1 //===-- ARMBaseInstrInfo.cpp - ARM Instruction Information ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the Base ARM implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "ARMBaseInstrInfo.h"
16 #include "ARMBaseRegisterInfo.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMHazardRecognizer.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "MCTargetDesc/ARMAddressingModes.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/CodeGen/LiveVariables.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineJumpTableInfo.h"
27 #include "llvm/CodeGen/MachineMemOperand.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/SelectionDAGNodes.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/GlobalValue.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/Support/BranchProbability.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
39 #define GET_INSTRINFO_CTOR
40 #include "ARMGenInstrInfo.inc"
45 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
46 cl::desc("Enable ARM 2-addr to 3-addr conv"));
49 WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true),
50 cl::desc("Widen ARM vmovs to vmovd when possible"));
52 static cl::opt<unsigned>
53 SwiftPartialUpdateClearance("swift-partial-update-clearance",
54 cl::Hidden, cl::init(12),
55 cl::desc("Clearance before partial register updates"));
57 /// ARM_MLxEntry - Record information about MLA / MLS instructions.
59 uint16_t MLxOpc; // MLA / MLS opcode
60 uint16_t MulOpc; // Expanded multiplication opcode
61 uint16_t AddSubOpc; // Expanded add / sub opcode
62 bool NegAcc; // True if the acc is negated before the add / sub.
63 bool HasLane; // True if instruction has an extra "lane" operand.
66 static const ARM_MLxEntry ARM_MLxTable[] = {
67 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
69 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
70 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
71 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
72 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
73 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
74 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
75 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
76 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
79 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
80 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
81 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
82 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
83 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
84 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
85 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
86 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
89 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
90 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP),
92 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
93 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
94 assert(false && "Duplicated entries?");
95 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
96 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
100 // Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl
101 // currently defaults to no prepass hazard recognizer.
102 ScheduleHazardRecognizer *ARMBaseInstrInfo::
103 CreateTargetHazardRecognizer(const TargetMachine *TM,
104 const ScheduleDAG *DAG) const {
105 if (usePreRAHazardRecognizer()) {
106 const InstrItineraryData *II = TM->getInstrItineraryData();
107 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched");
109 return TargetInstrInfo::CreateTargetHazardRecognizer(TM, DAG);
112 ScheduleHazardRecognizer *ARMBaseInstrInfo::
113 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
114 const ScheduleDAG *DAG) const {
115 if (Subtarget.isThumb2() || Subtarget.hasVFP2())
116 return (ScheduleHazardRecognizer *)new ARMHazardRecognizer(II, DAG);
117 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
121 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
122 MachineBasicBlock::iterator &MBBI,
123 LiveVariables *LV) const {
124 // FIXME: Thumb2 support.
129 MachineInstr *MI = MBBI;
130 MachineFunction &MF = *MI->getParent()->getParent();
131 uint64_t TSFlags = MI->getDesc().TSFlags;
133 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
134 default: return NULL;
135 case ARMII::IndexModePre:
138 case ARMII::IndexModePost:
142 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
144 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
148 MachineInstr *UpdateMI = NULL;
149 MachineInstr *MemMI = NULL;
150 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
151 const MCInstrDesc &MCID = MI->getDesc();
152 unsigned NumOps = MCID.getNumOperands();
153 bool isLoad = !MI->mayStore();
154 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
155 const MachineOperand &Base = MI->getOperand(2);
156 const MachineOperand &Offset = MI->getOperand(NumOps-3);
157 unsigned WBReg = WB.getReg();
158 unsigned BaseReg = Base.getReg();
159 unsigned OffReg = Offset.getReg();
160 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
161 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
163 default: llvm_unreachable("Unknown indexed op!");
164 case ARMII::AddrMode2: {
165 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
166 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
168 if (ARM_AM::getSOImmVal(Amt) == -1)
169 // Can't encode it in a so_imm operand. This transformation will
170 // add more than 1 instruction. Abandon!
172 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
173 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
174 .addReg(BaseReg).addImm(Amt)
175 .addImm(Pred).addReg(0).addReg(0);
176 } else if (Amt != 0) {
177 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
178 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
179 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
180 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
181 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
182 .addImm(Pred).addReg(0).addReg(0);
184 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
185 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
186 .addReg(BaseReg).addReg(OffReg)
187 .addImm(Pred).addReg(0).addReg(0);
190 case ARMII::AddrMode3 : {
191 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
192 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
194 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
195 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
196 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
197 .addReg(BaseReg).addImm(Amt)
198 .addImm(Pred).addReg(0).addReg(0);
200 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
201 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
202 .addReg(BaseReg).addReg(OffReg)
203 .addImm(Pred).addReg(0).addReg(0);
208 std::vector<MachineInstr*> NewMIs;
211 MemMI = BuildMI(MF, MI->getDebugLoc(),
212 get(MemOpc), MI->getOperand(0).getReg())
213 .addReg(WBReg).addImm(0).addImm(Pred);
215 MemMI = BuildMI(MF, MI->getDebugLoc(),
216 get(MemOpc)).addReg(MI->getOperand(1).getReg())
217 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
218 NewMIs.push_back(MemMI);
219 NewMIs.push_back(UpdateMI);
222 MemMI = BuildMI(MF, MI->getDebugLoc(),
223 get(MemOpc), MI->getOperand(0).getReg())
224 .addReg(BaseReg).addImm(0).addImm(Pred);
226 MemMI = BuildMI(MF, MI->getDebugLoc(),
227 get(MemOpc)).addReg(MI->getOperand(1).getReg())
228 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
230 UpdateMI->getOperand(0).setIsDead();
231 NewMIs.push_back(UpdateMI);
232 NewMIs.push_back(MemMI);
235 // Transfer LiveVariables states, kill / dead info.
237 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
238 MachineOperand &MO = MI->getOperand(i);
239 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
240 unsigned Reg = MO.getReg();
242 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
244 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
246 LV->addVirtualRegisterDead(Reg, NewMI);
248 if (MO.isUse() && MO.isKill()) {
249 for (unsigned j = 0; j < 2; ++j) {
250 // Look at the two new MI's in reverse order.
251 MachineInstr *NewMI = NewMIs[j];
252 if (!NewMI->readsRegister(Reg))
254 LV->addVirtualRegisterKilled(Reg, NewMI);
255 if (VI.removeKill(MI))
256 VI.Kills.push_back(NewMI);
264 MFI->insert(MBBI, NewMIs[1]);
265 MFI->insert(MBBI, NewMIs[0]);
271 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
272 MachineBasicBlock *&FBB,
273 SmallVectorImpl<MachineOperand> &Cond,
274 bool AllowModify) const {
275 // If the block has no terminators, it just falls into the block after it.
276 MachineBasicBlock::iterator I = MBB.end();
277 if (I == MBB.begin())
280 while (I->isDebugValue()) {
281 if (I == MBB.begin())
286 // Get the last instruction in the block.
287 MachineInstr *LastInst = I;
288 unsigned LastOpc = LastInst->getOpcode();
290 // Check if it's an indirect branch first, this should return 'unanalyzable'
291 // even if it's predicated.
292 if (isIndirectBranchOpcode(LastOpc))
295 if (!isUnpredicatedTerminator(I))
298 // If there is only one terminator instruction, process it.
299 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
300 if (isUncondBranchOpcode(LastOpc)) {
301 TBB = LastInst->getOperand(0).getMBB();
304 if (isCondBranchOpcode(LastOpc)) {
305 // Block ends with fall-through condbranch.
306 TBB = LastInst->getOperand(0).getMBB();
307 Cond.push_back(LastInst->getOperand(1));
308 Cond.push_back(LastInst->getOperand(2));
311 return true; // Can't handle indirect branch.
314 // Get the instruction before it if it is a terminator.
315 MachineInstr *SecondLastInst = I;
316 unsigned SecondLastOpc = SecondLastInst->getOpcode();
318 // If AllowModify is true and the block ends with two or more unconditional
319 // branches, delete all but the first unconditional branch.
320 if (AllowModify && isUncondBranchOpcode(LastOpc)) {
321 while (isUncondBranchOpcode(SecondLastOpc)) {
322 LastInst->eraseFromParent();
323 LastInst = SecondLastInst;
324 LastOpc = LastInst->getOpcode();
325 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
326 // Return now the only terminator is an unconditional branch.
327 TBB = LastInst->getOperand(0).getMBB();
331 SecondLastOpc = SecondLastInst->getOpcode();
336 // If there are three terminators, we don't know what sort of block this is.
337 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
340 // If the block ends with a B and a Bcc, handle it.
341 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
342 TBB = SecondLastInst->getOperand(0).getMBB();
343 Cond.push_back(SecondLastInst->getOperand(1));
344 Cond.push_back(SecondLastInst->getOperand(2));
345 FBB = LastInst->getOperand(0).getMBB();
349 // If the block ends with two unconditional branches, handle it. The second
350 // one is not executed, so remove it.
351 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
352 TBB = SecondLastInst->getOperand(0).getMBB();
355 I->eraseFromParent();
359 // ...likewise if it ends with a branch table followed by an unconditional
360 // branch. The branch folder can create these, and we must get rid of them for
361 // correctness of Thumb constant islands.
362 if ((isJumpTableBranchOpcode(SecondLastOpc) ||
363 isIndirectBranchOpcode(SecondLastOpc)) &&
364 isUncondBranchOpcode(LastOpc)) {
367 I->eraseFromParent();
371 // Otherwise, can't handle this.
376 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
377 MachineBasicBlock::iterator I = MBB.end();
378 if (I == MBB.begin()) return 0;
380 while (I->isDebugValue()) {
381 if (I == MBB.begin())
385 if (!isUncondBranchOpcode(I->getOpcode()) &&
386 !isCondBranchOpcode(I->getOpcode()))
389 // Remove the branch.
390 I->eraseFromParent();
394 if (I == MBB.begin()) return 1;
396 if (!isCondBranchOpcode(I->getOpcode()))
399 // Remove the branch.
400 I->eraseFromParent();
405 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
406 MachineBasicBlock *FBB,
407 const SmallVectorImpl<MachineOperand> &Cond,
409 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
410 int BOpc = !AFI->isThumbFunction()
411 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
412 int BccOpc = !AFI->isThumbFunction()
413 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
414 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
416 // Shouldn't be a fall through.
417 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
418 assert((Cond.size() == 2 || Cond.size() == 0) &&
419 "ARM branch conditions have two components!");
422 if (Cond.empty()) { // Unconditional branch?
424 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
426 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
428 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
429 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
433 // Two-way conditional branch.
434 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
435 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
437 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
439 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
443 bool ARMBaseInstrInfo::
444 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
445 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
446 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
450 bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
451 if (MI->isBundle()) {
452 MachineBasicBlock::const_instr_iterator I = MI;
453 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
454 while (++I != E && I->isInsideBundle()) {
455 int PIdx = I->findFirstPredOperandIdx();
456 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
462 int PIdx = MI->findFirstPredOperandIdx();
463 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
466 bool ARMBaseInstrInfo::
467 PredicateInstruction(MachineInstr *MI,
468 const SmallVectorImpl<MachineOperand> &Pred) const {
469 unsigned Opc = MI->getOpcode();
470 if (isUncondBranchOpcode(Opc)) {
471 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
472 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
473 .addImm(Pred[0].getImm())
474 .addReg(Pred[1].getReg());
478 int PIdx = MI->findFirstPredOperandIdx();
480 MachineOperand &PMO = MI->getOperand(PIdx);
481 PMO.setImm(Pred[0].getImm());
482 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
488 bool ARMBaseInstrInfo::
489 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
490 const SmallVectorImpl<MachineOperand> &Pred2) const {
491 if (Pred1.size() > 2 || Pred2.size() > 2)
494 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
495 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
505 return CC2 == ARMCC::HI;
507 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
509 return CC2 == ARMCC::GT;
511 return CC2 == ARMCC::LT;
515 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
516 std::vector<MachineOperand> &Pred) const {
518 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
519 const MachineOperand &MO = MI->getOperand(i);
520 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
521 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) {
530 /// isPredicable - Return true if the specified instruction can be predicated.
531 /// By default, this returns true for every instruction with a
532 /// PredicateOperand.
533 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
534 if (!MI->isPredicable())
537 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
538 ARMFunctionInfo *AFI =
539 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
540 return AFI->isThumb2Function();
545 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
546 LLVM_ATTRIBUTE_NOINLINE
547 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
549 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
551 assert(JTI < JT.size());
552 return JT[JTI].MBBs.size();
555 /// GetInstSize - Return the size of the specified MachineInstr.
557 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
558 const MachineBasicBlock &MBB = *MI->getParent();
559 const MachineFunction *MF = MBB.getParent();
560 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
562 const MCInstrDesc &MCID = MI->getDesc();
564 return MCID.getSize();
566 // If this machine instr is an inline asm, measure it.
