UseMCID.getOpcode()) {
3203 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
3207 case ARM::VSTMDIA_UPD:
3208 case ARM::VSTMDDB_UPD:
3210 case ARM::VSTMSIA_UPD:
3211 case ARM::VSTMSDB_UPD:
3212 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3219 case ARM::STMIA_UPD:
3220 case ARM::STMDA_UPD:
3221 case ARM::STMDB_UPD:
3222 case ARM::STMIB_UPD:
3223 case ARM::tSTMIA_UPD:
3228 case ARM::t2STMIA_UPD:
3229 case ARM::t2STMDB_UPD:
3230 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3235 // Assume it's read in the first stage.
3238 UseCycle = DefCycle - UseCycle + 1;
3241 // It's a variable_ops instruction so we can't use DefIdx here. Just use
3242 // first def operand.
3243 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
3246 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
3247 UseClass, UseIdx)) {
3255 static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
3256 const MachineInstr *MI, unsigned Reg,
3257 unsigned &DefIdx, unsigned &Dist) {
3260 MachineBasicBlock::const_iterator I = MI; ++I;
3261 MachineBasicBlock::const_instr_iterator II = std::prev(I.getInstrIterator());
3262 assert(II->isInsideBundle() && "Empty bundle?");
3265 while (II->isInsideBundle()) {
3266 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
3273 assert(Idx != -1 && "Cannot find bundled definition!");
3278 static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
3279 const MachineInstr *MI, unsigned Reg,
3280 unsigned &UseIdx, unsigned &Dist) {
3283 MachineBasicBlock::const_instr_iterator II = MI; ++II;
3284 assert(II->isInsideBundle() && "Empty bundle?");
3285 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3287 // FIXME: This doesn't properly handle multiple uses.
3289 while (II != E && II->isInsideBundle()) {
3290 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
3293 if (II->getOpcode() != ARM::t2IT)
3307 /// Return the number of cycles to add to (or subtract from) the static
3308 /// itinerary based on the def opcode and alignment. The caller will ensure that
3309 /// adjusted latency is at least one cycle.
3310 static int adjustDefLatency(const ARMSubtarget &Subtarget,
3311 const MachineInstr *DefMI,
3312 const MCInstrDesc *DefMCID, unsigned DefAlign) {
3314 if (Subtarget.isCortexA8() || Subtarget.isLikeA9() || Subtarget.isCortexA7()) {
3315 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3316 // variants are one cycle cheaper.
3317 switch (DefMCID->getOpcode()) {
3321 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3322 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3324 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3331 case ARM::t2LDRSHs: {
3332 // Thumb2 mode: lsl only.
3333 unsigned ShAmt = DefMI->getOperand(3).getImm();
3334 if (ShAmt == 0 || ShAmt == 2)
3339 } else if (Subtarget.isSwift()) {
3340 // FIXME: Properly handle all of the latency adjustments for address
3342 switch (DefMCID->getOpcode()) {
3346 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3347 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
3348 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3351 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3352 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
3355 ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3362 case ARM::t2LDRSHs: {
3363 // Thumb2 mode: lsl only.
3364 unsigned ShAmt = DefMI->getOperand(3).getImm();
3365 if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3)
3372 if (DefAlign < 8 && Subtarget.isLikeA9()) {
3373 switch (DefMCID->getOpcode()) {
3379 case ARM::VLD1q8wb_fixed:
3380 case ARM::VLD1q16wb_fixed:
3381 case ARM::VLD1q32wb_fixed:
3382 case ARM::VLD1q64wb_fixed:
3383 case ARM::VLD1q8wb_register:
3384 case ARM::VLD1q16wb_register:
3385 case ARM::VLD1q32wb_register:
3386 case ARM::VLD1q64wb_register:
3393 case ARM::VLD2d8wb_fixed:
3394 case ARM::VLD2d16wb_fixed:
3395 case ARM::VLD2d32wb_fixed:
3396 case ARM::VLD2q8wb_fixed:
3397 case ARM::VLD2q16wb_fixed:
3398 case ARM::VLD2q32wb_fixed:
3399 case ARM::VLD2d8wb_register:
3400 case ARM::VLD2d16wb_register:
3401 case ARM::VLD2d32wb_register:
3402 case ARM::VLD2q8wb_register:
3403 case ARM::VLD2q16wb_register:
3404 case ARM::VLD2q32wb_register:
3409 case ARM::VLD3d8_UPD:
3410 case ARM::VLD3d16_UPD:
3411 case ARM::VLD3d32_UPD:
3412 case ARM::VLD1d64Twb_fixed:
3413 case ARM::VLD1d64Twb_register:
3414 case ARM::VLD3q8_UPD:
3415 case ARM::VLD3q16_UPD:
3416 case ARM::VLD3q32_UPD:
3421 case ARM::VLD4d8_UPD:
3422 case ARM::VLD4d16_UPD:
3423 case ARM::VLD4d32_UPD:
3424 case ARM::VLD1d64Qwb_fixed:
3425 case ARM::VLD1d64Qwb_register:
3426 case ARM::VLD4q8_UPD:
3427 case ARM::VLD4q16_UPD:
3428 case ARM::VLD4q32_UPD:
3429 case ARM::VLD1DUPq8:
3430 case ARM::VLD1DUPq16:
3431 case ARM::VLD1DUPq32:
3432 case ARM::VLD1DUPq8wb_fixed:
3433 case ARM::VLD1DUPq16wb_fixed:
3434 case ARM::VLD1DUPq32wb_fixed:
3435 case ARM::VLD1DUPq8wb_register:
3436 case ARM::VLD1DUPq16wb_register:
3437 case ARM::VLD1DUPq32wb_register:
3438 case ARM::VLD2DUPd8:
3439 case ARM::VLD2DUPd16:
3440 case ARM::VLD2DUPd32:
3441 case ARM::VLD2DUPd8wb_fixed:
3442 case ARM::VLD2DUPd16wb_fixed:
3443 case ARM::VLD2DUPd32wb_fixed:
3444 case ARM::VLD2DUPd8wb_register:
3445 case ARM::VLD2DUPd16wb_register:
3446 case ARM::VLD2DUPd32wb_register:
3447 case ARM::VLD4DUPd8:
3448 case ARM::VLD4DUPd16:
3449 case ARM::VLD4DUPd32:
3450 case ARM::VLD4DUPd8_UPD:
3451 case ARM::VLD4DUPd16_UPD:
3452 case ARM::VLD4DUPd32_UPD:
3454 case ARM::VLD1LNd16:
3455 case ARM::VLD1LNd32:
3456 case ARM::VLD1LNd8_UPD:
3457 case ARM::VLD1LNd16_UPD:
3458 case ARM::VLD1LNd32_UPD:
3460 case ARM::VLD2LNd16:
3461 case ARM::VLD2LNd32:
3462 case ARM::VLD2LNq16:
3463 case ARM::VLD2LNq32:
3464 case ARM::VLD2LNd8_UPD:
3465 case ARM::VLD2LNd16_UPD:
3466 case ARM::VLD2LNd32_UPD:
3467 case ARM::VLD2LNq16_UPD:
3468 case ARM::VLD2LNq32_UPD:
3470 case ARM::VLD4LNd16:
3471 case ARM::VLD4LNd32:
3472 case ARM::VLD4LNq16:
3473 case ARM::VLD4LNq32:
3474 case ARM::VLD4LNd8_UPD:
3475 case ARM::VLD4LNd16_UPD:
3476 case ARM::VLD4LNd32_UPD:
3477 case ARM::VLD4LNq16_UPD:
3478 case ARM::VLD4LNq32_UPD:
3479 // If the address is not 64-bit aligned, the latencies of these
3480 // instructions increases by one.
3491 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3492 const MachineInstr *DefMI, unsigned DefIdx,
3493 const MachineInstr *UseMI,
3494 unsigned UseIdx) const {
3495 // No operand latency. The caller may fall back to getInstrLatency.
3496 if (!ItinData || ItinData->isEmpty())
3499 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
3500 unsigned Reg = DefMO.getReg();
3501 const MCInstrDesc *DefMCID = &DefMI->getDesc();
3502 const MCInstrDesc *UseMCID = &UseMI->getDesc();
3504 unsigned DefAdj = 0;
3505 if (DefMI->isBundle()) {
3506 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
3507 DefMCID = &DefMI->getDesc();
3509 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
3510 DefMI->isRegSequence() || DefMI->isImplicitDef()) {
3514 unsigned UseAdj = 0;
3515 if (UseMI->isBundle()) {
3517 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
3518 Reg, NewUseIdx, UseAdj);
3524 UseMCID = &UseMI->getDesc();
3527 if (Reg == ARM::CPSR) {
3528 if (DefMI->getOpcode() == ARM::FMSTAT) {
3529 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
3530 return Subtarget.isLikeA9() ? 1 : 20;
3533 // CPSR set and branch can be paired in the same cycle.
3534 if (UseMI->isBranch())
3537 // Otherwise it takes the instruction latency (generally one).
3538 unsigned Latency = getInstrLatency(ItinData, DefMI);
3540 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
3541 // its uses. Instructions which are otherwise scheduled between them may
3542 // incur a code size penalty (not able to use the CPSR setting 16-bit
3544 if (Latency > 0 && Subtarget.isThumb2()) {
3545 const MachineFunction *MF = DefMI->getParent()->getParent();
3546 if (MF->getFunction()->getAttributes().
3547 hasAttribute(AttributeSet::FunctionIndex,
3548 Attribute::OptimizeForSize))
3554 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
3557 unsigned DefAlign = DefMI->hasOneMemOperand()
3558 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
3559 unsigned UseAlign = UseMI->hasOneMemOperand()
3560 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
3562 // Get the itinerary's latency if possible, and handle variable_ops.
3563 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
3564 *UseMCID, UseIdx, UseAlign);
3565 // Unable to find operand latency. The caller may resort to getInstrLatency.
3569 // Adjust for IT block position.
3570 int Adj = DefAdj + UseAdj;
3572 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3573 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
3574 if (Adj >= 0 || (int)Latency > -Adj) {
3575 return Latency + Adj;
3577 // Return the itinerary latency, which may be zero but not less than zero.
3582 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3583 SDNode *DefNode, unsigned DefIdx,
3584 SDNode *UseNode, unsigned UseIdx) const {
3585 if (!DefNode->isMachineOpcode())
3588 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
3590 if (isZeroCost(DefMCID.Opcode))
3593 if (!ItinData || ItinData->isEmpty())
3594 return DefMCID.mayLoad() ? 3 : 1;
3596 if (!UseNode->isMachineOpcode()) {
3597 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
3598 if (Subtarget.isLikeA9() || Subtarget.isSwift())
3599 return Latency <= 2 ? 1 : Latency - 1;
3601 return Latency <= 3 ? 1 : Latency - 2;
3604 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
3605 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
3606 unsigned DefAlign = !DefMN->memoperands_empty()
3607 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
3608 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
3609 unsigned UseAlign = !UseMN->memoperands_empty()
3610 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
3611 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
3612 UseMCID, UseIdx, UseAlign);
3615 (Subtarget.isCortexA8() || Subtarget.isLikeA9() ||
3616 Subtarget.isCortexA7())) {
3617 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3618 // variants are one cycle cheaper.
3619 switch (DefMCID.getOpcode()) {
3624 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3625 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3627 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3634 case ARM::t2LDRSHs: {
3635 // Thumb2 mode: lsl only.
3637 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3638 if (ShAmt == 0 || ShAmt == 2)
3643 } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) {
3644 // FIXME: Properly handle all of the latency adjustments for address
3646 switch (DefMCID.getOpcode()) {
3651 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3652 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3654 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3655 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3657 else if (ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3664 case ARM::t2LDRSHs: {
3665 // Thumb2 mode: lsl 0-3 only.
3672 if (DefAlign < 8 && Subtarget.isLikeA9())
3673 switch (DefMCID.getOpcode()) {
3679 case ARM::VLD1q8wb_register:
3680 case ARM::VLD1q16wb_register:
3681 case ARM::VLD1q32wb_register:
3682 case ARM::VLD1q64wb_register:
3683 case ARM::VLD1q8wb_fixed:
3684 case ARM::VLD1q16wb_fixed:
3685 case ARM::VLD1q32wb_fixed:
3686 case ARM::VLD1q64wb_fixed:
3690 case ARM::VLD2q8Pseudo:
3691 case ARM::VLD2q16Pseudo:
3692 case ARM::VLD2q32Pseudo:
3693 case ARM::VLD2d8wb_fixed:
3694 case ARM::VLD2d16wb_fixed:
3695 case ARM::VLD2d32wb_fixed:
3696 case ARM::VLD2q8PseudoWB_fixed:
3697 case ARM::VLD2q16PseudoWB_fixed:
3698 case ARM::VLD2q32PseudoWB_fixed:
3699 case ARM::VLD2d8wb_register:
3700 case ARM::VLD2d16wb_register:
3701 case ARM::VLD2d32wb_register:
3702 case ARM::VLD2q8PseudoWB_register:
3703 case ARM::VLD2q16PseudoWB_register:
3704 case ARM::VLD2q32PseudoWB_register:
3705 case ARM::VLD3d8Pseudo:
3706 case ARM::VLD3d16Pseudo:
3707 case ARM::VLD3d32Pseudo:
3708 case ARM::VLD1d64TPseudo:
3709 case ARM::VLD1d64TPseudoWB_fixed:
3710 case ARM::VLD3d8Pseudo_UPD:
3711 case ARM::VLD3d16Pseudo_UPD:
3712 case ARM::VLD3d32Pseudo_UPD:
3713 case ARM::VLD3q8Pseudo_UPD:
3714 case ARM::VLD3q16Pseudo_UPD:
3715 case ARM::VLD3q32Pseudo_UPD:
3716 case ARM::VLD3q8oddPseudo:
3717 case ARM::VLD3q16oddPseudo:
3718 case ARM::VLD3q32oddPseudo:
3719 case ARM::VLD3q8oddPseudo_UPD:
3720 case ARM::VLD3q16oddPseudo_UPD:
3721 case ARM::VLD3q32oddPseudo_UPD:
3722 case ARM::VLD4d8Pseudo:
3723 case ARM::VLD4d16Pseudo:
3724 case ARM::VLD4d32Pseudo:
3725 case ARM::VLD1d64QPseudo:
3726 case ARM::VLD1d64QPseudoWB_fixed:
3727 case ARM::VLD4d8Pseudo_UPD:
3728 case ARM::VLD4d16Pseudo_UPD:
3729 case ARM::VLD4d32Pseudo_UPD:
3730 case ARM::VLD4q8Pseudo_UPD:
3731 case ARM::VLD4q16Pseudo_UPD:
3732 case ARM::VLD4q32Pseudo_UPD:
3733 case ARM::VLD4q8oddPseudo:
3734 case ARM::VLD4q16oddPseudo:
3735 case ARM::VLD4q32oddPseudo:
3736 case ARM::VLD4q8oddPseudo_UPD:
3737 case ARM::VLD4q16oddPseudo_UPD:
3738 case ARM::VLD4q32oddPseudo_UPD:
3739 case ARM::VLD1DUPq8:
3740 case ARM::VLD1DUPq16:
3741 case ARM::VLD1DUPq32:
3742 case ARM::VLD1DUPq8wb_fixed:
3743 case ARM::VLD1DUPq16wb_fixed:
3744 case ARM::VLD1DUPq32wb_fixed:
3745 case ARM::VLD1DUPq8wb_register:
3746 case ARM::VLD1DUPq16wb_register:
3747 case ARM::VLD1DUPq32wb_register:
3748 case ARM::VLD2DUPd8:
3749 case ARM::VLD2DUPd16:
3750 case ARM::VLD2DUPd32:
3751 case ARM::VLD2DUPd8wb_fixed:
3752 case ARM::VLD2DUPd16wb_fixed:
3753 case ARM::VLD2DUPd32wb_fixed:
3754 case ARM::VLD2DUPd8wb_register:
3755 case ARM::VLD2DUPd16wb_register:
3756 case ARM::VLD2DUPd32wb_register:
3757 case ARM::VLD4DUPd8Pseudo:
3758 case ARM::VLD4DUPd16Pseudo:
3759 case ARM::VLD4DUPd32Pseudo:
3760 case ARM::VLD4DUPd8Pseudo_UPD:
3761 case ARM::VLD4DUPd16Pseudo_UPD:
3762 case ARM::VLD4DUPd32Pseudo_UPD:
3763 case ARM::VLD1LNq8Pseudo:
3764 case ARM::VLD1LNq16Pseudo:
3765 case ARM::VLD1LNq32Pseudo:
3766 case ARM::VLD1LNq8Pseudo_UPD:
3767 case ARM::VLD1LNq16Pseudo_UPD:
3768 case ARM::VLD1LNq32Pseudo_UPD:
3769 case ARM::VLD2LNd8Pseudo:
3770 case ARM::VLD2LNd16Pseudo:
3771 case ARM::VLD2LNd32Pseudo:
3772 case ARM::VLD2LNq16Pseudo:
3773 case ARM::VLD2LNq32Pseudo:
3774 case ARM::VLD2LNd8Pseudo_UPD:
3775 case ARM::VLD2LNd16Pseudo_UPD:
3776 case ARM::VLD2LNd32Pseudo_UPD:
3777 case ARM::VLD2LNq16Pseudo_UPD:
3778 case ARM::VLD2LNq32Pseudo_UPD:
3779 case ARM::VLD4LNd8Pseudo:
3780 case ARM::VLD4LNd16Pseudo:
3781 case ARM::VLD4LNd32Pseudo:
3782 case ARM::VLD4LNq16Pseudo:
3783 case ARM::VLD4LNq32Pseudo:
3784 case ARM::VLD4LNd8Pseudo_UPD:
3785 case ARM::VLD4LNd16Pseudo_UPD:
3786 case ARM::VLD4LNd32Pseudo_UPD:
3787 case ARM::VLD4LNq16Pseudo_UPD:
3788 case ARM::VLD4LNq32Pseudo_UPD:
3789 // If the address is not 64-bit aligned, the latencies of these
3790 // instructions increases by one.
3798 unsigned ARMBaseInstrInfo::getPredicationCost(const MachineInstr *MI) const {
3799 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3800 MI->isRegSequence() || MI->isImplicitDef())
3806 const MCInstrDesc &MCID = MI->getDesc();
3808 if (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR)) {
3809 // When predicated, CPSR is an additional source operand for CPSR updating
3810 // instructions, this apparently increases their latencies.
3816 unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3817 const MachineInstr *MI,
3818 unsigned *PredCost) const {
3819 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3820 MI->isRegSequence() || MI->isImplicitDef())
3823 // An instruction scheduler typically runs on unbundled instructions, however
3824 // other passes may query the latency of a bundled instruction.
3825 if (MI->isBundle()) {
3826 unsigned Latency = 0;
3827 MachineBasicBlock::const_instr_iterator I = MI;
3828 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3829 while (++I != E && I->isInsideBundle()) {
3830 if (I->getOpcode() != ARM::t2IT)
3831 Latency += getInstrLatency(ItinData, I, PredCost);
3836 const MCInstrDesc &MCID = MI->getDesc();
3837 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) {
3838 // When predicated, CPSR is an additional source operand for CPSR updating
3839 // instructions, this apparently increases their latencies.
3842 // Be sure to call getStageLatency for an empty itinerary in case it has a
3843 // valid MinLatency property.
3845 return MI->mayLoad() ? 3 : 1;
3847 unsigned Class = MCID.getSchedClass();
3849 // For instructions with variable uops, use uops as latency.
3850 if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0)
3851 return getNumMicroOps(ItinData, MI);
3853 // For the common case, fall back on the itinerary's latency.
3854 unsigned Latency = ItinData->getStageLatency(Class);
3856 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3857 unsigned DefAlign = MI->hasOneMemOperand()
3858 ? (*MI->memoperands_begin())->getAlignment() : 0;
3859 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
3860 if (Adj >= 0 || (int)Latency > -Adj) {
3861 return Latency + Adj;
3866 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3867 SDNode *Node) const {
3868 if (!Node->isMachineOpcode())
3871 if (!ItinData || ItinData->isEmpty())
3874 unsigned Opcode = Node->getMachineOpcode();
3877 return ItinData->getStageLatency(get(Opcode).getSchedClass());
3884 bool ARMBaseInstrInfo::
3885 hasHighOperandLatency(const InstrItineraryData *ItinData,
3886 const MachineRegisterInfo *MRI,
3887 const MachineInstr *DefMI, unsigned DefIdx,
3888 const MachineInstr *UseMI, unsigned UseIdx) const {
3889 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3890 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
3891 if (Subtarget.isCortexA8() &&
3892 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
3893 // CortexA8 VFP instructions are not pipelined.
3896 // Hoist VFP / NEON instructions with 4 or higher latency.
3897 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
3899 Latency = getInstrLatency(ItinData, DefMI);
3902 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
3903 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
3906 bool ARMBaseInstrInfo::
3907 hasLowDefLatency(const InstrItineraryData *ItinData,
3908 const MachineInstr *DefMI, unsigned DefIdx) const {
3909 if (!ItinData || ItinData->isEmpty())
3912 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3913 if (DDomain == ARMII::DomainGeneral) {
3914 unsigned DefClass = DefMI->getDesc().getSchedClass();
3915 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3916 return (DefCycle != -1 && DefCycle <= 2);
3921 bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
3922 StringRef &ErrInfo) const {
3923 if (convertAddSubFlagsOpcode(MI->getOpcode())) {
3924 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
3931 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
3932 unsigned &AddSubOpc,
3933 bool &NegAcc, bool &HasLane) const {
3934 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
3935 if (I == MLxEntryMap.end())
3938 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
3939 MulOpc = Entry.MulOpc;
3940 AddSubOpc = Entry.AddSubOpc;
3941 NegAcc = Entry.NegAcc;
3942 HasLane = Entry.HasLane;
3946 //===----------------------------------------------------------------------===//
3947 // Execution domains.
3948 //===----------------------------------------------------------------------===//
3950 // Some instructions go down the NEON pipeline, some go down the VFP pipeline,
3951 // and some can go down both. The vmov instructions go down the VFP pipeline,
3952 // but they can be changed to vorr equivalents that are executed by the NEON
3955 // We use the following execution domain numbering:
3963 // Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
3965 std::pair<uint16_t, uint16_t>
3966 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
3967 // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
3968 // if they are not predicated.
3969 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
3970 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3972 // CortexA9 is particularly picky about mixing the two and wants these
3974 if (Subtarget.isCortexA9() && !isPredicated(MI) &&
3975 (MI->getOpcode() == ARM::VMOVRS ||
3976 MI->getOpcode() == ARM::VMOVSR ||
3977 MI->getOpcode() == ARM::VMOVS))
3978 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3980 // No other instructions can be swizzled, so just determine their domain.
3981 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
3983 if (Domain & ARMII::DomainNEON)
3984 return std::make_pair(ExeNEON, 0);
3986 // Certain instructions can go either way on Cortex-A8.
3987 // Treat them as NEON instructions.
3988 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
3989 return std::make_pair(ExeNEON, 0);
3991 if (Domain & ARMII::DomainVFP)
3992 return std::make_pair(ExeVFP, 0);
3994 return std::make_pair(ExeGeneric, 0);
3997 static unsigned getCorrespondingDRegAndLane(const TargetRegisterInfo *TRI,
3998 unsigned SReg, unsigned &Lane) {
3999 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_0, &ARM::DPRRegClass);
4002 if (DReg != ARM::NoRegister)
4006 DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1, &ARM::DPRRegClass);
4008 assert(DReg && "S-register with no D super-register?");
4012 /// getImplicitSPRUseForDPRUse - Given a use of a DPR register and lane,
4013 /// set ImplicitSReg to a register number that must be marked as implicit-use or
4014 /// zero if no register needs to be defined as implicit-use.
4016 /// If the function cannot determine if an SPR should be marked implicit use or
4017 /// not, it returns false.
4019 /// This function handles cases where an instruction is being modified from taking
4020 /// an SPR to a DPR[Lane]. A use of the DPR is being added, which may conflict
4021 /// with an earlier def of an SPR corresponding to DPR[Lane^1] (i.e. the other
4022 /// lane of the DPR).
4024 /// If the other SPR is defined, an implicit-use of it should be added. Else,
4025 /// (including the case where the DPR itself is defined), it should not.
4027 static bool getImplicitSPRUseForDPRUse(const TargetRegisterInfo *TRI,
4029 unsigned DReg, unsigned Lane,
4030 unsigned &ImplicitSReg) {
4031 // If the DPR is defined or used already, the other SPR lane will be chained
4032 // correctly, so there is nothing to be done.
4033 if (MI->definesRegister(DReg, TRI) || MI->readsRegister(DReg, TRI)) {
4038 // Otherwise we need to go searching to see if the SPR is set explicitly.
4039 ImplicitSReg = TRI->getSubReg(DReg,
4040 (Lane & 1) ? ARM::ssub_0 : ARM::ssub_1);
4041 MachineBasicBlock::LivenessQueryResult LQR =
4042 MI->getParent()->computeRegisterLiveness(TRI, ImplicitSReg, MI);
4044 if (LQR == MachineBasicBlock::LQR_Live)
4046 else if (LQR == MachineBasicBlock::LQR_Unknown)
4049 // If the register is known not to be live, there is no need to add an
4056 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
4057 unsigned DstReg, SrcReg, DReg;
4059 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
4060 const TargetRegisterInfo *TRI = &getRegisterInfo();
4061 switch (MI->getOpcode()) {
4063 llvm_unreachable("cannot handle opcode!");
4066 if (Domain != ExeNEON)
4069 // Zap the predicate operands.
4070 assert(!isPredicated(MI) && "Cannot predicate a VORRd");
4072 // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
4073 DstReg = MI->getOperand(0).getReg();
4074 SrcReg = MI->getOperand(1).getReg();
4076 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4077 MI->RemoveOperand(i-1);
4079 // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits)
4080 MI->setDesc(get(ARM::VORRd));
4081 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
4086 if (Domain != ExeNEON)
4088 assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
4090 // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits)
4091 DstReg = MI->getOperand(0).getReg();
4092 SrcReg = MI->getOperand(1).getReg();
4094 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4095 MI->RemoveOperand(i-1);
4097 DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane);
4099 // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps)
4100 // Note that DSrc has been widened and the other lane may be undef, which
4101 // contaminates the entire register.
4102 MI->setDesc(get(ARM::VGETLNi32));
4103 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
4104 .addReg(DReg, RegState::Undef)
4107 // The old source should be an implicit use, otherwise we might think it
4108 // was dead before here.
4109 MIB.addReg(SrcReg, RegState::Implicit);
4112 if (Domain != ExeNEON)
4114 assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
4116 // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits)
4117 DstReg = MI->getOperand(0).getReg();
4118 SrcReg = MI->getOperand(1).getReg();
4120 DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane);
4122 unsigned ImplicitSReg;
4123 if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg))
4126 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4127 MI->RemoveOperand(i-1);
4129 // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps)
4130 // Again DDst may be undefined at the beginning of this instruction.
4131 MI->setDesc(get(ARM::VSETLNi32));
4132 MIB.addReg(DReg, RegState::Define)
4133 .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI)))
4136 AddDefaultPred(MIB);
4138 // The narrower destination must be marked as set to keep previous chains
4140 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4141 if (ImplicitSReg != 0)
4142 MIB.addReg(ImplicitSReg, RegState::Implicit);
4146 if (Domain != ExeNEON)
4149 // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits)
4150 DstReg = MI->getOperand(0).getReg();
4151 SrcReg = MI->getOperand(1).getReg();
4153 unsigned DstLane = 0, SrcLane = 0, DDst, DSrc;
4154 DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane);
4155 DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane);
4157 unsigned ImplicitSReg;
4158 if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg))
4161 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
4162 MI->RemoveOperand(i-1);
4165 // Destination can be:
4166 // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits)
4167 MI->setDesc(get(ARM::VDUPLN32d));
4168 MIB.addReg(DDst, RegState::Define)
4169 .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI)))
4171 AddDefaultPred(MIB);
4173 // Neither the source or the destination are naturally represented any
4174 // more, so add them in manually.
4175 MIB.addReg(DstReg, RegState::Implicit | RegState::Define);
4176 MIB.addReg(SrcReg, RegState::Implicit);
4177 if (ImplicitSReg != 0)
4178 MIB.addReg(ImplicitSReg, RegState::Implicit);
4182 // In general there's no single instruction that can perform an S <-> S
4183 // move in NEON space, but a pair of VEXT instructions *can* do the
4184 // job. It turns out that the VEXTs needed will only use DSrc once, with
4185 // the position based purely on the combination of lane-0 and lane-1
4186 // involved. For example
4187 // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1
4188 // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1
4189 // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1
4190 // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1
4192 // Pattern of the MachineInstrs is:
4193 // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits)
4194 MachineInstrBuilder NewMIB;
4195 NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4196 get(ARM::VEXTd32), DDst);
4198 // On the first instruction, both DSrc and DDst may be <undef> if present.
4199 // Specifically when the original instruction didn't have them as an
4201 unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst;
4202 bool CurUndef = !MI->readsRegister(CurReg, TRI);
4203 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
4205 CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst;
4206 CurUndef = !MI->readsRegister(CurReg, TRI);
4207 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
4210 AddDefaultPred(NewMIB);
4212 if (SrcLane == DstLane)
4213 NewMIB.addReg(SrcReg, RegState::Implicit);
4215 MI->setDesc(get(ARM::VEXTd32));
4216 MIB.addReg(DDst, RegState::Define);
4218 // On the second instruction, DDst has definitely been defined above, so
4219 // it is not <undef>. DSrc, if present, can be <undef> as above.
4220 CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst;
4221 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4222 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4224 CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst;
4225 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4226 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4229 AddDefaultPred(MIB);
4231 if (SrcLane != DstLane)
4232 MIB.addReg(SrcReg, RegState::Implicit);
4234 // As before, the original destination is no longer represented, add it
4236 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4237 if (ImplicitSReg != 0)
4238 MIB.addReg(ImplicitSReg, RegState::Implicit);
4245 //===----------------------------------------------------------------------===//
4246 // Partial register updates
4247 //===----------------------------------------------------------------------===//
4249 // Swift renames NEON registers with 64-bit granularity. That means any
4250 // instruction writing an S-reg implicitly reads the containing D-reg. The
4251 // problem is mostly avoided by translating f32 operations to v2f32 operations
4252 // on D-registers, but f32 loads are still a problem.
4254 // These instructions can load an f32 into a NEON register:
4256 // VLDRS - Only writes S, partial D update.
4257 // VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops.
4258 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
4260 // FCONSTD can be used as a dependency-breaking instruction.
4261 unsigned ARMBaseInstrInfo::
4262 getPartialRegUpdateClearance(const MachineInstr *MI,
4264 const TargetRegisterInfo *TRI) const {
4265 if (!SwiftPartialUpdateClearance ||
4266 !(Subtarget.isSwift() || Subtarget.isCortexA15()))
4269 assert(TRI && "Need TRI instance");
4271 const MachineOperand &MO = MI->getOperand(OpNum);
4274 unsigned Reg = MO.getReg();
4277 switch(MI->getOpcode()) {
4278 // Normal instructions writing only an S-register.
4283 case ARM::VMOVv4i16:
4284 case ARM::VMOVv2i32:
4285 case ARM::VMOVv2f32:
4286 case ARM::VMOVv1i64:
4287 UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI);
4290 // Explicitly reads the dependency.
4291 case ARM::VLD1LNd32:
4298 // If this instruction actually reads a value from Reg, there is no unwanted
4300 if (UseOp != -1 && MI->getOperand(UseOp).readsReg())
4303 // We must be able to clobber the whole D-reg.
4304 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4305 // Virtual register must be a foo:ssub_0<def,undef> operand.
4306 if (!MO.getSubReg() || MI->readsVirtualRegister(Reg))
4308 } else if (ARM::SPRRegClass.contains(Reg)) {
4309 // Physical register: MI must define the full D-reg.
4310 unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0,
4312 if (!DReg || !MI->definesRegister(DReg, TRI))
4316 // MI has an unwanted D-register dependency.
4317 // Avoid defs in the previous N instructrions.
4318 return SwiftPartialUpdateClearance;
4321 // Break a partial register dependency after getPartialRegUpdateClearance
4322 // returned non-zero.
4323 void ARMBaseInstrInfo::
4324 breakPartialRegDependency(MachineBasicBlock::iterator MI,
4326 const TargetRegisterInfo *TRI) const {
4327 assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def");
4328 assert(TRI && "Need TRI instance");
4330 const MachineOperand &MO = MI->getOperand(OpNum);
4331 unsigned Reg = MO.getReg();
4332 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
4333 "Can't break virtual register dependencies.");
4334 unsigned DReg = Reg;
4336 // If MI defines an S-reg, find the corresponding D super-register.
4337 if (ARM::SPRRegClass.contains(Reg)) {
4338 DReg = ARM::D0 + (Reg - ARM::S0) / 2;
4339 assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken");
4342 assert(ARM::DPRRegClass.contains(DReg) && "Can only break D-reg deps");
4343 assert(MI->definesRegister(DReg, TRI) && "MI doesn't clobber full D-reg");
4345 // FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines
4346 // the full D-register by loading the same value to both lanes. The
4347 // instruction is micro-coded with 2 uops, so don't do this until we can
4348 // properly schedule micro-coded instructions. The dispatcher stalls cause
4349 // too big regressions.
4351 // Insert the dependency-breaking FCONSTD before MI.
4352 // 96 is the encoding of 0.5, but the actual value doesn't matter here.
4353 AddDefaultPred(BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4354 get(ARM::FCONSTD), DReg).addImm(96));
4355 MI->addRegisterKilled(DReg, TRI, true);
4358 bool ARMBaseInstrInfo::hasNOP() const {
4359 return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
4362 bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
4363 if (MI->getNumOperands() < 4)
4365 unsigned ShOpVal = MI->getOperand(3).getImm();
4366 unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
4367 // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
4368 if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) ||
4369 ((ShImm == 1 || ShImm == 2) &&
4370 ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl))