1 //===-- SystemZInstrInfo.cpp - SystemZ 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 SystemZ implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "SystemZInstrInfo.h"
15 #include "SystemZInstrBuilder.h"
16 #include "SystemZTargetMachine.h"
17 #include "llvm/CodeGen/LiveVariables.h"
18 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #define GET_INSTRINFO_CTOR_DTOR
24 #define GET_INSTRMAP_INFO
25 #include "SystemZGenInstrInfo.inc"
27 // Return a mask with Count low bits set.
28 static uint64_t allOnes(unsigned int Count) {
29 return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1;
32 // Reg should be a 32-bit GPR. Return true if it is a high register rather
33 // than a low register.
34 static bool isHighReg(unsigned int Reg) {
35 if (SystemZ::GRH32BitRegClass.contains(Reg))
37 assert(SystemZ::GR32BitRegClass.contains(Reg) && "Invalid GRX32");
41 // Pin the vtable to this file.
42 void SystemZInstrInfo::anchor() {}
44 SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti)
45 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP),
49 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores,
50 // each having the opcode given by NewOpcode.
51 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI,
52 unsigned NewOpcode) const {
53 MachineBasicBlock *MBB = MI->getParent();
54 MachineFunction &MF = *MBB->getParent();
56 // Get two load or store instructions. Use the original instruction for one
57 // of them (arbitrarily the second here) and create a clone for the other.
58 MachineInstr *EarlierMI = MF.CloneMachineInstr(&*MI);
59 MBB->insert(MI, EarlierMI);
61 // Set up the two 64-bit registers.
62 MachineOperand &HighRegOp = EarlierMI->getOperand(0);
63 MachineOperand &LowRegOp = MI->getOperand(0);
64 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_h64));
65 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_l64));
67 // The address in the first (high) instruction is already correct.
68 // Adjust the offset in the second (low) instruction.
69 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2);
70 MachineOperand &LowOffsetOp = MI->getOperand(2);
71 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8);
73 // Clear the kill flags for the base and index registers in the first
75 EarlierMI->getOperand(1).setIsKill(false);
76 EarlierMI->getOperand(3).setIsKill(false);
79 unsigned HighOpcode = getOpcodeForOffset(NewOpcode, HighOffsetOp.getImm());
80 unsigned LowOpcode = getOpcodeForOffset(NewOpcode, LowOffsetOp.getImm());
81 assert(HighOpcode && LowOpcode && "Both offsets should be in range");
83 EarlierMI->setDesc(get(HighOpcode));
84 MI->setDesc(get(LowOpcode));
87 // Split ADJDYNALLOC instruction MI.
88 void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const {
89 MachineBasicBlock *MBB = MI->getParent();
90 MachineFunction &MF = *MBB->getParent();
91 MachineFrameInfo &MFFrame = MF.getFrameInfo();
92 MachineOperand &OffsetMO = MI->getOperand(2);
94 uint64_t Offset = (MFFrame.getMaxCallFrameSize() +
95 SystemZMC::CallFrameSize +
97 unsigned NewOpcode = getOpcodeForOffset(SystemZ::LA, Offset);
98 assert(NewOpcode && "No support for huge argument lists yet");
99 MI->setDesc(get(NewOpcode));
100 OffsetMO.setImm(Offset);
103 // MI is an RI-style pseudo instruction. Replace it with LowOpcode
104 // if the first operand is a low GR32 and HighOpcode if the first operand
105 // is a high GR32. ConvertHigh is true if LowOpcode takes a signed operand
106 // and HighOpcode takes an unsigned 32-bit operand. In those cases,
107 // MI has the same kind of operand as LowOpcode, so needs to be converted
108 // if HighOpcode is used.
109 void SystemZInstrInfo::expandRIPseudo(MachineInstr &MI, unsigned LowOpcode,
111 bool ConvertHigh) const {
112 unsigned Reg = MI.getOperand(0).getReg();
113 bool IsHigh = isHighReg(Reg);
114 MI.setDesc(get(IsHigh ? HighOpcode : LowOpcode));
115 if (IsHigh && ConvertHigh)
116 MI.getOperand(1).setImm(uint32_t(MI.getOperand(1).getImm()));
119 // MI is a three-operand RIE-style pseudo instruction. Replace it with
120 // LowOpcodeK if the registers are both low GR32s, otherwise use a move
121 // followed by HighOpcode or LowOpcode, depending on whether the target
122 // is a high or low GR32.
123 void SystemZInstrInfo::expandRIEPseudo(MachineInstr &MI, unsigned LowOpcode,
125 unsigned HighOpcode) const {
126 unsigned DestReg = MI.getOperand(0).getReg();
127 unsigned SrcReg = MI.getOperand(1).getReg();
128 bool DestIsHigh = isHighReg(DestReg);
129 bool SrcIsHigh = isHighReg(SrcReg);
130 if (!DestIsHigh && !SrcIsHigh)
131 MI.setDesc(get(LowOpcodeK));
133 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, SrcReg,
134 SystemZ::LR, 32, MI.getOperand(1).isKill());
135 MI.setDesc(get(DestIsHigh ? HighOpcode : LowOpcode));
136 MI.getOperand(1).setReg(DestReg);
137 MI.tieOperands(0, 1);
141 // MI is an RXY-style pseudo instruction. Replace it with LowOpcode
142 // if the first operand is a low GR32 and HighOpcode if the first operand
144 void SystemZInstrInfo::expandRXYPseudo(MachineInstr &MI, unsigned LowOpcode,
145 unsigned HighOpcode) const {
146 unsigned Reg = MI.getOperand(0).getReg();
147 unsigned Opcode = getOpcodeForOffset(isHighReg(Reg) ? HighOpcode : LowOpcode,
148 MI.getOperand(2).getImm());
149 MI.setDesc(get(Opcode));
152 // MI is a load-on-condition pseudo instruction with a single register
153 // (source or destination) operand. Replace it with LowOpcode if the
154 // register is a low GR32 and HighOpcode if the register is a high GR32.
155 void SystemZInstrInfo::expandLOCPseudo(MachineInstr &MI, unsigned LowOpcode,
156 unsigned HighOpcode) const {
157 unsigned Reg = MI.getOperand(0).getReg();
158 unsigned Opcode = isHighReg(Reg) ? HighOpcode : LowOpcode;
159 MI.setDesc(get(Opcode));
162 // MI is a load-register-on-condition pseudo instruction. Replace it with
163 // LowOpcode if source and destination are both low GR32s and HighOpcode if
164 // source and destination are both high GR32s.
165 void SystemZInstrInfo::expandLOCRPseudo(MachineInstr &MI, unsigned LowOpcode,
166 unsigned HighOpcode) const {
167 unsigned DestReg = MI.getOperand(0).getReg();
168 unsigned SrcReg = MI.getOperand(2).getReg();
169 bool DestIsHigh = isHighReg(DestReg);
170 bool SrcIsHigh = isHighReg(SrcReg);
172 if (!DestIsHigh && !SrcIsHigh)
173 MI.setDesc(get(LowOpcode));
174 else if (DestIsHigh && SrcIsHigh)
175 MI.setDesc(get(HighOpcode));
177 // If we were unable to implement the pseudo with a single instruction, we
178 // need to convert it back into a branch sequence. This cannot be done here
179 // since the caller of expandPostRAPseudo does not handle changes to the CFG
180 // correctly. This change is defered to the SystemZExpandPseudo pass.
183 // MI is an RR-style pseudo instruction that zero-extends the low Size bits
184 // of one GRX32 into another. Replace it with LowOpcode if both operands
185 // are low registers, otherwise use RISB[LH]G.
186 void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode,
187 unsigned Size) const {
188 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(),
189 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode,
190 Size, MI.getOperand(1).isKill());
191 MI.eraseFromParent();
194 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const {
195 MachineBasicBlock *MBB = MI->getParent();
196 MachineFunction &MF = *MBB->getParent();
197 const unsigned Reg = MI->getOperand(0).getReg();
199 // Conveniently, all 4 instructions are cloned from LOAD_STACK_GUARD,
200 // so they already have operand 0 set to reg.
203 MachineInstr *Ear1MI = MF.CloneMachineInstr(MI);
204 MBB->insert(MI, Ear1MI);
205 Ear1MI->setDesc(get(SystemZ::EAR));
206 MachineInstrBuilder(MF, Ear1MI).addReg(SystemZ::A0);
208 // sllg <reg>, <reg>, 32
209 MachineInstr *SllgMI = MF.CloneMachineInstr(MI);
210 MBB->insert(MI, SllgMI);
211 SllgMI->setDesc(get(SystemZ::SLLG));
212 MachineInstrBuilder(MF, SllgMI).addReg(Reg).addReg(0).addImm(32);
215 MachineInstr *Ear2MI = MF.CloneMachineInstr(MI);
216 MBB->insert(MI, Ear2MI);
217 Ear2MI->setDesc(get(SystemZ::EAR));
218 MachineInstrBuilder(MF, Ear2MI).addReg(SystemZ::A1);
220 // lg <reg>, 40(<reg>)
221 MI->setDesc(get(SystemZ::LG));
222 MachineInstrBuilder(MF, MI).addReg(Reg).addImm(40).addReg(0);
225 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR
226 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg
227 // are low registers, otherwise use RISB[LH]G. Size is the number of bits
228 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR).
229 // KillSrc is true if this move is the last use of SrcReg.
230 void SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB,
231 MachineBasicBlock::iterator MBBI,
232 const DebugLoc &DL, unsigned DestReg,
233 unsigned SrcReg, unsigned LowLowOpcode,
234 unsigned Size, bool KillSrc) const {
236 bool DestIsHigh = isHighReg(DestReg);
237 bool SrcIsHigh = isHighReg(SrcReg);
238 if (DestIsHigh && SrcIsHigh)
239 Opcode = SystemZ::RISBHH;
240 else if (DestIsHigh && !SrcIsHigh)
241 Opcode = SystemZ::RISBHL;
242 else if (!DestIsHigh && SrcIsHigh)
243 Opcode = SystemZ::RISBLH;
245 BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg)
246 .addReg(SrcReg, getKillRegState(KillSrc));
249 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0);
250 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg)
251 .addReg(DestReg, RegState::Undef)
252 .addReg(SrcReg, getKillRegState(KillSrc))
253 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate);
257 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI,
260 unsigned OpIdx2) const {
261 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
263 return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
267 switch (MI.getOpcode()) {
268 case SystemZ::LOCRMux:
269 case SystemZ::LOCFHR:
271 case SystemZ::LOCGR: {
272 auto &WorkingMI = cloneIfNew(MI);
274 unsigned CCValid = WorkingMI.getOperand(3).getImm();
275 unsigned CCMask = WorkingMI.getOperand(4).getImm();
276 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid);
277 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
281 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
286 // If MI is a simple load or store for a frame object, return the register
287 // it loads or stores and set FrameIndex to the index of the frame object.
288 // Return 0 otherwise.
290 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
291 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex,
293 const MCInstrDesc &MCID = MI.getDesc();
294 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() &&
295 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) {
296 FrameIndex = MI.getOperand(1).getIndex();
297 return MI.getOperand(0).getReg();
302 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
303 int &FrameIndex) const {
304 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad);
307 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
308 int &FrameIndex) const {
309 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore);
312 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI,
314 int &SrcFrameIndex) const {
315 // Check for MVC 0(Length,FI1),0(FI2)
316 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo();
317 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() ||
318 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() ||
319 MI.getOperand(4).getImm() != 0)
322 // Check that Length covers the full slots.
323 int64_t Length = MI.getOperand(2).getImm();
324 unsigned FI1 = MI.getOperand(0).getIndex();
325 unsigned FI2 = MI.getOperand(3).getIndex();
326 if (MFI.getObjectSize(FI1) != Length ||
327 MFI.getObjectSize(FI2) != Length)
330 DestFrameIndex = FI1;
335 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
336 MachineBasicBlock *&TBB,
337 MachineBasicBlock *&FBB,
338 SmallVectorImpl<MachineOperand> &Cond,
339 bool AllowModify) const {
340 // Most of the code and comments here are boilerplate.
342 // Start from the bottom of the block and work up, examining the
343 // terminator instructions.
344 MachineBasicBlock::iterator I = MBB.end();
345 while (I != MBB.begin()) {
347 if (I->isDebugValue())
350 // Working from the bottom, when we see a non-terminator instruction, we're
352 if (!isUnpredicatedTerminator(*I))
355 // A terminator that isn't a branch can't easily be handled by this
360 // Can't handle indirect branches.
361 SystemZII::Branch Branch(getBranchInfo(*I));
362 if (!Branch.Target->isMBB())
365 // Punt on compound branches.
366 if (Branch.Type != SystemZII::BranchNormal)
369 if (Branch.CCMask == SystemZ::CCMASK_ANY) {
370 // Handle unconditional branches.
372 TBB = Branch.Target->getMBB();
376 // If the block has any instructions after a JMP, delete them.
377 while (std::next(I) != MBB.end())
378 std::next(I)->eraseFromParent();
383 // Delete the JMP if it's equivalent to a fall-through.
384 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) {
386 I->eraseFromParent();
391 // TBB is used to indicate the unconditinal destination.
392 TBB = Branch.Target->getMBB();
396 // Working from the bottom, handle the first conditional branch.
398 // FIXME: add X86-style branch swap
400 TBB = Branch.Target->getMBB();
401 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid));
402 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask));
406 // Handle subsequent conditional branches.
407 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch");
409 // Only handle the case where all conditional branches branch to the same
411 if (TBB != Branch.Target->getMBB())
414 // If the conditions are the same, we can leave them alone.
415 unsigned OldCCValid = Cond[0].getImm();
416 unsigned OldCCMask = Cond[1].getImm();
417 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask)
420 // FIXME: Try combining conditions like X86 does. Should be easy on Z!
427 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB,
428 int *BytesRemoved) const {
429 assert(!BytesRemoved && "code size not handled");
431 // Most of the code and comments here are boilerplate.
432 MachineBasicBlock::iterator I = MBB.end();
435 while (I != MBB.begin()) {
437 if (I->isDebugValue())
441 if (!getBranchInfo(*I).Target->isMBB())
443 // Remove the branch.
444 I->eraseFromParent();
452 bool SystemZInstrInfo::
453 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
454 assert(Cond.size() == 2 && "Invalid condition");
455 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm());
459 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB,
460 MachineBasicBlock *TBB,
461 MachineBasicBlock *FBB,
462 ArrayRef<MachineOperand> Cond,
464 int *BytesAdded) const {
465 // In this function we output 32-bit branches, which should always
466 // have enough range. They can be shortened and relaxed by later code
467 // in the pipeline, if desired.
469 // Shouldn't be a fall through.
470 assert(TBB && "insertBranch must not be told to insert a fallthrough");
471 assert((Cond.size() == 2 || Cond.size() == 0) &&
472 "SystemZ branch conditions have one component!");
473 assert(!BytesAdded && "code size not handled");
476 // Unconditional branch?
477 assert(!FBB && "Unconditional branch with multiple successors!");
478 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB);
482 // Conditional branch.
484 unsigned CCValid = Cond[0].getImm();
485 unsigned CCMask = Cond[1].getImm();
486 BuildMI(&MBB, DL, get(SystemZ::BRC))
487 .addImm(CCValid).addImm(CCMask).addMBB(TBB);
491 // Two-way Conditional branch. Insert the second branch.
492 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB);
498 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
499 unsigned &SrcReg2, int &Mask,
501 assert(MI.isCompare() && "Caller should have checked for a comparison");
503 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() &&
504 MI.getOperand(1).isImm()) {
505 SrcReg = MI.getOperand(0).getReg();
507 Value = MI.getOperand(1).getImm();
515 // If Reg is a virtual register, return its definition, otherwise return null.
516 static MachineInstr *getDef(unsigned Reg,
517 const MachineRegisterInfo *MRI) {
518 if (TargetRegisterInfo::isPhysicalRegister(Reg))
520 return MRI->getUniqueVRegDef(Reg);
523 // Return true if MI is a shift of type Opcode by Imm bits.
524 static bool isShift(MachineInstr *MI, unsigned Opcode, int64_t Imm) {
525 return (MI->getOpcode() == Opcode &&
526 !MI->getOperand(2).getReg() &&
527 MI->getOperand(3).getImm() == Imm);
530 // If the destination of MI has no uses, delete it as dead.
531 static void eraseIfDead(MachineInstr *MI, const MachineRegisterInfo *MRI) {
532 if (MRI->use_nodbg_empty(MI->getOperand(0).getReg()))
533 MI->eraseFromParent();
536 // Compare compares SrcReg against zero. Check whether SrcReg contains
537 // the result of an IPM sequence whose input CC survives until Compare,
538 // and whether Compare is therefore redundant. Delete it and return
540 static bool removeIPMBasedCompare(MachineInstr &Compare, unsigned SrcReg,
541 const MachineRegisterInfo *MRI,
542 const TargetRegisterInfo *TRI) {
543 MachineInstr *LGFR = nullptr;
544 MachineInstr *RLL = getDef(SrcReg, MRI);
545 if (RLL && RLL->getOpcode() == SystemZ::LGFR) {
547 RLL = getDef(LGFR->getOperand(1).getReg(), MRI);
549 if (!RLL || !isShift(RLL, SystemZ::RLL, 31))
552 MachineInstr *SRL = getDef(RLL->getOperand(1).getReg(), MRI);
553 if (!SRL || !isShift(SRL, SystemZ::SRL, SystemZ::IPM_CC))
556 MachineInstr *IPM = getDef(SRL->getOperand(1).getReg(), MRI);
557 if (!IPM || IPM->getOpcode() != SystemZ::IPM)
560 // Check that there are no assignments to CC between the IPM and Compare,
561 if (IPM->getParent() != Compare.getParent())
563 MachineBasicBlock::iterator MBBI = IPM, MBBE = Compare.getIterator();
564 for (++MBBI; MBBI != MBBE; ++MBBI) {
565 MachineInstr &MI = *MBBI;
566 if (MI.modifiesRegister(SystemZ::CC, TRI))
570 Compare.eraseFromParent();
572 eraseIfDead(LGFR, MRI);
573 eraseIfDead(RLL, MRI);
574 eraseIfDead(SRL, MRI);
575 eraseIfDead(IPM, MRI);
580 bool SystemZInstrInfo::optimizeCompareInstr(
581 MachineInstr &Compare, unsigned SrcReg, unsigned SrcReg2, int Mask,
582 int Value, const MachineRegisterInfo *MRI) const {
583 assert(!SrcReg2 && "Only optimizing constant comparisons so far");
584 bool IsLogical = (Compare.getDesc().TSFlags & SystemZII::IsLogical) != 0;
585 return Value == 0 && !IsLogical &&
586 removeIPMBasedCompare(Compare, SrcReg, MRI, &RI);
590 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
591 ArrayRef<MachineOperand> Pred,
592 unsigned TrueReg, unsigned FalseReg,
593 int &CondCycles, int &TrueCycles,
594 int &FalseCycles) const {
595 // Not all subtargets have LOCR instructions.
596 if (!STI.hasLoadStoreOnCond())
598 if (Pred.size() != 2)
601 // Check register classes.
602 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
603 const TargetRegisterClass *RC =
604 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
608 // We have LOCR instructions for 32 and 64 bit general purpose registers.
609 if ((STI.hasLoadStoreOnCond2() &&
610 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) ||
611 SystemZ::GR32BitRegClass.hasSubClassEq(RC) ||
612 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) {
619 // Can't do anything else.
623 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB,
624 MachineBasicBlock::iterator I,
625 const DebugLoc &DL, unsigned DstReg,
626 ArrayRef<MachineOperand> Pred,
628 unsigned FalseReg) const {
629 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
630 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
632 assert(Pred.size() == 2 && "Invalid condition");
633 unsigned CCValid = Pred[0].getImm();
634 unsigned CCMask = Pred[1].getImm();
637 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) {
638 if (STI.hasLoadStoreOnCond2())
639 Opc = SystemZ::LOCRMux;
642 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass);
644 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC))
645 Opc = SystemZ::LOCGR;
647 llvm_unreachable("Invalid register class");
649 BuildMI(MBB, I, DL, get(Opc), DstReg)
650 .addReg(FalseReg).addReg(TrueReg)
651 .addImm(CCValid).addImm(CCMask);
654 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
656 MachineRegisterInfo *MRI) const {
657 unsigned DefOpc = DefMI.getOpcode();
658 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI &&
659 DefOpc != SystemZ::LGHI)
661 if (DefMI.getOperand(0).getReg() != Reg)
663 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm();
665 unsigned UseOpc = UseMI.getOpcode();
670 case SystemZ::LOCRMux:
671 if (!STI.hasLoadStoreOnCond2())
673 NewUseOpc = SystemZ::LOCHIMux;
674 if (UseMI.getOperand(2).getReg() == Reg)
676 else if (UseMI.getOperand(1).getReg() == Reg)
677 UseIdx = 2, CommuteIdx = 1;
682 if (!STI.hasLoadStoreOnCond2())
684 NewUseOpc = SystemZ::LOCGHI;
685 if (UseMI.getOperand(2).getReg() == Reg)
687 else if (UseMI.getOperand(1).getReg() == Reg)
688 UseIdx = 2, CommuteIdx = 1;
696 if (CommuteIdx != -1)
697 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx))
700 bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
701 UseMI.setDesc(get(NewUseOpc));
702 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal);
704 DefMI.eraseFromParent();
709 bool SystemZInstrInfo::isPredicable(MachineInstr &MI) const {
710 unsigned Opcode = MI.getOpcode();
711 if (Opcode == SystemZ::Return ||
712 Opcode == SystemZ::Trap ||
713 Opcode == SystemZ::CallJG ||
714 Opcode == SystemZ::CallBR)
719 bool SystemZInstrInfo::
720 isProfitableToIfCvt(MachineBasicBlock &MBB,
721 unsigned NumCycles, unsigned ExtraPredCycles,
722 BranchProbability Probability) const {
723 // Avoid using conditional returns at the end of a loop (since then
724 // we'd need to emit an unconditional branch to the beginning anyway,
725 // making the loop body longer). This doesn't apply for low-probability
726 // loops (eg. compare-and-swap retry), so just decide based on branch
727 // probability instead of looping structure.
728 // However, since Compare and Trap instructions cost the same as a regular
729 // Compare instruction, we should allow the if conversion to convert this
730 // into a Conditional Compare regardless of the branch probability.
731 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap &&
732 MBB.succ_empty() && Probability < BranchProbability(1, 8))
734 // For now only convert single instructions.
735 return NumCycles == 1;
738 bool SystemZInstrInfo::
739 isProfitableToIfCvt(MachineBasicBlock &TMBB,
740 unsigned NumCyclesT, unsigned ExtraPredCyclesT,
741 MachineBasicBlock &FMBB,
742 unsigned NumCyclesF, unsigned ExtraPredCyclesF,
743 BranchProbability Probability) const {
744 // For now avoid converting mutually-exclusive cases.
748 bool SystemZInstrInfo::
749 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
750 BranchProbability Probability) const {
751 // For now only duplicate single instructions.
752 return NumCycles == 1;
755 bool SystemZInstrInfo::PredicateInstruction(
756 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const {
757 assert(Pred.size() == 2 && "Invalid condition");
758 unsigned CCValid = Pred[0].getImm();
759 unsigned CCMask = Pred[1].getImm();
760 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate");
761 unsigned Opcode = MI.getOpcode();
762 if (Opcode == SystemZ::Trap) {
763 MI.setDesc(get(SystemZ::CondTrap));
764 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
765 .addImm(CCValid).addImm(CCMask)
766 .addReg(SystemZ::CC, RegState::Implicit);
769 if (Opcode == SystemZ::Return) {
770 MI.setDesc(get(SystemZ::CondReturn));
771 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
772 .addImm(CCValid).addImm(CCMask)
773 .addReg(SystemZ::CC, RegState::Implicit);
776 if (Opcode == SystemZ::CallJG) {
777 MachineOperand FirstOp = MI.getOperand(0);
778 const uint32_t *RegMask = MI.getOperand(1).getRegMask();
781 MI.setDesc(get(SystemZ::CallBRCL));
782 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
783 .addImm(CCValid).addImm(CCMask)
786 .addReg(SystemZ::CC, RegState::Implicit);
789 if (Opcode == SystemZ::CallBR) {
790 const uint32_t *RegMask = MI.getOperand(0).getRegMask();
792 MI.setDesc(get(SystemZ::CallBCR));
793 MachineInstrBuilder(*MI.getParent()->getParent(), MI)
794 .addImm(CCValid).addImm(CCMask)
796 .addReg(SystemZ::CC, RegState::Implicit);
802 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
803 MachineBasicBlock::iterator MBBI,
804 const DebugLoc &DL, unsigned DestReg,
805 unsigned SrcReg, bool KillSrc) const {
806 // Split 128-bit GPR moves into two 64-bit moves. This handles ADDR128 too.
807 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) {
808 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64),
809 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc);
810 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64),
811 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc);
815 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) {
816 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc);
820 // Everything else needs only one instruction.
822 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg))
823 Opcode = SystemZ::LGR;
824 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg))
825 // For z13 we prefer LDR over LER to avoid partial register dependencies.
826 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER;
827 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg))
828 Opcode = SystemZ::LDR;
829 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg))
830 Opcode = SystemZ::LXR;
831 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg))
832 Opcode = SystemZ::VLR32;
833 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg))
834 Opcode = SystemZ::VLR64;
835 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg))
836 Opcode = SystemZ::VLR;
837 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg))
838 Opcode = SystemZ::CPYA;
839 else if (SystemZ::AR32BitRegClass.contains(DestReg) &&
840 SystemZ::GR32BitRegClass.contains(SrcReg))
841 Opcode = SystemZ::SAR;
842 else if (SystemZ::GR32BitRegClass.contains(DestReg) &&
843 SystemZ::AR32BitRegClass.contains(SrcReg))
844 Opcode = SystemZ::EAR;
846 llvm_unreachable("Impossible reg-to-reg copy");
848 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg)
849 .addReg(SrcReg, getKillRegState(KillSrc));
852 void SystemZInstrInfo::storeRegToStackSlot(
853 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg,
854 bool isKill, int FrameIdx, const TargetRegisterClass *RC,
855 const TargetRegisterInfo *TRI) const {
856 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
858 // Callers may expect a single instruction, so keep 128-bit moves
859 // together for now and lower them after register allocation.
860 unsigned LoadOpcode, StoreOpcode;
861 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
862 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode))
863 .addReg(SrcReg, getKillRegState(isKill)),
867 void SystemZInstrInfo::loadRegFromStackSlot(
868 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg,
869 int FrameIdx, const TargetRegisterClass *RC,
870 const TargetRegisterInfo *TRI) const {
871 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
873 // Callers may expect a single instruction, so keep 128-bit moves
874 // together for now and lower them after register allocation.
875 unsigned LoadOpcode, StoreOpcode;
876 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode);
877 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg),
881 // Return true if MI is a simple load or store with a 12-bit displacement
882 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores.
883 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) {
884 const MCInstrDesc &MCID = MI->getDesc();
885 return ((MCID.TSFlags & Flag) &&
886 isUInt<12>(MI->getOperand(2).getImm()) &&
887 MI->getOperand(3).getReg() == 0);
892 LogicOp() : RegSize(0), ImmLSB(0), ImmSize(0) {}
893 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize)
894 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {}
896 explicit operator bool() const { return RegSize; }
898 unsigned RegSize, ImmLSB, ImmSize;
900 } // end anonymous namespace
902 static LogicOp interpretAndImmediate(unsigned Opcode) {
904 case SystemZ::NILMux: return LogicOp(32, 0, 16);
905 case SystemZ::NIHMux: return LogicOp(32, 16, 16);
906 case SystemZ::NILL64: return LogicOp(64, 0, 16);
907 case SystemZ::NILH64: return LogicOp(64, 16, 16);
908 case SystemZ::NIHL64: return LogicOp(64, 32, 16);
909 case SystemZ::NIHH64: return LogicOp(64, 48, 16);
910 case SystemZ::NIFMux: return LogicOp(32, 0, 32);
911 case SystemZ::NILF64: return LogicOp(64, 0, 32);
912 case SystemZ::NIHF64: return LogicOp(64, 32, 32);
913 default: return LogicOp();
917 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) {
918 if (OldMI->registerDefIsDead(SystemZ::CC)) {
919 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC);
920 if (CCDef != nullptr)
921 CCDef->setIsDead(true);
925 // Used to return from convertToThreeAddress after replacing two-address
926 // instruction OldMI with three-address instruction NewMI.
927 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI,
931 unsigned NumOps = OldMI->getNumOperands();
932 for (unsigned I = 1; I < NumOps; ++I) {
933 MachineOperand &Op = OldMI->getOperand(I);
934 if (Op.isReg() && Op.isKill())
935 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI);
938 transferDeadCC(OldMI, NewMI);
942 MachineInstr *SystemZInstrInfo::convertToThreeAddress(
943 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const {
944 MachineBasicBlock *MBB = MI.getParent();
945 MachineFunction *MF = MBB->getParent();
946 MachineRegisterInfo &MRI = MF->getRegInfo();
948 unsigned Opcode = MI.getOpcode();
949 unsigned NumOps = MI.getNumOperands();
951 // Try to convert something like SLL into SLLK, if supported.
952 // We prefer to keep the two-operand form where possible both
953 // because it tends to be shorter and because some instructions
954 // have memory forms that can be used during spilling.
955 if (STI.hasDistinctOps()) {
956 MachineOperand &Dest = MI.getOperand(0);
957 MachineOperand &Src = MI.getOperand(1);
958 unsigned DestReg = Dest.getReg();
959 unsigned SrcReg = Src.getReg();
960 // AHIMux is only really a three-operand instruction when both operands
961 // are low registers. Try to constrain both operands to be low if
963 if (Opcode == SystemZ::AHIMux &&
964 TargetRegisterInfo::isVirtualRegister(DestReg) &&
965 TargetRegisterInfo::isVirtualRegister(SrcReg) &&
966 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) &&
967 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) {
968 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass);
969 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass);
971 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode);
972 if (ThreeOperandOpcode >= 0) {
973 // Create three address instruction without adding the implicit
974 // operands. Those will instead be copied over from the original
975 // instruction by the loop below.
976 MachineInstrBuilder MIB(
977 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(),
978 /*NoImplicit=*/true));
979 MIB.addOperand(Dest);
980 // Keep the kill state, but drop the tied flag.
981 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg());
982 // Keep the remaining operands as-is.
983 for (unsigned I = 2; I < NumOps; ++I)
984 MIB.addOperand(MI.getOperand(I));
985 MBB->insert(MI, MIB);
986 return finishConvertToThreeAddress(&MI, MIB, LV);
990 // Try to convert an AND into an RISBG-type instruction.
991 if (LogicOp And = interpretAndImmediate(Opcode)) {
992 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB;
993 // AND IMMEDIATE leaves the other bits of the register unchanged.
994 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB);
996 if (isRxSBGMask(Imm, And.RegSize, Start, End)) {
998 if (And.RegSize == 64) {
999 NewOpcode = SystemZ::RISBG;
1000 // Prefer RISBGN if available, since it does not clobber CC.
1001 if (STI.hasMiscellaneousExtensions())
1002 NewOpcode = SystemZ::RISBGN;
1004 NewOpcode = SystemZ::RISBMux;
1008 MachineOperand &Dest = MI.getOperand(0);
1009 MachineOperand &Src = MI.getOperand(1);
1010 MachineInstrBuilder MIB =
1011 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode))
1014 .addReg(Src.getReg(), getKillRegState(Src.isKill()),
1019 return finishConvertToThreeAddress(&MI, MIB, LV);
1025 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl(
1026 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
1027 MachineBasicBlock::iterator InsertPt, int FrameIndex,
1028 LiveIntervals *LIS) const {
1029 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1030 const MachineFrameInfo &MFI = MF.getFrameInfo();
1031 unsigned Size = MFI.getObjectSize(FrameIndex);
1032 unsigned Opcode = MI.getOpcode();
1034 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1035 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) &&
1036 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) {
1038 // Check CC liveness, since new instruction introduces a dead
1040 MCRegUnitIterator CCUnit(SystemZ::CC, TRI);
1041 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit);
1043 assert (!CCUnit.isValid() && "CC only has one reg unit.");
1045 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot();
1046 if (!CCLiveRange.liveAt(MISlot)) {
1047 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST
1048 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt,
1049 MI.getDebugLoc(), get(SystemZ::AGSI))
1050 .addFrameIndex(FrameIndex)
1052 .addImm(MI.getOperand(2).getImm());
1053 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true);
1054 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator());
1061 // All other cases require a single operand.
1062 if (Ops.size() != 1)
1065 unsigned OpNum = Ops[0];
1068 .getRegClass(MI.getOperand(OpNum).getReg())
1070 "Invalid size combination");
1072 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 &&
1073 isInt<8>(MI.getOperand(2).getImm())) {
1074 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST
1075 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI);
1076 MachineInstr *BuiltMI =
1077 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode))
1078 .addFrameIndex(FrameIndex)
1080 .addImm(MI.getOperand(2).getImm());
1081 transferDeadCC(&MI, BuiltMI);
1085 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) {
1086 bool Op0IsGPR = (Opcode == SystemZ::LGDR);
1087 bool Op1IsGPR = (Opcode == SystemZ::LDGR);
1088 // If we're spilling the destination of an LDGR or LGDR, store the
1089 // source register instead.
1091 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD;
1092 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1094 .addOperand(MI.getOperand(1))
1095 .addFrameIndex(FrameIndex)
1099 // If we're spilling the source of an LDGR or LGDR, load the
1100 // destination register instead.
1102 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD;
1103 unsigned Dest = MI.getOperand(0).getReg();
1104 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1105 get(LoadOpcode), Dest)
1106 .addFrameIndex(FrameIndex)
1112 // Look for cases where the source of a simple store or the destination
1113 // of a simple load is being spilled. Try to use MVC instead.
1115 // Although MVC is in practice a fast choice in these cases, it is still
1116 // logically a bytewise copy. This means that we cannot use it if the
1117 // load or store is volatile. We also wouldn't be able to use MVC if
1118 // the two memories partially overlap, but that case cannot occur here,
1119 // because we know that one of the memories is a full frame index.
1121 // For performance reasons, we also want to avoid using MVC if the addresses
1122 // might be equal. We don't worry about that case here, because spill slot
1123 // coloring happens later, and because we have special code to remove
1124 // MVCs that turn out to be redundant.
1125 if (OpNum == 0 && MI.hasOneMemOperand()) {
1126 MachineMemOperand *MMO = *MI.memoperands_begin();
1127 if (MMO->getSize() == Size && !MMO->isVolatile()) {
1128 // Handle conversion of loads.
1129 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) {
1130 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1132 .addFrameIndex(FrameIndex)
1135 .addOperand(MI.getOperand(1))
1136 .addImm(MI.getOperand(2).getImm())
1137 .addMemOperand(MMO);
1139 // Handle conversion of stores.
1140 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) {
1141 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(),
1143 .addOperand(MI.getOperand(1))
1144 .addImm(MI.getOperand(2).getImm())
1146 .addFrameIndex(FrameIndex)
1148 .addMemOperand(MMO);
1153 // If the spilled operand is the final one, try to change <INSN>R
1155 int MemOpcode = SystemZ::getMemOpcode(Opcode);
1156 if (MemOpcode >= 0) {
1157 unsigned NumOps = MI.getNumExplicitOperands();
1158 if (OpNum == NumOps - 1) {
1159 const MCInstrDesc &MemDesc = get(MemOpcode);
1160 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags);
1161 assert(AccessBytes != 0 && "Size of access should be known");
1162 assert(AccessBytes <= Size && "Access outside the frame index");
1163 uint64_t Offset = Size - AccessBytes;
1164 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
1165 MI.getDebugLoc(), get(MemOpcode));
1166 for (unsigned I = 0; I < OpNum; ++I)
1167 MIB.addOperand(MI.getOperand(I));
1168 MIB.addFrameIndex(FrameIndex).addImm(Offset);
1169 if (MemDesc.TSFlags & SystemZII::HasIndex)
1171 transferDeadCC(&MI, MIB);
1179 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl(
1180 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
1181 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
1182 LiveIntervals *LIS) const {
1186 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
1187 switch (MI.getOpcode()) {
1189 splitMove(MI, SystemZ::LG);
1192 case SystemZ::ST128:
1193 splitMove(MI, SystemZ::STG);
1197 splitMove(MI, SystemZ::LD);
1201 splitMove(MI, SystemZ::STD);
1204 case SystemZ::LBMux:
1205 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH);
1208 case SystemZ::LHMux:
1209 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH);
1212 case SystemZ::LLCRMux:
1213 expandZExtPseudo(MI, SystemZ::LLCR, 8);
1216 case SystemZ::LLHRMux:
1217 expandZExtPseudo(MI, SystemZ::LLHR, 16);
1220 case SystemZ::LLCMux:
1221 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH);
1224 case SystemZ::LLHMux:
1225 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH);
1229 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH);
1232 case SystemZ::LOCMux:
1233 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH);
1236 case SystemZ::LOCHIMux:
1237 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI);
1240 case SystemZ::LOCRMux:
1241 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR);
1244 case SystemZ::STCMux:
1245 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH);
1248 case SystemZ::STHMux:
1249 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH);
1252 case SystemZ::STMux:
1253 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH);
1256 case SystemZ::STOCMux:
1257 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH);
1260 case SystemZ::LHIMux:
1261 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true);
1264 case SystemZ::IIFMux:
1265 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false);
1268 case SystemZ::IILMux:
1269 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false);
1272 case SystemZ::IIHMux:
1273 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false);
1276 case SystemZ::NIFMux:
1277 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false);
1280 case SystemZ::NILMux:
1281 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false);
1284 case SystemZ::NIHMux:
1285 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false);
1288 case SystemZ::OIFMux:
1289 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false);
1292 case SystemZ::OILMux:
1293 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false);
1296 case SystemZ::OIHMux:
1297 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false);
1300 case SystemZ::XIFMux:
1301 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false);
1304 case SystemZ::TMLMux:
1305 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false);
1308 case SystemZ::TMHMux:
1309 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false);
1312 case SystemZ::AHIMux:
1313 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false);
1316 case SystemZ::AHIMuxK:
1317 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH);
1320 case SystemZ::AFIMux:
1321 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false);
1324 case SystemZ::CHIMux:
1325 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false);
1328 case SystemZ::CFIMux:
1329 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false);
1332 case SystemZ::CLFIMux:
1333 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false);
1337 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF);
1340 case SystemZ::CLMux:
1341 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF);
1344 case SystemZ::RISBMux: {
1345 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg());
1346 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg());
1347 if (SrcIsHigh == DestIsHigh)
1348 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL));
1350 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH));
1351 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32);
1356 case SystemZ::ADJDYNALLOC:
1357 splitAdjDynAlloc(MI);
1360 case TargetOpcode::LOAD_STACK_GUARD:
1361 expandLoadStackGuard(&MI);
1369 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
1370 if (MI.getOpcode() == TargetOpcode::INLINEASM) {
1371 const MachineFunction *MF = MI.getParent()->getParent();
1372 const char *AsmStr = MI.getOperand(0).getSymbolName();
1373 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
1375 return MI.getDesc().getSize();
1379 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const {
1380 switch (MI.getOpcode()) {
1384 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY,
1385 SystemZ::CCMASK_ANY, &MI.getOperand(0));
1389 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(),
1390 MI.getOperand(1).getImm(), &MI.getOperand(2));
1393 case SystemZ::BRCTH:
1394 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP,
1395 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2));
1397 case SystemZ::BRCTG:
1398 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP,
1399 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2));
1403 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP,
1404 MI.getOperand(2).getImm(), &MI.getOperand(3));
1408 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP,
1409 MI.getOperand(2).getImm(), &MI.getOperand(3));
1413 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP,
1414 MI.getOperand(2).getImm(), &MI.getOperand(3));
1416 case SystemZ::CLGIJ:
1417 case SystemZ::CLGRJ:
1418 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP,
1419 MI.getOperand(2).getImm(), &MI.getOperand(3));
1422 llvm_unreachable("Unrecognized branch opcode");
1426 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC,
1427 unsigned &LoadOpcode,
1428 unsigned &StoreOpcode) const {
1429 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) {
1430 LoadOpcode = SystemZ::L;
1431 StoreOpcode = SystemZ::ST;
1432 } else if (RC == &SystemZ::GRH32BitRegClass) {
1433 LoadOpcode = SystemZ::LFH;
1434 StoreOpcode = SystemZ::STFH;
1435 } else if (RC == &SystemZ::GRX32BitRegClass) {
1436 LoadOpcode = SystemZ::LMux;
1437 StoreOpcode = SystemZ::STMux;
1438 } else if (RC == &SystemZ::GR64BitRegClass ||
1439 RC == &SystemZ::ADDR64BitRegClass) {
1440 LoadOpcode = SystemZ::LG;
1441 StoreOpcode = SystemZ::STG;
1442 } else if (RC == &SystemZ::GR128BitRegClass ||
1443 RC == &SystemZ::ADDR128BitRegClass) {
1444 LoadOpcode = SystemZ::L128;
1445 StoreOpcode = SystemZ::ST128;
1446 } else if (RC == &SystemZ::FP32BitRegClass) {
1447 LoadOpcode = SystemZ::LE;
1448 StoreOpcode = SystemZ::STE;
1449 } else if (RC == &SystemZ::FP64BitRegClass) {
1450 LoadOpcode = SystemZ::LD;
1451 StoreOpcode = SystemZ::STD;
1452 } else if (RC == &SystemZ::FP128BitRegClass) {
1453 LoadOpcode = SystemZ::LX;
1454 StoreOpcode = SystemZ::STX;
1455 } else if (RC == &SystemZ::VR32BitRegClass) {
1456 LoadOpcode = SystemZ::VL32;
1457 StoreOpcode = SystemZ::VST32;
1458 } else if (RC == &SystemZ::VR64BitRegClass) {
1459 LoadOpcode = SystemZ::VL64;
1460 StoreOpcode = SystemZ::VST64;
1461 } else if (RC == &SystemZ::VF128BitRegClass ||
1462 RC == &SystemZ::VR128BitRegClass) {
1463 LoadOpcode = SystemZ::VL;
1464 StoreOpcode = SystemZ::VST;
1466 llvm_unreachable("Unsupported regclass to load or store");
1469 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode,
1470 int64_t Offset) const {
1471 const MCInstrDesc &MCID = get(Opcode);
1472 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset);
1473 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) {
1474 // Get the instruction to use for unsigned 12-bit displacements.
1475 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode);
1476 if (Disp12Opcode >= 0)
1477 return Disp12Opcode;
1479 // All address-related instructions can use unsigned 12-bit
1483 if (isInt<20>(Offset) && isInt<20>(Offset2)) {
1484 // Get the instruction to use for signed 20-bit displacements.
1485 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode);
1486 if (Disp20Opcode >= 0)
1487 return Disp20Opcode;
1489 // Check whether Opcode allows signed 20-bit displacements.
1490 if (MCID.TSFlags & SystemZII::Has20BitOffset)
1496 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const {
1498 case SystemZ::L: return SystemZ::LT;
1499 case SystemZ::LY: return SystemZ::LT;
1500 case SystemZ::LG: return SystemZ::LTG;
1501 case SystemZ::LGF: return SystemZ::LTGF;
1502 case SystemZ::LR: return SystemZ::LTR;
1503 case SystemZ::LGFR: return SystemZ::LTGFR;
1504 case SystemZ::LGR: return SystemZ::LTGR;
1505 case SystemZ::LER: return SystemZ::LTEBR;
1506 case SystemZ::LDR: return SystemZ::LTDBR;
1507 case SystemZ::LXR: return SystemZ::LTXBR;
1508 case SystemZ::LCDFR: return SystemZ::LCDBR;
1509 case SystemZ::LPDFR: return SystemZ::LPDBR;
1510 case SystemZ::LNDFR: return SystemZ::LNDBR;
1511 case SystemZ::LCDFR_32: return SystemZ::LCEBR;
1512 case SystemZ::LPDFR_32: return SystemZ::LPEBR;
1513 case SystemZ::LNDFR_32: return SystemZ::LNEBR;
1514 // On zEC12 we prefer to use RISBGN. But if there is a chance to
1515 // actually use the condition code, we may turn it back into RISGB.
1516 // Note that RISBG is not really a "load-and-test" instruction,
1517 // but sets the same condition code values, so is OK to use here.
1518 case SystemZ::RISBGN: return SystemZ::RISBG;
1523 // Return true if Mask matches the regexp 0*1+0*, given that zero masks
1524 // have already been filtered out. Store the first set bit in LSB and
1525 // the number of set bits in Length if so.
1526 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) {
1527 unsigned First = findFirstSet(Mask);
1528 uint64_t Top = (Mask >> First) + 1;
1529 if ((Top & -Top) == Top) {
1531 Length = findFirstSet(Top);
1537 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize,
1538 unsigned &Start, unsigned &End) const {
1539 // Reject trivial all-zero masks.
1540 Mask &= allOnes(BitSize);
1544 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of
1545 // the msb and End specifies the index of the lsb.
1546 unsigned LSB, Length;
1547 if (isStringOfOnes(Mask, LSB, Length)) {
1548 Start = 63 - (LSB + Length - 1);
1553 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb
1554 // of the low 1s and End specifies the lsb of the high 1s.
1555 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) {
1556 assert(LSB > 0 && "Bottom bit must be set");
1557 assert(LSB + Length < BitSize && "Top bit must be set");
1558 Start = 63 - (LSB - 1);
1559 End = 63 - (LSB + Length);
1566 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode,
1567 SystemZII::FusedCompareType Type,
1568 const MachineInstr *MI) const {
1572 if (!(MI && isInt<8>(MI->getOperand(1).getImm())))
1576 case SystemZ::CLGFI:
1577 if (!(MI && isUInt<8>(MI->getOperand(1).getImm())))
1582 if (!STI.hasMiscellaneousExtensions())
1584 if (!(MI && MI->getOperand(3).getReg() == 0))
1589 case SystemZII::CompareAndBranch:
1592 return SystemZ::CRJ;
1594 return SystemZ::CGRJ;
1596 return SystemZ::CIJ;
1598 return SystemZ::CGIJ;
1600 return SystemZ::CLRJ;
1602 return SystemZ::CLGRJ;
1604 return SystemZ::CLIJ;
1605 case SystemZ::CLGFI:
1606 return SystemZ::CLGIJ;
1610 case SystemZII::CompareAndReturn:
1613 return SystemZ::CRBReturn;
1615 return SystemZ::CGRBReturn;
1617 return SystemZ::CIBReturn;
1619 return SystemZ::CGIBReturn;
1621 return SystemZ::CLRBReturn;
1623 return SystemZ::CLGRBReturn;
1625 return SystemZ::CLIBReturn;
1626 case SystemZ::CLGFI:
1627 return SystemZ::CLGIBReturn;
1631 case SystemZII::CompareAndSibcall:
1634 return SystemZ::CRBCall;
1636 return SystemZ::CGRBCall;
1638 return SystemZ::CIBCall;
1640 return SystemZ::CGIBCall;
1642 return SystemZ::CLRBCall;
1644 return SystemZ::CLGRBCall;
1646 return SystemZ::CLIBCall;
1647 case SystemZ::CLGFI:
1648 return SystemZ::CLGIBCall;
1652 case SystemZII::CompareAndTrap:
1655 return SystemZ::CRT;
1657 return SystemZ::CGRT;
1659 return SystemZ::CIT;
1661 return SystemZ::CGIT;
1663 return SystemZ::CLRT;
1665 return SystemZ::CLGRT;
1667 return SystemZ::CLFIT;
1668 case SystemZ::CLGFI:
1669 return SystemZ::CLGIT;
1671 return SystemZ::CLT;
1673 return SystemZ::CLGT;
1681 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const {
1682 if (!STI.hasLoadAndTrap())
1687 return SystemZ::LAT;
1689 return SystemZ::LGAT;
1691 return SystemZ::LFHAT;
1693 return SystemZ::LLGFAT;
1695 return SystemZ::LLGTAT;
1700 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB,
1701 MachineBasicBlock::iterator MBBI,
1702 unsigned Reg, uint64_t Value) const {
1703 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
1705 if (isInt<16>(Value))
1706 Opcode = SystemZ::LGHI;
1707 else if (SystemZ::isImmLL(Value))
1708 Opcode = SystemZ::LLILL;
1709 else if (SystemZ::isImmLH(Value)) {
1710 Opcode = SystemZ::LLILH;
1713 assert(isInt<32>(Value) && "Huge values not handled yet");
1714 Opcode = SystemZ::LGFI;
1716 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value);
1719 bool SystemZInstrInfo::
1720 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb,
1721 AliasAnalysis *AA) const {
1723 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand())
1726 // If mem-operands show that the same address Value is used by both
1727 // instructions, check for non-overlapping offsets and widths. Not
1728 // sure if a register based analysis would be an improvement...
1730 MachineMemOperand *MMOa = *MIa.memoperands_begin();
1731 MachineMemOperand *MMOb = *MIb.memoperands_begin();
1732 const Value *VALa = MMOa->getValue();
1733 const Value *VALb = MMOb->getValue();
1734 bool SameVal = (VALa && VALb && (VALa == VALb));
1736 const PseudoSourceValue *PSVa = MMOa->getPseudoValue();
1737 const PseudoSourceValue *PSVb = MMOb->getPseudoValue();
1738 if (PSVa && PSVb && (PSVa == PSVb))
1742 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset();
1743 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize();
1744 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
1745 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
1746 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
1747 if (LowOffset + LowWidth <= HighOffset)