567 if (MI->getOpcode() == ARM::INLINEASM)
568 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
571 unsigned Opc = MI->getOpcode();
573 case TargetOpcode::IMPLICIT_DEF:
574 case TargetOpcode::KILL:
575 case TargetOpcode::PROLOG_LABEL:
576 case TargetOpcode::EH_LABEL:
577 case TargetOpcode::DBG_VALUE:
579 case TargetOpcode::BUNDLE:
580 return getInstBundleLength(MI);
581 case ARM::MOVi16_ga_pcrel:
582 case ARM::MOVTi16_ga_pcrel:
583 case ARM::t2MOVi16_ga_pcrel:
584 case ARM::t2MOVTi16_ga_pcrel:
587 case ARM::t2MOVi32imm:
589 case ARM::CONSTPOOL_ENTRY:
590 // If this machine instr is a constant pool entry, its size is recorded as
592 return MI->getOperand(2).getImm();
593 case ARM::Int_eh_sjlj_longjmp:
595 case ARM::tInt_eh_sjlj_longjmp:
597 case ARM::Int_eh_sjlj_setjmp:
598 case ARM::Int_eh_sjlj_setjmp_nofp:
600 case ARM::tInt_eh_sjlj_setjmp:
601 case ARM::t2Int_eh_sjlj_setjmp:
602 case ARM::t2Int_eh_sjlj_setjmp_nofp:
610 case ARM::t2TBH_JT: {
611 // These are jumptable branches, i.e. a branch followed by an inlined
612 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
613 // entry is one byte; TBH two byte each.
614 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
615 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
616 unsigned NumOps = MCID.getNumOperands();
617 MachineOperand JTOP =
618 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
619 unsigned JTI = JTOP.getIndex();
620 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
622 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
623 assert(JTI < JT.size());
624 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
625 // 4 aligned. The assembler / linker may add 2 byte padding just before
626 // the JT entries. The size does not include this padding; the
627 // constant islands pass does separate bookkeeping for it.
628 // FIXME: If we know the size of the function is less than (1 << 16) *2
629 // bytes, we can use 16-bit entries instead. Then there won't be an
631 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
632 unsigned NumEntries = getNumJTEntries(JT, JTI);
633 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
634 // Make sure the instruction that follows TBB is 2-byte aligned.
635 // FIXME: Constant island pass should insert an "ALIGN" instruction
638 return NumEntries * EntrySize + InstSize;
641 // Otherwise, pseudo-instruction sizes are zero.
646 unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
648 MachineBasicBlock::const_instr_iterator I = MI;
649 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
650 while (++I != E && I->isInsideBundle()) {
651 assert(!I->isBundle() && "No nested bundle!");
652 Size += GetInstSizeInBytes(&*I);
657 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
658 MachineBasicBlock::iterator I, DebugLoc DL,
659 unsigned DestReg, unsigned SrcReg,
660 bool KillSrc) const {
661 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
662 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
664 if (GPRDest && GPRSrc) {
665 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
666 .addReg(SrcReg, getKillRegState(KillSrc))));
670 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
671 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
674 if (SPRDest && SPRSrc)
676 else if (GPRDest && SPRSrc)
678 else if (SPRDest && GPRSrc)
680 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
682 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
686 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
687 MIB.addReg(SrcReg, getKillRegState(KillSrc));
688 if (Opc == ARM::VORRq)
689 MIB.addReg(SrcReg, getKillRegState(KillSrc));
694 // Handle register classes that require multiple instructions.
695 unsigned BeginIdx = 0;
696 unsigned SubRegs = 0;
699 // Use VORRq when possible.
700 if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
701 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 2;
702 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
703 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 4;
704 // Fall back to VMOVD.
705 else if (ARM::DPairRegClass.contains(DestReg, SrcReg))
706 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2;
707 else if (ARM::DTripleRegClass.contains(DestReg, SrcReg))
708 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3;
709 else if (ARM::DQuadRegClass.contains(DestReg, SrcReg))
710 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4;
711 else if (ARM::GPRPairRegClass.contains(DestReg, SrcReg))
712 Opc = ARM::MOVr, BeginIdx = ARM::gsub_0, SubRegs = 2;
714 else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg))
715 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2, Spacing = 2;
716 else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg))
717 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3, Spacing = 2;
718 else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg))
719 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4, Spacing = 2;
721 assert(Opc && "Impossible reg-to-reg copy");
723 const TargetRegisterInfo *TRI = &getRegisterInfo();
724 MachineInstrBuilder Mov;
726 // Copy register tuples backward when the first Dest reg overlaps with SrcReg.
727 if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) {
728 BeginIdx = BeginIdx + ((SubRegs-1)*Spacing);
732 SmallSet<unsigned, 4> DstRegs;
734 for (unsigned i = 0; i != SubRegs; ++i) {
735 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i*Spacing);
736 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i*Spacing);
737 assert(Dst && Src && "Bad sub-register");
739 assert(!DstRegs.count(Src) && "destructive vector copy");
742 Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst)
744 // VORR takes two source operands.
745 if (Opc == ARM::VORRq)
747 Mov = AddDefaultPred(Mov);
749 if (Opc == ARM::MOVr)
750 Mov = AddDefaultCC(Mov);
752 // Add implicit super-register defs and kills to the last instruction.
753 Mov->addRegisterDefined(DestReg, TRI);
755 Mov->addRegisterKilled(SrcReg, TRI);
758 const MachineInstrBuilder &
759 ARMBaseInstrInfo::AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
760 unsigned SubIdx, unsigned State,
761 const TargetRegisterInfo *TRI) const {
763 return MIB.addReg(Reg, State);
765 if (TargetRegisterInfo::isPhysicalRegister(Reg))
766 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
767 return MIB.addReg(Reg, State, SubIdx);
770 void ARMBaseInstrInfo::
771 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
772 unsigned SrcReg, bool isKill, int FI,
773 const TargetRegisterClass *RC,
774 const TargetRegisterInfo *TRI) const {
776 if (I != MBB.end()) DL = I->getDebugLoc();
777 MachineFunction &MF = *MBB.getParent();
778 MachineFrameInfo &MFI = *MF.getFrameInfo();
779 unsigned Align = MFI.getObjectAlignment(FI);
781 MachineMemOperand *MMO =
782 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
783 MachineMemOperand::MOStore,
784 MFI.getObjectSize(FI),
787 switch (RC->getSize()) {
789 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
790 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
791 .addReg(SrcReg, getKillRegState(isKill))
792 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
793 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
794 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
795 .addReg(SrcReg, getKillRegState(isKill))
796 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
798 llvm_unreachable("Unknown reg class!");
801 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
802 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
803 .addReg(SrcReg, getKillRegState(isKill))
804 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
805 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
806 if (Subtarget.hasV5TEOps()) {
807 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::STRD));
808 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
809 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
810 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
814 // Fallback to STM instruction, which has existed since the dawn of
816 MachineInstrBuilder MIB =
817 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STMIA))
818 .addFrameIndex(FI).addMemOperand(MMO));
819 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
820 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
823 llvm_unreachable("Unknown reg class!");
826 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
827 // Use aligned spills if the stack can be realigned.
828 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
829 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
830 .addFrameIndex(FI).addImm(16)
831 .addReg(SrcReg, getKillRegState(isKill))
832 .addMemOperand(MMO));
834 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
835 .addReg(SrcReg, getKillRegState(isKill))
837 .addMemOperand(MMO));
840 llvm_unreachable("Unknown reg class!");
843 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
844 // Use aligned spills if the stack can be realigned.
845 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
846 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
847 .addFrameIndex(FI).addImm(16)
848 .addReg(SrcReg, getKillRegState(isKill))
849 .addMemOperand(MMO));
851 MachineInstrBuilder MIB =
852 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
855 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
856 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
857 AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
860 llvm_unreachable("Unknown reg class!");
863 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
864 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
865 // FIXME: It's possible to only store part of the QQ register if the
866 // spilled def has a sub-register index.
867 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
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 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
879 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
882 llvm_unreachable("Unknown reg class!");
885 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
886 MachineInstrBuilder MIB =
887 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
890 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
891 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
892 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
893 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
894 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
895 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
896 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
897 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
899 llvm_unreachable("Unknown reg class!");
902 llvm_unreachable("Unknown reg class!");
907 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
908 int &FrameIndex) const {
909 switch (MI->getOpcode()) {
912 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
913 if (MI->getOperand(1).isFI() &&
914 MI->getOperand(2).isReg() &&
915 MI->getOperand(3).isImm() &&
916 MI->getOperand(2).getReg() == 0 &&
917 MI->getOperand(3).getImm() == 0) {
918 FrameIndex = MI->getOperand(1).getIndex();
919 return MI->getOperand(0).getReg();
927 if (MI->getOperand(1).isFI() &&
928 MI->getOperand(2).isImm() &&
929 MI->getOperand(2).getImm() == 0) {
930 FrameIndex = MI->getOperand(1).getIndex();
931 return MI->getOperand(0).getReg();
935 case ARM::VST1d64TPseudo:
936 case ARM::VST1d64QPseudo:
937 if (MI->getOperand(0).isFI() &&
938 MI->getOperand(2).getSubReg() == 0) {
939 FrameIndex = MI->getOperand(0).getIndex();
940 return MI->getOperand(2).getReg();
944 if (MI->getOperand(1).isFI() &&
945 MI->getOperand(0).getSubReg() == 0) {
946 FrameIndex = MI->getOperand(1).getIndex();
947 return MI->getOperand(0).getReg();
955 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
956 int &FrameIndex) const {
957 const MachineMemOperand *Dummy;
958 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
961 void ARMBaseInstrInfo::
962 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
963 unsigned DestReg, int FI,
964 const TargetRegisterClass *RC,
965 const TargetRegisterInfo *TRI) const {
967 if (I != MBB.end()) DL = I->getDebugLoc();
968 MachineFunction &MF = *MBB.getParent();
969 MachineFrameInfo &MFI = *MF.getFrameInfo();
970 unsigned Align = MFI.getObjectAlignment(FI);
971 MachineMemOperand *MMO =
972 MF.getMachineMemOperand(
973 MachinePointerInfo::getFixedStack(FI),
974 MachineMemOperand::MOLoad,
975 MFI.getObjectSize(FI),
978 switch (RC->getSize()) {
980 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
981 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
982 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
984 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
985 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
986 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
988 llvm_unreachable("Unknown reg class!");
991 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
992 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
993 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
994 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
995 MachineInstrBuilder MIB;
997 if (Subtarget.hasV5TEOps()) {
998 MIB = BuildMI(MBB, I, DL, get(ARM::LDRD));
999 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1000 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1001 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
1003 AddDefaultPred(MIB);
1005 // Fallback to LDM instruction, which has existed since the dawn of
1007 MIB = AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDMIA))
1008 .addFrameIndex(FI).addMemOperand(MMO));
1009 MIB = AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1010 MIB = AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1013 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1014 MIB.addReg(DestReg, RegState::ImplicitDefine);
1016 llvm_unreachable("Unknown reg class!");
1019 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
1020 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1021 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
1022 .addFrameIndex(FI).addImm(16)
1023 .addMemOperand(MMO));
1025 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
1027 .addMemOperand(MMO));
1030 llvm_unreachable("Unknown reg class!");
1033 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
1034 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1035 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
1036 .addFrameIndex(FI).addImm(16)
1037 .addMemOperand(MMO));
1039 MachineInstrBuilder MIB =
1040 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1042 .addMemOperand(MMO));
1043 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1044 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1045 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1046 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1047 MIB.addReg(DestReg, RegState::ImplicitDefine);
1050 llvm_unreachable("Unknown reg class!");
1053 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
1054 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1055 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
1056 .addFrameIndex(FI).addImm(16)
1057 .addMemOperand(MMO));
1059 MachineInstrBuilder MIB =
1060 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1062 .addMemOperand(MMO);
1063 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1064 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1065 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1066 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1067 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1068 MIB.addReg(DestReg, RegState::ImplicitDefine);
1071 llvm_unreachable("Unknown reg class!");
1074 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
1075 MachineInstrBuilder MIB =
1076 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1078 .addMemOperand(MMO);
1079 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1080 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1081 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1082 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1083 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
1084 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
1085 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
1086 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
1087 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1088 MIB.addReg(DestReg, RegState::ImplicitDefine);
1090 llvm_unreachable("Unknown reg class!");
1093 llvm_unreachable("Unknown regclass!");
1098 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1099 int &FrameIndex) const {
1100 switch (MI->getOpcode()) {
1103 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
1104 if (MI->getOperand(1).isFI() &&
1105 MI->getOperand(2).isReg() &&
1106 MI->getOperand(3).isImm() &&
1107 MI->getOperand(2).getReg() == 0 &&
1108 MI->getOperand(3).getImm() == 0) {
1109 FrameIndex = MI->getOperand(1).getIndex();
1110 return MI->getOperand(0).getReg();
1118 if (MI->getOperand(1).isFI() &&
1119 MI->getOperand(2).isImm() &&
1120 MI->getOperand(2).getImm() == 0) {
1121 FrameIndex = MI->getOperand(1).getIndex();
1122 return MI->getOperand(0).getReg();
1126 case ARM::VLD1d64TPseudo:
1127 case ARM::VLD1d64QPseudo:
1128 if (MI->getOperand(1).isFI() &&
1129 MI->getOperand(0).getSubReg() == 0) {
1130 FrameIndex = MI->getOperand(1).getIndex();
1131 return MI->getOperand(0).getReg();
1135 if (MI->getOperand(1).isFI() &&
1136 MI->getOperand(0).getSubReg() == 0) {
1137 FrameIndex = MI->getOperand(1).getIndex();
1138 return MI->getOperand(0).getReg();
1146 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1147 int &FrameIndex) const {
1148 const MachineMemOperand *Dummy;
1149 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1152 bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
1153 // This hook gets to expand COPY instructions before they become
1154 // copyPhysReg() calls. Look for VMOVS instructions that can legally be
1155 // widened to VMOVD. We prefer the VMOVD when possible because it may be
1156 // changed into a VORR that can go down the NEON pipeline.
1157 if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15())
1160 // Look for a copy between even S-registers. That is where we keep floats
1161 // when using NEON v2f32 instructions for f32 arithmetic.
1162 unsigned DstRegS = MI->getOperand(0).getReg();
1163 unsigned SrcRegS = MI->getOperand(1).getReg();
1164 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
1167 const TargetRegisterInfo *TRI = &getRegisterInfo();
1168 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
1170 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
1172 if (!DstRegD || !SrcRegD)
1175 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
1176 // legal if the COPY already defines the full DstRegD, and it isn't a
1177 // sub-register insertion.
1178 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
1181 // A dead copy shouldn't show up here, but reject it just in case.
1182 if (MI->getOperand(0).isDead())
1185 // All clear, widen the COPY.
1186 DEBUG(dbgs() << "widening: " << *MI);
1187 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1189 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
1190 // or some other super-register.
1191 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
1192 if (ImpDefIdx != -1)
1193 MI->RemoveOperand(ImpDefIdx);
1195 // Change the opcode and operands.
1196 MI->setDesc(get(ARM::VMOVD));
1197 MI->getOperand(0).setReg(DstRegD);
1198 MI->getOperand(1).setReg(SrcRegD);
1199 AddDefaultPred(MIB);
1201 // We are now reading SrcRegD instead of SrcRegS. This may upset the
1202 // register scavenger and machine verifier, so we need to indicate that we
1203 // are reading an undefined value from SrcRegD, but a proper value from
1205 MI->getOperand(1).setIsUndef();
1206 MIB.addReg(SrcRegS, RegState::Implicit);
1208 // SrcRegD may actually contain an unrelated value in the ssub_1
1209 // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
1210 if (MI->getOperand(1).isKill()) {
1211 MI->getOperand(1).setIsKill(false);
1212 MI->addRegisterKilled(SrcRegS, TRI, true);
1215 DEBUG(dbgs() << "replaced by: " << *MI);
1219 /// Create a copy of a const pool value. Update CPI to the new index and return
1221 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1222 MachineConstantPool *MCP = MF.getConstantPool();
1223 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1225 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1226 assert(MCPE.isMachineConstantPoolEntry() &&
1227 "Expecting a machine constantpool entry!");
1228 ARMConstantPoolValue *ACPV =
1229 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1231 unsigned PCLabelId = AFI->createPICLabelUId();
1232 ARMConstantPoolValue *NewCPV = 0;
1233 // FIXME: The below assumes PIC relocation model and that the function
1234 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1235 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1236 // instructions, so that's probably OK, but is PIC always correct when
1238 if (ACPV->isGlobalValue())
1239 NewCPV = ARMConstantPoolConstant::
1240 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
1242 else if (ACPV->isExtSymbol())
1243 NewCPV = ARMConstantPoolSymbol::
1244 Create(MF.getFunction()->getContext(),
1245 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
1246 else if (ACPV->isBlockAddress())
1247 NewCPV = ARMConstantPoolConstant::
1248 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
1249 ARMCP::CPBlockAddress, 4);
1250 else if (ACPV->isLSDA())
1251 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
1253 else if (ACPV->isMachineBasicBlock())
1254 NewCPV = ARMConstantPoolMBB::
1255 Create(MF.getFunction()->getContext(),
1256 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
1258 llvm_unreachable("Unexpected ARM constantpool value type!!");
1259 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1263 void ARMBaseInstrInfo::
1264 reMaterialize(MachineBasicBlock &MBB,
1265 MachineBasicBlock::iterator I,
1266 unsigned DestReg, unsigned SubIdx,
1267 const MachineInstr *Orig,
1268 const TargetRegisterInfo &TRI) const {
1269 unsigned Opcode = Orig->getOpcode();
1272 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1273 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1277 case ARM::tLDRpci_pic:
1278 case ARM::t2LDRpci_pic: {
1279 MachineFunction &MF = *MBB.getParent();
1280 unsigned CPI = Orig->getOperand(1).getIndex();
1281 unsigned PCLabelId = duplicateCPV(MF, CPI);
1282 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1284 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1285 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1292 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1293 MachineInstr *MI = TargetInstrInfo::duplicate(Orig, MF);
1294 switch(Orig->getOpcode()) {
1295 case ARM::tLDRpci_pic:
1296 case ARM::t2LDRpci_pic: {
1297 unsigned CPI = Orig->getOperand(1).getIndex();
1298 unsigned PCLabelId = duplicateCPV(MF, CPI);
1299 Orig->getOperand(1).setIndex(CPI);
1300 Orig->getOperand(2).setImm(PCLabelId);
1307 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1308 const MachineInstr *MI1,
1309 const MachineRegisterInfo *MRI) const {
1310 int Opcode = MI0->getOpcode();
1311 if (Opcode == ARM::t2LDRpci ||
1312 Opcode == ARM::t2LDRpci_pic ||
1313 Opcode == ARM::tLDRpci ||
1314 Opcode == ARM::tLDRpci_pic ||
1315 Opcode == ARM::MOV_ga_dyn ||
1316 Opcode == ARM::MOV_ga_pcrel ||
1317 Opcode == ARM::MOV_ga_pcrel_ldr ||
1318 Opcode == ARM::t2MOV_ga_dyn ||
1319 Opcode == ARM::t2MOV_ga_pcrel) {
1320 if (MI1->getOpcode() != Opcode)
1322 if (MI0->getNumOperands() != MI1->getNumOperands())
1325 const MachineOperand &MO0 = MI0->getOperand(1);
1326 const MachineOperand &MO1 = MI1->getOperand(1);
1327 if (MO0.getOffset() != MO1.getOffset())
1330 if (Opcode == ARM::MOV_ga_dyn ||
1331 Opcode == ARM::MOV_ga_pcrel ||
1332 Opcode == ARM::MOV_ga_pcrel_ldr ||
1333 Opcode == ARM::t2MOV_ga_dyn ||
1334 Opcode == ARM::t2MOV_ga_pcrel)
1335 // Ignore the PC labels.
1336 return MO0.getGlobal() == MO1.getGlobal();
1338 const MachineFunction *MF = MI0->getParent()->getParent();
1339 const MachineConstantPool *MCP = MF->getConstantPool();
1340 int CPI0 = MO0.getIndex();
1341 int CPI1 = MO1.getIndex();
1342 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1343 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1344 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1345 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1346 if (isARMCP0 && isARMCP1) {
1347 ARMConstantPoolValue *ACPV0 =
1348 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1349 ARMConstantPoolValue *ACPV1 =
1350 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1351 return ACPV0->hasSameValue(ACPV1);
1352 } else if (!isARMCP0 && !isARMCP1) {
1353 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1356 } else if (Opcode == ARM::PICLDR) {
1357 if (MI1->getOpcode() != Opcode)
1359 if (MI0->getNumOperands() != MI1->getNumOperands())
1362 unsigned Addr0 = MI0->getOperand(1).getReg();
1363 unsigned Addr1 = MI1->getOperand(1).getReg();
1364 if (Addr0 != Addr1) {
1366 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1367 !TargetRegisterInfo::isVirtualRegister(Addr1))
1370 // This assumes SSA form.
1371 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1372 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1373 // Check if the loaded value, e.g. a constantpool of a global address, are
1375 if (!produceSameValue(Def0, Def1, MRI))
1379 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1380 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1381 const MachineOperand &MO0 = MI0->getOperand(i);
1382 const MachineOperand &MO1 = MI1->getOperand(i);
1383 if (!MO0.isIdenticalTo(MO1))
1389 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1392 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1393 /// determine if two loads are loading from the same base address. It should
1394 /// only return true if the base pointers are the same and the only differences
1395 /// between the two addresses is the offset. It also returns the offsets by
1398 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1399 /// is permanently disabled.
1400 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1402 int64_t &Offset2) const {
1403 // Don't worry about Thumb: just ARM and Thumb2.
1404 if (Subtarget.isThumb1Only()) return false;
1406 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1409 switch (Load1->getMachineOpcode()) {
1422 case ARM::t2LDRSHi8:
1424 case ARM::t2LDRSHi12:
1428 switch (Load2->getMachineOpcode()) {
1440 case ARM::t2LDRSHi8:
1442 case ARM::t2LDRSHi12:
1446 // Check if base addresses and chain operands match.
1447 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1448 Load1->getOperand(4) != Load2->getOperand(4))
1451 // Index should be Reg0.
1452 if (Load1->getOperand(3) != Load2->getOperand(3))
1455 // Determine the offsets.
1456 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1457 isa<ConstantSDNode>(Load2->getOperand(1))) {
1458 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1459 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1466 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1467 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1468 /// be scheduled togther. On some targets if two loads are loading from
1469 /// addresses in the same cache line, it's better if they are scheduled
1470 /// together. This function takes two integers that represent the load offsets
1471 /// from the common base address. It returns true if it decides it's desirable
1472 /// to schedule the two loads together. "NumLoads" is the number of loads that
1473 /// have already been scheduled after Load1.
1475 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1476 /// is permanently disabled.
1477 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1478 int64_t Offset1, int64_t Offset2,
1479 unsigned NumLoads) const {
1480 // Don't worry about Thumb: just ARM and Thumb2.
1481 if (Subtarget.isThumb1Only()) return false;
1483 assert(Offset2 > Offset1);
1485 if ((Offset2 - Offset1) / 8 > 64)
1488 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1489 return false; // FIXME: overly conservative?
1491 // Four loads in a row should be sufficient.
1498 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1499 const MachineBasicBlock *MBB,
1500 const MachineFunction &MF) const {
1501 // Debug info is never a scheduling boundary. It's necessary to be explicit
1502 // due to the special treatment of IT instructions below, otherwise a
1503 // dbg_value followed by an IT will result in the IT instruction being
1504 // considered a scheduling hazard, which is wrong. It should be the actual
1505 // instruction preceding the dbg_value instruction(s), just like it is
1506 // when debug info is not present.
1507 if (MI->isDebugValue())
1510 // Terminators and labels can't be scheduled around.
1511 if (MI->isTerminator() || MI->isLabel())
1514 // Treat the start of the IT block as a scheduling boundary, but schedule
1515 // t2IT along with all instructions following it.
1516 // FIXME: This is a big hammer. But the alternative is to add all potential
1517 // true and anti dependencies to IT block instructions as implicit operands
1518 // to the t2IT instruction. The added compile time and complexity does not
1520 MachineBasicBlock::const_iterator I = MI;
1521 // Make sure to skip any dbg_value instructions
1522 while (++I != MBB->end() && I->isDebugValue())
1524 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1527 // Don't attempt to schedule around any instruction that defines
1528 // a stack-oriented pointer, as it's unlikely to be profitable. This
1529 // saves compile time, because it doesn't require every single
1530 // stack slot reference to depend on the instruction that does the
1532 // Calls don't actually change the stack pointer, even if they have imp-defs.
1533 // No ARM calling conventions change the stack pointer. (X86 calling
1534 // conventions sometimes do).
1535 if (!MI->isCall() && MI->definesRegister(ARM::SP))
1541 bool ARMBaseInstrInfo::
1542 isProfitableToIfCvt(MachineBasicBlock &MBB,
1543 unsigned NumCycles, unsigned ExtraPredCycles,
1544 const BranchProbability &Probability) const {
1548 // Attempt to estimate the relative costs of predication versus branching.
1549 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1550 UnpredCost /= Probability.getDenominator();
1551 UnpredCost += 1; // The branch itself
1552 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1554 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1557 bool ARMBaseInstrInfo::
1558 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1559 unsigned TCycles, unsigned TExtra,
1560 MachineBasicBlock &FMBB,
1561 unsigned FCycles, unsigned FExtra,
1562 const BranchProbability &Probability) const {
1563 if (!TCycles || !FCycles)
1566 // Attempt to estimate the relative costs of predication versus branching.
1567 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1568 TUnpredCost /= Probability.getDenominator();
1570 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1571 unsigned FUnpredCost = Comp * FCycles;
1572 FUnpredCost /= Probability.getDenominator();
1574 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1575 UnpredCost += 1; // The branch itself
1576 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1578 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1582 ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
1583 MachineBasicBlock &FMBB) const {
1584 // Reduce false anti-dependencies to let Swift's out-of-order execution
1585 // engine do its thing.
1586 return Subtarget.isSwift();
1589 /// getInstrPredicate - If instruction is predicated, returns its predicate
1590 /// condition, otherwise returns AL. It also returns the condition code
1591 /// register by reference.
1593 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1594 int PIdx = MI->findFirstPredOperandIdx();
1600 PredReg = MI->getOperand(PIdx+1).getReg();
1601 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1605 int llvm::getMatchingCondBranchOpcode(int Opc) {
1610 if (Opc == ARM::t2B)
1613 llvm_unreachable("Unknown unconditional branch opcode!");
1616 /// commuteInstruction - Handle commutable instructions.
1618 ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1619 switch (MI->getOpcode()) {
1621 case ARM::t2MOVCCr: {
1622 // MOVCC can be commuted by inverting the condition.
1623 unsigned PredReg = 0;
1624 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
1625 // MOVCC AL can't be inverted. Shouldn't happen.
1626 if (CC == ARMCC::AL || PredReg != ARM::CPSR)
1628 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
1631 // After swapping the MOVCC operands, also invert the condition.
1632 MI->getOperand(MI->findFirstPredOperandIdx())
1633 .setImm(ARMCC::getOppositeCondition(CC));
1637 return TargetInstrInfo::commuteInstruction(MI, NewMI);
1640 /// Identify instructions that can be folded into a MOVCC instruction, and
1641 /// return the defining instruction.
1642 static MachineInstr *canFoldIntoMOVCC(unsigned Reg,
1643 const MachineRegisterInfo &MRI,
1644 const TargetInstrInfo *TII) {
1645 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1647 if (!MRI.hasOneNonDBGUse(Reg))
1649 MachineInstr *MI = MRI.getVRegDef(Reg);
1652 // MI is folded into the MOVCC by predicating it.
1653 if (!MI->isPredicable())
1655 // Check if MI has any non-dead defs or physreg uses. This also detects
1656 // predicated instructions which will be reading CPSR.
1657 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
1658 const MachineOperand &MO = MI->getOperand(i);
1659 // Reject frame index operands, PEI can't handle the predicated pseudos.
1660 if (MO.isFI() || MO.isCPI() || MO.isJTI())
1664 // MI can't have any tied operands, that would conflict with predication.
1667 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1669 if (MO.isDef() && !MO.isDead())
1672 bool DontMoveAcrossStores = true;
1673 if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ 0, DontMoveAcrossStores))
1678 bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
1679 SmallVectorImpl<MachineOperand> &Cond,
1680 unsigned &TrueOp, unsigned &FalseOp,
1681 bool &Optimizable) const {
1682 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1683 "Unknown select instruction");
1688 // 3: Condition code.
1692 Cond.push_back(MI->getOperand(3));
1693 Cond.push_back(MI->getOperand(4));
1694 // We can always fold a def.
1699 MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
1700 bool PreferFalse) const {
1701 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1702 "Unknown select instruction");
1703 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1704 MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
1705 bool Invert = !DefMI;
1707 DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this);
1711 // Create a new predicated version of DefMI.
1712 // Rfalse is the first use.
1713 MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1715 MI->getOperand(0).getReg());
1717 // Copy all the DefMI operands, excluding its (null) predicate.
1718 const MCInstrDesc &DefDesc = DefMI->getDesc();
1719 for (unsigned i = 1, e = DefDesc.getNumOperands();
1720 i != e && !DefDesc.OpInfo[i].isPredicate(); ++i)
1721 NewMI.addOperand(DefMI->getOperand(i));
1723 unsigned CondCode = MI->getOperand(3).getImm();
1725 NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode)));
1727 NewMI.addImm(CondCode);
1728 NewMI.addOperand(MI->getOperand(4));
1730 // DefMI is not the -S version that sets CPSR, so add an optional %noreg.
1731 if (NewMI->hasOptionalDef())
1732 AddDefaultCC(NewMI);
1734 // The output register value when the predicate is false is an implicit
1735 // register operand tied to the first def.
1736 // The tie makes the register allocator ensure the FalseReg is allocated the
1737 // same register as operand 0.
1738 MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
1739 FalseReg.setImplicit();
1740 NewMI.addOperand(FalseReg);
1741 NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
1743 // The caller will erase MI, but not DefMI.
1744 DefMI->eraseFromParent();
1748 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
1749 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
1752 /// This will go away once we can teach tblgen how to set the optional CPSR def
1754 struct AddSubFlagsOpcodePair {
1756 uint16_t MachineOpc;
1759 static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
1760 {ARM::ADDSri, ARM::ADDri},
1761 {ARM::ADDSrr, ARM::ADDrr},
1762 {ARM::ADDSrsi, ARM::ADDrsi},
1763 {ARM::ADDSrsr, ARM::ADDrsr},
1765 {ARM::SUBSri, ARM::SUBri},
1766 {ARM::SUBSrr, ARM::SUBrr},
1767 {ARM::SUBSrsi, ARM::SUBrsi},
1768 {ARM::SUBSrsr, ARM::SUBrsr},
1770 {ARM::RSBSri, ARM::RSBri},
1771 {ARM::RSBSrsi, ARM::RSBrsi},
1772 {ARM::RSBSrsr, ARM::RSBrsr},
1774 {ARM::t2ADDSri, ARM::t2ADDri},
1775 {ARM::t2ADDSrr, ARM::t2ADDrr},
1776 {ARM::t2ADDSrs, ARM::t2ADDrs},
1778 {ARM::t2SUBSri, ARM::t2SUBri},
1779 {ARM::t2SUBSrr, ARM::t2SUBrr},
1780 {ARM::t2SUBSrs, ARM::t2SUBrs},
1782 {ARM::t2RSBSri, ARM::t2RSBri},
1783 {ARM::t2RSBSrs, ARM::t2RSBrs},
1786 unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
1787 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
1788 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
1789 return AddSubFlagsOpcodeMap[i].MachineOpc;
1793 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1794 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1795 unsigned DestReg, unsigned BaseReg, int NumBytes,
1796 ARMCC::CondCodes Pred, unsigned PredReg,
1797 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1798 bool isSub = NumBytes < 0;
1799 if (isSub) NumBytes = -NumBytes;
1802 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1803 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1804 assert(ThisVal && "Didn't extract field correctly");
1806 // We will handle these bits from offset, clear them.
1807 NumBytes &= ~ThisVal;
1809 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1811 // Build the new ADD / SUB.
1812 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1813 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1814 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1815 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1816 .setMIFlags(MIFlags);
1821 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1822 unsigned FrameReg, int &Offset,
1823 const ARMBaseInstrInfo &TII) {
1824 unsigned Opcode = MI.getOpcode();
1825 const MCInstrDesc &Desc = MI.getDesc();
1826 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1829 // Memory operands in inline assembly always use AddrMode2.
1830 if (Opcode == ARM::INLINEASM)
1831 AddrMode = ARMII::AddrMode2;
1833 if (Opcode == ARM::ADDri) {
1834 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1836 // Turn it into a move.
1837 MI.setDesc(TII.get(ARM::MOVr));
1838 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1839 MI.RemoveOperand(FrameRegIdx+1);
1842 } else if (Offset < 0) {
1845 MI.setDesc(TII.get(ARM::SUBri));
1848 // Common case: small offset, fits into instruction.
1849 if (ARM_AM::getSOImmVal(Offset) != -1) {
1850 // Replace the FrameIndex with sp / fp
1851 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1852 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1857 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1859 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1860 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1862 // We will handle these bits from offset, clear them.
1863 Offset &= ~ThisImmVal;
1865 // Get the properly encoded SOImmVal field.
1866 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1867 "Bit extraction didn't work?");
1868 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1870 unsigned ImmIdx = 0;
1872 unsigned NumBits = 0;
1875 case ARMII::AddrMode_i12: {
1876 ImmIdx = FrameRegIdx + 1;
1877 InstrOffs = MI.getOperand(ImmIdx).getImm();
1881 case ARMII::AddrMode2: {
1882 ImmIdx = FrameRegIdx+2;
1883 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1884 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1889 case ARMII::AddrMode3: {
1890 ImmIdx = FrameRegIdx+2;
1891 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1892 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1897 case ARMII::AddrMode4:
1898 case ARMII::AddrMode6:
1899 // Can't fold any offset even if it's zero.
1901 case ARMII::AddrMode5: {
1902 ImmIdx = FrameRegIdx+1;
1903 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1904 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1911 llvm_unreachable("Unsupported addressing mode!");
1914 Offset += InstrOffs * Scale;
1915 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1921 // Attempt to fold address comp. if opcode has offset bits
1923 // Common case: small offset, fits into instruction.
1924 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1925 int ImmedOffset = Offset / Scale;
1926 unsigned Mask = (1 << NumBits) - 1;
1927 if ((unsigned)Offset <= Mask * Scale) {
1928 // Replace the FrameIndex with sp
1929 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1930 // FIXME: When addrmode2 goes away, this will simplify (like the
1931 // T2 version), as the LDR.i12 versions don't need the encoding
1932 // tricks for the offset value.
1934 if (AddrMode == ARMII::AddrMode_i12)
1935 ImmedOffset = -ImmedOffset;
1937 ImmedOffset |= 1 << NumBits;
1939 ImmOp.ChangeToImmediate(ImmedOffset);
1944 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1945 ImmedOffset = ImmedOffset & Mask;
1947 if (AddrMode == ARMII::AddrMode_i12)
1948 ImmedOffset = -ImmedOffset;
1950 ImmedOffset |= 1 << NumBits;
1952 ImmOp.ChangeToImmediate(ImmedOffset);
1953 Offset &= ~(Mask*Scale);
1957 Offset = (isSub) ? -Offset : Offset;
1961 /// analyzeCompare - For a comparison instruction, return the source registers
1962 /// in SrcReg and SrcReg2 if having two register operands, and the value it
1963 /// compares against in CmpValue. Return true if the comparison instruction
1964 /// can be analyzed.
1965 bool ARMBaseInstrInfo::
1966 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
1967 int &CmpMask, int &CmpValue) const {
1968 switch (MI->getOpcode()) {
1972 SrcReg = MI->getOperand(0).getReg();
1975 CmpValue = MI->getOperand(1).getImm();
1979 SrcReg = MI->getOperand(0).getReg();
1980 SrcReg2 = MI->getOperand(1).getReg();
1986 SrcReg = MI->getOperand(0).getReg();
1988 CmpMask = MI->getOperand(1).getImm();
1996 /// isSuitableForMask - Identify a suitable 'and' instruction that
1997 /// operates on the given source register and applies the same mask
1998 /// as a 'tst' instruction. Provide a limited look-through for copies.
1999 /// When successful, MI will hold the found instruction.
2000 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
2001 int CmpMask, bool CommonUse) {
2002 switch (MI->getOpcode()) {
2005 if (CmpMask != MI->getOperand(2).getImm())
2007 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
2011 // Walk down one instruction which is potentially an 'and'.
2012 const MachineInstr &Copy = *MI;
2013 MachineBasicBlock::iterator AND(
2014 llvm::next(MachineBasicBlock::iterator(MI)));
2015 if (AND == MI->getParent()->end()) return false;
2017 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
2025 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2026 /// the condition code if we modify the instructions such that flags are
2028 inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) {
2030 default: return ARMCC::AL;
2031 case ARMCC::EQ: return ARMCC::EQ;
2032 case ARMCC::NE: return ARMCC::NE;
2033 case ARMCC::HS: return ARMCC::LS;
2034 case ARMCC::LO: return ARMCC::HI;
2035 case ARMCC::HI: return ARMCC::LO;
2036 case ARMCC::LS: return ARMCC::HS;
2037 case ARMCC::GE: return ARMCC::LE;
2038 case ARMCC::LT: return ARMCC::GT;
2039 case ARMCC::GT: return ARMCC::LT;
2040 case ARMCC::LE: return ARMCC::GE;
2044 /// isRedundantFlagInstr - check whether the first instruction, whose only
2045 /// purpose is to update flags, can be made redundant.
2046 /// CMPrr can be made redundant by SUBrr if the operands are the same.
2047 /// CMPri can be made redundant by SUBri if the operands are the same.
2048 /// This function can be extended later on.
2049 inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg,
2050 unsigned SrcReg2, int ImmValue,
2052 if ((CmpI->getOpcode() == ARM::CMPrr ||
2053 CmpI->getOpcode() == ARM::t2CMPrr) &&
2054 (OI->getOpcode() == ARM::SUBrr ||
2055 OI->getOpcode() == ARM::t2SUBrr) &&
2056 ((OI->getOperand(1).getReg() == SrcReg &&
2057 OI->getOperand(2).getReg() == SrcReg2) ||
2058 (OI->getOperand(1).getReg() == SrcReg2 &&
2059 OI->getOperand(2).getReg() == SrcReg)))
2062 if ((CmpI->getOpcode() == ARM::CMPri ||
2063 CmpI->getOpcode() == ARM::t2CMPri) &&
2064 (OI->getOpcode() == ARM::SUBri ||
2065 OI->getOpcode() == ARM::t2SUBri) &&
2066 OI->getOperand(1).getReg() == SrcReg &&
2067 OI->getOperand(2).getImm() == ImmValue)
2072 /// optimizeCompareInstr - Convert the instruction supplying the argument to the
2073 /// comparison into one that sets the zero bit in the flags register;
2074 /// Remove a redundant Compare instruction if an earlier instruction can set the
2075 /// flags in the same way as Compare.
2076 /// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two
2077 /// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the
2078 /// condition code of instructions which use the flags.
2079 bool ARMBaseInstrInfo::
2080 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
2081 int CmpMask, int CmpValue,
2082 const MachineRegisterInfo *MRI) const {
2083 // Get the unique definition of SrcReg.
2084 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
2085 if (!MI) return false;
2087 // Masked compares sometimes use the same register as the corresponding 'and'.
2088 if (CmpMask != ~0) {
2089 if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) {
2091 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
2092 UE = MRI->use_end(); UI != UE; ++UI) {
2093 if (UI->getParent() != CmpInstr->getParent()) continue;
2094 MachineInstr *PotentialAND = &*UI;
2095 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) ||
2096 isPredicated(PotentialAND))
2101 if (!MI) return false;
2105 // Get ready to iterate backward from CmpInstr.
2106 MachineBasicBlock::iterator I = CmpInstr, E = MI,
2107 B = CmpInstr->getParent()->begin();
2109 // Early exit if CmpInstr is at the beginning of the BB.
2110 if (I == B) return false;
2112 // There are two possible candidates which can be changed to set CPSR:
2113 // One is MI, the other is a SUB instruction.
2114 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
2115 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
2116 MachineInstr *Sub = NULL;
2118 // MI is not a candidate for CMPrr.
2120 else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
2121 // Conservatively refuse to convert an instruction which isn't in the same
2122 // BB as the comparison.
2123 // For CMPri, we need to check Sub, thus we can't return here.
2124 if (CmpInstr->getOpcode() == ARM::CMPri ||
2125 CmpInstr->getOpcode() == ARM::t2CMPri)
2131 // Check that CPSR isn't set between the comparison instruction and the one we
2132 // want to change. At the same time, search for Sub.
2133 const TargetRegisterInfo *TRI = &getRegisterInfo();
2135 for (; I != E; --I) {
2136 const MachineInstr &Instr = *I;
2138 if (Instr.modifiesRegister(ARM::CPSR, TRI) ||
2139 Instr.readsRegister(ARM::CPSR, TRI))
2140 // This instruction modifies or uses CPSR after the one we want to
2141 // change. We can't do this transformation.
2144 // Check whether CmpInstr can be made redundant by the current instruction.
2145 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
2151 // The 'and' is below the comparison instruction.
2155 // Return false if no candidates exist.
2159 // The single candidate is called MI.
2162 // We can't use a predicated instruction - it doesn't always write the flags.
2163 if (isPredicated(MI))
2166 switch (MI->getOpcode()) {
2200 case ARM::t2EORri: {
2201 // Scan forward for the use of CPSR
2202 // When checking against MI: if it's a conditional code requires
2203 // checking of V bit, then this is not safe to do.
2204 // It is safe to remove CmpInstr if CPSR is redefined or killed.
2205 // If we are done with the basic block, we need to check whether CPSR is
2207 SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
2209 bool isSafe = false;
2211 E = CmpInstr->getParent()->end();
2212 while (!isSafe && ++I != E) {
2213 const MachineInstr &Instr = *I;
2214 for (unsigned IO = 0, EO = Instr.getNumOperands();
2215 !isSafe && IO != EO; ++IO) {
2216 const MachineOperand &MO = Instr.getOperand(IO);
2217 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
2221 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
2227 // Condition code is after the operand before CPSR.
2228 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
2230 ARMCC::CondCodes NewCC = getSwappedCondition(CC);
2231 if (NewCC == ARMCC::AL)
2233 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
2234 // on CMP needs to be updated to be based on SUB.
2235 // Push the condition code operands to OperandsToUpdate.
2236 // If it is safe to remove CmpInstr, the condition code of these
2237 // operands will be modified.
2238 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
2239 Sub->getOperand(2).getReg() == SrcReg)
2240 OperandsToUpdate.push_back(std::make_pair(&((*I).getOperand(IO-1)),
2246 // CPSR can be used multiple times, we should continue.
2259 // If CPSR is not killed nor re-defined, we should check whether it is
2260 // live-out. If it is live-out, do not optimize.
2262 MachineBasicBlock *MBB = CmpInstr->getParent();
2263 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
2264 SE = MBB->succ_end(); SI != SE; ++SI)
2265 if ((*SI)->isLiveIn(ARM::CPSR))
2269 // Toggle the optional operand to CPSR.
2270 MI->getOperand(5).setReg(ARM::CPSR);
2271 MI->getOperand(5).setIsDef(true);
2272 assert(!isPredicated(MI) && "Can't use flags from predicated instruction");
2273 CmpInstr->eraseFromParent();
2275 // Modify the condition code of operands in OperandsToUpdate.
2276 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2277 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2278 for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++)
2279 OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second);
2287 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
2288 MachineInstr *DefMI, unsigned Reg,
2289 MachineRegisterInfo *MRI) const {
2290 // Fold large immediates into add, sub, or, xor.
2291 unsigned DefOpc = DefMI->getOpcode();
2292 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
2294 if (!DefMI->getOperand(1).isImm())
2295 // Could be t2MOVi32imm <ga:xx>
2298 if (!MRI->hasOneNonDBGUse(Reg))
2301 const MCInstrDesc &DefMCID = DefMI->getDesc();
2302 if (DefMCID.hasOptionalDef()) {
2303 unsigned NumOps = DefMCID.getNumOperands();
2304 const MachineOperand &MO = DefMI->getOperand(NumOps-1);
2305 if (MO.getReg() == ARM::CPSR && !MO.isDead())
2306 // If DefMI defines CPSR and it is not dead, it's obviously not safe
2311 const MCInstrDesc &UseMCID = UseMI->getDesc();
2312 if (UseMCID.hasOptionalDef()) {
2313 unsigned NumOps = UseMCID.getNumOperands();
2314 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
2315 // If the instruction sets the flag, do not attempt this optimization
2316 // since it may change the semantics of the code.
2320 unsigned UseOpc = UseMI->getOpcode();
2321 unsigned NewUseOpc = 0;
2322 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
2323 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
2324 bool Commute = false;
2326 default: return false;
2334 case ARM::t2EORrr: {
2335 Commute = UseMI->getOperand(2).getReg() != Reg;
2342 NewUseOpc = ARM::SUBri;
2348 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
2350 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
2351 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
2354 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
2355 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
2356 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
2360 case ARM::t2SUBrr: {
2364 NewUseOpc = ARM::t2SUBri;
2369 case ARM::t2EORrr: {
2370 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
2372 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
2373 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
2376 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
2377 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
2378 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
2386 unsigned OpIdx = Commute ? 2 : 1;
2387 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
2388 bool isKill = UseMI->getOperand(OpIdx).isKill();
2389 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
2390 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
2391 UseMI, UseMI->getDebugLoc(),
2392 get(NewUseOpc), NewReg)
2393 .addReg(Reg1, getKillRegState(isKill))
2394 .addImm(SOImmValV1)));
2395 UseMI->setDesc(get(NewUseOpc));
2396 UseMI->getOperand(1).setReg(NewReg);
2397 UseMI->getOperand(1).setIsKill();
2398 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
2399 DefMI->eraseFromParent();
2403 static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData,
2404 const MachineInstr *MI) {
2405 switch (MI->getOpcode()) {
2407 const MCInstrDesc &Desc = MI->getDesc();
2408 int UOps = ItinData->getNumMicroOps(Desc.getSchedClass());
2409 assert(UOps >= 0 && "bad # UOps");
2417 unsigned ShOpVal = MI->getOperand(3).getImm();
2418 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2419 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2422 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2423 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2430 if (!MI->getOperand(2).getReg())
2433 unsigned ShOpVal = MI->getOperand(3).getImm();
2434 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2435 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2438 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2439 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2446 return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2;
2448 case ARM::LDRSB_POST:
2449 case ARM::LDRSH_POST: {
2450 unsigned Rt = MI->getOperand(0).getReg();
2451 unsigned Rm = MI->getOperand(3).getReg();
2452 return (Rt == Rm) ? 4 : 3;
2455 case ARM::LDR_PRE_REG:
2456 case ARM::LDRB_PRE_REG: {
2457 unsigned Rt = MI->getOperand(0).getReg();
2458 unsigned Rm = MI->getOperand(3).getReg();
2461 unsigned ShOpVal = MI->getOperand(4).getImm();
2462 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2463 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2466 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2467 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2472 case ARM::STR_PRE_REG:
2473 case ARM::STRB_PRE_REG: {
2474 unsigned ShOpVal = MI->getOperand(4).getImm();
2475 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2476 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2479 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2480 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2486 case ARM::STRH_PRE: {
2487 unsigned Rt = MI->getOperand(0).getReg();
2488 unsigned Rm = MI->getOperand(3).getReg();
2493 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub)
2497 case ARM::LDR_POST_REG:
2498 case ARM::LDRB_POST_REG:
2499 case ARM::LDRH_POST: {
2500 unsigned Rt = MI->getOperand(0).getReg();
2501 unsigned Rm = MI->getOperand(3).getReg();
2502 return (Rt == Rm) ? 3 : 2;
2505 case ARM::LDR_PRE_IMM:
2506 case ARM::LDRB_PRE_IMM:
2507 case ARM::LDR_POST_IMM:
2508 case ARM::LDRB_POST_IMM:
2509 case ARM::STRB_POST_IMM:
2510 case ARM::STRB_POST_REG:
2511 case ARM::STRB_PRE_IMM:
2512 case ARM::STRH_POST:
2513 case ARM::STR_POST_IMM:
2514 case ARM::STR_POST_REG:
2515 case ARM::STR_PRE_IMM:
2518 case ARM::LDRSB_PRE:
2519 case ARM::LDRSH_PRE: {
2520 unsigned Rm = MI->getOperand(3).getReg();
2523 unsigned Rt = MI->getOperand(0).getReg();
2526 unsigned ShOpVal = MI->getOperand(4).getImm();
2527 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2528 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2531 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2532 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2538 unsigned Rt = MI->getOperand(0).getReg();
2539 unsigned Rn = MI->getOperand(2).getReg();
2540 unsigned Rm = MI->getOperand(3).getReg();
2542 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2543 return (Rt == Rn) ? 3 : 2;
2547 unsigned Rm = MI->getOperand(3).getReg();
2549 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2553 case ARM::LDRD_POST:
2554 case ARM::t2LDRD_POST:
2557 case ARM::STRD_POST:
2558 case ARM::t2STRD_POST:
2561 case ARM::LDRD_PRE: {
2562 unsigned Rt = MI->getOperand(0).getReg();
2563 unsigned Rn = MI->getOperand(3).getReg();
2564 unsigned Rm = MI->getOperand(4).getReg();
2566 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2567 return (Rt == Rn) ? 4 : 3;
2570 case ARM::t2LDRD_PRE: {
2571 unsigned Rt = MI->getOperand(0).getReg();
2572 unsigned Rn = MI->getOperand(3).getReg();
2573 return (Rt == Rn) ? 4 : 3;
2576 case ARM::STRD_PRE: {
2577 unsigned Rm = MI->getOperand(4).getReg();
2579 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2583 case ARM::t2STRD_PRE:
2586 case ARM::t2LDR_POST:
2587 case ARM::t2LDRB_POST:
2588 case ARM::t2LDRB_PRE:
2589 case ARM::t2LDRSBi12:
2590 case ARM::t2LDRSBi8:
2591 case ARM::t2LDRSBpci:
2593 case ARM::t2LDRH_POST:
2594 case ARM::t2LDRH_PRE:
2596 case ARM::t2LDRSB_POST:
2597 case ARM::t2LDRSB_PRE:
2598 case ARM::t2LDRSH_POST:
2599 case ARM::t2LDRSH_PRE:
2600 case ARM::t2LDRSHi12:
2601 case ARM::t2LDRSHi8:
2602 case ARM::t2LDRSHpci:
2606 case ARM::t2LDRDi8: {
2607 unsigned Rt = MI->getOperand(0).getReg();
2608 unsigned Rn = MI->getOperand(2).getReg();
2609 return (Rt == Rn) ? 3 : 2;
2612 case ARM::t2STRB_POST:
2613 case ARM::t2STRB_PRE:
2616 case ARM::t2STRH_POST:
2617 case ARM::t2STRH_PRE:
2619 case ARM::t2STR_POST:
2620 case ARM::t2STR_PRE:
2626 // Return the number of 32-bit words loaded by LDM or stored by STM. If this
2627 // can't be easily determined return 0 (missing MachineMemOperand).
2629 // FIXME: The current MachineInstr design does not support relying on machine
2630 // mem operands to determine the width of a memory access. Instead, we expect
2631 // the target to provide this information based on the instruction opcode and
2632 // operands. However, using MachineMemOperand is a the best solution now for
2635 // 1) getNumMicroOps tries to infer LDM memory width from the total number of MI
2636 // operands. This is much more dangerous than using the MachineMemOperand
2637 // sizes because CodeGen passes can insert/remove optional machine operands. In
2638 // fact, it's totally incorrect for preRA passes and appears to be wrong for
2639 // postRA passes as well.
2641 // 2) getNumLDMAddresses is only used by the scheduling machine model and any
2642 // machine model that calls this should handle the unknown (zero size) case.
2644 // Long term, we should require a target hook that verifies MachineMemOperand
2645 // sizes during MC lowering. That target hook should be local to MC lowering
2646 // because we can't ensure that it is aware of other MI forms. Doing this will
2647 // ensure that MachineMemOperands are correctly propagated through all passes.
2648 unsigned ARMBaseInstrInfo::getNumLDMAddresses(const MachineInstr *MI) const {
2650 for (MachineInstr::mmo_iterator I = MI->memoperands_begin(),
2651 E = MI->memoperands_end(); I != E; ++I) {
2652 Size += (*I)->getSize();
2658 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
2659 const MachineInstr *MI) const {
2660 if (!ItinData || ItinData->isEmpty())
2663 const MCInstrDesc &Desc = MI->getDesc();
2664 unsigned Class = Desc.getSchedClass();
2665 int ItinUOps = ItinData->getNumMicroOps(Class);
2666 if (ItinUOps >= 0) {
2667 if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore()))
2668 return getNumMicroOpsSwiftLdSt(ItinData, MI);
2673 unsigned Opc = MI->getOpcode();
2676 llvm_unreachable("Unexpected multi-uops instruction!");
2681 // The number of uOps for load / store multiple are determined by the number
2684 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
2685 // same cycle. The scheduling for the first load / store must be done
2686 // separately by assuming the address is not 64-bit aligned.
2688 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
2689 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
2690 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
2692 case ARM::VLDMDIA_UPD:
2693 case ARM::VLDMDDB_UPD:
2695 case ARM::VLDMSIA_UPD:
2696 case ARM::VLDMSDB_UPD:
2698 case ARM::VSTMDIA_UPD:
2699 case ARM::VSTMDDB_UPD:
2701 case ARM::VSTMSIA_UPD:
2702 case ARM::VSTMSDB_UPD: {
2703 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
2704 return (NumRegs / 2) + (NumRegs % 2) + 1;
2707 case ARM::LDMIA_RET:
2712 case ARM::LDMIA_UPD:
2713 case ARM::LDMDA_UPD:
2714 case ARM::LDMDB_UPD:
2715 case ARM::LDMIB_UPD:
2720 case ARM::STMIA_UPD:
2721 case ARM::STMDA_UPD:
2722 case ARM::STMDB_UPD:
2723 case ARM::STMIB_UPD:
2725 case ARM::tLDMIA_UPD:
2726 case ARM::tSTMIA_UPD:
2730 case ARM::t2LDMIA_RET:
2733 case ARM::t2LDMIA_UPD:
2734 case ARM::t2LDMDB_UPD:
2737 case ARM::t2STMIA_UPD:
2738 case ARM::t2STMDB_UPD: {
2739 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
2740 if (Subtarget.isSwift()) {
2741 int UOps = 1 + NumRegs; // One for address computation, one for each ld / st.
2744 case ARM::VLDMDIA_UPD:
2745 case ARM::VLDMDDB_UPD:
2746 case ARM::VLDMSIA_UPD:
2747 case ARM::VLDMSDB_UPD:
2748 case ARM::VSTMDIA_UPD:
2749 case ARM::VSTMDDB_UPD:
2750 case ARM::VSTMSIA_UPD:
2751 case ARM::VSTMSDB_UPD:
2752 case ARM::LDMIA_UPD:
2753 case ARM::LDMDA_UPD:
2754 case ARM::LDMDB_UPD:
2755 case ARM::LDMIB_UPD:
2756 case ARM::STMIA_UPD:
2757 case ARM::STMDA_UPD:
2758 case ARM::STMDB_UPD:
2759 case ARM::STMIB_UPD:
2760 case ARM::tLDMIA_UPD:
2761 case ARM::tSTMIA_UPD:
2762 case ARM::t2LDMIA_UPD:
2763 case ARM::t2LDMDB_UPD:
2764 case ARM::t2STMIA_UPD:
2765 case ARM::t2STMDB_UPD:
2766 ++UOps; // One for base register writeback.
2768 case ARM::LDMIA_RET:
2770 case ARM::t2LDMIA_RET:
2771 UOps += 2; // One for base reg wb, one for write to pc.
2775 } else if (Subtarget.isCortexA8()) {
2778 // 4 registers would be issued: 2, 2.
2779 // 5 registers would be issued: 2, 2, 1.
2780 int A8UOps = (NumRegs / 2);
2784 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2785 int A9UOps = (NumRegs / 2);
2786 // If there are odd number of registers or if it's not 64-bit aligned,
2787 // then it takes an extra AGU (Address Generation Unit) cycle.
2788 if ((NumRegs % 2) ||
2789 !MI->hasOneMemOperand() ||
2790 (*MI->memoperands_begin())->getAlignment() < 8)
2794 // Assume the worst.
2802 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
2803 const MCInstrDesc &DefMCID,
2805 unsigned DefIdx, unsigned DefAlign) const {
2806 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2808 // Def is the address writeback.
2809 return ItinData->getOperandCycle(DefClass, DefIdx);
2812 if (Subtarget.isCortexA8()) {
2813 // (regno / 2) + (regno % 2) + 1
2814 DefCycle = RegNo / 2 + 1;
2817 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2819 bool isSLoad = false;
2821 switch (DefMCID.getOpcode()) {
2824 case ARM::VLDMSIA_UPD:
2825 case ARM::VLDMSDB_UPD:
2830 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2831 // then it takes an extra cycle.
2832 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
2835 // Assume the worst.
2836 DefCycle = RegNo + 2;
2843 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
2844 const MCInstrDesc &DefMCID,
2846 unsigned DefIdx, unsigned DefAlign) const {
2847 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2849 // Def is the address writeback.
2850 return ItinData->getOperandCycle(DefClass, DefIdx);
2853 if (Subtarget.isCortexA8()) {
2854 // 4 registers would be issued: 1, 2, 1.
2855 // 5 registers would be issued: 1, 2, 2.
2856 DefCycle = RegNo / 2;
2859 // Result latency is issue cycle + 2: E2.
2861 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2862 DefCycle = (RegNo / 2);
2863 // If there are odd number of registers or if it's not 64-bit aligned,
2864 // then it takes an extra AGU (Address Generation Unit) cycle.
2865 if ((RegNo % 2) || DefAlign < 8)
2867 // Result latency is AGU cycles + 2.
2870 // Assume the worst.
2871 DefCycle = RegNo + 2;
2878 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
2879 const MCInstrDesc &UseMCID,
2881 unsigned UseIdx, unsigned UseAlign) const {
2882 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2884 return ItinData->getOperandCycle(UseClass, UseIdx);
2887 if (Subtarget.isCortexA8()) {
2888 // (regno / 2) + (regno % 2) + 1
2889 UseCycle = RegNo / 2 + 1;
2892 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2894 bool isSStore = false;
2896 switch (UseMCID.getOpcode()) {
2899 case ARM::VSTMSIA_UPD:
2900 case ARM::VSTMSDB_UPD:
2905 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2906 // then it takes an extra cycle.
2907 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2910 // Assume the worst.
2911 UseCycle = RegNo + 2;
2918 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2919 const MCInstrDesc &UseMCID,
2921 unsigned UseIdx, unsigned UseAlign) const {
2922 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2924 return ItinData->getOperandCycle(UseClass, UseIdx);
2927 if (Subtarget.isCortexA8()) {
2928 UseCycle = RegNo / 2;
2933 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2934 UseCycle = (RegNo / 2);
2935 // If there are odd number of registers or if it's not 64-bit aligned,
2936 // then it takes an extra AGU (Address Generation Unit) cycle.
2937 if ((RegNo % 2) || UseAlign < 8)
2940 // Assume the worst.
2947 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2948 const MCInstrDesc &DefMCID,
2949 unsigned DefIdx, unsigned DefAlign,
2950 const MCInstrDesc &UseMCID,
2951 unsigned UseIdx, unsigned UseAlign) const {
2952 unsigned DefClass = DefMCID.getSchedClass();
2953 unsigned UseClass = UseMCID.getSchedClass();
2955 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2956 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2958 // This may be a def / use of a variable_ops instruction, the operand
2959 // latency might be determinable dynamically. Let the target try to
2962 bool LdmBypass = false;
2963 switch (DefMCID.getOpcode()) {
2965 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2969 case ARM::VLDMDIA_UPD:
2970 case ARM::VLDMDDB_UPD:
2972 case ARM::VLDMSIA_UPD:
2973 case ARM::VLDMSDB_UPD:
2974 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2977 case ARM::LDMIA_RET:
2982 case ARM::LDMIA_UPD:
2983 case ARM::LDMDA_UPD:
2984 case ARM::LDMDB_UPD:
2985 case ARM::LDMIB_UPD:
2987 case ARM::tLDMIA_UPD:
2989 case ARM::t2LDMIA_RET:
2992 case ARM::t2LDMIA_UPD:
2993 case ARM::t2LDMDB_UPD:
2995 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
3000 // We can't seem to determine the result latency of the def, assume it's 2.
3004 switch (UseMCID.getOpcode()) {
3006 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
3010 case ARM::VSTMDIA_UPD:
3011 case ARM::VSTMDDB_UPD:
3013 case ARM::VSTMSIA_UPD:
3014 case ARM::VSTMSDB_UPD:
3015 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3022 case ARM::STMIA_UPD:
3023 case ARM::STMDA_UPD:
3024 case ARM::STMDB_UPD:
3025 case ARM::STMIB_UPD:
3026 case ARM::tSTMIA_UPD:
3031 case ARM::t2STMIA_UPD:
3032 case ARM::t2STMDB_UPD:
3033 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3038 // Assume it's read in the first stage.
3041 UseCycle = DefCycle - UseCycle + 1;
3044 // It's a variable_ops instruction so we can't use DefIdx here. Just use
3045 // first def operand.
3046 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
3049 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
3050 UseClass, UseIdx)) {
3058 static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
3059 const MachineInstr *MI, unsigned Reg,
3060 unsigned &DefIdx, unsigned &Dist) {
3063 MachineBasicBlock::const_iterator I = MI; ++I;
3064 MachineBasicBlock::const_instr_iterator II =
3065 llvm::prior(I.getInstrIterator());
3066 assert(II->isInsideBundle() && "Empty bundle?");
3069 while (II->isInsideBundle()) {
3070 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
3077 assert(Idx != -1 && "Cannot find bundled definition!");
3082 static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
3083 const MachineInstr *MI, unsigned Reg,
3084 unsigned &UseIdx, unsigned &Dist) {
3087 MachineBasicBlock::const_instr_iterator II = MI; ++II;
3088 assert(II->isInsideBundle() && "Empty bundle?");
3089 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3091 // FIXME: This doesn't properly handle multiple uses.
3093 while (II != E && II->isInsideBundle()) {
3094 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
3097 if (II->getOpcode() != ARM::t2IT)
3111 /// Return the number of cycles to add to (or subtract from) the static
3112 /// itinerary based on the def opcode and alignment. The caller will ensure that
3113 /// adjusted latency is at least one cycle.
3114 static int adjustDefLatency(const ARMSubtarget &Subtarget,
3115 const MachineInstr *DefMI,
3116 const MCInstrDesc *DefMCID, unsigned DefAlign) {
3118 if (Subtarget.isCortexA8() || Subtarget.isLikeA9()) {
3119 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3120 // variants are one cycle cheaper.
3121 switch (DefMCID->getOpcode()) {
3125 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3126 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3128 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3135 case ARM::t2LDRSHs: {
3136 // Thumb2 mode: lsl only.
3137 unsigned ShAmt = DefMI->getOperand(3).getImm();
3138 if (ShAmt == 0 || ShAmt == 2)
3143 } else if (Subtarget.isSwift()) {
3144 // FIXME: Properly handle all of the latency adjustments for address
3146 switch (DefMCID->getOpcode()) {
3150 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3151 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
3152 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3155 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3156 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
3159 ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3166 case ARM::t2LDRSHs: {
3167 // Thumb2 mode: lsl only.
3168 unsigned ShAmt = DefMI->getOperand(3).getImm();
3169 if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3)
3176 if (DefAlign < 8 && Subtarget.isLikeA9()) {
3177 switch (DefMCID->getOpcode()) {
3183 case ARM::VLD1q8wb_fixed:
3184 case ARM::VLD1q16wb_fixed:
3185 case ARM::VLD1q32wb_fixed:
3186 case ARM::VLD1q64wb_fixed:
3187 case ARM::VLD1q8wb_register:
3188 case ARM::VLD1q16wb_register:
3189 case ARM::VLD1q32wb_register:
3190 case ARM::VLD1q64wb_register:
3197 case ARM::VLD2d8wb_fixed:
3198 case ARM::VLD2d16wb_fixed:
3199 case ARM::VLD2d32wb_fixed:
3200 case ARM::VLD2q8wb_fixed:
3201 case ARM::VLD2q16wb_fixed:
3202 case ARM::VLD2q32wb_fixed:
3203 case ARM::VLD2d8wb_register:
3204 case ARM::VLD2d16wb_register:
3205 case ARM::VLD2d32wb_register:
3206 case ARM::VLD2q8wb_register:
3207 case ARM::VLD2q16wb_register:
3208 case ARM::VLD2q32wb_register:
3213 case ARM::VLD3d8_UPD:
3214 case ARM::VLD3d16_UPD:
3215 case ARM::VLD3d32_UPD:
3216 case ARM::VLD1d64Twb_fixed:
3217 case ARM::VLD1d64Twb_register:
3218 case ARM::VLD3q8_UPD:
3219 case ARM::VLD3q16_UPD:
3220 case ARM::VLD3q32_UPD:
3225 case ARM::VLD4d8_UPD:
3226 case ARM::VLD4d16_UPD:
3227 case ARM::VLD4d32_UPD:
3228 case ARM::VLD1d64Qwb_fixed:
3229 case ARM::VLD1d64Qwb_register:
3230 case ARM::VLD4q8_UPD:
3231 case ARM::VLD4q16_UPD:
3232 case ARM::VLD4q32_UPD:
3233 case ARM::VLD1DUPq8:
3234 case ARM::VLD1DUPq16:
3235 case ARM::VLD1DUPq32:
3236 case ARM::VLD1DUPq8wb_fixed:
3237 case ARM::VLD1DUPq16wb_fixed:
3238 case ARM::VLD1DUPq32wb_fixed:
3239 case ARM::VLD1DUPq8wb_register:
3240 case ARM::VLD1DUPq16wb_register:
3241 case ARM::VLD1DUPq32wb_register:
3242 case ARM::VLD2DUPd8:
3243 case ARM::VLD2DUPd16:
3244 case ARM::VLD2DUPd32:
3245 case ARM::VLD2DUPd8wb_fixed:
3246 case ARM::VLD2DUPd16wb_fixed:
3247 case ARM::VLD2DUPd32wb_fixed:
3248 case ARM::VLD2DUPd8wb_register:
3249 case ARM::VLD2DUPd16wb_register:
3250 case ARM::VLD2DUPd32wb_register:
3251 case ARM::VLD4DUPd8:
3252 case ARM::VLD4DUPd16:
3253 case ARM::VLD4DUPd32:
3254 case ARM::VLD4DUPd8_UPD:
3255 case ARM::VLD4DUPd16_UPD:
3256 case ARM::VLD4DUPd32_UPD:
3258 case ARM::VLD1LNd16:
3259 case ARM::VLD1LNd32:
3260 case ARM::VLD1LNd8_UPD:
3261 case ARM::VLD1LNd16_UPD:
3262 case ARM::VLD1LNd32_UPD:
3264 case ARM::VLD2LNd16:
3265 case ARM::VLD2LNd32:
3266 case ARM::VLD2LNq16:
3267 case ARM::VLD2LNq32:
3268 case ARM::VLD2LNd8_UPD:
3269 case ARM::VLD2LNd16_UPD:
3270 case ARM::VLD2LNd32_UPD:
3271 case ARM::VLD2LNq16_UPD:
3272 case ARM::VLD2LNq32_UPD:
3274 case ARM::VLD4LNd16:
3275 case ARM::VLD4LNd32:
3276 case ARM::VLD4LNq16:
3277 case ARM::VLD4LNq32:
3278 case ARM::VLD4LNd8_UPD:
3279 case ARM::VLD4LNd16_UPD:
3280 case ARM::VLD4LNd32_UPD:
3281 case ARM::VLD4LNq16_UPD:
3282 case ARM::VLD4LNq32_UPD:
3283 // If the address is not 64-bit aligned, the latencies of these
3284 // instructions increases by one.
3295 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3296 const MachineInstr *DefMI, unsigned DefIdx,
3297 const MachineInstr *UseMI,
3298 unsigned UseIdx) const {
3299 // No operand latency. The caller may fall back to getInstrLatency.
3300 if (!ItinData || ItinData->isEmpty())
3303 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
3304 unsigned Reg = DefMO.getReg();
3305 const MCInstrDesc *DefMCID = &DefMI->getDesc();
3306 const MCInstrDesc *UseMCID = &UseMI->getDesc();
3308 unsigned DefAdj = 0;
3309 if (DefMI->isBundle()) {
3310 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
3311 DefMCID = &DefMI->getDesc();
3313 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
3314 DefMI->isRegSequence() || DefMI->isImplicitDef()) {
3318 unsigned UseAdj = 0;
3319 if (UseMI->isBundle()) {
3321 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
3322 Reg, NewUseIdx, UseAdj);
3328 UseMCID = &UseMI->getDesc();
3331 if (Reg == ARM::CPSR) {
3332 if (DefMI->getOpcode() == ARM::FMSTAT) {
3333 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
3334 return Subtarget.isLikeA9() ? 1 : 20;
3337 // CPSR set and branch can be paired in the same cycle.
3338 if (UseMI->isBranch())
3341 // Otherwise it takes the instruction latency (generally one).
3342 unsigned Latency = getInstrLatency(ItinData, DefMI);
3344 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
3345 // its uses. Instructions which are otherwise scheduled between them may
3346 // incur a code size penalty (not able to use the CPSR setting 16-bit
3348 if (Latency > 0 && Subtarget.isThumb2()) {
3349 const MachineFunction *MF = DefMI->getParent()->getParent();
3350 if (MF->getFunction()->getAttributes().
3351 hasAttribute(AttributeSet::FunctionIndex,
3352 Attribute::OptimizeForSize))
3358 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
3361 unsigned DefAlign = DefMI->hasOneMemOperand()
3362 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
3363 unsigned UseAlign = UseMI->hasOneMemOperand()
3364 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
3366 // Get the itinerary's latency if possible, and handle variable_ops.
3367 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
3368 *UseMCID, UseIdx, UseAlign);
3369 // Unable to find operand latency. The caller may resort to getInstrLatency.
3373 // Adjust for IT block position.
3374 int Adj = DefAdj + UseAdj;
3376 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3377 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
3378 if (Adj >= 0 || (int)Latency > -Adj) {
3379 return Latency + Adj;
3381 // Return the itinerary latency, which may be zero but not less than zero.
3386 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3387 SDNode *DefNode, unsigned DefIdx,
3388 SDNode *UseNode, unsigned UseIdx) const {
3389 if (!DefNode->isMachineOpcode())
3392 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
3394 if (isZeroCost(DefMCID.Opcode))
3397 if (!ItinData || ItinData->isEmpty())
3398 return DefMCID.mayLoad() ? 3 : 1;
3400 if (!UseNode->isMachineOpcode()) {
3401 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
3402 if (Subtarget.isLikeA9() || Subtarget.isSwift())
3403 return Latency <= 2 ? 1 : Latency - 1;
3405 return Latency <= 3 ? 1 : Latency - 2;
3408 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
3409 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
3410 unsigned DefAlign = !DefMN->memoperands_empty()
3411 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
3412 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
3413 unsigned UseAlign = !UseMN->memoperands_empty()
3414 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
3415 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
3416 UseMCID, UseIdx, UseAlign);
3419 (Subtarget.isCortexA8() || Subtarget.isLikeA9())) {
3420 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3421 // variants are one cycle cheaper.
3422 switch (DefMCID.getOpcode()) {
3427 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3428 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3430 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3437 case ARM::t2LDRSHs: {
3438 // Thumb2 mode: lsl only.
3440 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3441 if (ShAmt == 0 || ShAmt == 2)
3446 } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) {
3447 // FIXME: Properly handle all of the latency adjustments for address
3449 switch (DefMCID.getOpcode()) {
3454 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3455 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3457 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3458 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3460 else if (ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3467 case ARM::t2LDRSHs: {
3468 // Thumb2 mode: lsl 0-3 only.
3475 if (DefAlign < 8 && Subtarget.isLikeA9())
3476 switch (DefMCID.getOpcode()) {
3482 case ARM::VLD1q8wb_register:
3483 case ARM::VLD1q16wb_register:
3484 case ARM::VLD1q32wb_register:
3485 case ARM::VLD1q64wb_register:
3486 case ARM::VLD1q8wb_fixed:
3487 case ARM::VLD1q16wb_fixed:
3488 case ARM::VLD1q32wb_fixed:
3489 case ARM::VLD1q64wb_fixed:
3493 case ARM::VLD2q8Pseudo:
3494 case ARM::VLD2q16Pseudo:
3495 case ARM::VLD2q32Pseudo:
3496 case ARM::VLD2d8wb_fixed:
3497 case ARM::VLD2d16wb_fixed:
3498 case ARM::VLD2d32wb_fixed:
3499 case ARM::VLD2q8PseudoWB_fixed:
3500 case ARM::VLD2q16PseudoWB_fixed:
3501 case ARM::VLD2q32PseudoWB_fixed:
3502 case ARM::VLD2d8wb_register:
3503 case ARM::VLD2d16wb_register:
3504 case ARM::VLD2d32wb_register:
3505 case ARM::VLD2q8PseudoWB_register:
3506 case ARM::VLD2q16PseudoWB_register:
3507 case ARM::VLD2q32PseudoWB_register:
3508 case ARM::VLD3d8Pseudo:
3509 case ARM::VLD3d16Pseudo:
3510 case ARM::VLD3d32Pseudo:
3511 case ARM::VLD1d64TPseudo:
3512 case ARM::VLD3d8Pseudo_UPD:
3513 case ARM::VLD3d16Pseudo_UPD:
3514 case ARM::VLD3d32Pseudo_UPD:
3515 case ARM::VLD3q8Pseudo_UPD:
3516 case ARM::VLD3q16Pseudo_UPD:
3517 case ARM::VLD3q32Pseudo_UPD:
3518 case ARM::VLD3q8oddPseudo:
3519 case ARM::VLD3q16oddPseudo:
3520 case ARM::VLD3q32oddPseudo:
3521 case ARM::VLD3q8oddPseudo_UPD:
3522 case ARM::VLD3q16oddPseudo_UPD:
3523 case ARM::VLD3q32oddPseudo_UPD:
3524 case ARM::VLD4d8Pseudo:
3525 case ARM::VLD4d16Pseudo:
3526 case ARM::VLD4d32Pseudo:
3527 case ARM::VLD1d64QPseudo:
3528 case ARM::VLD4d8Pseudo_UPD:
3529 case ARM::VLD4d16Pseudo_UPD:
3530 case ARM::VLD4d32Pseudo_UPD:
3531 case ARM::VLD4q8Pseudo_UPD:
3532 case ARM::VLD4q16Pseudo_UPD:
3533 case ARM::VLD4q32Pseudo_UPD:
3534 case ARM::VLD4q8oddPseudo:
3535 case ARM::VLD4q16oddPseudo:
3536 case ARM::VLD4q32oddPseudo:
3537 case ARM::VLD4q8oddPseudo_UPD:
3538 case ARM::VLD4q16oddPseudo_UPD:
3539 case ARM::VLD4q32oddPseudo_UPD:
3540 case ARM::VLD1DUPq8:
3541 case ARM::VLD1DUPq16:
3542 case ARM::VLD1DUPq32:
3543 case ARM::VLD1DUPq8wb_fixed:
3544 case ARM::VLD1DUPq16wb_fixed:
3545 case ARM::VLD1DUPq32wb_fixed:
3546 case ARM::VLD1DUPq8wb_register:
3547 case ARM::VLD1DUPq16wb_register:
3548 case ARM::VLD1DUPq32wb_register:
3549 case ARM::VLD2DUPd8:
3550 case ARM::VLD2DUPd16:
3551 case ARM::VLD2DUPd32:
3552 case ARM::VLD2DUPd8wb_fixed:
3553 case ARM::VLD2DUPd16wb_fixed:
3554 case ARM::VLD2DUPd32wb_fixed:
3555 case ARM::VLD2DUPd8wb_register:
3556 case ARM::VLD2DUPd16wb_register:
3557 case ARM::VLD2DUPd32wb_register:
3558 case ARM::VLD4DUPd8Pseudo:
3559 case ARM::VLD4DUPd16Pseudo:
3560 case ARM::VLD4DUPd32Pseudo:
3561 case ARM::VLD4DUPd8Pseudo_UPD:
3562 case ARM::VLD4DUPd16Pseudo_UPD:
3563 case ARM::VLD4DUPd32Pseudo_UPD:
3564 case ARM::VLD1LNq8Pseudo:
3565 case ARM::VLD1LNq16Pseudo:
3566 case ARM::VLD1LNq32Pseudo:
3567 case ARM::VLD1LNq8Pseudo_UPD:
3568 case ARM::VLD1LNq16Pseudo_UPD:
3569 case ARM::VLD1LNq32Pseudo_UPD:
3570 case ARM::VLD2LNd8Pseudo:
3571 case ARM::VLD2LNd16Pseudo:
3572 case ARM::VLD2LNd32Pseudo:
3573 case ARM::VLD2LNq16Pseudo:
3574 case ARM::VLD2LNq32Pseudo:
3575 case ARM::VLD2LNd8Pseudo_UPD:
3576 case ARM::VLD2LNd16Pseudo_UPD:
3577 case ARM::VLD2LNd32Pseudo_UPD:
3578 case ARM::VLD2LNq16Pseudo_UPD:
3579 case ARM::VLD2LNq32Pseudo_UPD:
3580 case ARM::VLD4LNd8Pseudo:
3581 case ARM::VLD4LNd16Pseudo:
3582 case ARM::VLD4LNd32Pseudo:
3583 case ARM::VLD4LNq16Pseudo:
3584 case ARM::VLD4LNq32Pseudo:
3585 case ARM::VLD4LNd8Pseudo_UPD:
3586 case ARM::VLD4LNd16Pseudo_UPD:
3587 case ARM::VLD4LNd32Pseudo_UPD:
3588 case ARM::VLD4LNq16Pseudo_UPD:
3589 case ARM::VLD4LNq32Pseudo_UPD:
3590 // If the address is not 64-bit aligned, the latencies of these
3591 // instructions increases by one.
3599 unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3600 const MachineInstr *MI,
3601 unsigned *PredCost) const {
3602 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3603 MI->isRegSequence() || MI->isImplicitDef())
3606 // An instruction scheduler typically runs on unbundled instructions, however
3607 // other passes may query the latency of a bundled instruction.
3608 if (MI->isBundle()) {
3609 unsigned Latency = 0;
3610 MachineBasicBlock::const_instr_iterator I = MI;
3611 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3612 while (++I != E && I->isInsideBundle()) {
3613 if (I->getOpcode() != ARM::t2IT)
3614 Latency += getInstrLatency(ItinData, I, PredCost);
3619 const MCInstrDesc &MCID = MI->getDesc();
3620 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) {
3621 // When predicated, CPSR is an additional source operand for CPSR updating
3622 // instructions, this apparently increases their latencies.
3625 // Be sure to call getStageLatency for an empty itinerary in case it has a
3626 // valid MinLatency property.
3628 return MI->mayLoad() ? 3 : 1;
3630 unsigned Class = MCID.getSchedClass();
3632 // For instructions with variable uops, use uops as latency.
3633 if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0)
3634 return getNumMicroOps(ItinData, MI);
3636 // For the common case, fall back on the itinerary's latency.
3637 unsigned Latency = ItinData->getStageLatency(Class);
3639 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3640 unsigned DefAlign = MI->hasOneMemOperand()
3641 ? (*MI->memoperands_begin())->getAlignment() : 0;
3642 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
3643 if (Adj >= 0 || (int)Latency > -Adj) {
3644 return Latency + Adj;
3649 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3650 SDNode *Node) const {
3651 if (!Node->isMachineOpcode())
3654 if (!ItinData || ItinData->isEmpty())
3657 unsigned Opcode = Node->getMachineOpcode();
3660 return ItinData->getStageLatency(get(Opcode).getSchedClass());
3667 bool ARMBaseInstrInfo::
3668 hasHighOperandLatency(const InstrItineraryData *ItinData,
3669 const MachineRegisterInfo *MRI,
3670 const MachineInstr *DefMI, unsigned DefIdx,
3671 const MachineInstr *UseMI, unsigned UseIdx) const {
3672 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3673 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
3674 if (Subtarget.isCortexA8() &&
3675 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
3676 // CortexA8 VFP instructions are not pipelined.
3679 // Hoist VFP / NEON instructions with 4 or higher latency.
3680 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
3682 Latency = getInstrLatency(ItinData, DefMI);
3685 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
3686 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
3689 bool ARMBaseInstrInfo::
3690 hasLowDefLatency(const InstrItineraryData *ItinData,
3691 const MachineInstr *DefMI, unsigned DefIdx) const {
3692 if (!ItinData || ItinData->isEmpty())
3695 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3696 if (DDomain == ARMII::DomainGeneral) {
3697 unsigned DefClass = DefMI->getDesc().getSchedClass();
3698 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3699 return (DefCycle != -1 && DefCycle <= 2);
3704 bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
3705 StringRef &ErrInfo) const {
3706 if (convertAddSubFlagsOpcode(MI->getOpcode())) {
3707 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
3714 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
3715 unsigned &AddSubOpc,
3716 bool &NegAcc, bool &HasLane) const {
3717 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
3718 if (I == MLxEntryMap.end())
3721 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
3722 MulOpc = Entry.MulOpc;
3723 AddSubOpc = Entry.AddSubOpc;
3724 NegAcc = Entry.NegAcc;
3725 HasLane = Entry.HasLane;
3729 //===----------------------------------------------------------------------===//
3730 // Execution domains.
3731 //===----------------------------------------------------------------------===//
3733 // Some instructions go down the NEON pipeline, some go down the VFP pipeline,
3734 // and some can go down both. The vmov instructions go down the VFP pipeline,
3735 // but they can be changed to vorr equivalents that are executed by the NEON
3738 // We use the following execution domain numbering:
3746 // Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
3748 std::pair<uint16_t, uint16_t>
3749 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
3750 // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
3751 // if they are not predicated.
3752 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
3753 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3755 // CortexA9 is particularly picky about mixing the two and wants these
3757 if (Subtarget.isCortexA9() && !isPredicated(MI) &&
3758 (MI->getOpcode() == ARM::VMOVRS ||
3759 MI->getOpcode() == ARM::VMOVSR ||
3760 MI->getOpcode() == ARM::VMOVS))
3761 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3763 // No other instructions can be swizzled, so just determine their domain.
3764 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
3766 if (Domain & ARMII::DomainNEON)
3767 return std::make_pair(ExeNEON, 0);
3769 // Certain instructions can go either way on Cortex-A8.
3770 // Treat them as NEON instructions.
3771 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
3772 return std::make_pair(ExeNEON, 0);
3774 if (Domain & ARMII::DomainVFP)
3775 return std::make_pair(ExeVFP, 0);
3777 return std::make_pair(ExeGeneric, 0);
3780 static unsigned getCorrespondingDRegAndLane(const TargetRegisterInfo *TRI,
3781 unsigned SReg, unsigned &Lane) {
3782 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_0, &ARM::DPRRegClass);
3785 if (DReg != ARM::NoRegister)
3789 DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1, &ARM::DPRRegClass);
3791 assert(DReg && "S-register with no D super-register?");
3795 /// getImplicitSPRUseForDPRUse - Given a use of a DPR register and lane,
3796 /// set ImplicitSReg to a register number that must be marked as implicit-use or
3797 /// zero if no register needs to be defined as implicit-use.
3799 /// If the function cannot determine if an SPR should be marked implicit use or
3800 /// not, it returns false.
3802 /// This function handles cases where an instruction is being modified from taking
3803 /// an SPR to a DPR[Lane]. A use of the DPR is being added, which may conflict
3804 /// with an earlier def of an SPR corresponding to DPR[Lane^1] (i.e. the other
3805 /// lane of the DPR).
3807 /// If the other SPR is defined, an implicit-use of it should be added. Else,
3808 /// (including the case where the DPR itself is defined), it should not.
3810 static bool getImplicitSPRUseForDPRUse(const TargetRegisterInfo *TRI,
3812 unsigned DReg, unsigned Lane,
3813 unsigned &ImplicitSReg) {
3814 // If the DPR is defined or used already, the other SPR lane will be chained
3815 // correctly, so there is nothing to be done.
3816 if (MI->definesRegister(DReg, TRI) || MI->readsRegister(DReg, TRI)) {
3821 // Otherwise we need to go searching to see if the SPR is set explicitly.
3822 ImplicitSReg = TRI->getSubReg(DReg,
3823 (Lane & 1) ? ARM::ssub_0 : ARM::ssub_1);
3824 MachineBasicBlock::LivenessQueryResult LQR =
3825 MI->getParent()->computeRegisterLiveness(TRI, ImplicitSReg, MI);
3827 if (LQR == MachineBasicBlock::LQR_Live)
3829 else if (LQR == MachineBasicBlock::LQR_Unknown)
3832 // If the register is known not to be live, there is no need to add an
3839 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
3840 unsigned DstReg, SrcReg, DReg;
3842 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3843 const TargetRegisterInfo *TRI = &getRegisterInfo();
3844 switch (MI->getOpcode()) {
3846 llvm_unreachable("cannot handle opcode!");
3849 if (Domain != ExeNEON)
3852 // Zap the predicate operands.
3853 assert(!isPredicated(MI) && "Cannot predicate a VORRd");
3855 // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
3856 DstReg = MI->getOperand(0).getReg();
3857 SrcReg = MI->getOperand(1).getReg();
3859 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3860 MI->RemoveOperand(i-1);
3862 // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits)
3863 MI->setDesc(get(ARM::VORRd));
3864 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3869 if (Domain != ExeNEON)
3871 assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
3873 // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits)
3874 DstReg = MI->getOperand(0).getReg();
3875 SrcReg = MI->getOperand(1).getReg();
3877 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3878 MI->RemoveOperand(i-1);
3880 DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane);
3882 // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps)
3883 // Note that DSrc has been widened and the other lane may be undef, which
3884 // contaminates the entire register.
3885 MI->setDesc(get(ARM::VGETLNi32));
3886 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3887 .addReg(DReg, RegState::Undef)
3890 // The old source should be an implicit use, otherwise we might think it
3891 // was dead before here.
3892 MIB.addReg(SrcReg, RegState::Implicit);
3895 if (Domain != ExeNEON)
3897 assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
3899 // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits)
3900 DstReg = MI->getOperand(0).getReg();
3901 SrcReg = MI->getOperand(1).getReg();
3903 DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane);
3905 unsigned ImplicitSReg;
3906 if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg))
3909 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3910 MI->RemoveOperand(i-1);
3912 // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps)
3913 // Again DDst may be undefined at the beginning of this instruction.
3914 MI->setDesc(get(ARM::VSETLNi32));
3915 MIB.addReg(DReg, RegState::Define)
3916 .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI)))
3919 AddDefaultPred(MIB);
3921 // The narrower destination must be marked as set to keep previous chains
3923 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
3924 if (ImplicitSReg != 0)
3925 MIB.addReg(ImplicitSReg, RegState::Implicit);
3929 if (Domain != ExeNEON)
3932 // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits)
3933 DstReg = MI->getOperand(0).getReg();
3934 SrcReg = MI->getOperand(1).getReg();
3936 unsigned DstLane = 0, SrcLane = 0, DDst, DSrc;
3937 DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane);
3938 DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane);
3940 unsigned ImplicitSReg;
3941 if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg))
3944 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3945 MI->RemoveOperand(i-1);
3948 // Destination can be:
3949 // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits)
3950 MI->setDesc(get(ARM::VDUPLN32d));
3951 MIB.addReg(DDst, RegState::Define)
3952 .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI)))
3954 AddDefaultPred(MIB);
3956 // Neither the source or the destination are naturally represented any
3957 // more, so add them in manually.
3958 MIB.addReg(DstReg, RegState::Implicit | RegState::Define);
3959 MIB.addReg(SrcReg, RegState::Implicit);
3960 if (ImplicitSReg != 0)
3961 MIB.addReg(ImplicitSReg, RegState::Implicit);
3965 // In general there's no single instruction that can perform an S <-> S
3966 // move in NEON space, but a pair of VEXT instructions *can* do the
3967 // job. It turns out that the VEXTs needed will only use DSrc once, with
3968 // the position based purely on the combination of lane-0 and lane-1
3969 // involved. For example
3970 // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1
3971 // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1
3972 // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1
3973 // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1
3975 // Pattern of the MachineInstrs is:
3976 // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits)
3977 MachineInstrBuilder NewMIB;
3978 NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
3979 get(ARM::VEXTd32), DDst);
3981 // On the first instruction, both DSrc and DDst may be <undef> if present.
3982 // Specifically when the original instruction didn't have them as an
3984 unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst;
3985 bool CurUndef = !MI->readsRegister(CurReg, TRI);
3986 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
3988 CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst;
3989 CurUndef = !MI->readsRegister(CurReg, TRI);
3990 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
3993 AddDefaultPred(NewMIB);
3995 if (SrcLane == DstLane)
3996 NewMIB.addReg(SrcReg, RegState::Implicit);
3998 MI->setDesc(get(ARM::VEXTd32));
3999 MIB.addReg(DDst, RegState::Define);
4001 // On the second instruction, DDst has definitely been defined above, so
4002 // it is not <undef>. DSrc, if present, can be <undef> as above.
4003 CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst;
4004 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4005 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4007 CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst;
4008 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4009 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4012 AddDefaultPred(MIB);
4014 if (SrcLane != DstLane)
4015 MIB.addReg(SrcReg, RegState::Implicit);
4017 // As before, the original destination is no longer represented, add it
4019 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4020 if (ImplicitSReg != 0)
4021 MIB.addReg(ImplicitSReg, RegState::Implicit);
4028 //===----------------------------------------------------------------------===//
4029 // Partial register updates
4030 //===----------------------------------------------------------------------===//
4032 // Swift renames NEON registers with 64-bit granularity. That means any
4033 // instruction writing an S-reg implicitly reads the containing D-reg. The
4034 // problem is mostly avoided by translating f32 operations to v2f32 operations
4035 // on D-registers, but f32 loads are still a problem.
4037 // These instructions can load an f32 into a NEON register:
4039 // VLDRS - Only writes S, partial D update.
4040 // VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops.
4041 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
4043 // FCONSTD can be used as a dependency-breaking instruction.
4044 unsigned ARMBaseInstrInfo::
4045 getPartialRegUpdateClearance(const MachineInstr *MI,
4047 const TargetRegisterInfo *TRI) const {
4048 if (!SwiftPartialUpdateClearance ||
4049 !(Subtarget.isSwift() || Subtarget.isCortexA15()))
4052 assert(TRI && "Need TRI instance");
4054 const MachineOperand &MO = MI->getOperand(OpNum);
4057 unsigned Reg = MO.getReg();
4060 switch(MI->getOpcode()) {
4061 // Normal instructions writing only an S-register.
4066 case ARM::VMOVv4i16:
4067 case ARM::VMOVv2i32:
4068 case ARM::VMOVv2f32:
4069 case ARM::VMOVv1i64:
4070 UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI);
4073 // Explicitly reads the dependency.
4074 case ARM::VLD1LNd32:
4081 // If this instruction actually reads a value from Reg, there is no unwanted
4083 if (UseOp != -1 && MI->getOperand(UseOp).readsReg())
4086 // We must be able to clobber the whole D-reg.
4087 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4088 // Virtual register must be a foo:ssub_0<def,undef> operand.
4089 if (!MO.getSubReg() || MI->readsVirtualRegister(Reg))
4091 } else if (ARM::SPRRegClass.contains(Reg)) {
4092 // Physical register: MI must define the full D-reg.
4093 unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0,
4095 if (!DReg || !MI->definesRegister(DReg, TRI))
4099 // MI has an unwanted D-register dependency.
4100 // Avoid defs in the previous N instructrions.
4101 return SwiftPartialUpdateClearance;
4104 // Break a partial register dependency after getPartialRegUpdateClearance
4105 // returned non-zero.
4106 void ARMBaseInstrInfo::
4107 breakPartialRegDependency(MachineBasicBlock::iterator MI,
4109 const TargetRegisterInfo *TRI) const {
4110 assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def");
4111 assert(TRI && "Need TRI instance");
4113 const MachineOperand &MO = MI->getOperand(OpNum);
4114 unsigned Reg = MO.getReg();
4115 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
4116 "Can't break virtual register dependencies.");
4117 unsigned DReg = Reg;
4119 // If MI defines an S-reg, find the corresponding D super-register.
4120 if (ARM::SPRRegClass.contains(Reg)) {
4121 DReg = ARM::D0 + (Reg - ARM::S0) / 2;
4122 assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken");
4125 assert(ARM::DPRRegClass.contains(DReg) && "Can only break D-reg deps");
4126 assert(MI->definesRegister(DReg, TRI) && "MI doesn't clobber full D-reg");
4128 // FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines
4129 // the full D-register by loading the same value to both lanes. The
4130 // instruction is micro-coded with 2 uops, so don't do this until we can
4131 // properly schedule micro-coded instuctions. The dispatcher stalls cause
4132 // too big regressions.
4134 // Insert the dependency-breaking FCONSTD before MI.
4135 // 96 is the encoding of 0.5, but the actual value doesn't matter here.
4136 AddDefaultPred(BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4137 get(ARM::FCONSTD), DReg).addImm(96));
4138 MI->addRegisterKilled(DReg, TRI, true);
4141 bool ARMBaseInstrInfo::hasNOP() const {
4142 return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
4145 bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
4146 if (MI->getNumOperands() < 4)
4148 unsigned ShOpVal = MI->getOperand(3).getImm();
4149 unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
4150 // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
4151 if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) ||
4152 ((ShImm == 1 || ShImm == 2) &&
4153 ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl))