1 //===- ARM64InstrInfo.cpp - ARM64 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 ARM64 implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "ARM64InstrInfo.h"
15 #include "ARM64Subtarget.h"
16 #include "MCTargetDesc/ARM64AddressingModes.h"
17 #include "llvm/CodeGen/MachineFrameInfo.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/MachineMemOperand.h"
20 #include "llvm/CodeGen/MachineRegisterInfo.h"
21 #include "llvm/CodeGen/PseudoSourceValue.h"
22 #include "llvm/MC/MCInst.h"
23 #include "llvm/Support/ErrorHandling.h"
24 #include "llvm/Support/TargetRegistry.h"
28 #define GET_INSTRINFO_CTOR_DTOR
29 #include "ARM64GenInstrInfo.inc"
31 ARM64InstrInfo::ARM64InstrInfo(const ARM64Subtarget &STI)
32 : ARM64GenInstrInfo(ARM64::ADJCALLSTACKDOWN, ARM64::ADJCALLSTACKUP),
33 RI(this, &STI), Subtarget(STI) {}
35 /// GetInstSize - Return the number of bytes of code the specified
36 /// instruction may be. This returns the maximum number of bytes.
37 unsigned ARM64InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
38 const MCInstrDesc &Desc = MI->getDesc();
40 switch (Desc.getOpcode()) {
42 // Anything not explicitly designated otherwise is a nomal 4-byte insn.
44 case TargetOpcode::DBG_VALUE:
45 case TargetOpcode::EH_LABEL:
46 case TargetOpcode::IMPLICIT_DEF:
47 case TargetOpcode::KILL:
51 llvm_unreachable("GetInstSizeInBytes()- Unable to determin insn size");
54 static void parseCondBranch(MachineInstr *LastInst, MachineBasicBlock *&Target,
55 SmallVectorImpl<MachineOperand> &Cond) {
56 // Block ends with fall-through condbranch.
57 switch (LastInst->getOpcode()) {
59 llvm_unreachable("Unknown branch instruction?");
61 Target = LastInst->getOperand(1).getMBB();
62 Cond.push_back(LastInst->getOperand(0));
68 Target = LastInst->getOperand(1).getMBB();
69 Cond.push_back(MachineOperand::CreateImm(-1));
70 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
71 Cond.push_back(LastInst->getOperand(0));
77 Target = LastInst->getOperand(2).getMBB();
78 Cond.push_back(MachineOperand::CreateImm(-1));
79 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
80 Cond.push_back(LastInst->getOperand(0));
81 Cond.push_back(LastInst->getOperand(1));
86 bool ARM64InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
87 MachineBasicBlock *&TBB,
88 MachineBasicBlock *&FBB,
89 SmallVectorImpl<MachineOperand> &Cond,
90 bool AllowModify) const {
91 // If the block has no terminators, it just falls into the block after it.
92 MachineBasicBlock::iterator I = MBB.end();
96 while (I->isDebugValue()) {
101 if (!isUnpredicatedTerminator(I))
104 // Get the last instruction in the block.
105 MachineInstr *LastInst = I;
107 // If there is only one terminator instruction, process it.
108 unsigned LastOpc = LastInst->getOpcode();
109 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
110 if (isUncondBranchOpcode(LastOpc)) {
111 TBB = LastInst->getOperand(0).getMBB();
114 if (isCondBranchOpcode(LastOpc)) {
115 // Block ends with fall-through condbranch.
116 parseCondBranch(LastInst, TBB, Cond);
119 return true; // Can't handle indirect branch.
122 // Get the instruction before it if it is a terminator.
123 MachineInstr *SecondLastInst = I;
124 unsigned SecondLastOpc = SecondLastInst->getOpcode();
126 // If AllowModify is true and the block ends with two or more unconditional
127 // branches, delete all but the first unconditional branch.
128 if (AllowModify && isUncondBranchOpcode(LastOpc)) {
129 while (isUncondBranchOpcode(SecondLastOpc)) {
130 LastInst->eraseFromParent();
131 LastInst = SecondLastInst;
132 LastOpc = LastInst->getOpcode();
133 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
134 // Return now the only terminator is an unconditional branch.
135 TBB = LastInst->getOperand(0).getMBB();
139 SecondLastOpc = SecondLastInst->getOpcode();
144 // If there are three terminators, we don't know what sort of block this is.
145 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I))
148 // If the block ends with a B and a Bcc, handle it.
149 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
150 parseCondBranch(SecondLastInst, TBB, Cond);
151 FBB = LastInst->getOperand(0).getMBB();
155 // If the block ends with two unconditional branches, handle it. The second
156 // one is not executed, so remove it.
157 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
158 TBB = SecondLastInst->getOperand(0).getMBB();
161 I->eraseFromParent();
165 // ...likewise if it ends with an indirect branch followed by an unconditional
167 if (isIndirectBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
170 I->eraseFromParent();
174 // Otherwise, can't handle this.
178 bool ARM64InstrInfo::ReverseBranchCondition(
179 SmallVectorImpl<MachineOperand> &Cond) const {
180 if (Cond[0].getImm() != -1) {
182 ARM64CC::CondCode CC = (ARM64CC::CondCode)(int)Cond[0].getImm();
183 Cond[0].setImm(ARM64CC::getInvertedCondCode(CC));
185 // Folded compare-and-branch
186 switch (Cond[1].getImm()) {
188 llvm_unreachable("Unknown conditional branch!");
190 Cond[1].setImm(ARM64::CBNZW);
193 Cond[1].setImm(ARM64::CBZW);
196 Cond[1].setImm(ARM64::CBNZX);
199 Cond[1].setImm(ARM64::CBZX);
202 Cond[1].setImm(ARM64::TBNZW);
205 Cond[1].setImm(ARM64::TBZW);
208 Cond[1].setImm(ARM64::TBNZX);
211 Cond[1].setImm(ARM64::TBZX);
219 unsigned ARM64InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
220 MachineBasicBlock::iterator I = MBB.end();
221 if (I == MBB.begin())
224 while (I->isDebugValue()) {
225 if (I == MBB.begin())
229 if (!isUncondBranchOpcode(I->getOpcode()) &&
230 !isCondBranchOpcode(I->getOpcode()))
233 // Remove the branch.
234 I->eraseFromParent();
238 if (I == MBB.begin())
241 if (!isCondBranchOpcode(I->getOpcode()))
244 // Remove the branch.
245 I->eraseFromParent();
249 void ARM64InstrInfo::instantiateCondBranch(
250 MachineBasicBlock &MBB, DebugLoc DL, MachineBasicBlock *TBB,
251 const SmallVectorImpl<MachineOperand> &Cond) const {
252 if (Cond[0].getImm() != -1) {
254 BuildMI(&MBB, DL, get(ARM64::Bcc)).addImm(Cond[0].getImm()).addMBB(TBB);
256 // Folded compare-and-branch
257 const MachineInstrBuilder MIB =
258 BuildMI(&MBB, DL, get(Cond[1].getImm())).addReg(Cond[2].getReg());
260 MIB.addImm(Cond[3].getImm());
265 unsigned ARM64InstrInfo::InsertBranch(
266 MachineBasicBlock &MBB, MachineBasicBlock *TBB, MachineBasicBlock *FBB,
267 const SmallVectorImpl<MachineOperand> &Cond, DebugLoc DL) const {
268 // Shouldn't be a fall through.
269 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
272 if (Cond.empty()) // Unconditional branch?
273 BuildMI(&MBB, DL, get(ARM64::B)).addMBB(TBB);
275 instantiateCondBranch(MBB, DL, TBB, Cond);
279 // Two-way conditional branch.
280 instantiateCondBranch(MBB, DL, TBB, Cond);
281 BuildMI(&MBB, DL, get(ARM64::B)).addMBB(FBB);
285 // Find the original register that VReg is copied from.
286 static unsigned removeCopies(const MachineRegisterInfo &MRI, unsigned VReg) {
287 while (TargetRegisterInfo::isVirtualRegister(VReg)) {
288 const MachineInstr *DefMI = MRI.getVRegDef(VReg);
289 if (!DefMI->isFullCopy())
291 VReg = DefMI->getOperand(1).getReg();
296 // Determine if VReg is defined by an instruction that can be folded into a
297 // csel instruction. If so, return the folded opcode, and the replacement
299 static unsigned canFoldIntoCSel(const MachineRegisterInfo &MRI, unsigned VReg,
300 unsigned *NewVReg = nullptr) {
301 VReg = removeCopies(MRI, VReg);
302 if (!TargetRegisterInfo::isVirtualRegister(VReg))
305 bool Is64Bit = ARM64::GPR64allRegClass.hasSubClassEq(MRI.getRegClass(VReg));
306 const MachineInstr *DefMI = MRI.getVRegDef(VReg);
308 unsigned SrcOpNum = 0;
309 switch (DefMI->getOpcode()) {
312 // if NZCV is used, do not fold.
313 if (DefMI->findRegisterDefOperandIdx(ARM64::NZCV, true) == -1)
315 // fall-through to ADDXri and ADDWri.
318 // add x, 1 -> csinc.
319 if (!DefMI->getOperand(2).isImm() || DefMI->getOperand(2).getImm() != 1 ||
320 DefMI->getOperand(3).getImm() != 0)
323 Opc = Is64Bit ? ARM64::CSINCXr : ARM64::CSINCWr;
327 case ARM64::ORNWrr: {
328 // not x -> csinv, represented as orn dst, xzr, src.
329 unsigned ZReg = removeCopies(MRI, DefMI->getOperand(1).getReg());
330 if (ZReg != ARM64::XZR && ZReg != ARM64::WZR)
333 Opc = Is64Bit ? ARM64::CSINVXr : ARM64::CSINVWr;
339 // if NZCV is used, do not fold.
340 if (DefMI->findRegisterDefOperandIdx(ARM64::NZCV, true) == -1)
342 // fall-through to SUBXrr and SUBWrr.
344 case ARM64::SUBWrr: {
345 // neg x -> csneg, represented as sub dst, xzr, src.
346 unsigned ZReg = removeCopies(MRI, DefMI->getOperand(1).getReg());
347 if (ZReg != ARM64::XZR && ZReg != ARM64::WZR)
350 Opc = Is64Bit ? ARM64::CSNEGXr : ARM64::CSNEGWr;
356 assert(Opc && SrcOpNum && "Missing parameters");
359 *NewVReg = DefMI->getOperand(SrcOpNum).getReg();
363 bool ARM64InstrInfo::canInsertSelect(
364 const MachineBasicBlock &MBB, const SmallVectorImpl<MachineOperand> &Cond,
365 unsigned TrueReg, unsigned FalseReg, int &CondCycles, int &TrueCycles,
366 int &FalseCycles) const {
367 // Check register classes.
368 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
369 const TargetRegisterClass *RC =
370 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
374 // Expanding cbz/tbz requires an extra cycle of latency on the condition.
375 unsigned ExtraCondLat = Cond.size() != 1;
377 // GPRs are handled by csel.
378 // FIXME: Fold in x+1, -x, and ~x when applicable.
379 if (ARM64::GPR64allRegClass.hasSubClassEq(RC) ||
380 ARM64::GPR32allRegClass.hasSubClassEq(RC)) {
381 // Single-cycle csel, csinc, csinv, and csneg.
382 CondCycles = 1 + ExtraCondLat;
383 TrueCycles = FalseCycles = 1;
384 if (canFoldIntoCSel(MRI, TrueReg))
386 else if (canFoldIntoCSel(MRI, FalseReg))
391 // Scalar floating point is handled by fcsel.
392 // FIXME: Form fabs, fmin, and fmax when applicable.
393 if (ARM64::FPR64RegClass.hasSubClassEq(RC) ||
394 ARM64::FPR32RegClass.hasSubClassEq(RC)) {
395 CondCycles = 5 + ExtraCondLat;
396 TrueCycles = FalseCycles = 2;
404 void ARM64InstrInfo::insertSelect(MachineBasicBlock &MBB,
405 MachineBasicBlock::iterator I, DebugLoc DL,
407 const SmallVectorImpl<MachineOperand> &Cond,
408 unsigned TrueReg, unsigned FalseReg) const {
409 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
411 // Parse the condition code, see parseCondBranch() above.
412 ARM64CC::CondCode CC;
413 switch (Cond.size()) {
415 llvm_unreachable("Unknown condition opcode in Cond");
417 CC = ARM64CC::CondCode(Cond[0].getImm());
419 case 3: { // cbz/cbnz
420 // We must insert a compare against 0.
422 switch (Cond[1].getImm()) {
424 llvm_unreachable("Unknown branch opcode in Cond");
442 unsigned SrcReg = Cond[2].getReg();
444 // cmp reg, #0 is actually subs xzr, reg, #0.
445 MRI.constrainRegClass(SrcReg, &ARM64::GPR64spRegClass);
446 BuildMI(MBB, I, DL, get(ARM64::SUBSXri), ARM64::XZR)
451 MRI.constrainRegClass(SrcReg, &ARM64::GPR32spRegClass);
452 BuildMI(MBB, I, DL, get(ARM64::SUBSWri), ARM64::WZR)
459 case 4: { // tbz/tbnz
460 // We must insert a tst instruction.
461 switch (Cond[1].getImm()) {
463 llvm_unreachable("Unknown branch opcode in Cond");
473 // cmp reg, #foo is actually ands xzr, reg, #1<<foo.
474 if (Cond[1].getImm() == ARM64::TBZW || Cond[1].getImm() == ARM64::TBNZW)
475 BuildMI(MBB, I, DL, get(ARM64::ANDSWri), ARM64::WZR)
476 .addReg(Cond[2].getReg())
477 .addImm(ARM64_AM::encodeLogicalImmediate(1ull << Cond[3].getImm(), 32));
479 BuildMI(MBB, I, DL, get(ARM64::ANDSXri), ARM64::XZR)
480 .addReg(Cond[2].getReg())
481 .addImm(ARM64_AM::encodeLogicalImmediate(1ull << Cond[3].getImm(), 64));
487 const TargetRegisterClass *RC = nullptr;
488 bool TryFold = false;
489 if (MRI.constrainRegClass(DstReg, &ARM64::GPR64RegClass)) {
490 RC = &ARM64::GPR64RegClass;
493 } else if (MRI.constrainRegClass(DstReg, &ARM64::GPR32RegClass)) {
494 RC = &ARM64::GPR32RegClass;
497 } else if (MRI.constrainRegClass(DstReg, &ARM64::FPR64RegClass)) {
498 RC = &ARM64::FPR64RegClass;
499 Opc = ARM64::FCSELDrrr;
500 } else if (MRI.constrainRegClass(DstReg, &ARM64::FPR32RegClass)) {
501 RC = &ARM64::FPR32RegClass;
502 Opc = ARM64::FCSELSrrr;
504 assert(RC && "Unsupported regclass");
506 // Try folding simple instructions into the csel.
508 unsigned NewVReg = 0;
509 unsigned FoldedOpc = canFoldIntoCSel(MRI, TrueReg, &NewVReg);
511 // The folded opcodes csinc, csinc and csneg apply the operation to
512 // FalseReg, so we need to invert the condition.
513 CC = ARM64CC::getInvertedCondCode(CC);
516 FoldedOpc = canFoldIntoCSel(MRI, FalseReg, &NewVReg);
518 // Fold the operation. Leave any dead instructions for DCE to clean up.
522 // The extends the live range of NewVReg.
523 MRI.clearKillFlags(NewVReg);
527 // Pull all virtual register into the appropriate class.
528 MRI.constrainRegClass(TrueReg, RC);
529 MRI.constrainRegClass(FalseReg, RC);
532 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(TrueReg).addReg(FalseReg).addImm(
536 bool ARM64InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
537 unsigned &SrcReg, unsigned &DstReg,
538 unsigned &SubIdx) const {
539 switch (MI.getOpcode()) {
542 case ARM64::SBFMXri: // aka sxtw
543 case ARM64::UBFMXri: // aka uxtw
544 // Check for the 32 -> 64 bit extension case, these instructions can do
546 if (MI.getOperand(2).getImm() != 0 || MI.getOperand(3).getImm() != 31)
548 // This is a signed or unsigned 32 -> 64 bit extension.
549 SrcReg = MI.getOperand(1).getReg();
550 DstReg = MI.getOperand(0).getReg();
551 SubIdx = ARM64::sub_32;
556 /// analyzeCompare - For a comparison instruction, return the source registers
557 /// in SrcReg and SrcReg2, and the value it compares against in CmpValue.
558 /// Return true if the comparison instruction can be analyzed.
559 bool ARM64InstrInfo::analyzeCompare(const MachineInstr *MI, unsigned &SrcReg,
560 unsigned &SrcReg2, int &CmpMask,
561 int &CmpValue) const {
562 switch (MI->getOpcode()) {
577 // Replace SUBSWrr with SUBWrr if NZCV is not used.
578 SrcReg = MI->getOperand(1).getReg();
579 SrcReg2 = MI->getOperand(2).getReg();
587 SrcReg = MI->getOperand(1).getReg();
590 CmpValue = MI->getOperand(2).getImm();
594 // ANDS does not use the same encoding scheme as the others xxxS
596 SrcReg = MI->getOperand(1).getReg();
599 CmpValue = ARM64_AM::decodeLogicalImmediate(
600 MI->getOperand(2).getImm(),
601 MI->getOpcode() == ARM64::ANDSWri ? 32 : 64);
608 static bool UpdateOperandRegClass(MachineInstr *Instr) {
609 MachineBasicBlock *MBB = Instr->getParent();
610 assert(MBB && "Can't get MachineBasicBlock here");
611 MachineFunction *MF = MBB->getParent();
612 assert(MF && "Can't get MachineFunction here");
613 const TargetMachine *TM = &MF->getTarget();
614 const TargetInstrInfo *TII = TM->getInstrInfo();
615 const TargetRegisterInfo *TRI = TM->getRegisterInfo();
616 MachineRegisterInfo *MRI = &MF->getRegInfo();
618 for (unsigned OpIdx = 0, EndIdx = Instr->getNumOperands(); OpIdx < EndIdx;
620 MachineOperand &MO = Instr->getOperand(OpIdx);
621 const TargetRegisterClass *OpRegCstraints =
622 Instr->getRegClassConstraint(OpIdx, TII, TRI);
624 // If there's no constraint, there's nothing to do.
627 // If the operand is a frame index, there's nothing to do here.
628 // A frame index operand will resolve correctly during PEI.
633 "Operand has register constraints without being a register!");
635 unsigned Reg = MO.getReg();
636 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
637 if (!OpRegCstraints->contains(Reg))
639 } else if (!OpRegCstraints->hasSubClassEq(MRI->getRegClass(Reg)) &&
640 !MRI->constrainRegClass(Reg, OpRegCstraints))
647 /// optimizeCompareInstr - Convert the instruction supplying the argument to the
648 /// comparison into one that sets the zero bit in the flags register.
649 bool ARM64InstrInfo::optimizeCompareInstr(
650 MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2, int CmpMask,
651 int CmpValue, const MachineRegisterInfo *MRI) const {
653 // Replace SUBSWrr with SUBWrr if NZCV is not used.
654 int Cmp_NZCV = CmpInstr->findRegisterDefOperandIdx(ARM64::NZCV, true);
655 if (Cmp_NZCV != -1) {
657 switch (CmpInstr->getOpcode()) {
660 case ARM64::ADDSWrr: NewOpc = ARM64::ADDWrr; break;
661 case ARM64::ADDSWri: NewOpc = ARM64::ADDWri; break;
662 case ARM64::ADDSWrs: NewOpc = ARM64::ADDWrs; break;
663 case ARM64::ADDSWrx: NewOpc = ARM64::ADDWrx; break;
664 case ARM64::ADDSXrr: NewOpc = ARM64::ADDXrr; break;
665 case ARM64::ADDSXri: NewOpc = ARM64::ADDXri; break;
666 case ARM64::ADDSXrs: NewOpc = ARM64::ADDXrs; break;
667 case ARM64::ADDSXrx: NewOpc = ARM64::ADDXrx; break;
668 case ARM64::SUBSWrr: NewOpc = ARM64::SUBWrr; break;
669 case ARM64::SUBSWri: NewOpc = ARM64::SUBWri; break;
670 case ARM64::SUBSWrs: NewOpc = ARM64::SUBWrs; break;
671 case ARM64::SUBSWrx: NewOpc = ARM64::SUBWrx; break;
672 case ARM64::SUBSXrr: NewOpc = ARM64::SUBXrr; break;
673 case ARM64::SUBSXri: NewOpc = ARM64::SUBXri; break;
674 case ARM64::SUBSXrs: NewOpc = ARM64::SUBXrs; break;
675 case ARM64::SUBSXrx: NewOpc = ARM64::SUBXrx; break;
678 const MCInstrDesc &MCID = get(NewOpc);
679 CmpInstr->setDesc(MCID);
680 CmpInstr->RemoveOperand(Cmp_NZCV);
681 bool succeeded = UpdateOperandRegClass(CmpInstr);
683 assert(succeeded && "Some operands reg class are incompatible!");
687 // Continue only if we have a "ri" where immediate is zero.
688 if (CmpValue != 0 || SrcReg2 != 0)
691 // CmpInstr is a Compare instruction if destination register is not used.
692 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
695 // Get the unique definition of SrcReg.
696 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
700 // We iterate backward, starting from the instruction before CmpInstr and
701 // stop when reaching the definition of the source register or done with the
702 // basic block, to check whether NZCV is used or modified in between.
703 MachineBasicBlock::iterator I = CmpInstr, E = MI,
704 B = CmpInstr->getParent()->begin();
706 // Early exit if CmpInstr is at the beginning of the BB.
710 // Check whether the definition of SrcReg is in the same basic block as
711 // Compare. If not, we can't optimize away the Compare.
712 if (MI->getParent() != CmpInstr->getParent())
715 // Check that NZCV isn't set between the comparison instruction and the one we
717 const TargetRegisterInfo *TRI = &getRegisterInfo();
718 for (--I; I != E; --I) {
719 const MachineInstr &Instr = *I;
721 if (Instr.modifiesRegister(ARM64::NZCV, TRI) ||
722 Instr.readsRegister(ARM64::NZCV, TRI))
723 // This instruction modifies or uses NZCV after the one we want to
724 // change. We can't do this transformation.
727 // The 'and' is below the comparison instruction.
731 unsigned NewOpc = MI->getOpcode();
732 switch (MI->getOpcode()) {
744 case ARM64::ADDWrr: NewOpc = ARM64::ADDSWrr; break;
745 case ARM64::ADDWri: NewOpc = ARM64::ADDSWri; break;
746 case ARM64::ADDXrr: NewOpc = ARM64::ADDSXrr; break;
747 case ARM64::ADDXri: NewOpc = ARM64::ADDSXri; break;
748 case ARM64::ADCWr: NewOpc = ARM64::ADCSWr; break;
749 case ARM64::ADCXr: NewOpc = ARM64::ADCSXr; break;
750 case ARM64::SUBWrr: NewOpc = ARM64::SUBSWrr; break;
751 case ARM64::SUBWri: NewOpc = ARM64::SUBSWri; break;
752 case ARM64::SUBXrr: NewOpc = ARM64::SUBSXrr; break;
753 case ARM64::SUBXri: NewOpc = ARM64::SUBSXri; break;
754 case ARM64::SBCWr: NewOpc = ARM64::SBCSWr; break;
755 case ARM64::SBCXr: NewOpc = ARM64::SBCSXr; break;
756 case ARM64::ANDWri: NewOpc = ARM64::ANDSWri; break;
757 case ARM64::ANDXri: NewOpc = ARM64::ANDSXri; break;
760 // Scan forward for the use of NZCV.
761 // When checking against MI: if it's a conditional code requires
762 // checking of V bit, then this is not safe to do.
763 // It is safe to remove CmpInstr if NZCV is redefined or killed.
764 // If we are done with the basic block, we need to check whether NZCV is
767 for (MachineBasicBlock::iterator I = CmpInstr,
768 E = CmpInstr->getParent()->end();
769 !IsSafe && ++I != E;) {
770 const MachineInstr &Instr = *I;
771 for (unsigned IO = 0, EO = Instr.getNumOperands(); !IsSafe && IO != EO;
773 const MachineOperand &MO = Instr.getOperand(IO);
774 if (MO.isRegMask() && MO.clobbersPhysReg(ARM64::NZCV)) {
778 if (!MO.isReg() || MO.getReg() != ARM64::NZCV)
785 // Decode the condition code.
786 unsigned Opc = Instr.getOpcode();
787 ARM64CC::CondCode CC;
792 CC = (ARM64CC::CondCode)Instr.getOperand(IO - 2).getImm();
802 case ARM64::FCSELSrrr:
803 case ARM64::FCSELDrrr:
804 CC = (ARM64CC::CondCode)Instr.getOperand(IO - 1).getImm();
808 // It is not safe to remove Compare instruction if Overflow(V) is used.
811 // NZCV can be used multiple times, we should continue.
824 // If NZCV is not killed nor re-defined, we should check whether it is
825 // live-out. If it is live-out, do not optimize.
827 MachineBasicBlock *ParentBlock = CmpInstr->getParent();
828 for (auto *MBB : ParentBlock->successors())
829 if (MBB->isLiveIn(ARM64::NZCV))
833 // Update the instruction to set NZCV.
834 MI->setDesc(get(NewOpc));
835 CmpInstr->eraseFromParent();
836 bool succeeded = UpdateOperandRegClass(MI);
838 assert(succeeded && "Some operands reg class are incompatible!");
839 MI->addRegisterDefined(ARM64::NZCV, TRI);
843 /// Return true if this is this instruction has a non-zero immediate
844 bool ARM64InstrInfo::hasShiftedReg(const MachineInstr *MI) const {
845 switch (MI->getOpcode()) {
860 case ARM64::CRC32Brr:
861 case ARM64::CRC32CBrr:
862 case ARM64::CRC32CHrr:
863 case ARM64::CRC32CWrr:
864 case ARM64::CRC32CXrr:
865 case ARM64::CRC32Hrr:
866 case ARM64::CRC32Wrr:
867 case ARM64::CRC32Xrr:
880 if (MI->getOperand(3).isImm()) {
881 unsigned val = MI->getOperand(3).getImm();
889 /// Return true if this is this instruction has a non-zero immediate
890 bool ARM64InstrInfo::hasExtendedReg(const MachineInstr *MI) const {
891 switch (MI->getOpcode()) {
896 case ARM64::ADDSXrx64:
899 case ARM64::ADDXrx64:
902 case ARM64::SUBSXrx64:
905 case ARM64::SUBXrx64:
906 if (MI->getOperand(3).isImm()) {
907 unsigned val = MI->getOperand(3).getImm();
916 // Return true if this instruction simply sets its single destination register
917 // to zero. This is equivalent to a register rename of the zero-register.
918 bool ARM64InstrInfo::isGPRZero(const MachineInstr *MI) const {
919 switch (MI->getOpcode()) {
923 case ARM64::MOVZXi: // movz Rd, #0 (LSL #0)
924 if (MI->getOperand(1).isImm() && MI->getOperand(1).getImm() == 0) {
925 assert(MI->getDesc().getNumOperands() == 3 &&
926 MI->getOperand(2).getImm() == 0 && "invalid MOVZi operands");
930 case ARM64::ANDWri: // and Rd, Rzr, #imm
931 return MI->getOperand(1).getReg() == ARM64::WZR;
933 return MI->getOperand(1).getReg() == ARM64::XZR;
934 case TargetOpcode::COPY:
935 return MI->getOperand(1).getReg() == ARM64::WZR;
940 // Return true if this instruction simply renames a general register without
942 bool ARM64InstrInfo::isGPRCopy(const MachineInstr *MI) const {
943 switch (MI->getOpcode()) {
946 case TargetOpcode::COPY: {
947 // GPR32 copies will by lowered to ORRXrs
948 unsigned DstReg = MI->getOperand(0).getReg();
949 return (ARM64::GPR32RegClass.contains(DstReg) ||
950 ARM64::GPR64RegClass.contains(DstReg));
952 case ARM64::ORRXrs: // orr Xd, Xzr, Xm (LSL #0)
953 if (MI->getOperand(1).getReg() == ARM64::XZR) {
954 assert(MI->getDesc().getNumOperands() == 4 &&
955 MI->getOperand(3).getImm() == 0 && "invalid ORRrs operands");
958 case ARM64::ADDXri: // add Xd, Xn, #0 (LSL #0)
959 if (MI->getOperand(2).getImm() == 0) {
960 assert(MI->getDesc().getNumOperands() == 4 &&
961 MI->getOperand(3).getImm() == 0 && "invalid ADDXri operands");
968 // Return true if this instruction simply renames a general register without
970 bool ARM64InstrInfo::isFPRCopy(const MachineInstr *MI) const {
971 switch (MI->getOpcode()) {
974 case TargetOpcode::COPY: {
975 // FPR64 copies will by lowered to ORR.16b
976 unsigned DstReg = MI->getOperand(0).getReg();
977 return (ARM64::FPR64RegClass.contains(DstReg) ||
978 ARM64::FPR128RegClass.contains(DstReg));
980 case ARM64::ORRv16i8:
981 if (MI->getOperand(1).getReg() == MI->getOperand(2).getReg()) {
982 assert(MI->getDesc().getNumOperands() == 3 && MI->getOperand(0).isReg() &&
983 "invalid ORRv16i8 operands");
990 unsigned ARM64InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
991 int &FrameIndex) const {
992 switch (MI->getOpcode()) {
1002 if (MI->getOperand(0).getSubReg() == 0 && MI->getOperand(1).isFI() &&
1003 MI->getOperand(2).isImm() && MI->getOperand(2).getImm() == 0) {
1004 FrameIndex = MI->getOperand(1).getIndex();
1005 return MI->getOperand(0).getReg();
1013 unsigned ARM64InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1014 int &FrameIndex) const {
1015 switch (MI->getOpcode()) {
1025 if (MI->getOperand(0).getSubReg() == 0 && MI->getOperand(1).isFI() &&
1026 MI->getOperand(2).isImm() && MI->getOperand(2).getImm() == 0) {
1027 FrameIndex = MI->getOperand(1).getIndex();
1028 return MI->getOperand(0).getReg();
1035 /// Return true if this is load/store scales or extends its register offset.
1036 /// This refers to scaling a dynamic index as opposed to scaled immediates.
1037 /// MI should be a memory op that allows scaled addressing.
1038 bool ARM64InstrInfo::isScaledAddr(const MachineInstr *MI) const {
1039 switch (MI->getOpcode()) {
1042 case ARM64::LDRBBro:
1045 case ARM64::LDRHHro:
1048 case ARM64::LDRSBWro:
1049 case ARM64::LDRSBXro:
1050 case ARM64::LDRSHWro:
1051 case ARM64::LDRSHXro:
1052 case ARM64::LDRSWro:
1056 case ARM64::STRBBro:
1059 case ARM64::STRHHro:
1065 unsigned Val = MI->getOperand(3).getImm();
1066 ARM64_AM::ShiftExtendType ExtType = ARM64_AM::getMemExtendType(Val);
1067 return (ExtType != ARM64_AM::UXTX) || ARM64_AM::getMemDoShift(Val);
1072 /// Check all MachineMemOperands for a hint to suppress pairing.
1073 bool ARM64InstrInfo::isLdStPairSuppressed(const MachineInstr *MI) const {
1074 assert(MOSuppressPair < (1 << MachineMemOperand::MOTargetNumBits) &&
1075 "Too many target MO flags");
1076 for (auto *MM : MI->memoperands()) {
1077 if (MM->getFlags() &
1078 (MOSuppressPair << MachineMemOperand::MOTargetStartBit)) {
1085 /// Set a flag on the first MachineMemOperand to suppress pairing.
1086 void ARM64InstrInfo::suppressLdStPair(MachineInstr *MI) const {
1087 if (MI->memoperands_empty())
1090 assert(MOSuppressPair < (1 << MachineMemOperand::MOTargetNumBits) &&
1091 "Too many target MO flags");
1092 (*MI->memoperands_begin())
1093 ->setFlags(MOSuppressPair << MachineMemOperand::MOTargetStartBit);
1096 bool ARM64InstrInfo::getLdStBaseRegImmOfs(MachineInstr *LdSt, unsigned &BaseReg,
1098 const TargetRegisterInfo *TRI) const {
1099 switch (LdSt->getOpcode()) {
1112 if (!LdSt->getOperand(1).isReg() || !LdSt->getOperand(2).isImm())
1114 BaseReg = LdSt->getOperand(1).getReg();
1115 MachineFunction &MF = *LdSt->getParent()->getParent();
1116 unsigned Width = getRegClass(LdSt->getDesc(), 0, TRI, MF)->getSize();
1117 Offset = LdSt->getOperand(2).getImm() * Width;
1122 /// Detect opportunities for ldp/stp formation.
1124 /// Only called for LdSt for which getLdStBaseRegImmOfs returns true.
1125 bool ARM64InstrInfo::shouldClusterLoads(MachineInstr *FirstLdSt,
1126 MachineInstr *SecondLdSt,
1127 unsigned NumLoads) const {
1128 // Only cluster up to a single pair.
1131 if (FirstLdSt->getOpcode() != SecondLdSt->getOpcode())
1133 // getLdStBaseRegImmOfs guarantees that oper 2 isImm.
1134 unsigned Ofs1 = FirstLdSt->getOperand(2).getImm();
1135 // Allow 6 bits of positive range.
1138 // The caller should already have ordered First/SecondLdSt by offset.
1139 unsigned Ofs2 = SecondLdSt->getOperand(2).getImm();
1140 return Ofs1 + 1 == Ofs2;
1143 bool ARM64InstrInfo::shouldScheduleAdjacent(MachineInstr *First,
1144 MachineInstr *Second) const {
1145 // Cyclone can fuse CMN, CMP followed by Bcc.
1147 // FIXME: B0 can also fuse:
1148 // AND, BIC, ORN, ORR, or EOR (optional S) followed by Bcc or CBZ or CBNZ.
1149 if (Second->getOpcode() != ARM64::Bcc)
1151 switch (First->getOpcode()) {
1154 case ARM64::SUBSWri:
1155 case ARM64::ADDSWri:
1156 case ARM64::ANDSWri:
1157 case ARM64::SUBSXri:
1158 case ARM64::ADDSXri:
1159 case ARM64::ANDSXri:
1164 MachineInstr *ARM64InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
1167 const MDNode *MDPtr,
1168 DebugLoc DL) const {
1169 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM64::DBG_VALUE))
1170 .addFrameIndex(FrameIx)
1173 .addMetadata(MDPtr);
1177 static const MachineInstrBuilder &AddSubReg(const MachineInstrBuilder &MIB,
1178 unsigned Reg, unsigned SubIdx,
1180 const TargetRegisterInfo *TRI) {
1182 return MIB.addReg(Reg, State);
1184 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1185 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State);
1186 return MIB.addReg(Reg, State, SubIdx);
1189 static bool forwardCopyWillClobberTuple(unsigned DestReg, unsigned SrcReg,
1191 // We really want the positive remainder mod 32 here, that happens to be
1192 // easily obtainable with a mask.
1193 return ((DestReg - SrcReg) & 0x1f) < NumRegs;
1196 void ARM64InstrInfo::copyPhysRegTuple(MachineBasicBlock &MBB,
1197 MachineBasicBlock::iterator I,
1198 DebugLoc DL, unsigned DestReg,
1199 unsigned SrcReg, bool KillSrc,
1201 llvm::ArrayRef<unsigned> Indices) const {
1202 assert(getSubTarget().hasNEON() &&
1203 "Unexpected register copy without NEON");
1204 const TargetRegisterInfo *TRI = &getRegisterInfo();
1205 uint16_t DestEncoding = TRI->getEncodingValue(DestReg);
1206 uint16_t SrcEncoding = TRI->getEncodingValue(SrcReg);
1207 unsigned NumRegs = Indices.size();
1209 int SubReg = 0, End = NumRegs, Incr = 1;
1210 if (forwardCopyWillClobberTuple(DestEncoding, SrcEncoding, NumRegs)) {
1211 SubReg = NumRegs - 1;
1216 for (; SubReg != End; SubReg += Incr) {
1217 const MachineInstrBuilder &MIB = BuildMI(MBB, I, DL, get(Opcode));
1218 AddSubReg(MIB, DestReg, Indices[SubReg], RegState::Define, TRI);
1219 AddSubReg(MIB, SrcReg, Indices[SubReg], 0, TRI);
1220 AddSubReg(MIB, SrcReg, Indices[SubReg], getKillRegState(KillSrc), TRI);
1224 void ARM64InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
1225 MachineBasicBlock::iterator I, DebugLoc DL,
1226 unsigned DestReg, unsigned SrcReg,
1227 bool KillSrc) const {
1228 if (ARM64::GPR32spRegClass.contains(DestReg) &&
1229 (ARM64::GPR32spRegClass.contains(SrcReg) || SrcReg == ARM64::WZR)) {
1230 const TargetRegisterInfo *TRI = &getRegisterInfo();
1232 if (DestReg == ARM64::WSP || SrcReg == ARM64::WSP) {
1233 // If either operand is WSP, expand to ADD #0.
1234 if (Subtarget.hasZeroCycleRegMove()) {
1235 // Cyclone recognizes "ADD Xd, Xn, #0" as a zero-cycle register move.
1236 unsigned DestRegX = TRI->getMatchingSuperReg(DestReg, ARM64::sub_32,
1237 &ARM64::GPR64spRegClass);
1238 unsigned SrcRegX = TRI->getMatchingSuperReg(SrcReg, ARM64::sub_32,
1239 &ARM64::GPR64spRegClass);
1240 // This instruction is reading and writing X registers. This may upset
1241 // the register scavenger and machine verifier, so we need to indicate
1242 // that we are reading an undefined value from SrcRegX, but a proper
1243 // value from SrcReg.
1244 BuildMI(MBB, I, DL, get(ARM64::ADDXri), DestRegX)
1245 .addReg(SrcRegX, RegState::Undef)
1247 .addImm(ARM64_AM::getShifterImm(ARM64_AM::LSL, 0))
1248 .addReg(SrcReg, RegState::Implicit | getKillRegState(KillSrc));
1250 BuildMI(MBB, I, DL, get(ARM64::ADDWri), DestReg)
1251 .addReg(SrcReg, getKillRegState(KillSrc))
1253 .addImm(ARM64_AM::getShifterImm(ARM64_AM::LSL, 0));
1255 } else if (SrcReg == ARM64::WZR && Subtarget.hasZeroCycleZeroing()) {
1256 BuildMI(MBB, I, DL, get(ARM64::MOVZWi), DestReg).addImm(0).addImm(
1257 ARM64_AM::getShifterImm(ARM64_AM::LSL, 0));
1259 if (Subtarget.hasZeroCycleRegMove()) {
1260 // Cyclone recognizes "ORR Xd, XZR, Xm" as a zero-cycle register move.
1261 unsigned DestRegX = TRI->getMatchingSuperReg(DestReg, ARM64::sub_32,
1262 &ARM64::GPR64spRegClass);
1263 unsigned SrcRegX = TRI->getMatchingSuperReg(SrcReg, ARM64::sub_32,
1264 &ARM64::GPR64spRegClass);
1265 // This instruction is reading and writing X registers. This may upset
1266 // the register scavenger and machine verifier, so we need to indicate
1267 // that we are reading an undefined value from SrcRegX, but a proper
1268 // value from SrcReg.
1269 BuildMI(MBB, I, DL, get(ARM64::ORRXrr), DestRegX)
1271 .addReg(SrcRegX, RegState::Undef)
1272 .addReg(SrcReg, RegState::Implicit | getKillRegState(KillSrc));
1274 // Otherwise, expand to ORR WZR.
1275 BuildMI(MBB, I, DL, get(ARM64::ORRWrr), DestReg)
1277 .addReg(SrcReg, getKillRegState(KillSrc));
1283 if (ARM64::GPR64spRegClass.contains(DestReg) &&
1284 (ARM64::GPR64spRegClass.contains(SrcReg) || SrcReg == ARM64::XZR)) {
1285 if (DestReg == ARM64::SP || SrcReg == ARM64::SP) {
1286 // If either operand is SP, expand to ADD #0.
1287 BuildMI(MBB, I, DL, get(ARM64::ADDXri), DestReg)
1288 .addReg(SrcReg, getKillRegState(KillSrc))
1290 .addImm(ARM64_AM::getShifterImm(ARM64_AM::LSL, 0));
1291 } else if (SrcReg == ARM64::XZR && Subtarget.hasZeroCycleZeroing()) {
1292 BuildMI(MBB, I, DL, get(ARM64::MOVZXi), DestReg).addImm(0).addImm(
1293 ARM64_AM::getShifterImm(ARM64_AM::LSL, 0));
1295 // Otherwise, expand to ORR XZR.
1296 BuildMI(MBB, I, DL, get(ARM64::ORRXrr), DestReg)
1298 .addReg(SrcReg, getKillRegState(KillSrc));
1303 // Copy a DDDD register quad by copying the individual sub-registers.
1304 if (ARM64::DDDDRegClass.contains(DestReg) &&
1305 ARM64::DDDDRegClass.contains(SrcReg)) {
1306 static const unsigned Indices[] = { ARM64::dsub0, ARM64::dsub1,
1307 ARM64::dsub2, ARM64::dsub3 };
1308 copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, ARM64::ORRv8i8,
1313 // Copy a DDD register triple by copying the individual sub-registers.
1314 if (ARM64::DDDRegClass.contains(DestReg) &&
1315 ARM64::DDDRegClass.contains(SrcReg)) {
1316 static const unsigned Indices[] = { ARM64::dsub0, ARM64::dsub1,
1318 copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, ARM64::ORRv8i8,
1323 // Copy a DD register pair by copying the individual sub-registers.
1324 if (ARM64::DDRegClass.contains(DestReg) &&
1325 ARM64::DDRegClass.contains(SrcReg)) {
1326 static const unsigned Indices[] = { ARM64::dsub0, ARM64::dsub1 };
1327 copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, ARM64::ORRv8i8,
1332 // Copy a QQQQ register quad by copying the individual sub-registers.
1333 if (ARM64::QQQQRegClass.contains(DestReg) &&
1334 ARM64::QQQQRegClass.contains(SrcReg)) {
1335 static const unsigned Indices[] = { ARM64::qsub0, ARM64::qsub1,
1336 ARM64::qsub2, ARM64::qsub3 };
1337 copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, ARM64::ORRv16i8,
1342 // Copy a QQQ register triple by copying the individual sub-registers.
1343 if (ARM64::QQQRegClass.contains(DestReg) &&
1344 ARM64::QQQRegClass.contains(SrcReg)) {
1345 static const unsigned Indices[] = { ARM64::qsub0, ARM64::qsub1,
1347 copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, ARM64::ORRv16i8,
1352 // Copy a QQ register pair by copying the individual sub-registers.
1353 if (ARM64::QQRegClass.contains(DestReg) &&
1354 ARM64::QQRegClass.contains(SrcReg)) {
1355 static const unsigned Indices[] = { ARM64::qsub0, ARM64::qsub1 };
1356 copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, ARM64::ORRv16i8,
1361 if (ARM64::FPR128RegClass.contains(DestReg) &&
1362 ARM64::FPR128RegClass.contains(SrcReg)) {
1363 if(getSubTarget().hasNEON()) {
1364 BuildMI(MBB, I, DL, get(ARM64::ORRv16i8), DestReg).addReg(SrcReg).addReg(
1365 SrcReg, getKillRegState(KillSrc));
1367 BuildMI(MBB, I, DL, get(ARM64::STRQpre))
1368 .addReg(SrcReg, getKillRegState(KillSrc))
1371 BuildMI(MBB, I, DL, get(ARM64::LDRQpre))
1372 .addReg(DestReg, RegState::Define)
1379 if (ARM64::FPR64RegClass.contains(DestReg) &&
1380 ARM64::FPR64RegClass.contains(SrcReg)) {
1381 if(getSubTarget().hasNEON()) {
1383 RI.getMatchingSuperReg(DestReg, ARM64::dsub, &ARM64::FPR128RegClass);
1385 RI.getMatchingSuperReg(SrcReg, ARM64::dsub, &ARM64::FPR128RegClass);
1386 BuildMI(MBB, I, DL, get(ARM64::ORRv16i8), DestReg).addReg(SrcReg).addReg(
1387 SrcReg, getKillRegState(KillSrc));
1389 BuildMI(MBB, I, DL, get(ARM64::FMOVDr), DestReg)
1390 .addReg(SrcReg, getKillRegState(KillSrc));
1395 if (ARM64::FPR32RegClass.contains(DestReg) &&
1396 ARM64::FPR32RegClass.contains(SrcReg)) {
1397 if(getSubTarget().hasNEON()) {
1399 RI.getMatchingSuperReg(DestReg, ARM64::ssub, &ARM64::FPR128RegClass);
1401 RI.getMatchingSuperReg(SrcReg, ARM64::ssub, &ARM64::FPR128RegClass);
1402 BuildMI(MBB, I, DL, get(ARM64::ORRv16i8), DestReg).addReg(SrcReg).addReg(
1403 SrcReg, getKillRegState(KillSrc));
1405 BuildMI(MBB, I, DL, get(ARM64::FMOVSr), DestReg)
1406 .addReg(SrcReg, getKillRegState(KillSrc));
1411 if (ARM64::FPR16RegClass.contains(DestReg) &&
1412 ARM64::FPR16RegClass.contains(SrcReg)) {
1413 if(getSubTarget().hasNEON()) {
1415 RI.getMatchingSuperReg(DestReg, ARM64::hsub, &ARM64::FPR128RegClass);
1417 RI.getMatchingSuperReg(SrcReg, ARM64::hsub, &ARM64::FPR128RegClass);
1418 BuildMI(MBB, I, DL, get(ARM64::ORRv16i8), DestReg).addReg(SrcReg).addReg(
1419 SrcReg, getKillRegState(KillSrc));
1422 RI.getMatchingSuperReg(DestReg, ARM64::hsub, &ARM64::FPR32RegClass);
1424 RI.getMatchingSuperReg(SrcReg, ARM64::hsub, &ARM64::FPR32RegClass);
1425 BuildMI(MBB, I, DL, get(ARM64::FMOVSr), DestReg)
1426 .addReg(SrcReg, getKillRegState(KillSrc));
1431 if (ARM64::FPR8RegClass.contains(DestReg) &&
1432 ARM64::FPR8RegClass.contains(SrcReg)) {
1433 if(getSubTarget().hasNEON()) {
1435 RI.getMatchingSuperReg(DestReg, ARM64::bsub, &ARM64::FPR128RegClass);
1437 RI.getMatchingSuperReg(SrcReg, ARM64::bsub, &ARM64::FPR128RegClass);
1438 BuildMI(MBB, I, DL, get(ARM64::ORRv16i8), DestReg).addReg(SrcReg).addReg(
1439 SrcReg, getKillRegState(KillSrc));
1442 RI.getMatchingSuperReg(DestReg, ARM64::bsub, &ARM64::FPR32RegClass);
1444 RI.getMatchingSuperReg(SrcReg, ARM64::bsub, &ARM64::FPR32RegClass);
1445 BuildMI(MBB, I, DL, get(ARM64::FMOVSr), DestReg)
1446 .addReg(SrcReg, getKillRegState(KillSrc));
1451 // Copies between GPR64 and FPR64.
1452 if (ARM64::FPR64RegClass.contains(DestReg) &&
1453 ARM64::GPR64RegClass.contains(SrcReg)) {
1454 BuildMI(MBB, I, DL, get(ARM64::FMOVXDr), DestReg)
1455 .addReg(SrcReg, getKillRegState(KillSrc));
1458 if (ARM64::GPR64RegClass.contains(DestReg) &&
1459 ARM64::FPR64RegClass.contains(SrcReg)) {
1460 BuildMI(MBB, I, DL, get(ARM64::FMOVDXr), DestReg)
1461 .addReg(SrcReg, getKillRegState(KillSrc));
1464 // Copies between GPR32 and FPR32.
1465 if (ARM64::FPR32RegClass.contains(DestReg) &&
1466 ARM64::GPR32RegClass.contains(SrcReg)) {
1467 BuildMI(MBB, I, DL, get(ARM64::FMOVWSr), DestReg)
1468 .addReg(SrcReg, getKillRegState(KillSrc));
1471 if (ARM64::GPR32RegClass.contains(DestReg) &&
1472 ARM64::FPR32RegClass.contains(SrcReg)) {
1473 BuildMI(MBB, I, DL, get(ARM64::FMOVSWr), DestReg)
1474 .addReg(SrcReg, getKillRegState(KillSrc));
1478 assert(0 && "unimplemented reg-to-reg copy");
1481 void ARM64InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1482 MachineBasicBlock::iterator MBBI,
1483 unsigned SrcReg, bool isKill, int FI,
1484 const TargetRegisterClass *RC,
1485 const TargetRegisterInfo *TRI) const {
1487 if (MBBI != MBB.end())
1488 DL = MBBI->getDebugLoc();
1489 MachineFunction &MF = *MBB.getParent();
1490 MachineFrameInfo &MFI = *MF.getFrameInfo();
1491 unsigned Align = MFI.getObjectAlignment(FI);
1493 MachinePointerInfo PtrInfo(PseudoSourceValue::getFixedStack(FI));
1494 MachineMemOperand *MMO = MF.getMachineMemOperand(
1495 PtrInfo, MachineMemOperand::MOStore, MFI.getObjectSize(FI), Align);
1498 switch (RC->getSize()) {
1500 if (ARM64::FPR8RegClass.hasSubClassEq(RC))
1501 Opc = ARM64::STRBui;
1504 if (ARM64::FPR16RegClass.hasSubClassEq(RC))
1505 Opc = ARM64::STRHui;
1508 if (ARM64::GPR32allRegClass.hasSubClassEq(RC)) {
1509 Opc = ARM64::STRWui;
1510 if (TargetRegisterInfo::isVirtualRegister(SrcReg))
1511 MF.getRegInfo().constrainRegClass(SrcReg, &ARM64::GPR32RegClass);
1513 assert(SrcReg != ARM64::WSP);
1514 } else if (ARM64::FPR32RegClass.hasSubClassEq(RC))
1515 Opc = ARM64::STRSui;
1518 if (ARM64::GPR64allRegClass.hasSubClassEq(RC)) {
1519 Opc = ARM64::STRXui;
1520 if (TargetRegisterInfo::isVirtualRegister(SrcReg))
1521 MF.getRegInfo().constrainRegClass(SrcReg, &ARM64::GPR64RegClass);
1523 assert(SrcReg != ARM64::SP);
1524 } else if (ARM64::FPR64RegClass.hasSubClassEq(RC))
1525 Opc = ARM64::STRDui;
1528 if (ARM64::FPR128RegClass.hasSubClassEq(RC))
1529 Opc = ARM64::STRQui;
1530 else if (ARM64::DDRegClass.hasSubClassEq(RC)) {
1531 assert(getSubTarget().hasNEON() &&
1532 "Unexpected register store without NEON");
1533 Opc = ARM64::ST1Twov1d, Offset = false;
1537 if (ARM64::DDDRegClass.hasSubClassEq(RC)) {
1538 assert(getSubTarget().hasNEON() &&
1539 "Unexpected register store without NEON");
1540 Opc = ARM64::ST1Threev1d, Offset = false;
1544 if (ARM64::DDDDRegClass.hasSubClassEq(RC)) {
1545 assert(getSubTarget().hasNEON() &&
1546 "Unexpected register store without NEON");
1547 Opc = ARM64::ST1Fourv1d, Offset = false;
1548 } else if (ARM64::QQRegClass.hasSubClassEq(RC)) {
1549 assert(getSubTarget().hasNEON() &&
1550 "Unexpected register store without NEON");
1551 Opc = ARM64::ST1Twov2d, Offset = false;
1555 if (ARM64::QQQRegClass.hasSubClassEq(RC)) {
1556 assert(getSubTarget().hasNEON() &&
1557 "Unexpected register store without NEON");
1558 Opc = ARM64::ST1Threev2d, Offset = false;
1562 if (ARM64::QQQQRegClass.hasSubClassEq(RC)) {
1563 assert(getSubTarget().hasNEON() &&
1564 "Unexpected register store without NEON");
1565 Opc = ARM64::ST1Fourv2d, Offset = false;
1569 assert(Opc && "Unknown register class");
1571 const MachineInstrBuilder &MI = BuildMI(MBB, MBBI, DL, get(Opc))
1572 .addReg(SrcReg, getKillRegState(isKill))
1577 MI.addMemOperand(MMO);
1580 void ARM64InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1581 MachineBasicBlock::iterator MBBI,
1582 unsigned DestReg, int FI,
1583 const TargetRegisterClass *RC,
1584 const TargetRegisterInfo *TRI) const {
1586 if (MBBI != MBB.end())
1587 DL = MBBI->getDebugLoc();
1588 MachineFunction &MF = *MBB.getParent();
1589 MachineFrameInfo &MFI = *MF.getFrameInfo();
1590 unsigned Align = MFI.getObjectAlignment(FI);
1591 MachinePointerInfo PtrInfo(PseudoSourceValue::getFixedStack(FI));
1592 MachineMemOperand *MMO = MF.getMachineMemOperand(
1593 PtrInfo, MachineMemOperand::MOLoad, MFI.getObjectSize(FI), Align);
1597 switch (RC->getSize()) {
1599 if (ARM64::FPR8RegClass.hasSubClassEq(RC))
1600 Opc = ARM64::LDRBui;
1603 if (ARM64::FPR16RegClass.hasSubClassEq(RC))
1604 Opc = ARM64::LDRHui;
1607 if (ARM64::GPR32allRegClass.hasSubClassEq(RC)) {
1608 Opc = ARM64::LDRWui;
1609 if (TargetRegisterInfo::isVirtualRegister(DestReg))
1610 MF.getRegInfo().constrainRegClass(DestReg, &ARM64::GPR32RegClass);
1612 assert(DestReg != ARM64::WSP);
1613 } else if (ARM64::FPR32RegClass.hasSubClassEq(RC))
1614 Opc = ARM64::LDRSui;
1617 if (ARM64::GPR64allRegClass.hasSubClassEq(RC)) {
1618 Opc = ARM64::LDRXui;
1619 if (TargetRegisterInfo::isVirtualRegister(DestReg))
1620 MF.getRegInfo().constrainRegClass(DestReg, &ARM64::GPR64RegClass);
1622 assert(DestReg != ARM64::SP);
1623 } else if (ARM64::FPR64RegClass.hasSubClassEq(RC))
1624 Opc = ARM64::LDRDui;
1627 if (ARM64::FPR128RegClass.hasSubClassEq(RC))
1628 Opc = ARM64::LDRQui;
1629 else if (ARM64::DDRegClass.hasSubClassEq(RC)) {
1630 assert(getSubTarget().hasNEON() &&
1631 "Unexpected register load without NEON");
1632 Opc = ARM64::LD1Twov1d, Offset = false;
1636 if (ARM64::DDDRegClass.hasSubClassEq(RC)) {
1637 assert(getSubTarget().hasNEON() &&
1638 "Unexpected register load without NEON");
1639 Opc = ARM64::LD1Threev1d, Offset = false;
1643 if (ARM64::DDDDRegClass.hasSubClassEq(RC)) {
1644 assert(getSubTarget().hasNEON() &&
1645 "Unexpected register load without NEON");
1646 Opc = ARM64::LD1Fourv1d, Offset = false;
1647 } else if (ARM64::QQRegClass.hasSubClassEq(RC)) {
1648 assert(getSubTarget().hasNEON() &&
1649 "Unexpected register load without NEON");
1650 Opc = ARM64::LD1Twov2d, Offset = false;
1654 if (ARM64::QQQRegClass.hasSubClassEq(RC)) {
1655 assert(getSubTarget().hasNEON() &&
1656 "Unexpected register load without NEON");
1657 Opc = ARM64::LD1Threev2d, Offset = false;
1661 if (ARM64::QQQQRegClass.hasSubClassEq(RC)) {
1662 assert(getSubTarget().hasNEON() &&
1663 "Unexpected register load without NEON");
1664 Opc = ARM64::LD1Fourv2d, Offset = false;
1668 assert(Opc && "Unknown register class");
1670 const MachineInstrBuilder &MI = BuildMI(MBB, MBBI, DL, get(Opc))
1671 .addReg(DestReg, getDefRegState(true))
1675 MI.addMemOperand(MMO);
1678 void llvm::emitFrameOffset(MachineBasicBlock &MBB,
1679 MachineBasicBlock::iterator MBBI, DebugLoc DL,
1680 unsigned DestReg, unsigned SrcReg, int Offset,
1681 const ARM64InstrInfo *TII, MachineInstr::MIFlag Flag,
1683 if (DestReg == SrcReg && Offset == 0)
1686 bool isSub = Offset < 0;
1690 // FIXME: If the offset won't fit in 24-bits, compute the offset into a
1691 // scratch register. If DestReg is a virtual register, use it as the
1692 // scratch register; otherwise, create a new virtual register (to be
1693 // replaced by the scavenger at the end of PEI). That case can be optimized
1694 // slightly if DestReg is SP which is always 16-byte aligned, so the scratch
1695 // register can be loaded with offset%8 and the add/sub can use an extending
1696 // instruction with LSL#3.
1697 // Currently the function handles any offsets but generates a poor sequence
1699 // assert(Offset < (1 << 24) && "unimplemented reg plus immediate");
1703 Opc = isSub ? ARM64::SUBSXri : ARM64::ADDSXri;
1705 Opc = isSub ? ARM64::SUBXri : ARM64::ADDXri;
1706 const unsigned MaxEncoding = 0xfff;
1707 const unsigned ShiftSize = 12;
1708 const unsigned MaxEncodableValue = MaxEncoding << ShiftSize;
1709 while (((unsigned)Offset) >= (1 << ShiftSize)) {
1711 if (((unsigned)Offset) > MaxEncodableValue) {
1712 ThisVal = MaxEncodableValue;
1714 ThisVal = Offset & MaxEncodableValue;
1716 assert((ThisVal >> ShiftSize) <= MaxEncoding &&
1717 "Encoding cannot handle value that big");
1718 BuildMI(MBB, MBBI, DL, TII->get(Opc), DestReg)
1720 .addImm(ThisVal >> ShiftSize)
1721 .addImm(ARM64_AM::getShifterImm(ARM64_AM::LSL, ShiftSize))
1729 BuildMI(MBB, MBBI, DL, TII->get(Opc), DestReg)
1732 .addImm(ARM64_AM::getShifterImm(ARM64_AM::LSL, 0))
1737 ARM64InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI,
1738 const SmallVectorImpl<unsigned> &Ops,
1739 int FrameIndex) const {
1740 // This is a bit of a hack. Consider this instruction:
1742 // %vreg0<def> = COPY %SP; GPR64all:%vreg0
1744 // We explicitly chose GPR64all for the virtual register so such a copy might
1745 // be eliminated by RegisterCoalescer. However, that may not be possible, and
1746 // %vreg0 may even spill. We can't spill %SP, and since it is in the GPR64all
1747 // register class, TargetInstrInfo::foldMemoryOperand() is going to try.
1749 // To prevent that, we are going to constrain the %vreg0 register class here.
1751 // <rdar://problem/11522048>
1754 unsigned DstReg = MI->getOperand(0).getReg();
1755 unsigned SrcReg = MI->getOperand(1).getReg();
1756 if (SrcReg == ARM64::SP && TargetRegisterInfo::isVirtualRegister(DstReg)) {
1757 MF.getRegInfo().constrainRegClass(DstReg, &ARM64::GPR64RegClass);
1760 if (DstReg == ARM64::SP && TargetRegisterInfo::isVirtualRegister(SrcReg)) {
1761 MF.getRegInfo().constrainRegClass(SrcReg, &ARM64::GPR64RegClass);
1770 int llvm::isARM64FrameOffsetLegal(const MachineInstr &MI, int &Offset,
1771 bool *OutUseUnscaledOp,
1772 unsigned *OutUnscaledOp,
1773 int *EmittableOffset) {
1775 bool IsSigned = false;
1776 // The ImmIdx should be changed case by case if it is not 2.
1777 unsigned ImmIdx = 2;
1778 unsigned UnscaledOp = 0;
1779 // Set output values in case of early exit.
1780 if (EmittableOffset)
1781 *EmittableOffset = 0;
1782 if (OutUseUnscaledOp)
1783 *OutUseUnscaledOp = false;
1786 switch (MI.getOpcode()) {
1788 assert(0 && "unhandled opcode in rewriteARM64FrameIndex");
1789 // Vector spills/fills can't take an immediate offset.
1790 case ARM64::LD1Twov2d:
1791 case ARM64::LD1Threev2d:
1792 case ARM64::LD1Fourv2d:
1793 case ARM64::LD1Twov1d:
1794 case ARM64::LD1Threev1d:
1795 case ARM64::LD1Fourv1d:
1796 case ARM64::ST1Twov2d:
1797 case ARM64::ST1Threev2d:
1798 case ARM64::ST1Fourv2d:
1799 case ARM64::ST1Twov1d:
1800 case ARM64::ST1Threev1d:
1801 case ARM64::ST1Fourv1d:
1802 return ARM64FrameOffsetCannotUpdate;
1805 UnscaledOp = ARM64::PRFUMi;
1809 UnscaledOp = ARM64::LDURXi;
1813 UnscaledOp = ARM64::LDURWi;
1817 UnscaledOp = ARM64::LDURBi;
1821 UnscaledOp = ARM64::LDURHi;
1825 UnscaledOp = ARM64::LDURSi;
1829 UnscaledOp = ARM64::LDURDi;
1833 UnscaledOp = ARM64::LDURQi;
1835 case ARM64::LDRBBui:
1837 UnscaledOp = ARM64::LDURBBi;
1839 case ARM64::LDRHHui:
1841 UnscaledOp = ARM64::LDURHHi;
1843 case ARM64::LDRSBXui:
1845 UnscaledOp = ARM64::LDURSBXi;
1847 case ARM64::LDRSBWui:
1849 UnscaledOp = ARM64::LDURSBWi;
1851 case ARM64::LDRSHXui:
1853 UnscaledOp = ARM64::LDURSHXi;
1855 case ARM64::LDRSHWui:
1857 UnscaledOp = ARM64::LDURSHWi;
1859 case ARM64::LDRSWui:
1861 UnscaledOp = ARM64::LDURSWi;
1866 UnscaledOp = ARM64::STURXi;
1870 UnscaledOp = ARM64::STURWi;
1874 UnscaledOp = ARM64::STURBi;
1878 UnscaledOp = ARM64::STURHi;
1882 UnscaledOp = ARM64::STURSi;
1886 UnscaledOp = ARM64::STURDi;
1890 UnscaledOp = ARM64::STURQi;
1892 case ARM64::STRBBui:
1894 UnscaledOp = ARM64::STURBBi;
1896 case ARM64::STRHHui:
1898 UnscaledOp = ARM64::STURHHi;
1928 case ARM64::LDURHHi:
1929 case ARM64::LDURBBi:
1930 case ARM64::LDURSBXi:
1931 case ARM64::LDURSBWi:
1932 case ARM64::LDURSHXi:
1933 case ARM64::LDURSHWi:
1934 case ARM64::LDURSWi:
1942 case ARM64::STURBBi:
1943 case ARM64::STURHHi:
1948 Offset += MI.getOperand(ImmIdx).getImm() * Scale;
1950 bool useUnscaledOp = false;
1951 // If the offset doesn't match the scale, we rewrite the instruction to
1952 // use the unscaled instruction instead. Likewise, if we have a negative
1953 // offset (and have an unscaled op to use).
1954 if ((Offset & (Scale - 1)) != 0 || (Offset < 0 && UnscaledOp != 0))
1955 useUnscaledOp = true;
1957 // Use an unscaled addressing mode if the instruction has a negative offset
1958 // (or if the instruction is already using an unscaled addressing mode).
1961 // ldp/stp instructions.
1964 } else if (UnscaledOp == 0 || useUnscaledOp) {
1974 // Attempt to fold address computation.
1975 int MaxOff = (1 << (MaskBits - IsSigned)) - 1;
1976 int MinOff = (IsSigned ? (-MaxOff - 1) : 0);
1977 if (Offset >= MinOff && Offset <= MaxOff) {
1978 if (EmittableOffset)
1979 *EmittableOffset = Offset;
1982 int NewOff = Offset < 0 ? MinOff : MaxOff;
1983 if (EmittableOffset)
1984 *EmittableOffset = NewOff;
1985 Offset = (Offset - NewOff) * Scale;
1987 if (OutUseUnscaledOp)
1988 *OutUseUnscaledOp = useUnscaledOp;
1990 *OutUnscaledOp = UnscaledOp;
1991 return ARM64FrameOffsetCanUpdate |
1992 (Offset == 0 ? ARM64FrameOffsetIsLegal : 0);
1995 bool llvm::rewriteARM64FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
1996 unsigned FrameReg, int &Offset,
1997 const ARM64InstrInfo *TII) {
1998 unsigned Opcode = MI.getOpcode();
1999 unsigned ImmIdx = FrameRegIdx + 1;
2001 if (Opcode == ARM64::ADDSXri || Opcode == ARM64::ADDXri) {
2002 Offset += MI.getOperand(ImmIdx).getImm();
2003 emitFrameOffset(*MI.getParent(), MI, MI.getDebugLoc(),
2004 MI.getOperand(0).getReg(), FrameReg, Offset, TII,
2005 MachineInstr::NoFlags, (Opcode == ARM64::ADDSXri));
2006 MI.eraseFromParent();
2012 unsigned UnscaledOp;
2014 int Status = isARM64FrameOffsetLegal(MI, Offset, &UseUnscaledOp, &UnscaledOp,
2016 if (Status & ARM64FrameOffsetCanUpdate) {
2017 if (Status & ARM64FrameOffsetIsLegal)
2018 // Replace the FrameIndex with FrameReg.
2019 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
2021 MI.setDesc(TII->get(UnscaledOp));
2023 MI.getOperand(ImmIdx).ChangeToImmediate(NewOffset);
2030 void ARM64InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
2031 NopInst.setOpcode(ARM64::HINT);
2032 NopInst.addOperand(MCOperand::CreateImm(0));