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 *)
117 new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget, DAG);
118 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
122 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
123 MachineBasicBlock::iterator &MBBI,
124 LiveVariables *LV) const {
125 // FIXME: Thumb2 support.
130 MachineInstr *MI = MBBI;
131 MachineFunction &MF = *MI->getParent()->getParent();
132 uint64_t TSFlags = MI->getDesc().TSFlags;
134 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
135 default: return NULL;
136 case ARMII::IndexModePre:
139 case ARMII::IndexModePost:
143 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
145 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
149 MachineInstr *UpdateMI = NULL;
150 MachineInstr *MemMI = NULL;
151 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
152 const MCInstrDesc &MCID = MI->getDesc();
153 unsigned NumOps = MCID.getNumOperands();
154 bool isLoad = !MI->mayStore();
155 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
156 const MachineOperand &Base = MI->getOperand(2);
157 const MachineOperand &Offset = MI->getOperand(NumOps-3);
158 unsigned WBReg = WB.getReg();
159 unsigned BaseReg = Base.getReg();
160 unsigned OffReg = Offset.getReg();
161 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
162 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
164 default: llvm_unreachable("Unknown indexed op!");
165 case ARMII::AddrMode2: {
166 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
167 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
169 if (ARM_AM::getSOImmVal(Amt) == -1)
170 // Can't encode it in a so_imm operand. This transformation will
171 // add more than 1 instruction. Abandon!
173 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
174 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
175 .addReg(BaseReg).addImm(Amt)
176 .addImm(Pred).addReg(0).addReg(0);
177 } else if (Amt != 0) {
178 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
179 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
180 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
181 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg)
182 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
183 .addImm(Pred).addReg(0).addReg(0);
185 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
186 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
187 .addReg(BaseReg).addReg(OffReg)
188 .addImm(Pred).addReg(0).addReg(0);
191 case ARMII::AddrMode3 : {
192 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
193 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
195 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
196 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
197 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
198 .addReg(BaseReg).addImm(Amt)
199 .addImm(Pred).addReg(0).addReg(0);
201 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
202 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
203 .addReg(BaseReg).addReg(OffReg)
204 .addImm(Pred).addReg(0).addReg(0);
209 std::vector<MachineInstr*> NewMIs;
212 MemMI = BuildMI(MF, MI->getDebugLoc(),
213 get(MemOpc), MI->getOperand(0).getReg())
214 .addReg(WBReg).addImm(0).addImm(Pred);
216 MemMI = BuildMI(MF, MI->getDebugLoc(),
217 get(MemOpc)).addReg(MI->getOperand(1).getReg())
218 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
219 NewMIs.push_back(MemMI);
220 NewMIs.push_back(UpdateMI);
223 MemMI = BuildMI(MF, MI->getDebugLoc(),
224 get(MemOpc), MI->getOperand(0).getReg())
225 .addReg(BaseReg).addImm(0).addImm(Pred);
227 MemMI = BuildMI(MF, MI->getDebugLoc(),
228 get(MemOpc)).addReg(MI->getOperand(1).getReg())
229 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
231 UpdateMI->getOperand(0).setIsDead();
232 NewMIs.push_back(UpdateMI);
233 NewMIs.push_back(MemMI);
236 // Transfer LiveVariables states, kill / dead info.
238 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
239 MachineOperand &MO = MI->getOperand(i);
240 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
241 unsigned Reg = MO.getReg();
243 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
245 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
247 LV->addVirtualRegisterDead(Reg, NewMI);
249 if (MO.isUse() && MO.isKill()) {
250 for (unsigned j = 0; j < 2; ++j) {
251 // Look at the two new MI's in reverse order.
252 MachineInstr *NewMI = NewMIs[j];
253 if (!NewMI->readsRegister(Reg))
255 LV->addVirtualRegisterKilled(Reg, NewMI);
256 if (VI.removeKill(MI))
257 VI.Kills.push_back(NewMI);
265 MFI->insert(MBBI, NewMIs[1]);
266 MFI->insert(MBBI, NewMIs[0]);
272 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
273 MachineBasicBlock *&FBB,
274 SmallVectorImpl<MachineOperand> &Cond,
275 bool AllowModify) const {
276 // If the block has no terminators, it just falls into the block after it.
277 MachineBasicBlock::iterator I = MBB.end();
278 if (I == MBB.begin())
281 while (I->isDebugValue()) {
282 if (I == MBB.begin())
286 if (!isUnpredicatedTerminator(I))
289 // Get the last instruction in the block.
290 MachineInstr *LastInst = I;
292 // If there is only one terminator instruction, process it.
293 unsigned LastOpc = LastInst->getOpcode();
294 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
295 if (isUncondBranchOpcode(LastOpc)) {
296 TBB = LastInst->getOperand(0).getMBB();
299 if (isCondBranchOpcode(LastOpc)) {
300 // Block ends with fall-through condbranch.
301 TBB = LastInst->getOperand(0).getMBB();
302 Cond.push_back(LastInst->getOperand(1));
303 Cond.push_back(LastInst->getOperand(2));
306 return true; // Can't handle indirect branch.
309 // Get the instruction before it if it is a terminator.
310 MachineInstr *SecondLastInst = I;
311 unsigned SecondLastOpc = SecondLastInst->getOpcode();
313 // If AllowModify is true and the block ends with two or more unconditional
314 // branches, delete all but the first unconditional branch.
315 if (AllowModify && isUncondBranchOpcode(LastOpc)) {
316 while (isUncondBranchOpcode(SecondLastOpc)) {
317 LastInst->eraseFromParent();
318 LastInst = SecondLastInst;
319 LastOpc = LastInst->getOpcode();
320 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
321 // Return now the only terminator is an unconditional branch.
322 TBB = LastInst->getOperand(0).getMBB();
326 SecondLastOpc = SecondLastInst->getOpcode();
331 // If there are three terminators, we don't know what sort of block this is.
332 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
335 // If the block ends with a B and a Bcc, handle it.
336 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
337 TBB = SecondLastInst->getOperand(0).getMBB();
338 Cond.push_back(SecondLastInst->getOperand(1));
339 Cond.push_back(SecondLastInst->getOperand(2));
340 FBB = LastInst->getOperand(0).getMBB();
344 // If the block ends with two unconditional branches, handle it. The second
345 // one is not executed, so remove it.
346 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
347 TBB = SecondLastInst->getOperand(0).getMBB();
350 I->eraseFromParent();
354 // ...likewise if it ends with a branch table followed by an unconditional
355 // branch. The branch folder can create these, and we must get rid of them for
356 // correctness of Thumb constant islands.
357 if ((isJumpTableBranchOpcode(SecondLastOpc) ||
358 isIndirectBranchOpcode(SecondLastOpc)) &&
359 isUncondBranchOpcode(LastOpc)) {
362 I->eraseFromParent();
366 // Otherwise, can't handle this.
371 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
372 MachineBasicBlock::iterator I = MBB.end();
373 if (I == MBB.begin()) return 0;
375 while (I->isDebugValue()) {
376 if (I == MBB.begin())
380 if (!isUncondBranchOpcode(I->getOpcode()) &&
381 !isCondBranchOpcode(I->getOpcode()))
384 // Remove the branch.
385 I->eraseFromParent();
389 if (I == MBB.begin()) return 1;
391 if (!isCondBranchOpcode(I->getOpcode()))
394 // Remove the branch.
395 I->eraseFromParent();
400 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
401 MachineBasicBlock *FBB,
402 const SmallVectorImpl<MachineOperand> &Cond,
404 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
405 int BOpc = !AFI->isThumbFunction()
406 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
407 int BccOpc = !AFI->isThumbFunction()
408 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
409 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function();
411 // Shouldn't be a fall through.
412 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
413 assert((Cond.size() == 2 || Cond.size() == 0) &&
414 "ARM branch conditions have two components!");
417 if (Cond.empty()) { // Unconditional branch?
419 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0);
421 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
423 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
424 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
428 // Two-way conditional branch.
429 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
430 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
432 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
434 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
438 bool ARMBaseInstrInfo::
439 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
440 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
441 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
445 bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const {
446 if (MI->isBundle()) {
447 MachineBasicBlock::const_instr_iterator I = MI;
448 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
449 while (++I != E && I->isInsideBundle()) {
450 int PIdx = I->findFirstPredOperandIdx();
451 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL)
457 int PIdx = MI->findFirstPredOperandIdx();
458 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
461 bool ARMBaseInstrInfo::
462 PredicateInstruction(MachineInstr *MI,
463 const SmallVectorImpl<MachineOperand> &Pred) const {
464 unsigned Opc = MI->getOpcode();
465 if (isUncondBranchOpcode(Opc)) {
466 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
467 MachineInstrBuilder(*MI->getParent()->getParent(), MI)
468 .addImm(Pred[0].getImm())
469 .addReg(Pred[1].getReg());
473 int PIdx = MI->findFirstPredOperandIdx();
475 MachineOperand &PMO = MI->getOperand(PIdx);
476 PMO.setImm(Pred[0].getImm());
477 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
483 bool ARMBaseInstrInfo::
484 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
485 const SmallVectorImpl<MachineOperand> &Pred2) const {
486 if (Pred1.size() > 2 || Pred2.size() > 2)
489 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
490 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
500 return CC2 == ARMCC::HI;
502 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
504 return CC2 == ARMCC::GT;
506 return CC2 == ARMCC::LT;
510 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
511 std::vector<MachineOperand> &Pred) const {
513 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
514 const MachineOperand &MO = MI->getOperand(i);
515 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) ||
516 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) {
525 /// isPredicable - Return true if the specified instruction can be predicated.
526 /// By default, this returns true for every instruction with a
527 /// PredicateOperand.
528 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
529 if (!MI->isPredicable())
532 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
533 ARMFunctionInfo *AFI =
534 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
535 return AFI->isThumb2Function();
540 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
541 LLVM_ATTRIBUTE_NOINLINE
542 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
544 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
546 assert(JTI < JT.size());
547 return JT[JTI].MBBs.size();
550 /// GetInstSize - Return the size of the specified MachineInstr.
552 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
553 const MachineBasicBlock &MBB = *MI->getParent();
554 const MachineFunction *MF = MBB.getParent();
555 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
557 const MCInstrDesc &MCID = MI->getDesc();
559 return MCID.getSize();
561 // If this machine instr is an inline asm, measure it.
562 if (MI->getOpcode() == ARM::INLINEASM)
563 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
566 unsigned Opc = MI->getOpcode();
568 case TargetOpcode::IMPLICIT_DEF:
569 case TargetOpcode::KILL:
570 case TargetOpcode::PROLOG_LABEL:
571 case TargetOpcode::EH_LABEL:
572 case TargetOpcode::DBG_VALUE:
574 case TargetOpcode::BUNDLE:
575 return getInstBundleLength(MI);
576 case ARM::MOVi16_ga_pcrel:
577 case ARM::MOVTi16_ga_pcrel:
578 case ARM::t2MOVi16_ga_pcrel:
579 case ARM::t2MOVTi16_ga_pcrel:
582 case ARM::t2MOVi32imm:
584 case ARM::CONSTPOOL_ENTRY:
585 // If this machine instr is a constant pool entry, its size is recorded as
587 return MI->getOperand(2).getImm();
588 case ARM::Int_eh_sjlj_longjmp:
590 case ARM::tInt_eh_sjlj_longjmp:
592 case ARM::Int_eh_sjlj_setjmp:
593 case ARM::Int_eh_sjlj_setjmp_nofp:
595 case ARM::tInt_eh_sjlj_setjmp:
596 case ARM::t2Int_eh_sjlj_setjmp:
597 case ARM::t2Int_eh_sjlj_setjmp_nofp:
605 case ARM::t2TBH_JT: {
606 // These are jumptable branches, i.e. a branch followed by an inlined
607 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
608 // entry is one byte; TBH two byte each.
609 unsigned EntrySize = (Opc == ARM::t2TBB_JT)
610 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
611 unsigned NumOps = MCID.getNumOperands();
612 MachineOperand JTOP =
613 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
614 unsigned JTI = JTOP.getIndex();
615 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
617 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
618 assert(JTI < JT.size());
619 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
620 // 4 aligned. The assembler / linker may add 2 byte padding just before
621 // the JT entries. The size does not include this padding; the
622 // constant islands pass does separate bookkeeping for it.
623 // FIXME: If we know the size of the function is less than (1 << 16) *2
624 // bytes, we can use 16-bit entries instead. Then there won't be an
626 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
627 unsigned NumEntries = getNumJTEntries(JT, JTI);
628 if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
629 // Make sure the instruction that follows TBB is 2-byte aligned.
630 // FIXME: Constant island pass should insert an "ALIGN" instruction
633 return NumEntries * EntrySize + InstSize;
636 // Otherwise, pseudo-instruction sizes are zero.
641 unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const {
643 MachineBasicBlock::const_instr_iterator I = MI;
644 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
645 while (++I != E && I->isInsideBundle()) {
646 assert(!I->isBundle() && "No nested bundle!");
647 Size += GetInstSizeInBytes(&*I);
652 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
653 MachineBasicBlock::iterator I, DebugLoc DL,
654 unsigned DestReg, unsigned SrcReg,
655 bool KillSrc) const {
656 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
657 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
659 if (GPRDest && GPRSrc) {
660 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
661 .addReg(SrcReg, getKillRegState(KillSrc))));
665 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
666 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
669 if (SPRDest && SPRSrc)
671 else if (GPRDest && SPRSrc)
673 else if (SPRDest && GPRSrc)
675 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
677 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
681 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
682 MIB.addReg(SrcReg, getKillRegState(KillSrc));
683 if (Opc == ARM::VORRq)
684 MIB.addReg(SrcReg, getKillRegState(KillSrc));
689 // Handle register classes that require multiple instructions.
690 unsigned BeginIdx = 0;
691 unsigned SubRegs = 0;
694 // Use VORRq when possible.
695 if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
696 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 2;
697 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
698 Opc = ARM::VORRq, BeginIdx = ARM::qsub_0, SubRegs = 4;
699 // Fall back to VMOVD.
700 else if (ARM::DPairRegClass.contains(DestReg, SrcReg))
701 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2;
702 else if (ARM::DTripleRegClass.contains(DestReg, SrcReg))
703 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3;
704 else if (ARM::DQuadRegClass.contains(DestReg, SrcReg))
705 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4;
706 else if (ARM::GPRPairRegClass.contains(DestReg, SrcReg))
707 Opc = ARM::MOVr, BeginIdx = ARM::gsub_0, SubRegs = 2;
709 else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg))
710 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 2, Spacing = 2;
711 else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg))
712 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 3, Spacing = 2;
713 else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg))
714 Opc = ARM::VMOVD, BeginIdx = ARM::dsub_0, SubRegs = 4, Spacing = 2;
716 assert(Opc && "Impossible reg-to-reg copy");
718 const TargetRegisterInfo *TRI = &getRegisterInfo();
719 MachineInstrBuilder Mov;
721 // Copy register tuples backward when the first Dest reg overlaps with SrcReg.
722 if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) {
723 BeginIdx = BeginIdx + ((SubRegs-1)*Spacing);
727 SmallSet<unsigned, 4> DstRegs;
729 for (unsigned i = 0; i != SubRegs; ++i) {
730 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i*Spacing);
731 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i*Spacing);
732 assert(Dst && Src && "Bad sub-register");
734 assert(!DstRegs.count(Src) && "destructive vector copy");
737 Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst)
739 // VORR takes two source operands.
740 if (Opc == ARM::VORRq)
742 Mov = AddDefaultPred(Mov);
744 // Add implicit super-register defs and kills to the last instruction.
745 Mov->addRegisterDefined(DestReg, TRI);
747 Mov->addRegisterKilled(SrcReg, TRI);
750 const MachineInstrBuilder &
751 ARMBaseInstrInfo::AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
752 unsigned SubIdx, unsigned State,
753 const TargetRegisterInfo *TRI) const {
755 return MIB.addReg(Reg, State);
757 if (TargetRegisterInfo::isPhysicalRegister(Reg))
758 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
759 return MIB.addReg(Reg, State, SubIdx);
762 void ARMBaseInstrInfo::
763 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
764 unsigned SrcReg, bool isKill, int FI,
765 const TargetRegisterClass *RC,
766 const TargetRegisterInfo *TRI) const {
768 if (I != MBB.end()) DL = I->getDebugLoc();
769 MachineFunction &MF = *MBB.getParent();
770 MachineFrameInfo &MFI = *MF.getFrameInfo();
771 unsigned Align = MFI.getObjectAlignment(FI);
773 MachineMemOperand *MMO =
774 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
775 MachineMemOperand::MOStore,
776 MFI.getObjectSize(FI),
779 switch (RC->getSize()) {
781 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
782 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12))
783 .addReg(SrcReg, getKillRegState(isKill))
784 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
785 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
786 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
787 .addReg(SrcReg, getKillRegState(isKill))
788 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
790 llvm_unreachable("Unknown reg class!");
793 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
794 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
795 .addReg(SrcReg, getKillRegState(isKill))
796 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
797 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
798 if (Subtarget.hasV5TEOps()) {
799 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::STRD));
800 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
801 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
802 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
806 // Fallback to STM instruction, which has existed since the dawn of
808 MachineInstrBuilder MIB =
809 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STMIA))
810 .addFrameIndex(FI).addMemOperand(MMO));
811 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
812 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
815 llvm_unreachable("Unknown reg class!");
818 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
819 // Use aligned spills if the stack can be realigned.
820 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
821 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64))
822 .addFrameIndex(FI).addImm(16)
823 .addReg(SrcReg, getKillRegState(isKill))
824 .addMemOperand(MMO));
826 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA))
827 .addReg(SrcReg, getKillRegState(isKill))
829 .addMemOperand(MMO));
832 llvm_unreachable("Unknown reg class!");
835 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
836 // Use aligned spills if the stack can be realigned.
837 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
838 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
839 .addFrameIndex(FI).addImm(16)
840 .addReg(SrcReg, getKillRegState(isKill))
841 .addMemOperand(MMO));
843 MachineInstrBuilder MIB =
844 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
847 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
848 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
849 AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
852 llvm_unreachable("Unknown reg class!");
855 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
856 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
857 // FIXME: It's possible to only store part of the QQ register if the
858 // spilled def has a sub-register index.
859 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo))
860 .addFrameIndex(FI).addImm(16)
861 .addReg(SrcReg, getKillRegState(isKill))
862 .addMemOperand(MMO));
864 MachineInstrBuilder MIB =
865 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
868 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
869 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
870 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
871 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
874 llvm_unreachable("Unknown reg class!");
877 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
878 MachineInstrBuilder MIB =
879 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
882 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
883 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
884 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
885 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
886 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
887 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
888 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
889 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
891 llvm_unreachable("Unknown reg class!");
894 llvm_unreachable("Unknown reg class!");
899 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
900 int &FrameIndex) const {
901 switch (MI->getOpcode()) {
904 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
905 if (MI->getOperand(1).isFI() &&
906 MI->getOperand(2).isReg() &&
907 MI->getOperand(3).isImm() &&
908 MI->getOperand(2).getReg() == 0 &&
909 MI->getOperand(3).getImm() == 0) {
910 FrameIndex = MI->getOperand(1).getIndex();
911 return MI->getOperand(0).getReg();
919 if (MI->getOperand(1).isFI() &&
920 MI->getOperand(2).isImm() &&
921 MI->getOperand(2).getImm() == 0) {
922 FrameIndex = MI->getOperand(1).getIndex();
923 return MI->getOperand(0).getReg();
927 case ARM::VST1d64TPseudo:
928 case ARM::VST1d64QPseudo:
929 if (MI->getOperand(0).isFI() &&
930 MI->getOperand(2).getSubReg() == 0) {
931 FrameIndex = MI->getOperand(0).getIndex();
932 return MI->getOperand(2).getReg();
936 if (MI->getOperand(1).isFI() &&
937 MI->getOperand(0).getSubReg() == 0) {
938 FrameIndex = MI->getOperand(1).getIndex();
939 return MI->getOperand(0).getReg();
947 unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
948 int &FrameIndex) const {
949 const MachineMemOperand *Dummy;
950 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex);
953 void ARMBaseInstrInfo::
954 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
955 unsigned DestReg, int FI,
956 const TargetRegisterClass *RC,
957 const TargetRegisterInfo *TRI) const {
959 if (I != MBB.end()) DL = I->getDebugLoc();
960 MachineFunction &MF = *MBB.getParent();
961 MachineFrameInfo &MFI = *MF.getFrameInfo();
962 unsigned Align = MFI.getObjectAlignment(FI);
963 MachineMemOperand *MMO =
964 MF.getMachineMemOperand(
965 MachinePointerInfo::getFixedStack(FI),
966 MachineMemOperand::MOLoad,
967 MFI.getObjectSize(FI),
970 switch (RC->getSize()) {
972 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
973 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
974 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
976 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
977 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
978 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
980 llvm_unreachable("Unknown reg class!");
983 if (ARM::DPRRegClass.hasSubClassEq(RC)) {
984 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
985 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
986 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
987 MachineInstrBuilder MIB;
989 if (Subtarget.hasV5TEOps()) {
990 MIB = BuildMI(MBB, I, DL, get(ARM::LDRD));
991 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
992 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
993 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO);
997 // Fallback to LDM instruction, which has existed since the dawn of
999 MIB = AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDMIA))
1000 .addFrameIndex(FI).addMemOperand(MMO));
1001 MIB = AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
1002 MIB = AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
1005 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1006 MIB.addReg(DestReg, RegState::ImplicitDefine);
1008 llvm_unreachable("Unknown reg class!");
1011 if (ARM::DPairRegClass.hasSubClassEq(RC)) {
1012 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1013 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg)
1014 .addFrameIndex(FI).addImm(16)
1015 .addMemOperand(MMO));
1017 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg)
1019 .addMemOperand(MMO));
1022 llvm_unreachable("Unknown reg class!");
1025 if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
1026 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1027 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
1028 .addFrameIndex(FI).addImm(16)
1029 .addMemOperand(MMO));
1031 MachineInstrBuilder MIB =
1032 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1034 .addMemOperand(MMO));
1035 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1036 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1037 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1038 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1039 MIB.addReg(DestReg, RegState::ImplicitDefine);
1042 llvm_unreachable("Unknown reg class!");
1045 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
1046 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
1047 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
1048 .addFrameIndex(FI).addImm(16)
1049 .addMemOperand(MMO));
1051 MachineInstrBuilder MIB =
1052 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1054 .addMemOperand(MMO);
1055 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1056 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1057 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1058 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1059 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1060 MIB.addReg(DestReg, RegState::ImplicitDefine);
1063 llvm_unreachable("Unknown reg class!");
1066 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) {
1067 MachineInstrBuilder MIB =
1068 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
1070 .addMemOperand(MMO);
1071 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
1072 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
1073 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
1074 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI);
1075 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI);
1076 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI);
1077 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI);
1078 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI);
1079 if (TargetRegisterInfo::isPhysicalRegister(DestReg))
1080 MIB.addReg(DestReg, RegState::ImplicitDefine);
1082 llvm_unreachable("Unknown reg class!");
1085 llvm_unreachable("Unknown regclass!");
1090 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1091 int &FrameIndex) const {
1092 switch (MI->getOpcode()) {
1095 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
1096 if (MI->getOperand(1).isFI() &&
1097 MI->getOperand(2).isReg() &&
1098 MI->getOperand(3).isImm() &&
1099 MI->getOperand(2).getReg() == 0 &&
1100 MI->getOperand(3).getImm() == 0) {
1101 FrameIndex = MI->getOperand(1).getIndex();
1102 return MI->getOperand(0).getReg();
1110 if (MI->getOperand(1).isFI() &&
1111 MI->getOperand(2).isImm() &&
1112 MI->getOperand(2).getImm() == 0) {
1113 FrameIndex = MI->getOperand(1).getIndex();
1114 return MI->getOperand(0).getReg();
1118 case ARM::VLD1d64TPseudo:
1119 case ARM::VLD1d64QPseudo:
1120 if (MI->getOperand(1).isFI() &&
1121 MI->getOperand(0).getSubReg() == 0) {
1122 FrameIndex = MI->getOperand(1).getIndex();
1123 return MI->getOperand(0).getReg();
1127 if (MI->getOperand(1).isFI() &&
1128 MI->getOperand(0).getSubReg() == 0) {
1129 FrameIndex = MI->getOperand(1).getIndex();
1130 return MI->getOperand(0).getReg();
1138 unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1139 int &FrameIndex) const {
1140 const MachineMemOperand *Dummy;
1141 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1144 bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
1145 // This hook gets to expand COPY instructions before they become
1146 // copyPhysReg() calls. Look for VMOVS instructions that can legally be
1147 // widened to VMOVD. We prefer the VMOVD when possible because it may be
1148 // changed into a VORR that can go down the NEON pipeline.
1149 if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15())
1152 // Look for a copy between even S-registers. That is where we keep floats
1153 // when using NEON v2f32 instructions for f32 arithmetic.
1154 unsigned DstRegS = MI->getOperand(0).getReg();
1155 unsigned SrcRegS = MI->getOperand(1).getReg();
1156 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS))
1159 const TargetRegisterInfo *TRI = &getRegisterInfo();
1160 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0,
1162 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0,
1164 if (!DstRegD || !SrcRegD)
1167 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only
1168 // legal if the COPY already defines the full DstRegD, and it isn't a
1169 // sub-register insertion.
1170 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI))
1173 // A dead copy shouldn't show up here, but reject it just in case.
1174 if (MI->getOperand(0).isDead())
1177 // All clear, widen the COPY.
1178 DEBUG(dbgs() << "widening: " << *MI);
1179 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1181 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg
1182 // or some other super-register.
1183 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD);
1184 if (ImpDefIdx != -1)
1185 MI->RemoveOperand(ImpDefIdx);
1187 // Change the opcode and operands.
1188 MI->setDesc(get(ARM::VMOVD));
1189 MI->getOperand(0).setReg(DstRegD);
1190 MI->getOperand(1).setReg(SrcRegD);
1191 AddDefaultPred(MIB);
1193 // We are now reading SrcRegD instead of SrcRegS. This may upset the
1194 // register scavenger and machine verifier, so we need to indicate that we
1195 // are reading an undefined value from SrcRegD, but a proper value from
1197 MI->getOperand(1).setIsUndef();
1198 MIB.addReg(SrcRegS, RegState::Implicit);
1200 // SrcRegD may actually contain an unrelated value in the ssub_1
1201 // sub-register. Don't kill it. Only kill the ssub_0 sub-register.
1202 if (MI->getOperand(1).isKill()) {
1203 MI->getOperand(1).setIsKill(false);
1204 MI->addRegisterKilled(SrcRegS, TRI, true);
1207 DEBUG(dbgs() << "replaced by: " << *MI);
1212 ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
1213 int FrameIx, uint64_t Offset,
1214 const MDNode *MDPtr,
1215 DebugLoc DL) const {
1216 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
1217 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
1221 /// Create a copy of a const pool value. Update CPI to the new index and return
1223 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1224 MachineConstantPool *MCP = MF.getConstantPool();
1225 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1227 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1228 assert(MCPE.isMachineConstantPoolEntry() &&
1229 "Expecting a machine constantpool entry!");
1230 ARMConstantPoolValue *ACPV =
1231 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1233 unsigned PCLabelId = AFI->createPICLabelUId();
1234 ARMConstantPoolValue *NewCPV = 0;
1235 // FIXME: The below assumes PIC relocation model and that the function
1236 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and
1237 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR
1238 // instructions, so that's probably OK, but is PIC always correct when
1240 if (ACPV->isGlobalValue())
1241 NewCPV = ARMConstantPoolConstant::
1242 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId,
1244 else if (ACPV->isExtSymbol())
1245 NewCPV = ARMConstantPoolSymbol::
1246 Create(MF.getFunction()->getContext(),
1247 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4);
1248 else if (ACPV->isBlockAddress())
1249 NewCPV = ARMConstantPoolConstant::
1250 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId,
1251 ARMCP::CPBlockAddress, 4);
1252 else if (ACPV->isLSDA())
1253 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId,
1255 else if (ACPV->isMachineBasicBlock())
1256 NewCPV = ARMConstantPoolMBB::
1257 Create(MF.getFunction()->getContext(),
1258 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4);
1260 llvm_unreachable("Unexpected ARM constantpool value type!!");
1261 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1265 void ARMBaseInstrInfo::
1266 reMaterialize(MachineBasicBlock &MBB,
1267 MachineBasicBlock::iterator I,
1268 unsigned DestReg, unsigned SubIdx,
1269 const MachineInstr *Orig,
1270 const TargetRegisterInfo &TRI) const {
1271 unsigned Opcode = Orig->getOpcode();
1274 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1275 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1279 case ARM::tLDRpci_pic:
1280 case ARM::t2LDRpci_pic: {
1281 MachineFunction &MF = *MBB.getParent();
1282 unsigned CPI = Orig->getOperand(1).getIndex();
1283 unsigned PCLabelId = duplicateCPV(MF, CPI);
1284 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1286 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1287 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1294 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1295 MachineInstr *MI = TargetInstrInfo::duplicate(Orig, MF);
1296 switch(Orig->getOpcode()) {
1297 case ARM::tLDRpci_pic:
1298 case ARM::t2LDRpci_pic: {
1299 unsigned CPI = Orig->getOperand(1).getIndex();
1300 unsigned PCLabelId = duplicateCPV(MF, CPI);
1301 Orig->getOperand(1).setIndex(CPI);
1302 Orig->getOperand(2).setImm(PCLabelId);
1309 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1310 const MachineInstr *MI1,
1311 const MachineRegisterInfo *MRI) const {
1312 int Opcode = MI0->getOpcode();
1313 if (Opcode == ARM::t2LDRpci ||
1314 Opcode == ARM::t2LDRpci_pic ||
1315 Opcode == ARM::tLDRpci ||
1316 Opcode == ARM::tLDRpci_pic ||
1317 Opcode == ARM::MOV_ga_dyn ||
1318 Opcode == ARM::MOV_ga_pcrel ||
1319 Opcode == ARM::MOV_ga_pcrel_ldr ||
1320 Opcode == ARM::t2MOV_ga_dyn ||
1321 Opcode == ARM::t2MOV_ga_pcrel) {
1322 if (MI1->getOpcode() != Opcode)
1324 if (MI0->getNumOperands() != MI1->getNumOperands())
1327 const MachineOperand &MO0 = MI0->getOperand(1);
1328 const MachineOperand &MO1 = MI1->getOperand(1);
1329 if (MO0.getOffset() != MO1.getOffset())
1332 if (Opcode == ARM::MOV_ga_dyn ||
1333 Opcode == ARM::MOV_ga_pcrel ||
1334 Opcode == ARM::MOV_ga_pcrel_ldr ||
1335 Opcode == ARM::t2MOV_ga_dyn ||
1336 Opcode == ARM::t2MOV_ga_pcrel)
1337 // Ignore the PC labels.
1338 return MO0.getGlobal() == MO1.getGlobal();
1340 const MachineFunction *MF = MI0->getParent()->getParent();
1341 const MachineConstantPool *MCP = MF->getConstantPool();
1342 int CPI0 = MO0.getIndex();
1343 int CPI1 = MO1.getIndex();
1344 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1345 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1346 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry();
1347 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry();
1348 if (isARMCP0 && isARMCP1) {
1349 ARMConstantPoolValue *ACPV0 =
1350 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1351 ARMConstantPoolValue *ACPV1 =
1352 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1353 return ACPV0->hasSameValue(ACPV1);
1354 } else if (!isARMCP0 && !isARMCP1) {
1355 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal;
1358 } else if (Opcode == ARM::PICLDR) {
1359 if (MI1->getOpcode() != Opcode)
1361 if (MI0->getNumOperands() != MI1->getNumOperands())
1364 unsigned Addr0 = MI0->getOperand(1).getReg();
1365 unsigned Addr1 = MI1->getOperand(1).getReg();
1366 if (Addr0 != Addr1) {
1368 !TargetRegisterInfo::isVirtualRegister(Addr0) ||
1369 !TargetRegisterInfo::isVirtualRegister(Addr1))
1372 // This assumes SSA form.
1373 MachineInstr *Def0 = MRI->getVRegDef(Addr0);
1374 MachineInstr *Def1 = MRI->getVRegDef(Addr1);
1375 // Check if the loaded value, e.g. a constantpool of a global address, are
1377 if (!produceSameValue(Def0, Def1, MRI))
1381 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) {
1382 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg
1383 const MachineOperand &MO0 = MI0->getOperand(i);
1384 const MachineOperand &MO1 = MI1->getOperand(i);
1385 if (!MO0.isIdenticalTo(MO1))
1391 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1394 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1395 /// determine if two loads are loading from the same base address. It should
1396 /// only return true if the base pointers are the same and the only differences
1397 /// between the two addresses is the offset. It also returns the offsets by
1400 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1401 /// is permanently disabled.
1402 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1404 int64_t &Offset2) const {
1405 // Don't worry about Thumb: just ARM and Thumb2.
1406 if (Subtarget.isThumb1Only()) return false;
1408 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1411 switch (Load1->getMachineOpcode()) {
1424 case ARM::t2LDRSHi8:
1426 case ARM::t2LDRSHi12:
1430 switch (Load2->getMachineOpcode()) {
1442 case ARM::t2LDRSHi8:
1444 case ARM::t2LDRSHi12:
1448 // Check if base addresses and chain operands match.
1449 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1450 Load1->getOperand(4) != Load2->getOperand(4))
1453 // Index should be Reg0.
1454 if (Load1->getOperand(3) != Load2->getOperand(3))
1457 // Determine the offsets.
1458 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1459 isa<ConstantSDNode>(Load2->getOperand(1))) {
1460 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1461 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1468 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1469 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
1470 /// be scheduled togther. On some targets if two loads are loading from
1471 /// addresses in the same cache line, it's better if they are scheduled
1472 /// together. This function takes two integers that represent the load offsets
1473 /// from the common base address. It returns true if it decides it's desirable
1474 /// to schedule the two loads together. "NumLoads" is the number of loads that
1475 /// have already been scheduled after Load1.
1477 /// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched
1478 /// is permanently disabled.
1479 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1480 int64_t Offset1, int64_t Offset2,
1481 unsigned NumLoads) const {
1482 // Don't worry about Thumb: just ARM and Thumb2.
1483 if (Subtarget.isThumb1Only()) return false;
1485 assert(Offset2 > Offset1);
1487 if ((Offset2 - Offset1) / 8 > 64)
1490 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1491 return false; // FIXME: overly conservative?
1493 // Four loads in a row should be sufficient.
1500 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1501 const MachineBasicBlock *MBB,
1502 const MachineFunction &MF) const {
1503 // Debug info is never a scheduling boundary. It's necessary to be explicit
1504 // due to the special treatment of IT instructions below, otherwise a
1505 // dbg_value followed by an IT will result in the IT instruction being
1506 // considered a scheduling hazard, which is wrong. It should be the actual
1507 // instruction preceding the dbg_value instruction(s), just like it is
1508 // when debug info is not present.
1509 if (MI->isDebugValue())
1512 // Terminators and labels can't be scheduled around.
1513 if (MI->isTerminator() || MI->isLabel())
1516 // Treat the start of the IT block as a scheduling boundary, but schedule
1517 // t2IT along with all instructions following it.
1518 // FIXME: This is a big hammer. But the alternative is to add all potential
1519 // true and anti dependencies to IT block instructions as implicit operands
1520 // to the t2IT instruction. The added compile time and complexity does not
1522 MachineBasicBlock::const_iterator I = MI;
1523 // Make sure to skip any dbg_value instructions
1524 while (++I != MBB->end() && I->isDebugValue())
1526 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1529 // Don't attempt to schedule around any instruction that defines
1530 // a stack-oriented pointer, as it's unlikely to be profitable. This
1531 // saves compile time, because it doesn't require every single
1532 // stack slot reference to depend on the instruction that does the
1534 // Calls don't actually change the stack pointer, even if they have imp-defs.
1535 // No ARM calling conventions change the stack pointer. (X86 calling
1536 // conventions sometimes do).
1537 if (!MI->isCall() && MI->definesRegister(ARM::SP))
1543 bool ARMBaseInstrInfo::
1544 isProfitableToIfCvt(MachineBasicBlock &MBB,
1545 unsigned NumCycles, unsigned ExtraPredCycles,
1546 const BranchProbability &Probability) const {
1550 // Attempt to estimate the relative costs of predication versus branching.
1551 unsigned UnpredCost = Probability.getNumerator() * NumCycles;
1552 UnpredCost /= Probability.getDenominator();
1553 UnpredCost += 1; // The branch itself
1554 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1556 return (NumCycles + ExtraPredCycles) <= UnpredCost;
1559 bool ARMBaseInstrInfo::
1560 isProfitableToIfCvt(MachineBasicBlock &TMBB,
1561 unsigned TCycles, unsigned TExtra,
1562 MachineBasicBlock &FMBB,
1563 unsigned FCycles, unsigned FExtra,
1564 const BranchProbability &Probability) const {
1565 if (!TCycles || !FCycles)
1568 // Attempt to estimate the relative costs of predication versus branching.
1569 unsigned TUnpredCost = Probability.getNumerator() * TCycles;
1570 TUnpredCost /= Probability.getDenominator();
1572 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator();
1573 unsigned FUnpredCost = Comp * FCycles;
1574 FUnpredCost /= Probability.getDenominator();
1576 unsigned UnpredCost = TUnpredCost + FUnpredCost;
1577 UnpredCost += 1; // The branch itself
1578 UnpredCost += Subtarget.getMispredictionPenalty() / 10;
1580 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost;
1584 ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
1585 MachineBasicBlock &FMBB) const {
1586 // Reduce false anti-dependencies to let Swift's out-of-order execution
1587 // engine do its thing.
1588 return Subtarget.isSwift();
1591 /// getInstrPredicate - If instruction is predicated, returns its predicate
1592 /// condition, otherwise returns AL. It also returns the condition code
1593 /// register by reference.
1595 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1596 int PIdx = MI->findFirstPredOperandIdx();
1602 PredReg = MI->getOperand(PIdx+1).getReg();
1603 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1607 int llvm::getMatchingCondBranchOpcode(int Opc) {
1612 if (Opc == ARM::t2B)
1615 llvm_unreachable("Unknown unconditional branch opcode!");
1618 /// commuteInstruction - Handle commutable instructions.
1620 ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1621 switch (MI->getOpcode()) {
1623 case ARM::t2MOVCCr: {
1624 // MOVCC can be commuted by inverting the condition.
1625 unsigned PredReg = 0;
1626 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg);
1627 // MOVCC AL can't be inverted. Shouldn't happen.
1628 if (CC == ARMCC::AL || PredReg != ARM::CPSR)
1630 MI = TargetInstrInfo::commuteInstruction(MI, NewMI);
1633 // After swapping the MOVCC operands, also invert the condition.
1634 MI->getOperand(MI->findFirstPredOperandIdx())
1635 .setImm(ARMCC::getOppositeCondition(CC));
1639 return TargetInstrInfo::commuteInstruction(MI, NewMI);
1642 /// Identify instructions that can be folded into a MOVCC instruction, and
1643 /// return the defining instruction.
1644 static MachineInstr *canFoldIntoMOVCC(unsigned Reg,
1645 const MachineRegisterInfo &MRI,
1646 const TargetInstrInfo *TII) {
1647 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1649 if (!MRI.hasOneNonDBGUse(Reg))
1651 MachineInstr *MI = MRI.getVRegDef(Reg);
1654 // MI is folded into the MOVCC by predicating it.
1655 if (!MI->isPredicable())
1657 // Check if MI has any non-dead defs or physreg uses. This also detects
1658 // predicated instructions which will be reading CPSR.
1659 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
1660 const MachineOperand &MO = MI->getOperand(i);
1661 // Reject frame index operands, PEI can't handle the predicated pseudos.
1662 if (MO.isFI() || MO.isCPI() || MO.isJTI())
1666 // MI can't have any tied operands, that would conflict with predication.
1669 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
1671 if (MO.isDef() && !MO.isDead())
1674 bool DontMoveAcrossStores = true;
1675 if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ 0, DontMoveAcrossStores))
1680 bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
1681 SmallVectorImpl<MachineOperand> &Cond,
1682 unsigned &TrueOp, unsigned &FalseOp,
1683 bool &Optimizable) const {
1684 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1685 "Unknown select instruction");
1690 // 3: Condition code.
1694 Cond.push_back(MI->getOperand(3));
1695 Cond.push_back(MI->getOperand(4));
1696 // We can always fold a def.
1701 MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
1702 bool PreferFalse) const {
1703 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
1704 "Unknown select instruction");
1705 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
1706 MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
1707 bool Invert = !DefMI;
1709 DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this);
1713 // Create a new predicated version of DefMI.
1714 // Rfalse is the first use.
1715 MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1717 MI->getOperand(0).getReg());
1719 // Copy all the DefMI operands, excluding its (null) predicate.
1720 const MCInstrDesc &DefDesc = DefMI->getDesc();
1721 for (unsigned i = 1, e = DefDesc.getNumOperands();
1722 i != e && !DefDesc.OpInfo[i].isPredicate(); ++i)
1723 NewMI.addOperand(DefMI->getOperand(i));
1725 unsigned CondCode = MI->getOperand(3).getImm();
1727 NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode)));
1729 NewMI.addImm(CondCode);
1730 NewMI.addOperand(MI->getOperand(4));
1732 // DefMI is not the -S version that sets CPSR, so add an optional %noreg.
1733 if (NewMI->hasOptionalDef())
1734 AddDefaultCC(NewMI);
1736 // The output register value when the predicate is false is an implicit
1737 // register operand tied to the first def.
1738 // The tie makes the register allocator ensure the FalseReg is allocated the
1739 // same register as operand 0.
1740 MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
1741 FalseReg.setImplicit();
1742 NewMI.addOperand(FalseReg);
1743 NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
1745 // The caller will erase MI, but not DefMI.
1746 DefMI->eraseFromParent();
1750 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
1751 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
1754 /// This will go away once we can teach tblgen how to set the optional CPSR def
1756 struct AddSubFlagsOpcodePair {
1758 uint16_t MachineOpc;
1761 static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
1762 {ARM::ADDSri, ARM::ADDri},
1763 {ARM::ADDSrr, ARM::ADDrr},
1764 {ARM::ADDSrsi, ARM::ADDrsi},
1765 {ARM::ADDSrsr, ARM::ADDrsr},
1767 {ARM::SUBSri, ARM::SUBri},
1768 {ARM::SUBSrr, ARM::SUBrr},
1769 {ARM::SUBSrsi, ARM::SUBrsi},
1770 {ARM::SUBSrsr, ARM::SUBrsr},
1772 {ARM::RSBSri, ARM::RSBri},
1773 {ARM::RSBSrsi, ARM::RSBrsi},
1774 {ARM::RSBSrsr, ARM::RSBrsr},
1776 {ARM::t2ADDSri, ARM::t2ADDri},
1777 {ARM::t2ADDSrr, ARM::t2ADDrr},
1778 {ARM::t2ADDSrs, ARM::t2ADDrs},
1780 {ARM::t2SUBSri, ARM::t2SUBri},
1781 {ARM::t2SUBSrr, ARM::t2SUBrr},
1782 {ARM::t2SUBSrs, ARM::t2SUBrs},
1784 {ARM::t2RSBSri, ARM::t2RSBri},
1785 {ARM::t2RSBSrs, ARM::t2RSBrs},
1788 unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) {
1789 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i)
1790 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc)
1791 return AddSubFlagsOpcodeMap[i].MachineOpc;
1795 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1796 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1797 unsigned DestReg, unsigned BaseReg, int NumBytes,
1798 ARMCC::CondCodes Pred, unsigned PredReg,
1799 const ARMBaseInstrInfo &TII, unsigned MIFlags) {
1800 bool isSub = NumBytes < 0;
1801 if (isSub) NumBytes = -NumBytes;
1804 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1805 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1806 assert(ThisVal && "Didn't extract field correctly");
1808 // We will handle these bits from offset, clear them.
1809 NumBytes &= ~ThisVal;
1811 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1813 // Build the new ADD / SUB.
1814 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1815 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1816 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1817 .addImm((unsigned)Pred).addReg(PredReg).addReg(0)
1818 .setMIFlags(MIFlags);
1823 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1824 unsigned FrameReg, int &Offset,
1825 const ARMBaseInstrInfo &TII) {
1826 unsigned Opcode = MI.getOpcode();
1827 const MCInstrDesc &Desc = MI.getDesc();
1828 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1831 // Memory operands in inline assembly always use AddrMode2.
1832 if (Opcode == ARM::INLINEASM)
1833 AddrMode = ARMII::AddrMode2;
1835 if (Opcode == ARM::ADDri) {
1836 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1838 // Turn it into a move.
1839 MI.setDesc(TII.get(ARM::MOVr));
1840 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1841 MI.RemoveOperand(FrameRegIdx+1);
1844 } else if (Offset < 0) {
1847 MI.setDesc(TII.get(ARM::SUBri));
1850 // Common case: small offset, fits into instruction.
1851 if (ARM_AM::getSOImmVal(Offset) != -1) {
1852 // Replace the FrameIndex with sp / fp
1853 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1854 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1859 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1861 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1862 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1864 // We will handle these bits from offset, clear them.
1865 Offset &= ~ThisImmVal;
1867 // Get the properly encoded SOImmVal field.
1868 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1869 "Bit extraction didn't work?");
1870 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1872 unsigned ImmIdx = 0;
1874 unsigned NumBits = 0;
1877 case ARMII::AddrMode_i12: {
1878 ImmIdx = FrameRegIdx + 1;
1879 InstrOffs = MI.getOperand(ImmIdx).getImm();
1883 case ARMII::AddrMode2: {
1884 ImmIdx = FrameRegIdx+2;
1885 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1886 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1891 case ARMII::AddrMode3: {
1892 ImmIdx = FrameRegIdx+2;
1893 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1894 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1899 case ARMII::AddrMode4:
1900 case ARMII::AddrMode6:
1901 // Can't fold any offset even if it's zero.
1903 case ARMII::AddrMode5: {
1904 ImmIdx = FrameRegIdx+1;
1905 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1906 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1913 llvm_unreachable("Unsupported addressing mode!");
1916 Offset += InstrOffs * Scale;
1917 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1923 // Attempt to fold address comp. if opcode has offset bits
1925 // Common case: small offset, fits into instruction.
1926 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1927 int ImmedOffset = Offset / Scale;
1928 unsigned Mask = (1 << NumBits) - 1;
1929 if ((unsigned)Offset <= Mask * Scale) {
1930 // Replace the FrameIndex with sp
1931 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1932 // FIXME: When addrmode2 goes away, this will simplify (like the
1933 // T2 version), as the LDR.i12 versions don't need the encoding
1934 // tricks for the offset value.
1936 if (AddrMode == ARMII::AddrMode_i12)
1937 ImmedOffset = -ImmedOffset;
1939 ImmedOffset |= 1 << NumBits;
1941 ImmOp.ChangeToImmediate(ImmedOffset);
1946 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1947 ImmedOffset = ImmedOffset & Mask;
1949 if (AddrMode == ARMII::AddrMode_i12)
1950 ImmedOffset = -ImmedOffset;
1952 ImmedOffset |= 1 << NumBits;
1954 ImmOp.ChangeToImmediate(ImmedOffset);
1955 Offset &= ~(Mask*Scale);
1959 Offset = (isSub) ? -Offset : Offset;
1963 /// analyzeCompare - For a comparison instruction, return the source registers
1964 /// in SrcReg and SrcReg2 if having two register operands, and the value it
1965 /// compares against in CmpValue. Return true if the comparison instruction
1966 /// can be analyzed.
1967 bool ARMBaseInstrInfo::
1968 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
1969 int &CmpMask, int &CmpValue) const {
1970 switch (MI->getOpcode()) {
1974 SrcReg = MI->getOperand(0).getReg();
1977 CmpValue = MI->getOperand(1).getImm();
1981 SrcReg = MI->getOperand(0).getReg();
1982 SrcReg2 = MI->getOperand(1).getReg();
1988 SrcReg = MI->getOperand(0).getReg();
1990 CmpMask = MI->getOperand(1).getImm();
1998 /// isSuitableForMask - Identify a suitable 'and' instruction that
1999 /// operates on the given source register and applies the same mask
2000 /// as a 'tst' instruction. Provide a limited look-through for copies.
2001 /// When successful, MI will hold the found instruction.
2002 static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg,
2003 int CmpMask, bool CommonUse) {
2004 switch (MI->getOpcode()) {
2007 if (CmpMask != MI->getOperand(2).getImm())
2009 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
2013 // Walk down one instruction which is potentially an 'and'.
2014 const MachineInstr &Copy = *MI;
2015 MachineBasicBlock::iterator AND(
2016 llvm::next(MachineBasicBlock::iterator(MI)));
2017 if (AND == MI->getParent()->end()) return false;
2019 return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
2027 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2028 /// the condition code if we modify the instructions such that flags are
2030 inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) {
2032 default: return ARMCC::AL;
2033 case ARMCC::EQ: return ARMCC::EQ;
2034 case ARMCC::NE: return ARMCC::NE;
2035 case ARMCC::HS: return ARMCC::LS;
2036 case ARMCC::LO: return ARMCC::HI;
2037 case ARMCC::HI: return ARMCC::LO;
2038 case ARMCC::LS: return ARMCC::HS;
2039 case ARMCC::GE: return ARMCC::LE;
2040 case ARMCC::LT: return ARMCC::GT;
2041 case ARMCC::GT: return ARMCC::LT;
2042 case ARMCC::LE: return ARMCC::GE;
2046 /// isRedundantFlagInstr - check whether the first instruction, whose only
2047 /// purpose is to update flags, can be made redundant.
2048 /// CMPrr can be made redundant by SUBrr if the operands are the same.
2049 /// CMPri can be made redundant by SUBri if the operands are the same.
2050 /// This function can be extended later on.
2051 inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg,
2052 unsigned SrcReg2, int ImmValue,
2054 if ((CmpI->getOpcode() == ARM::CMPrr ||
2055 CmpI->getOpcode() == ARM::t2CMPrr) &&
2056 (OI->getOpcode() == ARM::SUBrr ||
2057 OI->getOpcode() == ARM::t2SUBrr) &&
2058 ((OI->getOperand(1).getReg() == SrcReg &&
2059 OI->getOperand(2).getReg() == SrcReg2) ||
2060 (OI->getOperand(1).getReg() == SrcReg2 &&
2061 OI->getOperand(2).getReg() == SrcReg)))
2064 if ((CmpI->getOpcode() == ARM::CMPri ||
2065 CmpI->getOpcode() == ARM::t2CMPri) &&
2066 (OI->getOpcode() == ARM::SUBri ||
2067 OI->getOpcode() == ARM::t2SUBri) &&
2068 OI->getOperand(1).getReg() == SrcReg &&
2069 OI->getOperand(2).getImm() == ImmValue)
2074 /// optimizeCompareInstr - Convert the instruction supplying the argument to the
2075 /// comparison into one that sets the zero bit in the flags register;
2076 /// Remove a redundant Compare instruction if an earlier instruction can set the
2077 /// flags in the same way as Compare.
2078 /// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two
2079 /// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the
2080 /// condition code of instructions which use the flags.
2081 bool ARMBaseInstrInfo::
2082 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
2083 int CmpMask, int CmpValue,
2084 const MachineRegisterInfo *MRI) const {
2085 // Get the unique definition of SrcReg.
2086 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
2087 if (!MI) return false;
2089 // Masked compares sometimes use the same register as the corresponding 'and'.
2090 if (CmpMask != ~0) {
2091 if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) {
2093 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SrcReg),
2094 UE = MRI->use_end(); UI != UE; ++UI) {
2095 if (UI->getParent() != CmpInstr->getParent()) continue;
2096 MachineInstr *PotentialAND = &*UI;
2097 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) ||
2098 isPredicated(PotentialAND))
2103 if (!MI) return false;
2107 // Get ready to iterate backward from CmpInstr.
2108 MachineBasicBlock::iterator I = CmpInstr, E = MI,
2109 B = CmpInstr->getParent()->begin();
2111 // Early exit if CmpInstr is at the beginning of the BB.
2112 if (I == B) return false;
2114 // There are two possible candidates which can be changed to set CPSR:
2115 // One is MI, the other is a SUB instruction.
2116 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
2117 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
2118 MachineInstr *Sub = NULL;
2120 // MI is not a candidate for CMPrr.
2122 else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
2123 // Conservatively refuse to convert an instruction which isn't in the same
2124 // BB as the comparison.
2125 // For CMPri, we need to check Sub, thus we can't return here.
2126 if (CmpInstr->getOpcode() == ARM::CMPri ||
2127 CmpInstr->getOpcode() == ARM::t2CMPri)
2133 // Check that CPSR isn't set between the comparison instruction and the one we
2134 // want to change. At the same time, search for Sub.
2135 const TargetRegisterInfo *TRI = &getRegisterInfo();
2137 for (; I != E; --I) {
2138 const MachineInstr &Instr = *I;
2140 if (Instr.modifiesRegister(ARM::CPSR, TRI) ||
2141 Instr.readsRegister(ARM::CPSR, TRI))
2142 // This instruction modifies or uses CPSR after the one we want to
2143 // change. We can't do this transformation.
2146 // Check whether CmpInstr can be made redundant by the current instruction.
2147 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) {
2153 // The 'and' is below the comparison instruction.
2157 // Return false if no candidates exist.
2161 // The single candidate is called MI.
2164 // We can't use a predicated instruction - it doesn't always write the flags.
2165 if (isPredicated(MI))
2168 switch (MI->getOpcode()) {
2202 case ARM::t2EORri: {
2203 // Scan forward for the use of CPSR
2204 // When checking against MI: if it's a conditional code requires
2205 // checking of V bit, then this is not safe to do.
2206 // It is safe to remove CmpInstr if CPSR is redefined or killed.
2207 // If we are done with the basic block, we need to check whether CPSR is
2209 SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
2211 bool isSafe = false;
2213 E = CmpInstr->getParent()->end();
2214 while (!isSafe && ++I != E) {
2215 const MachineInstr &Instr = *I;
2216 for (unsigned IO = 0, EO = Instr.getNumOperands();
2217 !isSafe && IO != EO; ++IO) {
2218 const MachineOperand &MO = Instr.getOperand(IO);
2219 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) {
2223 if (!MO.isReg() || MO.getReg() != ARM::CPSR)
2229 // Condition code is after the operand before CPSR.
2230 ARMCC::CondCodes CC = (ARMCC::CondCodes)Instr.getOperand(IO-1).getImm();
2232 ARMCC::CondCodes NewCC = getSwappedCondition(CC);
2233 if (NewCC == ARMCC::AL)
2235 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based
2236 // on CMP needs to be updated to be based on SUB.
2237 // Push the condition code operands to OperandsToUpdate.
2238 // If it is safe to remove CmpInstr, the condition code of these
2239 // operands will be modified.
2240 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
2241 Sub->getOperand(2).getReg() == SrcReg)
2242 OperandsToUpdate.push_back(std::make_pair(&((*I).getOperand(IO-1)),
2248 // CPSR can be used multiple times, we should continue.
2261 // If CPSR is not killed nor re-defined, we should check whether it is
2262 // live-out. If it is live-out, do not optimize.
2264 MachineBasicBlock *MBB = CmpInstr->getParent();
2265 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
2266 SE = MBB->succ_end(); SI != SE; ++SI)
2267 if ((*SI)->isLiveIn(ARM::CPSR))
2271 // Toggle the optional operand to CPSR.
2272 MI->getOperand(5).setReg(ARM::CPSR);
2273 MI->getOperand(5).setIsDef(true);
2274 assert(!isPredicated(MI) && "Can't use flags from predicated instruction");
2275 CmpInstr->eraseFromParent();
2277 // Modify the condition code of operands in OperandsToUpdate.
2278 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to
2279 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
2280 for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++)
2281 OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second);
2289 bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI,
2290 MachineInstr *DefMI, unsigned Reg,
2291 MachineRegisterInfo *MRI) const {
2292 // Fold large immediates into add, sub, or, xor.
2293 unsigned DefOpc = DefMI->getOpcode();
2294 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm)
2296 if (!DefMI->getOperand(1).isImm())
2297 // Could be t2MOVi32imm <ga:xx>
2300 if (!MRI->hasOneNonDBGUse(Reg))
2303 const MCInstrDesc &DefMCID = DefMI->getDesc();
2304 if (DefMCID.hasOptionalDef()) {
2305 unsigned NumOps = DefMCID.getNumOperands();
2306 const MachineOperand &MO = DefMI->getOperand(NumOps-1);
2307 if (MO.getReg() == ARM::CPSR && !MO.isDead())
2308 // If DefMI defines CPSR and it is not dead, it's obviously not safe
2313 const MCInstrDesc &UseMCID = UseMI->getDesc();
2314 if (UseMCID.hasOptionalDef()) {
2315 unsigned NumOps = UseMCID.getNumOperands();
2316 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR)
2317 // If the instruction sets the flag, do not attempt this optimization
2318 // since it may change the semantics of the code.
2322 unsigned UseOpc = UseMI->getOpcode();
2323 unsigned NewUseOpc = 0;
2324 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm();
2325 uint32_t SOImmValV1 = 0, SOImmValV2 = 0;
2326 bool Commute = false;
2328 default: return false;
2336 case ARM::t2EORrr: {
2337 Commute = UseMI->getOperand(2).getReg() != Reg;
2344 NewUseOpc = ARM::SUBri;
2350 if (!ARM_AM::isSOImmTwoPartVal(ImmVal))
2352 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal);
2353 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal);
2356 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break;
2357 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break;
2358 case ARM::EORrr: NewUseOpc = ARM::EORri; break;
2362 case ARM::t2SUBrr: {
2366 NewUseOpc = ARM::t2SUBri;
2371 case ARM::t2EORrr: {
2372 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal))
2374 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal);
2375 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal);
2378 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break;
2379 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break;
2380 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break;
2388 unsigned OpIdx = Commute ? 2 : 1;
2389 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg();
2390 bool isKill = UseMI->getOperand(OpIdx).isKill();
2391 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg));
2392 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(),
2393 UseMI, UseMI->getDebugLoc(),
2394 get(NewUseOpc), NewReg)
2395 .addReg(Reg1, getKillRegState(isKill))
2396 .addImm(SOImmValV1)));
2397 UseMI->setDesc(get(NewUseOpc));
2398 UseMI->getOperand(1).setReg(NewReg);
2399 UseMI->getOperand(1).setIsKill();
2400 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2);
2401 DefMI->eraseFromParent();
2405 static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData,
2406 const MachineInstr *MI) {
2407 switch (MI->getOpcode()) {
2409 const MCInstrDesc &Desc = MI->getDesc();
2410 int UOps = ItinData->getNumMicroOps(Desc.getSchedClass());
2411 assert(UOps >= 0 && "bad # UOps");
2419 unsigned ShOpVal = MI->getOperand(3).getImm();
2420 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2421 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2424 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2425 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2432 if (!MI->getOperand(2).getReg())
2435 unsigned ShOpVal = MI->getOperand(3).getImm();
2436 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2437 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2440 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2441 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2448 return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2;
2450 case ARM::LDRSB_POST:
2451 case ARM::LDRSH_POST: {
2452 unsigned Rt = MI->getOperand(0).getReg();
2453 unsigned Rm = MI->getOperand(3).getReg();
2454 return (Rt == Rm) ? 4 : 3;
2457 case ARM::LDR_PRE_REG:
2458 case ARM::LDRB_PRE_REG: {
2459 unsigned Rt = MI->getOperand(0).getReg();
2460 unsigned Rm = MI->getOperand(3).getReg();
2463 unsigned ShOpVal = MI->getOperand(4).getImm();
2464 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2465 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2468 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2469 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2474 case ARM::STR_PRE_REG:
2475 case ARM::STRB_PRE_REG: {
2476 unsigned ShOpVal = MI->getOperand(4).getImm();
2477 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2478 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2481 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2482 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2488 case ARM::STRH_PRE: {
2489 unsigned Rt = MI->getOperand(0).getReg();
2490 unsigned Rm = MI->getOperand(3).getReg();
2495 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub)
2499 case ARM::LDR_POST_REG:
2500 case ARM::LDRB_POST_REG:
2501 case ARM::LDRH_POST: {
2502 unsigned Rt = MI->getOperand(0).getReg();
2503 unsigned Rm = MI->getOperand(3).getReg();
2504 return (Rt == Rm) ? 3 : 2;
2507 case ARM::LDR_PRE_IMM:
2508 case ARM::LDRB_PRE_IMM:
2509 case ARM::LDR_POST_IMM:
2510 case ARM::LDRB_POST_IMM:
2511 case ARM::STRB_POST_IMM:
2512 case ARM::STRB_POST_REG:
2513 case ARM::STRB_PRE_IMM:
2514 case ARM::STRH_POST:
2515 case ARM::STR_POST_IMM:
2516 case ARM::STR_POST_REG:
2517 case ARM::STR_PRE_IMM:
2520 case ARM::LDRSB_PRE:
2521 case ARM::LDRSH_PRE: {
2522 unsigned Rm = MI->getOperand(3).getReg();
2525 unsigned Rt = MI->getOperand(0).getReg();
2528 unsigned ShOpVal = MI->getOperand(4).getImm();
2529 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
2530 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
2533 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
2534 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
2540 unsigned Rt = MI->getOperand(0).getReg();
2541 unsigned Rn = MI->getOperand(2).getReg();
2542 unsigned Rm = MI->getOperand(3).getReg();
2544 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2545 return (Rt == Rn) ? 3 : 2;
2549 unsigned Rm = MI->getOperand(3).getReg();
2551 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3;
2555 case ARM::LDRD_POST:
2556 case ARM::t2LDRD_POST:
2559 case ARM::STRD_POST:
2560 case ARM::t2STRD_POST:
2563 case ARM::LDRD_PRE: {
2564 unsigned Rt = MI->getOperand(0).getReg();
2565 unsigned Rn = MI->getOperand(3).getReg();
2566 unsigned Rm = MI->getOperand(4).getReg();
2568 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2569 return (Rt == Rn) ? 4 : 3;
2572 case ARM::t2LDRD_PRE: {
2573 unsigned Rt = MI->getOperand(0).getReg();
2574 unsigned Rn = MI->getOperand(3).getReg();
2575 return (Rt == Rn) ? 4 : 3;
2578 case ARM::STRD_PRE: {
2579 unsigned Rm = MI->getOperand(4).getReg();
2581 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4;
2585 case ARM::t2STRD_PRE:
2588 case ARM::t2LDR_POST:
2589 case ARM::t2LDRB_POST:
2590 case ARM::t2LDRB_PRE:
2591 case ARM::t2LDRSBi12:
2592 case ARM::t2LDRSBi8:
2593 case ARM::t2LDRSBpci:
2595 case ARM::t2LDRH_POST:
2596 case ARM::t2LDRH_PRE:
2598 case ARM::t2LDRSB_POST:
2599 case ARM::t2LDRSB_PRE:
2600 case ARM::t2LDRSH_POST:
2601 case ARM::t2LDRSH_PRE:
2602 case ARM::t2LDRSHi12:
2603 case ARM::t2LDRSHi8:
2604 case ARM::t2LDRSHpci:
2608 case ARM::t2LDRDi8: {
2609 unsigned Rt = MI->getOperand(0).getReg();
2610 unsigned Rn = MI->getOperand(2).getReg();
2611 return (Rt == Rn) ? 3 : 2;
2614 case ARM::t2STRB_POST:
2615 case ARM::t2STRB_PRE:
2618 case ARM::t2STRH_POST:
2619 case ARM::t2STRH_PRE:
2621 case ARM::t2STR_POST:
2622 case ARM::t2STR_PRE:
2628 // Return the number of 32-bit words loaded by LDM or stored by STM. If this
2629 // can't be easily determined return 0 (missing MachineMemOperand).
2631 // FIXME: The current MachineInstr design does not support relying on machine
2632 // mem operands to determine the width of a memory access. Instead, we expect
2633 // the target to provide this information based on the instruction opcode and
2634 // operands. However, using MachineMemOperand is a the best solution now for
2637 // 1) getNumMicroOps tries to infer LDM memory width from the total number of MI
2638 // operands. This is much more dangerous than using the MachineMemOperand
2639 // sizes because CodeGen passes can insert/remove optional machine operands. In
2640 // fact, it's totally incorrect for preRA passes and appears to be wrong for
2641 // postRA passes as well.
2643 // 2) getNumLDMAddresses is only used by the scheduling machine model and any
2644 // machine model that calls this should handle the unknown (zero size) case.
2646 // Long term, we should require a target hook that verifies MachineMemOperand
2647 // sizes during MC lowering. That target hook should be local to MC lowering
2648 // because we can't ensure that it is aware of other MI forms. Doing this will
2649 // ensure that MachineMemOperands are correctly propagated through all passes.
2650 unsigned ARMBaseInstrInfo::getNumLDMAddresses(const MachineInstr *MI) const {
2652 for (MachineInstr::mmo_iterator I = MI->memoperands_begin(),
2653 E = MI->memoperands_end(); I != E; ++I) {
2654 Size += (*I)->getSize();
2660 ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
2661 const MachineInstr *MI) const {
2662 if (!ItinData || ItinData->isEmpty())
2665 const MCInstrDesc &Desc = MI->getDesc();
2666 unsigned Class = Desc.getSchedClass();
2667 int ItinUOps = ItinData->getNumMicroOps(Class);
2668 if (ItinUOps >= 0) {
2669 if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore()))
2670 return getNumMicroOpsSwiftLdSt(ItinData, MI);
2675 unsigned Opc = MI->getOpcode();
2678 llvm_unreachable("Unexpected multi-uops instruction!");
2683 // The number of uOps for load / store multiple are determined by the number
2686 // On Cortex-A8, each pair of register loads / stores can be scheduled on the
2687 // same cycle. The scheduling for the first load / store must be done
2688 // separately by assuming the address is not 64-bit aligned.
2690 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
2691 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON
2692 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1.
2694 case ARM::VLDMDIA_UPD:
2695 case ARM::VLDMDDB_UPD:
2697 case ARM::VLDMSIA_UPD:
2698 case ARM::VLDMSDB_UPD:
2700 case ARM::VSTMDIA_UPD:
2701 case ARM::VSTMDDB_UPD:
2703 case ARM::VSTMSIA_UPD:
2704 case ARM::VSTMSDB_UPD: {
2705 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands();
2706 return (NumRegs / 2) + (NumRegs % 2) + 1;
2709 case ARM::LDMIA_RET:
2714 case ARM::LDMIA_UPD:
2715 case ARM::LDMDA_UPD:
2716 case ARM::LDMDB_UPD:
2717 case ARM::LDMIB_UPD:
2722 case ARM::STMIA_UPD:
2723 case ARM::STMDA_UPD:
2724 case ARM::STMDB_UPD:
2725 case ARM::STMIB_UPD:
2727 case ARM::tLDMIA_UPD:
2728 case ARM::tSTMIA_UPD:
2732 case ARM::t2LDMIA_RET:
2735 case ARM::t2LDMIA_UPD:
2736 case ARM::t2LDMDB_UPD:
2739 case ARM::t2STMIA_UPD:
2740 case ARM::t2STMDB_UPD: {
2741 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1;
2742 if (Subtarget.isSwift()) {
2743 int UOps = 1 + NumRegs; // One for address computation, one for each ld / st.
2746 case ARM::VLDMDIA_UPD:
2747 case ARM::VLDMDDB_UPD:
2748 case ARM::VLDMSIA_UPD:
2749 case ARM::VLDMSDB_UPD:
2750 case ARM::VSTMDIA_UPD:
2751 case ARM::VSTMDDB_UPD:
2752 case ARM::VSTMSIA_UPD:
2753 case ARM::VSTMSDB_UPD:
2754 case ARM::LDMIA_UPD:
2755 case ARM::LDMDA_UPD:
2756 case ARM::LDMDB_UPD:
2757 case ARM::LDMIB_UPD:
2758 case ARM::STMIA_UPD:
2759 case ARM::STMDA_UPD:
2760 case ARM::STMDB_UPD:
2761 case ARM::STMIB_UPD:
2762 case ARM::tLDMIA_UPD:
2763 case ARM::tSTMIA_UPD:
2764 case ARM::t2LDMIA_UPD:
2765 case ARM::t2LDMDB_UPD:
2766 case ARM::t2STMIA_UPD:
2767 case ARM::t2STMDB_UPD:
2768 ++UOps; // One for base register writeback.
2770 case ARM::LDMIA_RET:
2772 case ARM::t2LDMIA_RET:
2773 UOps += 2; // One for base reg wb, one for write to pc.
2777 } else if (Subtarget.isCortexA8()) {
2780 // 4 registers would be issued: 2, 2.
2781 // 5 registers would be issued: 2, 2, 1.
2782 int A8UOps = (NumRegs / 2);
2786 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2787 int A9UOps = (NumRegs / 2);
2788 // If there are odd number of registers or if it's not 64-bit aligned,
2789 // then it takes an extra AGU (Address Generation Unit) cycle.
2790 if ((NumRegs % 2) ||
2791 !MI->hasOneMemOperand() ||
2792 (*MI->memoperands_begin())->getAlignment() < 8)
2796 // Assume the worst.
2804 ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData,
2805 const MCInstrDesc &DefMCID,
2807 unsigned DefIdx, unsigned DefAlign) const {
2808 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2810 // Def is the address writeback.
2811 return ItinData->getOperandCycle(DefClass, DefIdx);
2814 if (Subtarget.isCortexA8()) {
2815 // (regno / 2) + (regno % 2) + 1
2816 DefCycle = RegNo / 2 + 1;
2819 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2821 bool isSLoad = false;
2823 switch (DefMCID.getOpcode()) {
2826 case ARM::VLDMSIA_UPD:
2827 case ARM::VLDMSDB_UPD:
2832 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2833 // then it takes an extra cycle.
2834 if ((isSLoad && (RegNo % 2)) || DefAlign < 8)
2837 // Assume the worst.
2838 DefCycle = RegNo + 2;
2845 ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData,
2846 const MCInstrDesc &DefMCID,
2848 unsigned DefIdx, unsigned DefAlign) const {
2849 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1;
2851 // Def is the address writeback.
2852 return ItinData->getOperandCycle(DefClass, DefIdx);
2855 if (Subtarget.isCortexA8()) {
2856 // 4 registers would be issued: 1, 2, 1.
2857 // 5 registers would be issued: 1, 2, 2.
2858 DefCycle = RegNo / 2;
2861 // Result latency is issue cycle + 2: E2.
2863 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2864 DefCycle = (RegNo / 2);
2865 // If there are odd number of registers or if it's not 64-bit aligned,
2866 // then it takes an extra AGU (Address Generation Unit) cycle.
2867 if ((RegNo % 2) || DefAlign < 8)
2869 // Result latency is AGU cycles + 2.
2872 // Assume the worst.
2873 DefCycle = RegNo + 2;
2880 ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData,
2881 const MCInstrDesc &UseMCID,
2883 unsigned UseIdx, unsigned UseAlign) const {
2884 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2886 return ItinData->getOperandCycle(UseClass, UseIdx);
2889 if (Subtarget.isCortexA8()) {
2890 // (regno / 2) + (regno % 2) + 1
2891 UseCycle = RegNo / 2 + 1;
2894 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2896 bool isSStore = false;
2898 switch (UseMCID.getOpcode()) {
2901 case ARM::VSTMSIA_UPD:
2902 case ARM::VSTMSDB_UPD:
2907 // If there are odd number of 'S' registers or if it's not 64-bit aligned,
2908 // then it takes an extra cycle.
2909 if ((isSStore && (RegNo % 2)) || UseAlign < 8)
2912 // Assume the worst.
2913 UseCycle = RegNo + 2;
2920 ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData,
2921 const MCInstrDesc &UseMCID,
2923 unsigned UseIdx, unsigned UseAlign) const {
2924 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1;
2926 return ItinData->getOperandCycle(UseClass, UseIdx);
2929 if (Subtarget.isCortexA8()) {
2930 UseCycle = RegNo / 2;
2935 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) {
2936 UseCycle = (RegNo / 2);
2937 // If there are odd number of registers or if it's not 64-bit aligned,
2938 // then it takes an extra AGU (Address Generation Unit) cycle.
2939 if ((RegNo % 2) || UseAlign < 8)
2942 // Assume the worst.
2949 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
2950 const MCInstrDesc &DefMCID,
2951 unsigned DefIdx, unsigned DefAlign,
2952 const MCInstrDesc &UseMCID,
2953 unsigned UseIdx, unsigned UseAlign) const {
2954 unsigned DefClass = DefMCID.getSchedClass();
2955 unsigned UseClass = UseMCID.getSchedClass();
2957 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands())
2958 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
2960 // This may be a def / use of a variable_ops instruction, the operand
2961 // latency might be determinable dynamically. Let the target try to
2964 bool LdmBypass = false;
2965 switch (DefMCID.getOpcode()) {
2967 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
2971 case ARM::VLDMDIA_UPD:
2972 case ARM::VLDMDDB_UPD:
2974 case ARM::VLDMSIA_UPD:
2975 case ARM::VLDMSDB_UPD:
2976 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
2979 case ARM::LDMIA_RET:
2984 case ARM::LDMIA_UPD:
2985 case ARM::LDMDA_UPD:
2986 case ARM::LDMDB_UPD:
2987 case ARM::LDMIB_UPD:
2989 case ARM::tLDMIA_UPD:
2991 case ARM::t2LDMIA_RET:
2994 case ARM::t2LDMIA_UPD:
2995 case ARM::t2LDMDB_UPD:
2997 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign);
3002 // We can't seem to determine the result latency of the def, assume it's 2.
3006 switch (UseMCID.getOpcode()) {
3008 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx);
3012 case ARM::VSTMDIA_UPD:
3013 case ARM::VSTMDDB_UPD:
3015 case ARM::VSTMSIA_UPD:
3016 case ARM::VSTMSDB_UPD:
3017 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3024 case ARM::STMIA_UPD:
3025 case ARM::STMDA_UPD:
3026 case ARM::STMDB_UPD:
3027 case ARM::STMIB_UPD:
3028 case ARM::tSTMIA_UPD:
3033 case ARM::t2STMIA_UPD:
3034 case ARM::t2STMDB_UPD:
3035 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign);
3040 // Assume it's read in the first stage.
3043 UseCycle = DefCycle - UseCycle + 1;
3046 // It's a variable_ops instruction so we can't use DefIdx here. Just use
3047 // first def operand.
3048 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1,
3051 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx,
3052 UseClass, UseIdx)) {
3060 static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI,
3061 const MachineInstr *MI, unsigned Reg,
3062 unsigned &DefIdx, unsigned &Dist) {
3065 MachineBasicBlock::const_iterator I = MI; ++I;
3066 MachineBasicBlock::const_instr_iterator II =
3067 llvm::prior(I.getInstrIterator());
3068 assert(II->isInsideBundle() && "Empty bundle?");
3071 while (II->isInsideBundle()) {
3072 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI);
3079 assert(Idx != -1 && "Cannot find bundled definition!");
3084 static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI,
3085 const MachineInstr *MI, unsigned Reg,
3086 unsigned &UseIdx, unsigned &Dist) {
3089 MachineBasicBlock::const_instr_iterator II = MI; ++II;
3090 assert(II->isInsideBundle() && "Empty bundle?");
3091 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3093 // FIXME: This doesn't properly handle multiple uses.
3095 while (II != E && II->isInsideBundle()) {
3096 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI);
3099 if (II->getOpcode() != ARM::t2IT)
3113 /// Return the number of cycles to add to (or subtract from) the static
3114 /// itinerary based on the def opcode and alignment. The caller will ensure that
3115 /// adjusted latency is at least one cycle.
3116 static int adjustDefLatency(const ARMSubtarget &Subtarget,
3117 const MachineInstr *DefMI,
3118 const MCInstrDesc *DefMCID, unsigned DefAlign) {
3120 if (Subtarget.isCortexA8() || Subtarget.isLikeA9()) {
3121 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3122 // variants are one cycle cheaper.
3123 switch (DefMCID->getOpcode()) {
3127 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3128 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3130 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3137 case ARM::t2LDRSHs: {
3138 // Thumb2 mode: lsl only.
3139 unsigned ShAmt = DefMI->getOperand(3).getImm();
3140 if (ShAmt == 0 || ShAmt == 2)
3145 } else if (Subtarget.isSwift()) {
3146 // FIXME: Properly handle all of the latency adjustments for address
3148 switch (DefMCID->getOpcode()) {
3152 unsigned ShOpVal = DefMI->getOperand(3).getImm();
3153 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub;
3154 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3157 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3158 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)))
3161 ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3168 case ARM::t2LDRSHs: {
3169 // Thumb2 mode: lsl only.
3170 unsigned ShAmt = DefMI->getOperand(3).getImm();
3171 if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3)
3178 if (DefAlign < 8 && Subtarget.isLikeA9()) {
3179 switch (DefMCID->getOpcode()) {
3185 case ARM::VLD1q8wb_fixed:
3186 case ARM::VLD1q16wb_fixed:
3187 case ARM::VLD1q32wb_fixed:
3188 case ARM::VLD1q64wb_fixed:
3189 case ARM::VLD1q8wb_register:
3190 case ARM::VLD1q16wb_register:
3191 case ARM::VLD1q32wb_register:
3192 case ARM::VLD1q64wb_register:
3199 case ARM::VLD2d8wb_fixed:
3200 case ARM::VLD2d16wb_fixed:
3201 case ARM::VLD2d32wb_fixed:
3202 case ARM::VLD2q8wb_fixed:
3203 case ARM::VLD2q16wb_fixed:
3204 case ARM::VLD2q32wb_fixed:
3205 case ARM::VLD2d8wb_register:
3206 case ARM::VLD2d16wb_register:
3207 case ARM::VLD2d32wb_register:
3208 case ARM::VLD2q8wb_register:
3209 case ARM::VLD2q16wb_register:
3210 case ARM::VLD2q32wb_register:
3215 case ARM::VLD3d8_UPD:
3216 case ARM::VLD3d16_UPD:
3217 case ARM::VLD3d32_UPD:
3218 case ARM::VLD1d64Twb_fixed:
3219 case ARM::VLD1d64Twb_register:
3220 case ARM::VLD3q8_UPD:
3221 case ARM::VLD3q16_UPD:
3222 case ARM::VLD3q32_UPD:
3227 case ARM::VLD4d8_UPD:
3228 case ARM::VLD4d16_UPD:
3229 case ARM::VLD4d32_UPD:
3230 case ARM::VLD1d64Qwb_fixed:
3231 case ARM::VLD1d64Qwb_register:
3232 case ARM::VLD4q8_UPD:
3233 case ARM::VLD4q16_UPD:
3234 case ARM::VLD4q32_UPD:
3235 case ARM::VLD1DUPq8:
3236 case ARM::VLD1DUPq16:
3237 case ARM::VLD1DUPq32:
3238 case ARM::VLD1DUPq8wb_fixed:
3239 case ARM::VLD1DUPq16wb_fixed:
3240 case ARM::VLD1DUPq32wb_fixed:
3241 case ARM::VLD1DUPq8wb_register:
3242 case ARM::VLD1DUPq16wb_register:
3243 case ARM::VLD1DUPq32wb_register:
3244 case ARM::VLD2DUPd8:
3245 case ARM::VLD2DUPd16:
3246 case ARM::VLD2DUPd32:
3247 case ARM::VLD2DUPd8wb_fixed:
3248 case ARM::VLD2DUPd16wb_fixed:
3249 case ARM::VLD2DUPd32wb_fixed:
3250 case ARM::VLD2DUPd8wb_register:
3251 case ARM::VLD2DUPd16wb_register:
3252 case ARM::VLD2DUPd32wb_register:
3253 case ARM::VLD4DUPd8:
3254 case ARM::VLD4DUPd16:
3255 case ARM::VLD4DUPd32:
3256 case ARM::VLD4DUPd8_UPD:
3257 case ARM::VLD4DUPd16_UPD:
3258 case ARM::VLD4DUPd32_UPD:
3260 case ARM::VLD1LNd16:
3261 case ARM::VLD1LNd32:
3262 case ARM::VLD1LNd8_UPD:
3263 case ARM::VLD1LNd16_UPD:
3264 case ARM::VLD1LNd32_UPD:
3266 case ARM::VLD2LNd16:
3267 case ARM::VLD2LNd32:
3268 case ARM::VLD2LNq16:
3269 case ARM::VLD2LNq32:
3270 case ARM::VLD2LNd8_UPD:
3271 case ARM::VLD2LNd16_UPD:
3272 case ARM::VLD2LNd32_UPD:
3273 case ARM::VLD2LNq16_UPD:
3274 case ARM::VLD2LNq32_UPD:
3276 case ARM::VLD4LNd16:
3277 case ARM::VLD4LNd32:
3278 case ARM::VLD4LNq16:
3279 case ARM::VLD4LNq32:
3280 case ARM::VLD4LNd8_UPD:
3281 case ARM::VLD4LNd16_UPD:
3282 case ARM::VLD4LNd32_UPD:
3283 case ARM::VLD4LNq16_UPD:
3284 case ARM::VLD4LNq32_UPD:
3285 // If the address is not 64-bit aligned, the latencies of these
3286 // instructions increases by one.
3297 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3298 const MachineInstr *DefMI, unsigned DefIdx,
3299 const MachineInstr *UseMI,
3300 unsigned UseIdx) const {
3301 // No operand latency. The caller may fall back to getInstrLatency.
3302 if (!ItinData || ItinData->isEmpty())
3305 const MachineOperand &DefMO = DefMI->getOperand(DefIdx);
3306 unsigned Reg = DefMO.getReg();
3307 const MCInstrDesc *DefMCID = &DefMI->getDesc();
3308 const MCInstrDesc *UseMCID = &UseMI->getDesc();
3310 unsigned DefAdj = 0;
3311 if (DefMI->isBundle()) {
3312 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj);
3313 DefMCID = &DefMI->getDesc();
3315 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() ||
3316 DefMI->isRegSequence() || DefMI->isImplicitDef()) {
3320 unsigned UseAdj = 0;
3321 if (UseMI->isBundle()) {
3323 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI,
3324 Reg, NewUseIdx, UseAdj);
3330 UseMCID = &UseMI->getDesc();
3333 if (Reg == ARM::CPSR) {
3334 if (DefMI->getOpcode() == ARM::FMSTAT) {
3335 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?)
3336 return Subtarget.isLikeA9() ? 1 : 20;
3339 // CPSR set and branch can be paired in the same cycle.
3340 if (UseMI->isBranch())
3343 // Otherwise it takes the instruction latency (generally one).
3344 unsigned Latency = getInstrLatency(ItinData, DefMI);
3346 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to
3347 // its uses. Instructions which are otherwise scheduled between them may
3348 // incur a code size penalty (not able to use the CPSR setting 16-bit
3350 if (Latency > 0 && Subtarget.isThumb2()) {
3351 const MachineFunction *MF = DefMI->getParent()->getParent();
3352 if (MF->getFunction()->getAttributes().
3353 hasAttribute(AttributeSet::FunctionIndex,
3354 Attribute::OptimizeForSize))
3360 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit())
3363 unsigned DefAlign = DefMI->hasOneMemOperand()
3364 ? (*DefMI->memoperands_begin())->getAlignment() : 0;
3365 unsigned UseAlign = UseMI->hasOneMemOperand()
3366 ? (*UseMI->memoperands_begin())->getAlignment() : 0;
3368 // Get the itinerary's latency if possible, and handle variable_ops.
3369 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign,
3370 *UseMCID, UseIdx, UseAlign);
3371 // Unable to find operand latency. The caller may resort to getInstrLatency.
3375 // Adjust for IT block position.
3376 int Adj = DefAdj + UseAdj;
3378 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3379 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign);
3380 if (Adj >= 0 || (int)Latency > -Adj) {
3381 return Latency + Adj;
3383 // Return the itinerary latency, which may be zero but not less than zero.
3388 ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
3389 SDNode *DefNode, unsigned DefIdx,
3390 SDNode *UseNode, unsigned UseIdx) const {
3391 if (!DefNode->isMachineOpcode())
3394 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode());
3396 if (isZeroCost(DefMCID.Opcode))
3399 if (!ItinData || ItinData->isEmpty())
3400 return DefMCID.mayLoad() ? 3 : 1;
3402 if (!UseNode->isMachineOpcode()) {
3403 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx);
3404 if (Subtarget.isLikeA9() || Subtarget.isSwift())
3405 return Latency <= 2 ? 1 : Latency - 1;
3407 return Latency <= 3 ? 1 : Latency - 2;
3410 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode());
3411 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode);
3412 unsigned DefAlign = !DefMN->memoperands_empty()
3413 ? (*DefMN->memoperands_begin())->getAlignment() : 0;
3414 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode);
3415 unsigned UseAlign = !UseMN->memoperands_empty()
3416 ? (*UseMN->memoperands_begin())->getAlignment() : 0;
3417 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign,
3418 UseMCID, UseIdx, UseAlign);
3421 (Subtarget.isCortexA8() || Subtarget.isLikeA9())) {
3422 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2]
3423 // variants are one cycle cheaper.
3424 switch (DefMCID.getOpcode()) {
3429 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3430 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3432 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3439 case ARM::t2LDRSHs: {
3440 // Thumb2 mode: lsl only.
3442 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3443 if (ShAmt == 0 || ShAmt == 2)
3448 } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) {
3449 // FIXME: Properly handle all of the latency adjustments for address
3451 switch (DefMCID.getOpcode()) {
3456 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue();
3457 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal);
3459 ((ShImm == 1 || ShImm == 2 || ShImm == 3) &&
3460 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))
3462 else if (ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr)
3469 case ARM::t2LDRSHs: {
3470 // Thumb2 mode: lsl 0-3 only.
3477 if (DefAlign < 8 && Subtarget.isLikeA9())
3478 switch (DefMCID.getOpcode()) {
3484 case ARM::VLD1q8wb_register:
3485 case ARM::VLD1q16wb_register:
3486 case ARM::VLD1q32wb_register:
3487 case ARM::VLD1q64wb_register:
3488 case ARM::VLD1q8wb_fixed:
3489 case ARM::VLD1q16wb_fixed:
3490 case ARM::VLD1q32wb_fixed:
3491 case ARM::VLD1q64wb_fixed:
3495 case ARM::VLD2q8Pseudo:
3496 case ARM::VLD2q16Pseudo:
3497 case ARM::VLD2q32Pseudo:
3498 case ARM::VLD2d8wb_fixed:
3499 case ARM::VLD2d16wb_fixed:
3500 case ARM::VLD2d32wb_fixed:
3501 case ARM::VLD2q8PseudoWB_fixed:
3502 case ARM::VLD2q16PseudoWB_fixed:
3503 case ARM::VLD2q32PseudoWB_fixed:
3504 case ARM::VLD2d8wb_register:
3505 case ARM::VLD2d16wb_register:
3506 case ARM::VLD2d32wb_register:
3507 case ARM::VLD2q8PseudoWB_register:
3508 case ARM::VLD2q16PseudoWB_register:
3509 case ARM::VLD2q32PseudoWB_register:
3510 case ARM::VLD3d8Pseudo:
3511 case ARM::VLD3d16Pseudo:
3512 case ARM::VLD3d32Pseudo:
3513 case ARM::VLD1d64TPseudo:
3514 case ARM::VLD3d8Pseudo_UPD:
3515 case ARM::VLD3d16Pseudo_UPD:
3516 case ARM::VLD3d32Pseudo_UPD:
3517 case ARM::VLD3q8Pseudo_UPD:
3518 case ARM::VLD3q16Pseudo_UPD:
3519 case ARM::VLD3q32Pseudo_UPD:
3520 case ARM::VLD3q8oddPseudo:
3521 case ARM::VLD3q16oddPseudo:
3522 case ARM::VLD3q32oddPseudo:
3523 case ARM::VLD3q8oddPseudo_UPD:
3524 case ARM::VLD3q16oddPseudo_UPD:
3525 case ARM::VLD3q32oddPseudo_UPD:
3526 case ARM::VLD4d8Pseudo:
3527 case ARM::VLD4d16Pseudo:
3528 case ARM::VLD4d32Pseudo:
3529 case ARM::VLD1d64QPseudo:
3530 case ARM::VLD4d8Pseudo_UPD:
3531 case ARM::VLD4d16Pseudo_UPD:
3532 case ARM::VLD4d32Pseudo_UPD:
3533 case ARM::VLD4q8Pseudo_UPD:
3534 case ARM::VLD4q16Pseudo_UPD:
3535 case ARM::VLD4q32Pseudo_UPD:
3536 case ARM::VLD4q8oddPseudo:
3537 case ARM::VLD4q16oddPseudo:
3538 case ARM::VLD4q32oddPseudo:
3539 case ARM::VLD4q8oddPseudo_UPD:
3540 case ARM::VLD4q16oddPseudo_UPD:
3541 case ARM::VLD4q32oddPseudo_UPD:
3542 case ARM::VLD1DUPq8:
3543 case ARM::VLD1DUPq16:
3544 case ARM::VLD1DUPq32:
3545 case ARM::VLD1DUPq8wb_fixed:
3546 case ARM::VLD1DUPq16wb_fixed:
3547 case ARM::VLD1DUPq32wb_fixed:
3548 case ARM::VLD1DUPq8wb_register:
3549 case ARM::VLD1DUPq16wb_register:
3550 case ARM::VLD1DUPq32wb_register:
3551 case ARM::VLD2DUPd8:
3552 case ARM::VLD2DUPd16:
3553 case ARM::VLD2DUPd32:
3554 case ARM::VLD2DUPd8wb_fixed:
3555 case ARM::VLD2DUPd16wb_fixed:
3556 case ARM::VLD2DUPd32wb_fixed:
3557 case ARM::VLD2DUPd8wb_register:
3558 case ARM::VLD2DUPd16wb_register:
3559 case ARM::VLD2DUPd32wb_register:
3560 case ARM::VLD4DUPd8Pseudo:
3561 case ARM::VLD4DUPd16Pseudo:
3562 case ARM::VLD4DUPd32Pseudo:
3563 case ARM::VLD4DUPd8Pseudo_UPD:
3564 case ARM::VLD4DUPd16Pseudo_UPD:
3565 case ARM::VLD4DUPd32Pseudo_UPD:
3566 case ARM::VLD1LNq8Pseudo:
3567 case ARM::VLD1LNq16Pseudo:
3568 case ARM::VLD1LNq32Pseudo:
3569 case ARM::VLD1LNq8Pseudo_UPD:
3570 case ARM::VLD1LNq16Pseudo_UPD:
3571 case ARM::VLD1LNq32Pseudo_UPD:
3572 case ARM::VLD2LNd8Pseudo:
3573 case ARM::VLD2LNd16Pseudo:
3574 case ARM::VLD2LNd32Pseudo:
3575 case ARM::VLD2LNq16Pseudo:
3576 case ARM::VLD2LNq32Pseudo:
3577 case ARM::VLD2LNd8Pseudo_UPD:
3578 case ARM::VLD2LNd16Pseudo_UPD:
3579 case ARM::VLD2LNd32Pseudo_UPD:
3580 case ARM::VLD2LNq16Pseudo_UPD:
3581 case ARM::VLD2LNq32Pseudo_UPD:
3582 case ARM::VLD4LNd8Pseudo:
3583 case ARM::VLD4LNd16Pseudo:
3584 case ARM::VLD4LNd32Pseudo:
3585 case ARM::VLD4LNq16Pseudo:
3586 case ARM::VLD4LNq32Pseudo:
3587 case ARM::VLD4LNd8Pseudo_UPD:
3588 case ARM::VLD4LNd16Pseudo_UPD:
3589 case ARM::VLD4LNd32Pseudo_UPD:
3590 case ARM::VLD4LNq16Pseudo_UPD:
3591 case ARM::VLD4LNq32Pseudo_UPD:
3592 // If the address is not 64-bit aligned, the latencies of these
3593 // instructions increases by one.
3601 unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3602 const MachineInstr *MI,
3603 unsigned *PredCost) const {
3604 if (MI->isCopyLike() || MI->isInsertSubreg() ||
3605 MI->isRegSequence() || MI->isImplicitDef())
3608 // An instruction scheduler typically runs on unbundled instructions, however
3609 // other passes may query the latency of a bundled instruction.
3610 if (MI->isBundle()) {
3611 unsigned Latency = 0;
3612 MachineBasicBlock::const_instr_iterator I = MI;
3613 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end();
3614 while (++I != E && I->isInsideBundle()) {
3615 if (I->getOpcode() != ARM::t2IT)
3616 Latency += getInstrLatency(ItinData, I, PredCost);
3621 const MCInstrDesc &MCID = MI->getDesc();
3622 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) {
3623 // When predicated, CPSR is an additional source operand for CPSR updating
3624 // instructions, this apparently increases their latencies.
3627 // Be sure to call getStageLatency for an empty itinerary in case it has a
3628 // valid MinLatency property.
3630 return MI->mayLoad() ? 3 : 1;
3632 unsigned Class = MCID.getSchedClass();
3634 // For instructions with variable uops, use uops as latency.
3635 if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0)
3636 return getNumMicroOps(ItinData, MI);
3638 // For the common case, fall back on the itinerary's latency.
3639 unsigned Latency = ItinData->getStageLatency(Class);
3641 // Adjust for dynamic def-side opcode variants not captured by the itinerary.
3642 unsigned DefAlign = MI->hasOneMemOperand()
3643 ? (*MI->memoperands_begin())->getAlignment() : 0;
3644 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign);
3645 if (Adj >= 0 || (int)Latency > -Adj) {
3646 return Latency + Adj;
3651 int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
3652 SDNode *Node) const {
3653 if (!Node->isMachineOpcode())
3656 if (!ItinData || ItinData->isEmpty())
3659 unsigned Opcode = Node->getMachineOpcode();
3662 return ItinData->getStageLatency(get(Opcode).getSchedClass());
3669 bool ARMBaseInstrInfo::
3670 hasHighOperandLatency(const InstrItineraryData *ItinData,
3671 const MachineRegisterInfo *MRI,
3672 const MachineInstr *DefMI, unsigned DefIdx,
3673 const MachineInstr *UseMI, unsigned UseIdx) const {
3674 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3675 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask;
3676 if (Subtarget.isCortexA8() &&
3677 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP))
3678 // CortexA8 VFP instructions are not pipelined.
3681 // Hoist VFP / NEON instructions with 4 or higher latency.
3682 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx,
3685 Latency = getInstrLatency(ItinData, DefMI);
3688 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
3689 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
3692 bool ARMBaseInstrInfo::
3693 hasLowDefLatency(const InstrItineraryData *ItinData,
3694 const MachineInstr *DefMI, unsigned DefIdx) const {
3695 if (!ItinData || ItinData->isEmpty())
3698 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
3699 if (DDomain == ARMII::DomainGeneral) {
3700 unsigned DefClass = DefMI->getDesc().getSchedClass();
3701 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
3702 return (DefCycle != -1 && DefCycle <= 2);
3707 bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI,
3708 StringRef &ErrInfo) const {
3709 if (convertAddSubFlagsOpcode(MI->getOpcode())) {
3710 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG";
3717 ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
3718 unsigned &AddSubOpc,
3719 bool &NegAcc, bool &HasLane) const {
3720 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
3721 if (I == MLxEntryMap.end())
3724 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
3725 MulOpc = Entry.MulOpc;
3726 AddSubOpc = Entry.AddSubOpc;
3727 NegAcc = Entry.NegAcc;
3728 HasLane = Entry.HasLane;
3732 //===----------------------------------------------------------------------===//
3733 // Execution domains.
3734 //===----------------------------------------------------------------------===//
3736 // Some instructions go down the NEON pipeline, some go down the VFP pipeline,
3737 // and some can go down both. The vmov instructions go down the VFP pipeline,
3738 // but they can be changed to vorr equivalents that are executed by the NEON
3741 // We use the following execution domain numbering:
3749 // Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h
3751 std::pair<uint16_t, uint16_t>
3752 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
3753 // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
3754 // if they are not predicated.
3755 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
3756 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3758 // CortexA9 is particularly picky about mixing the two and wants these
3760 if (Subtarget.isCortexA9() && !isPredicated(MI) &&
3761 (MI->getOpcode() == ARM::VMOVRS ||
3762 MI->getOpcode() == ARM::VMOVSR ||
3763 MI->getOpcode() == ARM::VMOVS))
3764 return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
3766 // No other instructions can be swizzled, so just determine their domain.
3767 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
3769 if (Domain & ARMII::DomainNEON)
3770 return std::make_pair(ExeNEON, 0);
3772 // Certain instructions can go either way on Cortex-A8.
3773 // Treat them as NEON instructions.
3774 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8())
3775 return std::make_pair(ExeNEON, 0);
3777 if (Domain & ARMII::DomainVFP)
3778 return std::make_pair(ExeVFP, 0);
3780 return std::make_pair(ExeGeneric, 0);
3783 static unsigned getCorrespondingDRegAndLane(const TargetRegisterInfo *TRI,
3784 unsigned SReg, unsigned &Lane) {
3785 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_0, &ARM::DPRRegClass);
3788 if (DReg != ARM::NoRegister)
3792 DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1, &ARM::DPRRegClass);
3794 assert(DReg && "S-register with no D super-register?");
3798 /// getImplicitSPRUseForDPRUse - Given a use of a DPR register and lane,
3799 /// set ImplicitSReg to a register number that must be marked as implicit-use or
3800 /// zero if no register needs to be defined as implicit-use.
3802 /// If the function cannot determine if an SPR should be marked implicit use or
3803 /// not, it returns false.
3805 /// This function handles cases where an instruction is being modified from taking
3806 /// an SPR to a DPR[Lane]. A use of the DPR is being added, which may conflict
3807 /// with an earlier def of an SPR corresponding to DPR[Lane^1] (i.e. the other
3808 /// lane of the DPR).
3810 /// If the other SPR is defined, an implicit-use of it should be added. Else,
3811 /// (including the case where the DPR itself is defined), it should not.
3813 static bool getImplicitSPRUseForDPRUse(const TargetRegisterInfo *TRI,
3815 unsigned DReg, unsigned Lane,
3816 unsigned &ImplicitSReg) {
3817 // If the DPR is defined or used already, the other SPR lane will be chained
3818 // correctly, so there is nothing to be done.
3819 if (MI->definesRegister(DReg, TRI) || MI->readsRegister(DReg, TRI)) {
3824 // Otherwise we need to go searching to see if the SPR is set explicitly.
3825 ImplicitSReg = TRI->getSubReg(DReg,
3826 (Lane & 1) ? ARM::ssub_0 : ARM::ssub_1);
3827 MachineBasicBlock::LivenessQueryResult LQR =
3828 MI->getParent()->computeRegisterLiveness(TRI, ImplicitSReg, MI);
3830 if (LQR == MachineBasicBlock::LQR_Live)
3832 else if (LQR == MachineBasicBlock::LQR_Unknown)
3835 // If the register is known not to be live, there is no need to add an
3842 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
3843 unsigned DstReg, SrcReg, DReg;
3845 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3846 const TargetRegisterInfo *TRI = &getRegisterInfo();
3847 switch (MI->getOpcode()) {
3849 llvm_unreachable("cannot handle opcode!");
3852 if (Domain != ExeNEON)
3855 // Zap the predicate operands.
3856 assert(!isPredicated(MI) && "Cannot predicate a VORRd");
3858 // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
3859 DstReg = MI->getOperand(0).getReg();
3860 SrcReg = MI->getOperand(1).getReg();
3862 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3863 MI->RemoveOperand(i-1);
3865 // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits)
3866 MI->setDesc(get(ARM::VORRd));
3867 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3872 if (Domain != ExeNEON)
3874 assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
3876 // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits)
3877 DstReg = MI->getOperand(0).getReg();
3878 SrcReg = MI->getOperand(1).getReg();
3880 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3881 MI->RemoveOperand(i-1);
3883 DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane);
3885 // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps)
3886 // Note that DSrc has been widened and the other lane may be undef, which
3887 // contaminates the entire register.
3888 MI->setDesc(get(ARM::VGETLNi32));
3889 AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
3890 .addReg(DReg, RegState::Undef)
3893 // The old source should be an implicit use, otherwise we might think it
3894 // was dead before here.
3895 MIB.addReg(SrcReg, RegState::Implicit);
3898 if (Domain != ExeNEON)
3900 assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
3902 // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits)
3903 DstReg = MI->getOperand(0).getReg();
3904 SrcReg = MI->getOperand(1).getReg();
3906 DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane);
3908 unsigned ImplicitSReg;
3909 if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg))
3912 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3913 MI->RemoveOperand(i-1);
3915 // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps)
3916 // Again DDst may be undefined at the beginning of this instruction.
3917 MI->setDesc(get(ARM::VSETLNi32));
3918 MIB.addReg(DReg, RegState::Define)
3919 .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI)))
3922 AddDefaultPred(MIB);
3924 // The narrower destination must be marked as set to keep previous chains
3926 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
3927 if (ImplicitSReg != 0)
3928 MIB.addReg(ImplicitSReg, RegState::Implicit);
3932 if (Domain != ExeNEON)
3935 // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits)
3936 DstReg = MI->getOperand(0).getReg();
3937 SrcReg = MI->getOperand(1).getReg();
3939 unsigned DstLane = 0, SrcLane = 0, DDst, DSrc;
3940 DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane);
3941 DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane);
3943 unsigned ImplicitSReg;
3944 if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg))
3947 for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
3948 MI->RemoveOperand(i-1);
3951 // Destination can be:
3952 // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits)
3953 MI->setDesc(get(ARM::VDUPLN32d));
3954 MIB.addReg(DDst, RegState::Define)
3955 .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI)))
3957 AddDefaultPred(MIB);
3959 // Neither the source or the destination are naturally represented any
3960 // more, so add them in manually.
3961 MIB.addReg(DstReg, RegState::Implicit | RegState::Define);
3962 MIB.addReg(SrcReg, RegState::Implicit);
3963 if (ImplicitSReg != 0)
3964 MIB.addReg(ImplicitSReg, RegState::Implicit);
3968 // In general there's no single instruction that can perform an S <-> S
3969 // move in NEON space, but a pair of VEXT instructions *can* do the
3970 // job. It turns out that the VEXTs needed will only use DSrc once, with
3971 // the position based purely on the combination of lane-0 and lane-1
3972 // involved. For example
3973 // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1
3974 // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1
3975 // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1
3976 // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1
3978 // Pattern of the MachineInstrs is:
3979 // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits)
3980 MachineInstrBuilder NewMIB;
3981 NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
3982 get(ARM::VEXTd32), DDst);
3984 // On the first instruction, both DSrc and DDst may be <undef> if present.
3985 // Specifically when the original instruction didn't have them as an
3987 unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst;
3988 bool CurUndef = !MI->readsRegister(CurReg, TRI);
3989 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
3991 CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst;
3992 CurUndef = !MI->readsRegister(CurReg, TRI);
3993 NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
3996 AddDefaultPred(NewMIB);
3998 if (SrcLane == DstLane)
3999 NewMIB.addReg(SrcReg, RegState::Implicit);
4001 MI->setDesc(get(ARM::VEXTd32));
4002 MIB.addReg(DDst, RegState::Define);
4004 // On the second instruction, DDst has definitely been defined above, so
4005 // it is not <undef>. DSrc, if present, can be <undef> as above.
4006 CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst;
4007 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4008 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4010 CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst;
4011 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
4012 MIB.addReg(CurReg, getUndefRegState(CurUndef));
4015 AddDefaultPred(MIB);
4017 if (SrcLane != DstLane)
4018 MIB.addReg(SrcReg, RegState::Implicit);
4020 // As before, the original destination is no longer represented, add it
4022 MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
4023 if (ImplicitSReg != 0)
4024 MIB.addReg(ImplicitSReg, RegState::Implicit);
4031 //===----------------------------------------------------------------------===//
4032 // Partial register updates
4033 //===----------------------------------------------------------------------===//
4035 // Swift renames NEON registers with 64-bit granularity. That means any
4036 // instruction writing an S-reg implicitly reads the containing D-reg. The
4037 // problem is mostly avoided by translating f32 operations to v2f32 operations
4038 // on D-registers, but f32 loads are still a problem.
4040 // These instructions can load an f32 into a NEON register:
4042 // VLDRS - Only writes S, partial D update.
4043 // VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops.
4044 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
4046 // FCONSTD can be used as a dependency-breaking instruction.
4047 unsigned ARMBaseInstrInfo::
4048 getPartialRegUpdateClearance(const MachineInstr *MI,
4050 const TargetRegisterInfo *TRI) const {
4051 if (!SwiftPartialUpdateClearance ||
4052 !(Subtarget.isSwift() || Subtarget.isCortexA15()))
4055 assert(TRI && "Need TRI instance");
4057 const MachineOperand &MO = MI->getOperand(OpNum);
4060 unsigned Reg = MO.getReg();
4063 switch(MI->getOpcode()) {
4064 // Normal instructions writing only an S-register.
4069 case ARM::VMOVv4i16:
4070 case ARM::VMOVv2i32:
4071 case ARM::VMOVv2f32:
4072 case ARM::VMOVv1i64:
4073 UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI);
4076 // Explicitly reads the dependency.
4077 case ARM::VLD1LNd32:
4084 // If this instruction actually reads a value from Reg, there is no unwanted
4086 if (UseOp != -1 && MI->getOperand(UseOp).readsReg())
4089 // We must be able to clobber the whole D-reg.
4090 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4091 // Virtual register must be a foo:ssub_0<def,undef> operand.
4092 if (!MO.getSubReg() || MI->readsVirtualRegister(Reg))
4094 } else if (ARM::SPRRegClass.contains(Reg)) {
4095 // Physical register: MI must define the full D-reg.
4096 unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0,
4098 if (!DReg || !MI->definesRegister(DReg, TRI))
4102 // MI has an unwanted D-register dependency.
4103 // Avoid defs in the previous N instructrions.
4104 return SwiftPartialUpdateClearance;
4107 // Break a partial register dependency after getPartialRegUpdateClearance
4108 // returned non-zero.
4109 void ARMBaseInstrInfo::
4110 breakPartialRegDependency(MachineBasicBlock::iterator MI,
4112 const TargetRegisterInfo *TRI) const {
4113 assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def");
4114 assert(TRI && "Need TRI instance");
4116 const MachineOperand &MO = MI->getOperand(OpNum);
4117 unsigned Reg = MO.getReg();
4118 assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
4119 "Can't break virtual register dependencies.");
4120 unsigned DReg = Reg;
4122 // If MI defines an S-reg, find the corresponding D super-register.
4123 if (ARM::SPRRegClass.contains(Reg)) {
4124 DReg = ARM::D0 + (Reg - ARM::S0) / 2;
4125 assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken");
4128 assert(ARM::DPRRegClass.contains(DReg) && "Can only break D-reg deps");
4129 assert(MI->definesRegister(DReg, TRI) && "MI doesn't clobber full D-reg");
4131 // FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines
4132 // the full D-register by loading the same value to both lanes. The
4133 // instruction is micro-coded with 2 uops, so don't do this until we can
4134 // properly schedule micro-coded instuctions. The dispatcher stalls cause
4135 // too big regressions.
4137 // Insert the dependency-breaking FCONSTD before MI.
4138 // 96 is the encoding of 0.5, but the actual value doesn't matter here.
4139 AddDefaultPred(BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
4140 get(ARM::FCONSTD), DReg).addImm(96));
4141 MI->addRegisterKilled(DReg, TRI, true);
4144 bool ARMBaseInstrInfo::hasNOP() const {
4145 return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
4148 bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {
4149 unsigned ShOpVal = MI->getOperand(3).getImm();
4150 unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal);
4151 // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1.
4152 if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) ||
4153 ((ShImm == 1 || ShImm == 2) &&
4154 ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl))