1 //===- ARMBaseInstrInfo.cpp - ARM Instruction Information -------*- C++ -*-===//
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 "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMGenInstrInfo.inc"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMRegisterInfo.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalValue.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/CodeGen/LiveVariables.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineJumpTableInfo.h"
30 #include "llvm/CodeGen/MachineMemOperand.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/PseudoSourceValue.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
40 EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
41 cl::desc("Enable ARM 2-addr to 3-addr conv"));
43 ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
44 : TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)),
49 ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
50 MachineBasicBlock::iterator &MBBI,
51 LiveVariables *LV) const {
52 // FIXME: Thumb2 support.
57 MachineInstr *MI = MBBI;
58 MachineFunction &MF = *MI->getParent()->getParent();
59 uint64_t TSFlags = MI->getDesc().TSFlags;
61 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) {
63 case ARMII::IndexModePre:
66 case ARMII::IndexModePost:
70 // Try splitting an indexed load/store to an un-indexed one plus an add/sub
72 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode());
76 MachineInstr *UpdateMI = NULL;
77 MachineInstr *MemMI = NULL;
78 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask);
79 const TargetInstrDesc &TID = MI->getDesc();
80 unsigned NumOps = TID.getNumOperands();
81 bool isLoad = !TID.mayStore();
82 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0);
83 const MachineOperand &Base = MI->getOperand(2);
84 const MachineOperand &Offset = MI->getOperand(NumOps-3);
85 unsigned WBReg = WB.getReg();
86 unsigned BaseReg = Base.getReg();
87 unsigned OffReg = Offset.getReg();
88 unsigned OffImm = MI->getOperand(NumOps-2).getImm();
89 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm();
92 assert(false && "Unknown indexed op!");
94 case ARMII::AddrMode2: {
95 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub;
96 unsigned Amt = ARM_AM::getAM2Offset(OffImm);
98 if (ARM_AM::getSOImmVal(Amt) == -1)
99 // Can't encode it in a so_imm operand. This transformation will
100 // add more than 1 instruction. Abandon!
102 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
103 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
104 .addReg(BaseReg).addImm(Amt)
105 .addImm(Pred).addReg(0).addReg(0);
106 } else if (Amt != 0) {
107 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm);
108 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt);
109 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
110 get(isSub ? ARM::SUBrs : ARM::ADDrs), WBReg)
111 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc)
112 .addImm(Pred).addReg(0).addReg(0);
114 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
115 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
116 .addReg(BaseReg).addReg(OffReg)
117 .addImm(Pred).addReg(0).addReg(0);
120 case ARMII::AddrMode3 : {
121 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub;
122 unsigned Amt = ARM_AM::getAM3Offset(OffImm);
124 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand.
125 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
126 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg)
127 .addReg(BaseReg).addImm(Amt)
128 .addImm(Pred).addReg(0).addReg(0);
130 UpdateMI = BuildMI(MF, MI->getDebugLoc(),
131 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg)
132 .addReg(BaseReg).addReg(OffReg)
133 .addImm(Pred).addReg(0).addReg(0);
138 std::vector<MachineInstr*> NewMIs;
141 MemMI = BuildMI(MF, MI->getDebugLoc(),
142 get(MemOpc), MI->getOperand(0).getReg())
143 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
145 MemMI = BuildMI(MF, MI->getDebugLoc(),
146 get(MemOpc)).addReg(MI->getOperand(1).getReg())
147 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred);
148 NewMIs.push_back(MemMI);
149 NewMIs.push_back(UpdateMI);
152 MemMI = BuildMI(MF, MI->getDebugLoc(),
153 get(MemOpc), MI->getOperand(0).getReg())
154 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
156 MemMI = BuildMI(MF, MI->getDebugLoc(),
157 get(MemOpc)).addReg(MI->getOperand(1).getReg())
158 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred);
160 UpdateMI->getOperand(0).setIsDead();
161 NewMIs.push_back(UpdateMI);
162 NewMIs.push_back(MemMI);
165 // Transfer LiveVariables states, kill / dead info.
167 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
168 MachineOperand &MO = MI->getOperand(i);
169 if (MO.isReg() && MO.getReg() &&
170 TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
171 unsigned Reg = MO.getReg();
173 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg);
175 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI;
177 LV->addVirtualRegisterDead(Reg, NewMI);
179 if (MO.isUse() && MO.isKill()) {
180 for (unsigned j = 0; j < 2; ++j) {
181 // Look at the two new MI's in reverse order.
182 MachineInstr *NewMI = NewMIs[j];
183 if (!NewMI->readsRegister(Reg))
185 LV->addVirtualRegisterKilled(Reg, NewMI);
186 if (VI.removeKill(MI))
187 VI.Kills.push_back(NewMI);
195 MFI->insert(MBBI, NewMIs[1]);
196 MFI->insert(MBBI, NewMIs[0]);
201 ARMBaseInstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
202 MachineBasicBlock::iterator MI,
203 const std::vector<CalleeSavedInfo> &CSI,
204 const TargetRegisterInfo *TRI) const {
209 if (MI != MBB.end()) DL = MI->getDebugLoc();
211 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
212 unsigned Reg = CSI[i].getReg();
215 // Add the callee-saved register as live-in unless it's LR and
216 // @llvm.returnaddress is called. If LR is returned for @llvm.returnaddress
217 // then it's already added to the function and entry block live-in sets.
218 if (Reg == ARM::LR) {
219 MachineFunction &MF = *MBB.getParent();
220 if (MF.getFrameInfo()->isReturnAddressTaken() &&
221 MF.getRegInfo().isLiveIn(Reg))
228 // Insert the spill to the stack frame. The register is killed at the spill
230 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
231 storeRegToStackSlot(MBB, MI, Reg, isKill,
232 CSI[i].getFrameIdx(), RC, TRI);
239 ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
240 MachineBasicBlock *&FBB,
241 SmallVectorImpl<MachineOperand> &Cond,
242 bool AllowModify) const {
243 // If the block has no terminators, it just falls into the block after it.
244 MachineBasicBlock::iterator I = MBB.end();
245 if (I == MBB.begin())
248 while (I->isDebugValue()) {
249 if (I == MBB.begin())
253 if (!isUnpredicatedTerminator(I))
256 // Get the last instruction in the block.
257 MachineInstr *LastInst = I;
259 // If there is only one terminator instruction, process it.
260 unsigned LastOpc = LastInst->getOpcode();
261 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
262 if (isUncondBranchOpcode(LastOpc)) {
263 TBB = LastInst->getOperand(0).getMBB();
266 if (isCondBranchOpcode(LastOpc)) {
267 // Block ends with fall-through condbranch.
268 TBB = LastInst->getOperand(0).getMBB();
269 Cond.push_back(LastInst->getOperand(1));
270 Cond.push_back(LastInst->getOperand(2));
273 return true; // Can't handle indirect branch.
276 // Get the instruction before it if it is a terminator.
277 MachineInstr *SecondLastInst = I;
279 // If there are three terminators, we don't know what sort of block this is.
280 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
283 // If the block ends with a B and a Bcc, handle it.
284 unsigned SecondLastOpc = SecondLastInst->getOpcode();
285 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
286 TBB = SecondLastInst->getOperand(0).getMBB();
287 Cond.push_back(SecondLastInst->getOperand(1));
288 Cond.push_back(SecondLastInst->getOperand(2));
289 FBB = LastInst->getOperand(0).getMBB();
293 // If the block ends with two unconditional branches, handle it. The second
294 // one is not executed, so remove it.
295 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
296 TBB = SecondLastInst->getOperand(0).getMBB();
299 I->eraseFromParent();
303 // ...likewise if it ends with a branch table followed by an unconditional
304 // branch. The branch folder can create these, and we must get rid of them for
305 // correctness of Thumb constant islands.
306 if ((isJumpTableBranchOpcode(SecondLastOpc) ||
307 isIndirectBranchOpcode(SecondLastOpc)) &&
308 isUncondBranchOpcode(LastOpc)) {
311 I->eraseFromParent();
315 // Otherwise, can't handle this.
320 unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
321 MachineBasicBlock::iterator I = MBB.end();
322 if (I == MBB.begin()) return 0;
324 while (I->isDebugValue()) {
325 if (I == MBB.begin())
329 if (!isUncondBranchOpcode(I->getOpcode()) &&
330 !isCondBranchOpcode(I->getOpcode()))
333 // Remove the branch.
334 I->eraseFromParent();
338 if (I == MBB.begin()) return 1;
340 if (!isCondBranchOpcode(I->getOpcode()))
343 // Remove the branch.
344 I->eraseFromParent();
349 ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
350 MachineBasicBlock *FBB,
351 const SmallVectorImpl<MachineOperand> &Cond,
353 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
354 int BOpc = !AFI->isThumbFunction()
355 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB);
356 int BccOpc = !AFI->isThumbFunction()
357 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc);
359 // Shouldn't be a fall through.
360 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
361 assert((Cond.size() == 2 || Cond.size() == 0) &&
362 "ARM branch conditions have two components!");
365 if (Cond.empty()) // Unconditional branch?
366 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
368 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
369 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
373 // Two-way conditional branch.
374 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
375 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
376 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
380 bool ARMBaseInstrInfo::
381 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
382 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm();
383 Cond[0].setImm(ARMCC::getOppositeCondition(CC));
387 bool ARMBaseInstrInfo::
388 PredicateInstruction(MachineInstr *MI,
389 const SmallVectorImpl<MachineOperand> &Pred) const {
390 unsigned Opc = MI->getOpcode();
391 if (isUncondBranchOpcode(Opc)) {
392 MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
393 MI->addOperand(MachineOperand::CreateImm(Pred[0].getImm()));
394 MI->addOperand(MachineOperand::CreateReg(Pred[1].getReg(), false));
398 int PIdx = MI->findFirstPredOperandIdx();
400 MachineOperand &PMO = MI->getOperand(PIdx);
401 PMO.setImm(Pred[0].getImm());
402 MI->getOperand(PIdx+1).setReg(Pred[1].getReg());
408 bool ARMBaseInstrInfo::
409 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
410 const SmallVectorImpl<MachineOperand> &Pred2) const {
411 if (Pred1.size() > 2 || Pred2.size() > 2)
414 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm();
415 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm();
425 return CC2 == ARMCC::HI;
427 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ;
429 return CC2 == ARMCC::GT;
431 return CC2 == ARMCC::LT;
435 bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI,
436 std::vector<MachineOperand> &Pred) const {
437 // FIXME: This confuses implicit_def with optional CPSR def.
438 const TargetInstrDesc &TID = MI->getDesc();
439 if (!TID.getImplicitDefs() && !TID.hasOptionalDef())
443 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
444 const MachineOperand &MO = MI->getOperand(i);
445 if (MO.isReg() && MO.getReg() == ARM::CPSR) {
454 /// isPredicable - Return true if the specified instruction can be predicated.
455 /// By default, this returns true for every instruction with a
456 /// PredicateOperand.
457 bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const {
458 const TargetInstrDesc &TID = MI->getDesc();
459 if (!TID.isPredicable())
462 if ((TID.TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) {
463 ARMFunctionInfo *AFI =
464 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>();
465 return AFI->isThumb2Function();
470 /// FIXME: Works around a gcc miscompilation with -fstrict-aliasing.
472 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
474 static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT,
476 assert(JTI < JT.size());
477 return JT[JTI].MBBs.size();
480 /// GetInstSize - Return the size of the specified MachineInstr.
482 unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
483 const MachineBasicBlock &MBB = *MI->getParent();
484 const MachineFunction *MF = MBB.getParent();
485 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
487 // Basic size info comes from the TSFlags field.
488 const TargetInstrDesc &TID = MI->getDesc();
489 uint64_t TSFlags = TID.TSFlags;
491 unsigned Opc = MI->getOpcode();
492 switch ((TSFlags & ARMII::SizeMask) >> ARMII::SizeShift) {
494 // If this machine instr is an inline asm, measure it.
495 if (MI->getOpcode() == ARM::INLINEASM)
496 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI);
501 llvm_unreachable("Unknown or unset size field for instr!");
502 case TargetOpcode::IMPLICIT_DEF:
503 case TargetOpcode::KILL:
504 case TargetOpcode::DBG_LABEL:
505 case TargetOpcode::EH_LABEL:
506 case TargetOpcode::DBG_VALUE:
511 case ARMII::Size8Bytes: return 8; // ARM instruction x 2.
512 case ARMII::Size4Bytes: return 4; // ARM / Thumb2 instruction.
513 case ARMII::Size2Bytes: return 2; // Thumb1 instruction.
514 case ARMII::SizeSpecial: {
516 case ARM::CONSTPOOL_ENTRY:
517 // If this machine instr is a constant pool entry, its size is recorded as
519 return MI->getOperand(2).getImm();
520 case ARM::Int_eh_sjlj_longjmp:
522 case ARM::tInt_eh_sjlj_longjmp:
524 case ARM::Int_eh_sjlj_setjmp:
525 case ARM::Int_eh_sjlj_setjmp_nofp:
527 case ARM::tInt_eh_sjlj_setjmp:
528 case ARM::t2Int_eh_sjlj_setjmp:
529 case ARM::t2Int_eh_sjlj_setjmp_nofp:
538 // These are jumptable branches, i.e. a branch followed by an inlined
539 // jumptable. The size is 4 + 4 * number of entries. For TBB, each
540 // entry is one byte; TBH two byte each.
541 unsigned EntrySize = (Opc == ARM::t2TBB)
542 ? 1 : ((Opc == ARM::t2TBH) ? 2 : 4);
543 unsigned NumOps = TID.getNumOperands();
544 MachineOperand JTOP =
545 MI->getOperand(NumOps - (TID.isPredicable() ? 3 : 2));
546 unsigned JTI = JTOP.getIndex();
547 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
549 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
550 assert(JTI < JT.size());
551 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
552 // 4 aligned. The assembler / linker may add 2 byte padding just before
553 // the JT entries. The size does not include this padding; the
554 // constant islands pass does separate bookkeeping for it.
555 // FIXME: If we know the size of the function is less than (1 << 16) *2
556 // bytes, we can use 16-bit entries instead. Then there won't be an
558 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
559 unsigned NumEntries = getNumJTEntries(JT, JTI);
560 if (Opc == ARM::t2TBB && (NumEntries & 1))
561 // Make sure the instruction that follows TBB is 2-byte aligned.
562 // FIXME: Constant island pass should insert an "ALIGN" instruction
565 return NumEntries * EntrySize + InstSize;
568 // Otherwise, pseudo-instruction sizes are zero.
573 return 0; // Not reached
576 /// Return true if the instruction is a register to register move and
577 /// leave the source and dest operands in the passed parameters.
580 ARMBaseInstrInfo::isMoveInstr(const MachineInstr &MI,
581 unsigned &SrcReg, unsigned &DstReg,
582 unsigned& SrcSubIdx, unsigned& DstSubIdx) const {
583 switch (MI.getOpcode()) {
590 SrcReg = MI.getOperand(1).getReg();
591 DstReg = MI.getOperand(0).getReg();
592 SrcSubIdx = MI.getOperand(1).getSubReg();
593 DstSubIdx = MI.getOperand(0).getSubReg();
599 case ARM::tMOVgpr2tgpr:
600 case ARM::tMOVtgpr2gpr:
601 case ARM::tMOVgpr2gpr:
603 assert(MI.getDesc().getNumOperands() >= 2 &&
604 MI.getOperand(0).isReg() &&
605 MI.getOperand(1).isReg() &&
606 "Invalid ARM MOV instruction");
607 SrcReg = MI.getOperand(1).getReg();
608 DstReg = MI.getOperand(0).getReg();
609 SrcSubIdx = MI.getOperand(1).getSubReg();
610 DstSubIdx = MI.getOperand(0).getSubReg();
619 ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
620 int &FrameIndex) const {
621 switch (MI->getOpcode()) {
624 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame.
625 if (MI->getOperand(1).isFI() &&
626 MI->getOperand(2).isReg() &&
627 MI->getOperand(3).isImm() &&
628 MI->getOperand(2).getReg() == 0 &&
629 MI->getOperand(3).getImm() == 0) {
630 FrameIndex = MI->getOperand(1).getIndex();
631 return MI->getOperand(0).getReg();
636 if (MI->getOperand(1).isFI() &&
637 MI->getOperand(2).isImm() &&
638 MI->getOperand(2).getImm() == 0) {
639 FrameIndex = MI->getOperand(1).getIndex();
640 return MI->getOperand(0).getReg();
645 if (MI->getOperand(1).isFI() &&
646 MI->getOperand(2).isImm() &&
647 MI->getOperand(2).getImm() == 0) {
648 FrameIndex = MI->getOperand(1).getIndex();
649 return MI->getOperand(0).getReg();
658 ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
659 int &FrameIndex) const {
660 switch (MI->getOpcode()) {
663 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame.
664 if (MI->getOperand(1).isFI() &&
665 MI->getOperand(2).isReg() &&
666 MI->getOperand(3).isImm() &&
667 MI->getOperand(2).getReg() == 0 &&
668 MI->getOperand(3).getImm() == 0) {
669 FrameIndex = MI->getOperand(1).getIndex();
670 return MI->getOperand(0).getReg();
675 if (MI->getOperand(1).isFI() &&
676 MI->getOperand(2).isImm() &&
677 MI->getOperand(2).getImm() == 0) {
678 FrameIndex = MI->getOperand(1).getIndex();
679 return MI->getOperand(0).getReg();
684 if (MI->getOperand(1).isFI() &&
685 MI->getOperand(2).isImm() &&
686 MI->getOperand(2).getImm() == 0) {
687 FrameIndex = MI->getOperand(1).getIndex();
688 return MI->getOperand(0).getReg();
696 void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
697 MachineBasicBlock::iterator I, DebugLoc DL,
698 unsigned DestReg, unsigned SrcReg,
699 bool KillSrc) const {
700 bool GPRDest = ARM::GPRRegClass.contains(DestReg);
701 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg);
703 if (GPRDest && GPRSrc) {
704 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg)
705 .addReg(SrcReg, getKillRegState(KillSrc))));
709 bool SPRDest = ARM::SPRRegClass.contains(DestReg);
710 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg);
713 if (SPRDest && SPRSrc)
715 else if (GPRDest && SPRSrc)
717 else if (SPRDest && GPRSrc)
719 else if (ARM::DPRRegClass.contains(DestReg, SrcReg))
721 else if (ARM::QPRRegClass.contains(DestReg, SrcReg))
723 else if (ARM::QQPRRegClass.contains(DestReg, SrcReg))
725 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg))
728 llvm_unreachable("Impossible reg-to-reg copy");
730 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg);
731 MIB.addReg(SrcReg, getKillRegState(KillSrc));
732 if (Opc != ARM::VMOVQQ && Opc != ARM::VMOVQQQQ)
737 MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB,
738 unsigned Reg, unsigned SubIdx, unsigned State,
739 const TargetRegisterInfo *TRI) {
741 return MIB.addReg(Reg, State);
743 if (TargetRegisterInfo::isPhysicalRegister(Reg))
744 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
745 return MIB.addReg(Reg, State, SubIdx);
748 void ARMBaseInstrInfo::
749 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
750 unsigned SrcReg, bool isKill, int FI,
751 const TargetRegisterClass *RC,
752 const TargetRegisterInfo *TRI) const {
754 if (I != MBB.end()) DL = I->getDebugLoc();
755 MachineFunction &MF = *MBB.getParent();
756 MachineFrameInfo &MFI = *MF.getFrameInfo();
757 unsigned Align = MFI.getObjectAlignment(FI);
759 MachineMemOperand *MMO =
760 MF.getMachineMemOperand(PseudoSourceValue::getFixedStack(FI),
761 MachineMemOperand::MOStore, 0,
762 MFI.getObjectSize(FI),
765 // tGPR is used sometimes in ARM instructions that need to avoid using
766 // certain registers. Just treat it as GPR here.
767 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass)
768 RC = ARM::GPRRegisterClass;
770 switch (RC->getID()) {
771 case ARM::GPRRegClassID:
772 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STR))
773 .addReg(SrcReg, getKillRegState(isKill))
774 .addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO));
776 case ARM::SPRRegClassID:
777 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS))
778 .addReg(SrcReg, getKillRegState(isKill))
779 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
781 case ARM::DPRRegClassID:
782 case ARM::DPR_VFP2RegClassID:
783 case ARM::DPR_8RegClassID:
784 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD))
785 .addReg(SrcReg, getKillRegState(isKill))
786 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
788 case ARM::QPRRegClassID:
789 case ARM::QPR_VFP2RegClassID:
790 case ARM::QPR_8RegClassID:
791 // FIXME: Neon instructions should support predicates
792 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
793 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q))
794 .addFrameIndex(FI).addImm(16)
795 .addReg(SrcReg, getKillRegState(isKill))
796 .addMemOperand(MMO));
798 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQ))
799 .addReg(SrcReg, getKillRegState(isKill))
801 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))
802 .addMemOperand(MMO));
805 case ARM::QQPRRegClassID:
806 case ARM::QQPR_VFP2RegClassID:
807 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
808 // FIXME: It's possible to only store part of the QQ register if the
809 // spilled def has a sub-register index.
810 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::VST1d64Q))
811 .addFrameIndex(FI).addImm(16);
812 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
813 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
814 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
815 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
816 AddDefaultPred(MIB.addMemOperand(MMO));
818 MachineInstrBuilder MIB =
819 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMD))
821 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4)))
823 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
824 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
825 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
826 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
829 case ARM::QQQQPRRegClassID: {
830 MachineInstrBuilder MIB =
831 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMD))
833 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4)))
835 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
836 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
837 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
838 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI);
839 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI);
840 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI);
841 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI);
842 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI);
846 llvm_unreachable("Unknown regclass!");
850 void ARMBaseInstrInfo::
851 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
852 unsigned DestReg, int FI,
853 const TargetRegisterClass *RC,
854 const TargetRegisterInfo *TRI) const {
856 if (I != MBB.end()) DL = I->getDebugLoc();
857 MachineFunction &MF = *MBB.getParent();
858 MachineFrameInfo &MFI = *MF.getFrameInfo();
859 unsigned Align = MFI.getObjectAlignment(FI);
860 MachineMemOperand *MMO =
861 MF.getMachineMemOperand(PseudoSourceValue::getFixedStack(FI),
862 MachineMemOperand::MOLoad, 0,
863 MFI.getObjectSize(FI),
866 // tGPR is used sometimes in ARM instructions that need to avoid using
867 // certain registers. Just treat it as GPR here.
868 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass)
869 RC = ARM::GPRRegisterClass;
871 switch (RC->getID()) {
872 case ARM::GPRRegClassID:
873 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDR), DestReg)
874 .addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO));
876 case ARM::SPRRegClassID:
877 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
878 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
880 case ARM::DPRRegClassID:
881 case ARM::DPR_VFP2RegClassID:
882 case ARM::DPR_8RegClassID:
883 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg)
884 .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
886 case ARM::QPRRegClassID:
887 case ARM::QPR_VFP2RegClassID:
888 case ARM::QPR_8RegClassID:
889 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
890 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q), DestReg)
891 .addFrameIndex(FI).addImm(16)
892 .addMemOperand(MMO));
894 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQ), DestReg)
896 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))
897 .addMemOperand(MMO));
900 case ARM::QQPRRegClassID:
901 case ARM::QQPR_VFP2RegClassID:
902 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
903 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::VLD1d64Q));
904 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
905 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
906 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
907 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
908 AddDefaultPred(MIB.addFrameIndex(FI).addImm(16).addMemOperand(MMO));
910 MachineInstrBuilder MIB =
911 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMD))
913 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4)))
915 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
916 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
917 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
918 AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
921 case ARM::QQQQPRRegClassID: {
922 MachineInstrBuilder MIB =
923 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMD))
925 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4)))
927 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI);
928 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI);
929 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI);
930 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI);
931 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI);
932 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI);
933 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI);
934 AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI);
938 llvm_unreachable("Unknown regclass!");
943 ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
944 int FrameIx, uint64_t Offset,
947 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE))
948 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr);
952 MachineInstr *ARMBaseInstrInfo::
953 foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI,
954 const SmallVectorImpl<unsigned> &Ops, int FI) const {
955 if (Ops.size() != 1) return NULL;
957 unsigned OpNum = Ops[0];
958 unsigned Opc = MI->getOpcode();
959 MachineInstr *NewMI = NULL;
960 if (Opc == ARM::MOVr || Opc == ARM::t2MOVr) {
961 // If it is updating CPSR, then it cannot be folded.
962 if (MI->getOperand(4).getReg() == ARM::CPSR && !MI->getOperand(4).isDead())
964 unsigned Pred = MI->getOperand(2).getImm();
965 unsigned PredReg = MI->getOperand(3).getReg();
966 if (OpNum == 0) { // move -> store
967 unsigned SrcReg = MI->getOperand(1).getReg();
968 unsigned SrcSubReg = MI->getOperand(1).getSubReg();
969 bool isKill = MI->getOperand(1).isKill();
970 bool isUndef = MI->getOperand(1).isUndef();
971 if (Opc == ARM::MOVr)
972 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::STR))
974 getKillRegState(isKill) | getUndefRegState(isUndef),
976 .addFrameIndex(FI).addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
978 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::t2STRi12))
980 getKillRegState(isKill) | getUndefRegState(isUndef),
982 .addFrameIndex(FI).addImm(0).addImm(Pred).addReg(PredReg);
983 } else { // move -> load
984 unsigned DstReg = MI->getOperand(0).getReg();
985 unsigned DstSubReg = MI->getOperand(0).getSubReg();
986 bool isDead = MI->getOperand(0).isDead();
987 bool isUndef = MI->getOperand(0).isUndef();
988 if (Opc == ARM::MOVr)
989 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::LDR))
992 getDeadRegState(isDead) |
993 getUndefRegState(isUndef), DstSubReg)
994 .addFrameIndex(FI).addReg(0).addImm(0).addImm(Pred).addReg(PredReg);
996 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::t2LDRi12))
999 getDeadRegState(isDead) |
1000 getUndefRegState(isUndef), DstSubReg)
1001 .addFrameIndex(FI).addImm(0).addImm(Pred).addReg(PredReg);
1003 } else if (Opc == ARM::tMOVgpr2gpr ||
1004 Opc == ARM::tMOVtgpr2gpr ||
1005 Opc == ARM::tMOVgpr2tgpr) {
1006 if (OpNum == 0) { // move -> store
1007 unsigned SrcReg = MI->getOperand(1).getReg();
1008 unsigned SrcSubReg = MI->getOperand(1).getSubReg();
1009 bool isKill = MI->getOperand(1).isKill();
1010 bool isUndef = MI->getOperand(1).isUndef();
1011 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::t2STRi12))
1013 getKillRegState(isKill) | getUndefRegState(isUndef),
1015 .addFrameIndex(FI).addImm(0).addImm(ARMCC::AL).addReg(0);
1016 } else { // move -> load
1017 unsigned DstReg = MI->getOperand(0).getReg();
1018 unsigned DstSubReg = MI->getOperand(0).getSubReg();
1019 bool isDead = MI->getOperand(0).isDead();
1020 bool isUndef = MI->getOperand(0).isUndef();
1021 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::t2LDRi12))
1024 getDeadRegState(isDead) |
1025 getUndefRegState(isUndef),
1027 .addFrameIndex(FI).addImm(0).addImm(ARMCC::AL).addReg(0);
1029 } else if (Opc == ARM::VMOVS) {
1030 unsigned Pred = MI->getOperand(2).getImm();
1031 unsigned PredReg = MI->getOperand(3).getReg();
1032 if (OpNum == 0) { // move -> store
1033 unsigned SrcReg = MI->getOperand(1).getReg();
1034 unsigned SrcSubReg = MI->getOperand(1).getSubReg();
1035 bool isKill = MI->getOperand(1).isKill();
1036 bool isUndef = MI->getOperand(1).isUndef();
1037 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VSTRS))
1038 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef),
1041 .addImm(0).addImm(Pred).addReg(PredReg);
1042 } else { // move -> load
1043 unsigned DstReg = MI->getOperand(0).getReg();
1044 unsigned DstSubReg = MI->getOperand(0).getSubReg();
1045 bool isDead = MI->getOperand(0).isDead();
1046 bool isUndef = MI->getOperand(0).isUndef();
1047 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VLDRS))
1050 getDeadRegState(isDead) |
1051 getUndefRegState(isUndef),
1053 .addFrameIndex(FI).addImm(0).addImm(Pred).addReg(PredReg);
1055 } else if (Opc == ARM::VMOVD || Opc == ARM::VMOVDneon) {
1056 unsigned Pred = MI->getOperand(2).getImm();
1057 unsigned PredReg = MI->getOperand(3).getReg();
1058 if (OpNum == 0) { // move -> store
1059 unsigned SrcReg = MI->getOperand(1).getReg();
1060 unsigned SrcSubReg = MI->getOperand(1).getSubReg();
1061 bool isKill = MI->getOperand(1).isKill();
1062 bool isUndef = MI->getOperand(1).isUndef();
1063 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VSTRD))
1065 getKillRegState(isKill) | getUndefRegState(isUndef),
1067 .addFrameIndex(FI).addImm(0).addImm(Pred).addReg(PredReg);
1068 } else { // move -> load
1069 unsigned DstReg = MI->getOperand(0).getReg();
1070 unsigned DstSubReg = MI->getOperand(0).getSubReg();
1071 bool isDead = MI->getOperand(0).isDead();
1072 bool isUndef = MI->getOperand(0).isUndef();
1073 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VLDRD))
1076 getDeadRegState(isDead) |
1077 getUndefRegState(isUndef),
1079 .addFrameIndex(FI).addImm(0).addImm(Pred).addReg(PredReg);
1081 } else if (Opc == ARM::VMOVQ) {
1082 MachineFrameInfo &MFI = *MF.getFrameInfo();
1083 unsigned Pred = MI->getOperand(2).getImm();
1084 unsigned PredReg = MI->getOperand(3).getReg();
1085 if (OpNum == 0) { // move -> store
1086 unsigned SrcReg = MI->getOperand(1).getReg();
1087 unsigned SrcSubReg = MI->getOperand(1).getSubReg();
1088 bool isKill = MI->getOperand(1).isKill();
1089 bool isUndef = MI->getOperand(1).isUndef();
1090 if (MFI.getObjectAlignment(FI) >= 16 &&
1091 getRegisterInfo().canRealignStack(MF)) {
1092 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VST1q))
1093 .addFrameIndex(FI).addImm(16)
1095 getKillRegState(isKill) | getUndefRegState(isUndef),
1097 .addImm(Pred).addReg(PredReg);
1099 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VSTMQ))
1101 getKillRegState(isKill) | getUndefRegState(isUndef),
1103 .addFrameIndex(FI).addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))
1104 .addImm(Pred).addReg(PredReg);
1106 } else { // move -> load
1107 unsigned DstReg = MI->getOperand(0).getReg();
1108 unsigned DstSubReg = MI->getOperand(0).getSubReg();
1109 bool isDead = MI->getOperand(0).isDead();
1110 bool isUndef = MI->getOperand(0).isUndef();
1111 if (MFI.getObjectAlignment(FI) >= 16 &&
1112 getRegisterInfo().canRealignStack(MF)) {
1113 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VLD1q))
1116 getDeadRegState(isDead) |
1117 getUndefRegState(isUndef),
1119 .addFrameIndex(FI).addImm(16).addImm(Pred).addReg(PredReg);
1121 NewMI = BuildMI(MF, MI->getDebugLoc(), get(ARM::VLDMQ))
1124 getDeadRegState(isDead) |
1125 getUndefRegState(isUndef),
1127 .addFrameIndex(FI).addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))
1128 .addImm(Pred).addReg(PredReg);
1137 ARMBaseInstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
1139 const SmallVectorImpl<unsigned> &Ops,
1140 MachineInstr* LoadMI) const {
1146 ARMBaseInstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
1147 const SmallVectorImpl<unsigned> &Ops) const {
1148 if (Ops.size() != 1) return false;
1150 unsigned Opc = MI->getOpcode();
1151 if (Opc == ARM::MOVr || Opc == ARM::t2MOVr) {
1152 // If it is updating CPSR, then it cannot be folded.
1153 return MI->getOperand(4).getReg() != ARM::CPSR ||
1154 MI->getOperand(4).isDead();
1155 } else if (Opc == ARM::tMOVgpr2gpr ||
1156 Opc == ARM::tMOVtgpr2gpr ||
1157 Opc == ARM::tMOVgpr2tgpr) {
1159 } else if (Opc == ARM::VMOVS || Opc == ARM::VMOVD ||
1160 Opc == ARM::VMOVDneon || Opc == ARM::VMOVQ) {
1164 // FIXME: VMOVQQ and VMOVQQQQ?
1166 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
1169 /// Create a copy of a const pool value. Update CPI to the new index and return
1171 static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) {
1172 MachineConstantPool *MCP = MF.getConstantPool();
1173 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1175 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
1176 assert(MCPE.isMachineConstantPoolEntry() &&
1177 "Expecting a machine constantpool entry!");
1178 ARMConstantPoolValue *ACPV =
1179 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
1181 unsigned PCLabelId = AFI->createConstPoolEntryUId();
1182 ARMConstantPoolValue *NewCPV = 0;
1183 if (ACPV->isGlobalValue())
1184 NewCPV = new ARMConstantPoolValue(ACPV->getGV(), PCLabelId,
1186 else if (ACPV->isExtSymbol())
1187 NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(),
1188 ACPV->getSymbol(), PCLabelId, 4);
1189 else if (ACPV->isBlockAddress())
1190 NewCPV = new ARMConstantPoolValue(ACPV->getBlockAddress(), PCLabelId,
1191 ARMCP::CPBlockAddress, 4);
1193 llvm_unreachable("Unexpected ARM constantpool value type!!");
1194 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
1198 void ARMBaseInstrInfo::
1199 reMaterialize(MachineBasicBlock &MBB,
1200 MachineBasicBlock::iterator I,
1201 unsigned DestReg, unsigned SubIdx,
1202 const MachineInstr *Orig,
1203 const TargetRegisterInfo &TRI) const {
1204 unsigned Opcode = Orig->getOpcode();
1207 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1208 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1212 case ARM::tLDRpci_pic:
1213 case ARM::t2LDRpci_pic: {
1214 MachineFunction &MF = *MBB.getParent();
1215 unsigned CPI = Orig->getOperand(1).getIndex();
1216 unsigned PCLabelId = duplicateCPV(MF, CPI);
1217 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
1219 .addConstantPoolIndex(CPI).addImm(PCLabelId);
1220 (*MIB).setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
1227 ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const {
1228 MachineInstr *MI = TargetInstrInfoImpl::duplicate(Orig, MF);
1229 switch(Orig->getOpcode()) {
1230 case ARM::tLDRpci_pic:
1231 case ARM::t2LDRpci_pic: {
1232 unsigned CPI = Orig->getOperand(1).getIndex();
1233 unsigned PCLabelId = duplicateCPV(MF, CPI);
1234 Orig->getOperand(1).setIndex(CPI);
1235 Orig->getOperand(2).setImm(PCLabelId);
1242 bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
1243 const MachineInstr *MI1) const {
1244 int Opcode = MI0->getOpcode();
1245 if (Opcode == ARM::t2LDRpci ||
1246 Opcode == ARM::t2LDRpci_pic ||
1247 Opcode == ARM::tLDRpci ||
1248 Opcode == ARM::tLDRpci_pic) {
1249 if (MI1->getOpcode() != Opcode)
1251 if (MI0->getNumOperands() != MI1->getNumOperands())
1254 const MachineOperand &MO0 = MI0->getOperand(1);
1255 const MachineOperand &MO1 = MI1->getOperand(1);
1256 if (MO0.getOffset() != MO1.getOffset())
1259 const MachineFunction *MF = MI0->getParent()->getParent();
1260 const MachineConstantPool *MCP = MF->getConstantPool();
1261 int CPI0 = MO0.getIndex();
1262 int CPI1 = MO1.getIndex();
1263 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0];
1264 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1];
1265 ARMConstantPoolValue *ACPV0 =
1266 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal);
1267 ARMConstantPoolValue *ACPV1 =
1268 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal);
1269 return ACPV0->hasSameValue(ACPV1);
1272 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
1275 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
1276 /// determine if two loads are loading from the same base address. It should
1277 /// only return true if the base pointers are the same and the only differences
1278 /// between the two addresses is the offset. It also returns the offsets by
1280 bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1282 int64_t &Offset2) const {
1283 // Don't worry about Thumb: just ARM and Thumb2.
1284 if (Subtarget.isThumb1Only()) return false;
1286 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
1289 switch (Load1->getMachineOpcode()) {
1302 case ARM::t2LDRSHi8:
1304 case ARM::t2LDRSHi12:
1308 switch (Load2->getMachineOpcode()) {
1321 case ARM::t2LDRSHi8:
1323 case ARM::t2LDRSHi12:
1327 // Check if base addresses and chain operands match.
1328 if (Load1->getOperand(0) != Load2->getOperand(0) ||
1329 Load1->getOperand(4) != Load2->getOperand(4))
1332 // Index should be Reg0.
1333 if (Load1->getOperand(3) != Load2->getOperand(3))
1336 // Determine the offsets.
1337 if (isa<ConstantSDNode>(Load1->getOperand(1)) &&
1338 isa<ConstantSDNode>(Load2->getOperand(1))) {
1339 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue();
1340 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue();
1347 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
1348 /// determine (in conjuction with areLoadsFromSameBasePtr) if two loads should
1349 /// be scheduled togther. On some targets if two loads are loading from
1350 /// addresses in the same cache line, it's better if they are scheduled
1351 /// together. This function takes two integers that represent the load offsets
1352 /// from the common base address. It returns true if it decides it's desirable
1353 /// to schedule the two loads together. "NumLoads" is the number of loads that
1354 /// have already been scheduled after Load1.
1355 bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1356 int64_t Offset1, int64_t Offset2,
1357 unsigned NumLoads) const {
1358 // Don't worry about Thumb: just ARM and Thumb2.
1359 if (Subtarget.isThumb1Only()) return false;
1361 assert(Offset2 > Offset1);
1363 if ((Offset2 - Offset1) / 8 > 64)
1366 if (Load1->getMachineOpcode() != Load2->getMachineOpcode())
1367 return false; // FIXME: overly conservative?
1369 // Four loads in a row should be sufficient.
1376 bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
1377 const MachineBasicBlock *MBB,
1378 const MachineFunction &MF) const {
1379 // Debug info is never a scheduling boundary. It's necessary to be explicit
1380 // due to the special treatment of IT instructions below, otherwise a
1381 // dbg_value followed by an IT will result in the IT instruction being
1382 // considered a scheduling hazard, which is wrong. It should be the actual
1383 // instruction preceding the dbg_value instruction(s), just like it is
1384 // when debug info is not present.
1385 if (MI->isDebugValue())
1388 // Terminators and labels can't be scheduled around.
1389 if (MI->getDesc().isTerminator() || MI->isLabel())
1392 // Treat the start of the IT block as a scheduling boundary, but schedule
1393 // t2IT along with all instructions following it.
1394 // FIXME: This is a big hammer. But the alternative is to add all potential
1395 // true and anti dependencies to IT block instructions as implicit operands
1396 // to the t2IT instruction. The added compile time and complexity does not
1398 MachineBasicBlock::const_iterator I = MI;
1399 // Make sure to skip any dbg_value instructions
1400 while (++I != MBB->end() && I->isDebugValue())
1402 if (I != MBB->end() && I->getOpcode() == ARM::t2IT)
1405 // Don't attempt to schedule around any instruction that defines
1406 // a stack-oriented pointer, as it's unlikely to be profitable. This
1407 // saves compile time, because it doesn't require every single
1408 // stack slot reference to depend on the instruction that does the
1410 if (MI->definesRegister(ARM::SP))
1416 bool ARMBaseInstrInfo::
1417 isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumInstrs) const {
1420 if (Subtarget.getCPUString() == "generic")
1421 // Generic (and overly aggressive) if-conversion limits for testing.
1422 return NumInstrs <= 10;
1423 else if (Subtarget.hasV7Ops())
1424 return NumInstrs <= 3;
1425 return NumInstrs <= 2;
1428 bool ARMBaseInstrInfo::
1429 isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
1430 MachineBasicBlock &FMBB, unsigned NumF) const {
1431 return NumT && NumF && NumT <= 2 && NumF <= 2;
1434 /// getInstrPredicate - If instruction is predicated, returns its predicate
1435 /// condition, otherwise returns AL. It also returns the condition code
1436 /// register by reference.
1438 llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) {
1439 int PIdx = MI->findFirstPredOperandIdx();
1445 PredReg = MI->getOperand(PIdx+1).getReg();
1446 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm();
1450 int llvm::getMatchingCondBranchOpcode(int Opc) {
1453 else if (Opc == ARM::tB)
1455 else if (Opc == ARM::t2B)
1458 llvm_unreachable("Unknown unconditional branch opcode!");
1463 void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB,
1464 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
1465 unsigned DestReg, unsigned BaseReg, int NumBytes,
1466 ARMCC::CondCodes Pred, unsigned PredReg,
1467 const ARMBaseInstrInfo &TII) {
1468 bool isSub = NumBytes < 0;
1469 if (isSub) NumBytes = -NumBytes;
1472 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes);
1473 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt);
1474 assert(ThisVal && "Didn't extract field correctly");
1476 // We will handle these bits from offset, clear them.
1477 NumBytes &= ~ThisVal;
1479 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?");
1481 // Build the new ADD / SUB.
1482 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri;
1483 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
1484 .addReg(BaseReg, RegState::Kill).addImm(ThisVal)
1485 .addImm((unsigned)Pred).addReg(PredReg).addReg(0);
1490 bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1491 unsigned FrameReg, int &Offset,
1492 const ARMBaseInstrInfo &TII) {
1493 unsigned Opcode = MI.getOpcode();
1494 const TargetInstrDesc &Desc = MI.getDesc();
1495 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
1498 // Memory operands in inline assembly always use AddrMode2.
1499 if (Opcode == ARM::INLINEASM)
1500 AddrMode = ARMII::AddrMode2;
1502 if (Opcode == ARM::ADDri) {
1503 Offset += MI.getOperand(FrameRegIdx+1).getImm();
1505 // Turn it into a move.
1506 MI.setDesc(TII.get(ARM::MOVr));
1507 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1508 MI.RemoveOperand(FrameRegIdx+1);
1511 } else if (Offset < 0) {
1514 MI.setDesc(TII.get(ARM::SUBri));
1517 // Common case: small offset, fits into instruction.
1518 if (ARM_AM::getSOImmVal(Offset) != -1) {
1519 // Replace the FrameIndex with sp / fp
1520 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1521 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
1526 // Otherwise, pull as much of the immedidate into this ADDri/SUBri
1528 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset);
1529 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt);
1531 // We will handle these bits from offset, clear them.
1532 Offset &= ~ThisImmVal;
1534 // Get the properly encoded SOImmVal field.
1535 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 &&
1536 "Bit extraction didn't work?");
1537 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
1539 unsigned ImmIdx = 0;
1541 unsigned NumBits = 0;
1544 case ARMII::AddrMode2: {
1545 ImmIdx = FrameRegIdx+2;
1546 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm());
1547 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1552 case ARMII::AddrMode3: {
1553 ImmIdx = FrameRegIdx+2;
1554 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm());
1555 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1560 case ARMII::AddrMode4:
1561 case ARMII::AddrMode6:
1562 // Can't fold any offset even if it's zero.
1564 case ARMII::AddrMode5: {
1565 ImmIdx = FrameRegIdx+1;
1566 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm());
1567 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub)
1574 llvm_unreachable("Unsupported addressing mode!");
1578 Offset += InstrOffs * Scale;
1579 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
1585 // Attempt to fold address comp. if opcode has offset bits
1587 // Common case: small offset, fits into instruction.
1588 MachineOperand &ImmOp = MI.getOperand(ImmIdx);
1589 int ImmedOffset = Offset / Scale;
1590 unsigned Mask = (1 << NumBits) - 1;
1591 if ((unsigned)Offset <= Mask * Scale) {
1592 // Replace the FrameIndex with sp
1593 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
1595 ImmedOffset |= 1 << NumBits;
1596 ImmOp.ChangeToImmediate(ImmedOffset);
1601 // Otherwise, it didn't fit. Pull in what we can to simplify the immed.
1602 ImmedOffset = ImmedOffset & Mask;
1604 ImmedOffset |= 1 << NumBits;
1605 ImmOp.ChangeToImmediate(ImmedOffset);
1606 Offset &= ~(Mask*Scale);
1610 Offset = (isSub) ? -Offset : Offset;